{"id": "2607.24148", "title": "A Motion-Aware Vector Quantization Framework with Centroid Reuse for Efficient VLA Inference", "authors": ["Zhuoran Song", "Haozhe Jiang", "Chunyu Qi", "Minnan Pei", "Gang Li", "Xiaoyao Liang", "Haibing Guan", "Min-Nan Pei"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.24148", "pdf_url": "https://arxiv.org/pdf/2607.24148", "code_url": "", "clusters": ["C1", "C2", "C3", "C5", "C9"], "keywords": ["vector quantization", "accelerator", "centroid reuse", "codebook", "motion-aware precision"], "model_class": "VLA", "techniques": ["vector quantization", "hardware co-design", "motion-aware vector quantization", "codebook-domain GEMM", "custom accelerator", "centroid reuse"], "base_model": "", "params": "", "hardware": "custom accelerator (vs A100)", "reported_metric": "6.5x speedup over A100 | 6.5x over A100, 2.8x over Dadu-Corki, 1.9x over LUT-DLA", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VQVLA quantizes VLA weights into codebooks whose precision adapts to robot motion state and builds an accelerator that computes directly on centroid indices with reuse.", "relevance": 5, "relevance_reason": "Dedicated VLA accelerator with memory-access reduction.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search: vision-language-action quantization; snowball from 2407.04292 (citations); venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2607.29169", "title": "ActFovea: Runtime Safeguarding for VLA Policies via Spatiotemporal Visual-Action Consistency", "authors": ["Wenda Yu", "Tianshi Wang", "Fengling Li", "Xin Li", "Jingjing Li", "Lei Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.29169", "pdf_url": "https://arxiv.org/pdf/2607.29169", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["ActFovea", "observation freshness", "stale observation detection", "safe failure"], "model_class": "VLA", "techniques": ["action-conditioned foveation", "consistency checks", "bounded safe-failure procedure"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "success under overlays 49.3% -> 90.3%; +9.8 pts under visual delay", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Plug-in safeguard that checks whether visual motion and observation freshness agree with kinematics and actions, recovering from disturbances or triggering safe failure on stale/replayed frames.", "relevance": 5, "relevance_reason": "Directly detects timing faults (visual delay, frozen observations).", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2608.00337", "title": "Action Chunk Scheduling for Batched Robot Policy Serving", "authors": ["Rohan Bansal", "David He", "Nadun Ranawaka Arachchige", "Zhenyang Chen", "Soobum Kim", "Kexin Rong", "Danfei Xu", "Bansal, Rohan", "He, David", "Arachchige, Nadun Ranawaka", "Chen, Zhenyang", "Kim, Soobum", "Rong, Kexin", "Xu, Danfei", "N. R. Arachchige", "Ke-Xin Rong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00337", "pdf_url": "https://arxiv.org/pdf/2608.00337", "code_url": "", "clusters": ["C1", "C3", "C4", "C5", "C6", "C7"], "keywords": ["batched serving", "chunk scheduling", "robot fleet", "Armory", "batched policy serving", "action chunk scheduling", "fleet", "policy serving", "batching", "multi-robot"], "model_class": "VLA", "techniques": ["serving scheduling", "batching", "chunk-aware scheduling", "chunk-aware batch scheduling", "remote GPU serving", "chunk-rate-aware scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 18% higher system throughput | up to 18% system throughput improvement in real-world experiments", "metric_measures": "throughput", "benchmarks": ["simulated and real robot fleets"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Formulates serving one VLA to many robots from a remote GPU as a scheduling problem and shows chunk-consumption-aware scheduling beats conventional batching when robots consume chunks at different rates.", "relevance": 5, "relevance_reason": "Formalizes multi-robot remote VLA serving as a scheduling problem.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA; WebSearch: action chunk size inference latency trade-off; snowball from 2602.18397 (citations); search query; snowball: citations of 2506.07339 (S2)"}
{"id": "2607.06370", "title": "ActionCache: Training-Free Acceleration for Vision-Language-Action Models with Action Caching and Refinement", "authors": ["Ryuji Oi", "Hikari Otsuka", "Kosuke Matsushima", "Yuki Ichikawa", "Masato Motomura", "Tatsuya Kaneko", "Daichi Fujiki"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.06370", "pdf_url": "https://arxiv.org/pdf/2607.06370", "code_url": "", "clusters": ["C1", "C3", "C9"], "keywords": ["action caching", "warm start", "flow matching", "training-free", "acceleration"], "model_class": "VLA", "techniques": ["action caching", "warm-start denoising", "action caching and refinement", "training-free"], "base_model": "pi0.5, GR00T N1.6", "params": "", "hardware": "", "reported_metric": "action-head speedup up to 10.44x (pi0.5) and 40.17x (GR00T N1.6)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Stores intermediate actions under compact multimodal keys and warm-starts flow denoising from retrieved similar contexts.", "relevance": 5, "relevance_reason": "Title unmistakably about VLA acceleration; unverified", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; venue sweep: S2 'vision-language-action real-time inference latency'; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2605.24011", "title": "ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models", "authors": ["Arash Akbari", "Arman Akbari", "Masih Eskandar", "Qitao Tan", "Yi-Xiao Chen", "Jing Luo", "Bertha Pangaribuan", "Li-Yun Zhang", "Jennifer G. Dy", "Geng Yuan", "Xuemin Lin", "Gao-Wen Liu", "Stratis Ioannidis", "Yanzhi Wang", "Yixiao Chen", "Jingwu Luo", "Liyun Zhang", "Jennifer Dy", "Xue Lin", "Gaowen Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24011", "pdf_url": "https://arxiv.org/pdf/2605.24011", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["sub-4-bit", "bits-per-weight", "C++ runtime", "UR3", "mixed precision PTQ", "C/C++ runtime", "action-guided PTQ"], "model_class": "VLA", "techniques": ["mixed-precision post-training quantization", "action-guided mixed-precision quantization", "low-bit kernels", "OmniModel.cpp runtime", "action-guided quantization"], "base_model": "OpenVLA-OFT; pi0.5", "params": "", "hardware": "", "reported_metric": "95.0% retained at <=3 bpw (OpenVLA-OFT); backbone 14.3 GB -> 2.7 GB at 2.5 bpw", "metric_measures": "relative memory", "benchmarks": ["LIBERO", "real robot", "real robot (UR3)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Allocates bit-widths per weight matrix by action sensitivity and tunes block scales, then ships models through a native C++ low-bit runtime; retains accuracy near 3 bits per weight.", "relevance": 5, "relevance_reason": "Quantization plus native edge runtime for deployment.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2603.07404", "title": "Adaptive Capacity Allocation for Vision Language Action Fine-tuning", "authors": ["Donghoon Kim", "Minji Bae", "Unghui Nam", "Gyeonghun Kim", "Suyun Lee", "Kyuhong Shim", "Byonghyo Shim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.07404", "pdf_url": "https://arxiv.org/pdf/2603.07404", "code_url": "", "clusters": ["C3"], "keywords": ["LoRA rank", "rank-adaptive LoRA", "LoRA-SP", "multi-task"], "model_class": "VLA", "techniques": ["LoRA", "rank-adaptive PEFT", "routing"], "base_model": "pi0; SmolVLA", "params": "", "hardware": "", "reported_metric": "up to +31.6% multi-task success over standard LoRA", "metric_measures": "other", "benchmarks": ["AgileX PiPER real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows VLA transfer needs much higher LoRA rank than LLMs and proposes input/layer-adaptive rank selection via a routed SVD-style vector bank.", "relevance": 5, "relevance_reason": "Core PEFT-for-VLA method with analysis of rank requirements.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: LoRA fine-tuning robot policy VLA adapter arXiv 2025"}
{"id": "2606.09811", "title": "AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing", "authors": ["Jisong Cai", "Long Ling", "Shiwei Chu", "Zhongshan Liu", "Jiayue Kang", "Zhixuan Liang", "Wenjie Xu", "Yinan Mao", "Weinan Zhang", "Xiaokang Yang", "Ru Ying", "Ran Zheng", "Yao Mu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09811", "pdf_url": "https://arxiv.org/pdf/2606.09811", "code_url": "", "clusters": ["C1", "C6", "C7", "C8", "C9"], "keywords": ["asynchronous WAM", "dual-rate", "dual DiT", "asynchronous horizon-adaptive", "context routing"], "model_class": "world-action model", "techniques": ["low-frequency video DiT planner with rolling KV memory", "high-frequency action DiT", "observation-guided context routing", "decoupled temporal resolutions", "async"], "base_model": "", "params": "", "hardware": "", "reported_metric": "24.17 Hz closed-loop control; 4.59x speedup over Fast-WAM", "metric_measures": "closed-loop control", "benchmarks": ["RoboTwin", "real robot (4 tasks)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Decouples the video world branch (slow, long-horizon context) from a fast action DiT that queries cached video context each step, so actions run at ~24 Hz without rerunning the video model.", "relevance": 5, "relevance_reason": "Dual-rate WAM reaching high closed-loop rates.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2608.15636", "title": "Algorithm-Architecture Co-Design for Efficient VLA Inference via Speculative Inference and Verification", "authors": ["Chunyu Qi", "Zhuoran Song", "Jian Weng", "Haozhe Jiang", "Xueyuan Liu", "Naifeng Jing", "Guanghui He", "Xiaoyao Liang", "Haibing Guan", "Xue-Yuan Liu", "Nai-Feng Jing", "Guang-Hui He"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15636", "pdf_url": "https://arxiv.org/pdf/2608.15636", "code_url": "", "clusters": ["C1", "C2", "C3", "C5", "C9"], "keywords": ["speculative inference", "verification model", "accelerator", "small verifier VLA", "mixed-precision", "heterogeneous hardware", "heterogeneous GPU + robotic module", "Dadu-Corki follow-up", "verification", "Dadu-Corki"], "model_class": "VLA", "techniques": ["speculative inference", "mixed-precision quantization", "hardware co-design", "speculative execution", "state-aware speculative long-horizon prediction", "small verification VLA", "block-wise mixed-precision quantization", "heterogeneous architecture", "speculative inference and verification"], "base_model": "OpenVLA, RDT", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "ManiSkill"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SpecVLA predicts long action sequences speculatively in low-stakes states and verifies them selectively with a small quantized model, running on a GPU plus a robot-specific hardware module.", "relevance": 5, "relevance_reason": "Uses quantized small-model construction; primarily a scheduling co-design.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search: vision-language-action quantization; snowball from 2407.04292 (citations); venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2602.02110", "title": "An Empirical Study of World Model Quantization", "authors": ["Zhongqian Fu", "Tianyi Zhao", "Kai Han", "Hang Zhou", "Xinghao Chen", "Yunhe Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.02110", "pdf_url": "https://arxiv.org/pdf/2602.02110", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["world model quantization", "PTQ study", "DINO-WM"], "model_class": "world model", "techniques": ["weight-only and weight-activation PTQ", "bit-width/granularity sweeps"], "base_model": "DINO-WM", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["visual planning tasks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Systematic study of post-training quantization of a latent planning world model (DINO-WM), showing how bit-width, granularity and planning horizon interact.", "relevance": 5, "relevance_reason": "First quantization study specifically for world-model planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.23048", "title": "Anatomy of a Closed-Loop Collapse: A Causal Case Study of a Compressed VLA Policy", "authors": ["Fengze Jia"], "year": 2026, "venue": "IROS 2026 Workshop (ScaleInfra)", "type": "paper", "url": "https://arxiv.org/abs/2609.23048", "pdf_url": "https://arxiv.org/pdf/2609.23048", "code_url": "https://github.com/Xanadum/closed-loop-collapse", "clusters": ["C2", "C4"], "keywords": ["distillation", "closed-loop collapse", "offline vs online eval"], "model_class": "VLA", "techniques": ["layer-reduction distillation"], "base_model": "Octo-Base", "params": "", "hardware": "", "reported_metric": "student 0/72 vs teacher 40/72 despite passing offline checks", "metric_measures": "other", "benchmarks": ["SimplerEnv WidowX"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Causal case study of an 8-layer Octo distillation that passes offline metrics but fails closed-loop; only training on deployment-distribution teacher rollouts repairs it.", "relevance": 5, "relevance_reason": "Direct evidence that compressed policies can silently break closed-loop.", "foundational_technique": false, "verified": true, "found_via": "search: distillation diffusion policy"}
{"id": "2609.33748", "title": "AnyStep-WAM: Budget-Aligned Distillation and Adaptive Inference for World Action Models", "authors": ["Rui Wang", "Xiangyu Wang", "Donglin Yang", "Yibo Li", "Canyang Chen", "Zhongrui Wang", "Xiaojuan Qi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33748", "pdf_url": "https://arxiv.org/pdf/2609.33748", "code_url": "", "clusters": ["C2", "C7", "C8"], "keywords": ["budget-aligned distillation", "adaptive step count", "risk-aware compute allocation"], "model_class": "world-action model", "techniques": ["interval-conditioned flow-map distillation with shared LoRA", "risk-based per-chunk step selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Distills a WAM so one model can generate actions with anywhere from one to many steps, then uses a light risk estimator to spend more steps only on precision-critical chunks.", "relevance": 5, "relevance_reason": "Adaptive-compute WAM inference tied to risk.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2603.16666 (citations)"}
{"id": "2604.24086", "title": "AsyncShield: A Plug-and-Play Edge Adapter for Asynchronous Cloud-based VLA Navigation", "authors": ["Kai Yang", "Zedong Chu", "Yingnan Guo", "Zhengbo Wang", "Shichao Xie", "Yanfen Shen", "Xiaolong Wu", "Xing Li", "Mu Xu", "Ying-Nan Guo", "Zheng-Bo Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.24086", "pdf_url": "https://arxiv.org/pdf/2604.24086", "code_url": "", "clusters": ["C5", "C6", "C7", "C8", "C9"], "keywords": ["cloud VLA navigation", "network jitter", "pose buffer", "CMDP", "AsyncShield", "cloud VLA", "navigation", "edge adapter", "jitter"], "model_class": "VLA", "techniques": ["kinematic lag compensation", "safe RL edge adapter (PPO-Lagrangian)", "kinematic spatial remapping of stale intents", "edge adapter", "plug-and-play edge adapter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Plug-and-play edge module that converts cloud-VLA latency into pose offsets using a pose buffer and a constrained RL adapter balancing VLA intent against LiDAR obstacle avoidance.", "relevance": 5, "relevance_reason": "Edge-side safety/latency compensation for cloud VLA.", "foundational_technique": false, "verified": true, "found_via": "search query; snowball: citations of 2506.07339 (S2); venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2602.13476", "title": "AsyncVLA: An Asynchronous VLA for Fast and Robust Navigation on the Edge", "authors": ["Noriaki Hirose", "Catherine Glossop", "Dhruv Shah", "Sergey Levine"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.13476", "pdf_url": "https://arxiv.org/pdf/2602.13476", "code_url": "", "clusters": ["C1", "C2", "C3", "C5", "C6", "C7", "C9"], "keywords": ["asynchronous VLA", "edge adapter", "navigation", "remote workstation", "communication delay", "AsyncVLA", "remote inference", "asynchronous inference", "remote foundation model"], "model_class": "VLA", "techniques": ["cloud-edge split", "asynchronous execution", "hierarchical remote/onboard architecture", "edge adapter", "asynchronous inference", "hierarchical async control", "onboard edge adapter", "edge-cloud split", "lightweight onboard adapter", "trajectory re-weighting"], "base_model": "OmniVLA", "params": "", "hardware": "remote workstation + onboard edge compute", "reported_metric": "40% higher success than baselines with delays up to 6 s | +40% success vs baselines with communication delays up to 6 s", "metric_measures": "other", "benchmarks": ["real-world visual navigation", "real robot navigation"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Runs a large navigation VLA remotely for high-level guidance while a small onboard Edge Adapter refines actions at high rate, fine-tuned end-to-end to tolerate asynchronous, delayed guidance.", "relevance": 5, "relevance_reason": "Direct edge-cloud hybrid VLA for real-time robot control by a workshop organizer.", "foundational_technique": false, "verified": true, "found_via": "search query: hierarchical VLA latency; search query; arXiv search: vision-language-action inference latency delay real-time; author sweep: Dhruv Shah"}
{"id": "2606.02775", "title": "AURA: Action-Gated Memory for Robot Policies at Constant VRAM", "authors": ["Josef Chen", "J. Chen"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.02775", "pdf_url": "https://arxiv.org/pdf/2606.02775", "code_url": "", "clusters": ["C1", "C2", "C5"], "keywords": ["constant memory", "KV cache", "edge memory", "write endurance", "constant VRAM", "edge memory writes"], "model_class": "VLA", "techniques": ["recurrent memory", "learned write gating", "gated recurrent memory", "write reduction"], "base_model": "OpenVLA-OFT 7B", "params": "7B", "hardware": "", "reported_metric": "fixed 4,224-byte inference state vs KV cache 6,061x larger at 100k steps; 7.0x fewer writes at matched LIBERO-Long success (0.233)", "metric_measures": "relative memory", "benchmarks": ["LIBERO-Long", "synthetic"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Replaces a growing KV cache with a fixed-size recurrent memory that only writes when the observation would change the next action, keeping long-episode memory constant on edge devices without hurting OpenVLA-OFT success.", "relevance": 5, "relevance_reason": "Addresses memory growth that matters for fitting 7B VLAs with history on fixed-memory Jetsons.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: arXiv abs:OpenVLA AND abs:memory AND abs:edge; recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\"); snowball round 2: cites 2205.14135, 2502.02175"}
{"id": "2609.36967", "title": "Beyond Token Importance: Preserving Spatial Scaffolds for Efficient Vision-Language-Action Inference", "authors": ["Jia-Yu Chen", "Shuyong Gao", "Jing-Kai Jia", "Xiao-Sheng Bu", "Ji-Yuan Fu", "Ling-Yi Hong", "Kai-Xun Jiang", "Yi-Pan Xu", "Wen-Qiang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36967", "pdf_url": "https://arxiv.org/pdf/2609.36967", "code_url": "", "clusters": ["C1"], "keywords": ["token pruning", "spatial coverage"], "model_class": "VLA", "techniques": ["visual token pruning", "training-free"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "1.78x prefill speedup with 20% of visual tokens", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "GeoScaffold shows semantic token pruning for VLAs neglects spatial coverage and proposes a training-free pruning that allocates per-region budgets and keeps spatially spread scaffold tokens.", "relevance": 5, "relevance_reason": "direct: VLA visual token pruning", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2608.03938", "title": "Bimanual Manipulation Within an 8 GB Budget: Zero-Copy Sensing and Quantized ACT on an Entry-Level Jetson", "authors": ["Ekansh Singh", "Eva Samuel", "Alessandra Reneau", "Ryan Schmeelk", "Yashvi Gandhi", "Y. Gandhi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.03938", "pdf_url": "https://arxiv.org/pdf/2608.03938", "code_url": "", "clusters": ["C1", "C2", "C3", "C5"], "keywords": ["Jetson Orin Nano", "TensorRT", "ACT", "INT8", "SO-101", "bimanual", "Jetson Orin Nano Super", "TensorRT INT8", "zero-copy GStreamer"], "model_class": "other", "techniques": ["quantization", "TensorRT", "zero-copy sensing", "INT8 quantization", "FP16", "TensorRT FP16/INT8", "zero-copy sensing (NVMM)"], "base_model": "ACT; Diffusion Policy", "params": "", "hardware": "Jetson Orin Nano Super 8 GB", "reported_metric": "ACT latency 114.02 ms -> 17.93 ms (FP16), 12.65 ms (INT8) | ACT latency 114.02 ms -> 12.65 ms with INT8 (9.0x), success 19/20", "metric_measures": "relative latency", "benchmarks": ["real robot", "real robot (bimanual SO-101)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Runs a bimanual ACT policy fully on an 8 GB Orin Nano, quantifying zero-copy camera capture gains and TensorRT FP16/INT8 speedups, and noting INT8 calibration skips all transformer layers.", "relevance": 5, "relevance_reason": "Detailed entry-level Jetson onboard deployment characterization.", "foundational_technique": false, "verified": true, "found_via": "search query: temporal ensembling; search: quantization imitation learning; web: Jetson diffusion policy real-time onboard"}
{"id": "2606.08962", "title": "C$^3$ache: Accelerating World Action Models with Cross Inference Chunk Cache", "authors": ["Weisen Zhao", "Lam Nguyen", "Zhicong Lu", "Yuzhang Shang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08962", "pdf_url": "https://arxiv.org/pdf/2606.08962", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["cross-chunk cache", "WAM acceleration", "training-free"], "model_class": "world-action model", "techniques": ["reuse denoising residuals across consecutive inference chunks"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Exploits that denoising residuals are highly correlated between consecutive WAM chunks during smooth behavior, caching them across chunks to cut inference cost.", "relevance": 5, "relevance_reason": "New axis of caching for WAM inference.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2603.16666 (citations)"}
{"id": "2609.31938", "title": "Cache-Aware Conv3D Lowering Across Embedded World-Model Decoders", "authors": ["Jiaming Zhang", "Wu Yang", "Shuai Tao", "Wulong Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.31938", "pdf_url": "https://arxiv.org/pdf/2609.31938", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["Jetson AGX Orin", "Conv3D lowering", "edge world model decoder"], "model_class": "world model", "techniques": ["cache-aware Conv3D-to-Conv2D lowering"], "base_model": "Cosmos3-Edge, Cosmos3-Nano", "params": "", "hardware": "NVIDIA Jetson AGX Orin 64GB", "reported_metric": "VAE decoding ~7x faster; full image-to-video generation latency >2x lower", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Runtime-level rewrite of causal Conv3D ops as batched Conv2D that speeds up Cosmos edge world-model decoders ~7x on Jetson AGX Orin.", "relevance": 5, "relevance_reason": "On-device world-model rollout acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.17496", "title": "Calibrated Predictive Safety for Heterogeneous Robots: An Action-Conditioned JEPA Framework with Model-Based Safety Shields", "authors": ["Kaiming Zhong", "Tianhua Liu", "Yue Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.17496", "pdf_url": "https://arxiv.org/pdf/2608.17496", "code_url": "", "clusters": ["C4", "C7", "C8", "C9"], "keywords": ["action-conditioned JEPA", "safety shield", "candidate chunk scoring", "risk prediction", "heterogeneous robots", "JEPA world model", "predictive safety", "calibration"], "model_class": "world model", "techniques": ["JEPA rollout risk heads", "deterministic safety shield", "latent rollout of candidate chunks", "calibrated risk/progress heads", "deterministic per-embodiment shield", "fallback ladder", "action-conditioned JEPA", "model-based safety check"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Long (sim only)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Receding-horizon pipeline that rolls candidate action chunks through an action-conditioned JEPA to score calibrated risk and progress, with a deterministic embodiment-specific shield providing hard guarantees.", "relevance": 5, "relevance_reason": "Latent world model as pre-execution safety monitor for VLA outputs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"shielding\" learned robot policy safety; search query; venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2609.21022", "title": "Catch Me If You Can: Real-Time Feedback Denoising for Responsive VLAs", "authors": ["Yiheng Ji", "Xingru Zhou", "Luis Sentis", "Mingyo Seo", "Ji, Yiheng", "Zhou, Xingru", "Sentis, Luis", "Seo, Mingyo"], "year": 2026, "venue": "CoRL 2026", "type": "preprint", "url": "https://arxiv.org/abs/2609.21022", "pdf_url": "https://arxiv.org/pdf/2609.21022", "code_url": "", "clusters": ["C1", "C2", "C4", "C7"], "keywords": ["two-timescale", "feedback denoising", "final denoising step", "dynamic tasks", "VLA-Feedback", "final denoising step as feedback", "real-time correction"], "model_class": "VLA", "techniques": ["dual-rate control", "feedback correction", "last-step feedback denoising", "retain last denoising step for per-action correction"], "base_model": "GR00T", "params": "", "hardware": "", "reported_metric": "dynamic sim success 27.5% to 85.0%; real 51% to 73%", "metric_measures": "other", "benchmarks": ["LIBERO", "real robot", "dynamic simulation tasks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Keeps the final denoising step of a diffusion VLA chunk as a lightweight feedback interface so each action is corrected with the latest observation right before execution, without rerunning the backbone.", "relevance": 5, "relevance_reason": "Shows static LIBERO parity masks dynamic-task differences.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; WebSearch: closed-loop vs open-loop control frequency VLA; arXiv search: real-time chunking"}
{"id": "2608.02453", "title": "Certifying Plans under Model Mismatch: A Trilemma for Reachability from Scarce Data", "authors": ["Yanliang Huang", "Zhen Zhang", "Ahmad Hafez", "Wenyuan Wu", "Peng Xie", "Zhuoqi Zeng", "Amr Alanwar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.02453", "pdf_url": "https://arxiv.org/pdf/2608.02453", "code_url": "", "clusters": ["C7"], "keywords": ["action chunk certification", "model mismatch"], "model_class": "other", "techniques": ["reachability", "certification", "Lipschitz bounds"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Studies pre-execution certification of a fixed control sequence such as a policy action chunk under sim-to-real model mismatch with scarce data, proving a trilemma between containment, finite tube width, and unrestricted model error, and offering a Lipschitz-based certifier.", "relevance": 5, "relevance_reason": "Explicitly frames certifying an action chunk before execution and gives impossibility results.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:reachability AND (\"visuomotor policy\" OR \"learned policy\" OR \"robot foundation model\" OR \"vision-language-action\")"}
{"id": "2604.24622", "title": "CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies", "authors": ["Fan Du", "Feng Yan", "Jianxiong Wu", "Xinrun Xu", "Weiye Zhang", "Weinong Wang", "Yu Guo", "Bin Qian", "Zhihai He", "Fei Wang", "Heng Yang"], "year": 2026, "venue": "ACM MM 2026", "type": "paper", "url": "https://arxiv.org/abs/2604.24622", "pdf_url": "https://arxiv.org/pdf/2604.24622", "code_url": "https://github.com/EmbodiedAI-RoboTron/CF-VLA", "clusters": ["C1"], "keywords": ["flow matching", "low NFE", "coarse-to-fine"], "model_class": "VLA", "techniques": ["few-step sampling", "learned initialization"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "75.4% lower action sampling latency; NFE=2", "metric_measures": "relative latency", "benchmarks": ["CALVIN", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Replaces Gaussian-noise starts with a learned action-aware coarse initialization followed by one refinement step, giving strong results at two function evaluations.", "relevance": 5, "relevance_reason": "Few-step flow VLA action generation.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient"}
{"id": "2604.24447", "title": "Characterizing Vision-Language-Action Models across XPUs: Constraints and Acceleration for On-Robot Deployment", "authors": ["Kaijun Zhou", "Qiwei Chen", "Da Peng", "Zhiyang Li", "Xijun Li", "Jinyu Gu", "Kai-Jun Zhou", "Qi-Wei Chen", "Dajiang Peng", "Zhi-Yang Li", "Xi-Jun Li", "Jin-Yu Gu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.24447", "pdf_url": "https://arxiv.org/pdf/2604.24447", "code_url": "", "clusters": ["C1", "C3", "C4", "C5", "C7", "C9"], "keywords": ["hardware characterization", "NPU", "energy", "pipeline parallelism", "cross-accelerator leaderboard", "cost energy time", "roofline", "two-phase inference", "XPU", "cost-energy-time", "leaderboard", "compute-bound vs memory-bound", "XPU characterization", "edge accelerators", "control-rate constraints"], "model_class": "VLA", "techniques": ["profiling", "diffusion caching", "asynchronous pipeline", "pipeline parallelism", "asynchronous execution", "hardware characterization", "DP-Cache", "V-AEFusion pipeline parallelism", "model-hardware co-characterization"], "base_model": "", "params": "", "hardware": "desktop GPUs, edge GPUs/NPUs", "reported_metric": "up to 2.9x (GPU) and 6x (edge NPU) speedup | up to 2.9x speedup on GPUs, 6x on edge NPUs", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "VLA leaderboard"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Builds a cross-accelerator VLA leaderboard scored by cost, energy and time, profiles a compute-bound VLM phase followed by a memory-bound action expert, and proposes DP-Cache and V-AEFusion to exploit it.", "relevance": 5, "relevance_reason": "Central model-hardware co-characterization across heterogeneous accelerators.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; search query: VLA acceleration evaluation; arXiv: vision-language-action AND Jetson/edge/onboard; snowball: citations of 2506.07339 (S2)"}
{"id": "2603.02271", "title": "Characterizing VLA Models: Identifying the Action Generation Bottleneck for Edge AI Architectures", "authors": ["Manoj Vishwanathan", "Suvinay Subramanian", "Anand Raghunathan"], "year": 2026, "venue": "workshop paper (arXiv)", "type": "preprint", "url": "https://arxiv.org/abs/2603.02271", "pdf_url": "https://arxiv.org/pdf/2603.02271", "code_url": "", "clusters": ["C1", "C3", "C4", "C5", "C6", "C9"], "keywords": ["edge characterization", "Jetson Orin/Thor", "memory-bound action generation", "PIM", "Jetson Orin", "Jetson Thor", "memory-bound", "processing-in-memory", "HBM", "100B scaling", "VLA characterization"], "model_class": "VLA", "techniques": ["profiling", "analytical modeling", "workload characterization", "PIM exploration", "hardware projection", "analytical hardware projection"], "base_model": "MolmoAct-7B", "params": "7B", "hardware": "NVIDIA Jetson Orin, Jetson Thor", "reported_metric": "action generation up to 75% of end-to-end latency | up to 75% of end-to-end latency in memory-bound action generation", "metric_measures": "other", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Profiles MolmoAct-7B on Jetson Orin and Thor, finds action generation is memory-bandwidth-bound, and projects edge hardware needs (HBM, processing-in-memory) for scaling VLAs to 100B parameters.", "relevance": 5, "relevance_reason": "Direct memory-bandwidth bottleneck analysis and architecture projection for edge VLA hardware.", "foundational_technique": false, "verified": true, "found_via": "search query: Jetson VLA; arXiv: vision-language-action AND Jetson/edge/onboard; search query; venue sweep: arXiv API VLA AND Jetson"}
{"id": "2609.07838", "title": "ComVLA: Communication-Aware Split Inference for VLA Models in 6G-Connected Robotics", "authors": ["Boliang Liu", "Wint Yi Poe", "Jingyun Di", "Riccardo Trivisonno", "Giuseppe Caire"], "year": 2026, "venue": "IEEE GLOBECOM 2026", "type": "preprint", "url": "https://arxiv.org/abs/2609.07838", "pdf_url": "https://arxiv.org/pdf/2609.07838", "code_url": "", "clusters": ["C1", "C2", "C4", "C5", "C6", "C7", "C9"], "keywords": ["split inference", "6G", "wireless", "token budget", "wireless uplink", "cloud offload", "channel-adaptive token budget", "language-guided token selection", "communication-aware"], "model_class": "VLA", "techniques": ["language-guided token pruning", "split inference", "channel-adaptive language-guided token selection", "split computing", "visual token pruning", "channel-aware adaptation", "language-guided token selection", "communication-aware offloading"], "base_model": "OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "74% less compute, 22% lower latency, 95.4% vs 96.9% success | 32 vs 512 tokens: -74% inference compute, -22% latency, 95.4% vs 96.9% LIBERO success", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Uses the language instruction to choose which visual tokens to send over a wireless link to a cloud VLA, matching the token budget to instantaneous channel capacity.", "relevance": 5, "relevance_reason": "Bandwidth-aware split inference co-designed with the wireless channel for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA latency; arXiv: offloading AND (VLA OR robot foundation model); search query; arXiv search: edge cloud offloading vision-language-action latency; venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2608.13438", "title": "ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models", "authors": ["Ge-Han Zheng", "Matthew Johnson-Roberson", "Wei-Ming Zhi", "Gehan Zheng", "Weiming Zhi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.13438", "pdf_url": "https://arxiv.org/pdf/2608.13438", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["ContactGuard", "pre-contact monitoring", "abort signal", "latent world model", "abort"], "model_class": "world model", "techniques": ["action-conditioned latent prediction", "failure probe", "latent multi-view embedding prediction under planned chunk", "lightweight failure probe"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real contact-rich manipulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Predicts the latent visual consequence of the planned action chunk before contact and aborts if predicted post-contact state looks like failure, without modifying the policy.", "relevance": 5, "relevance_reason": "Cheap latent imagination for pre-execution safety.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2); search query"}
{"id": "2604.11572", "title": "DA-PTQ: Drift-Aware Post-Training Quantization for Efficient Vision-Language-Action Models", "authors": ["Siyuan Xu", "Tian-Shi Wang", "Fengling Li", "Lei Zhu", "Heng Shen", "Tianshi Wang", "Heng Tao Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.11572", "pdf_url": "https://arxiv.org/pdf/2604.11572", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["kinematic drift", "error accumulation", "mixed precision", "drift-aware PTQ"], "model_class": "VLA", "techniques": ["post-training quantization", "mixed-precision quantization", "drift-aware post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Frames PTQ as minimizing trajectory drift over sequential control, compensating VL-to-action representation distortion and allocating bits by motion error.", "relevance": 5, "relevance_reason": "Addresses closed-loop error accumulation from quantization.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2602.01780", "title": "DDP-WM: Disentangled Dynamics Prediction for Efficient World Models", "authors": ["Shicheng Yin", "Kaixuan Yin", "Weixing Chen", "Yang Liu", "Guanbin Li", "Liang Lin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.01780", "pdf_url": "https://arxiv.org/pdf/2602.01780", "code_url": "", "clusters": ["C8"], "keywords": ["disentangled dynamics", "sparse primary dynamics", "efficient latent WM"], "model_class": "world model", "techniques": ["decompose foreground interaction vs background updates", "cross-attention background update", "MPC planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~9x inference speedup on Push-T; MPC success 90% -> 98%", "metric_measures": "relative latency", "benchmarks": ["Push-T", "navigation", "deformable/multi-body tasks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Splits latent dynamics into sparse interaction-driven tokens and cheap background updates, giving ~9x faster rollouts and better MPC success than dense DINO-WM-style models.", "relevance": 5, "relevance_reason": "Efficient latent world model for MPC with speedup numbers.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.29382", "title": "Decoupled Early Exits for Task-Dependent Compute Allocation in Flow-Matching VLAs", "authors": ["Riccardo Andrea Izzo", "Rimvydas Rubavicius", "Gianluca Bardaro", "Subramanian Ramamoorthy", "Matteo Matteucci", "Alessandro Suglia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.29382", "pdf_url": "https://arxiv.org/pdf/2609.29382", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C9"], "keywords": ["early exit", "action expert depth", "denoising steps", "compute allocation", "KV cache synthesis", "compute axes", "decoupled early exits", "backbone depth", "action-expert depth"], "model_class": "VLA", "techniques": ["early exit", "few-step sampling", "KV synthesis", "distillation to exits", "denoising step reduction", "KV cache synthesis", "exit transformers", "KV handling", "joint compute axes"], "base_model": "SmolVLA, pi0.5", "params": "", "hardware": "", "reported_metric": "79.2% lower latency, 31.8% fewer FLOPs", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "Meta-World"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds distilled exit heads to both the VLM backbone and action expert and treats backbone depth, expert depth, and denoising steps as jointly tunable compute axes.", "relevance": 5, "relevance_reason": "Joint configuration of multiple efficiency axes (composability).", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache; search: layer skipping VLA; search query: VLA latency; venue sweep: arXiv API VLA AND (MoE OR early exit)"}
{"id": "2609.18374", "title": "Decoupling Vision, Language, and Action for Efficient Multi-Task Robot Policies", "authors": ["Xiatao Sun", "Chen Liang", "Ziyao Zeng", "Qian Wang", "Haoyang Zhang", "Yue Sun", "Qiucheng Li", "Daniel Rakita"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18374", "pdf_url": "https://arxiv.org/pdf/2609.18374", "code_url": "", "clusters": ["C1"], "keywords": ["decoupled encoders", "MeanFlow head", "no VLM"], "model_class": "VLA", "techniques": ["LLM-free architecture", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "6.1 ms per forward pass; 162.7 policy calls/s", "metric_measures": "throughput", "benchmarks": ["RoboCasa", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Separate DINOv3 vision and NeoBERT language encoders feeding a one-step MeanFlow head nearly match GR00T N1.7 and pi0.5 at a fraction of latency.", "relevance": 5, "relevance_reason": "Challenges need for VLM backbones.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk latency VLA"}
{"id": "2605.19294", "title": "DEFLECT: Delay-Robust Execution via Flow-matching Likelihood-Estimated Counterfactual Tuning for VLA Policies", "authors": ["Yixiang Zhu", "Yonghao Chen", "Zijie Yang", "Yusong Hu", "Xinyu Chen", "Zhu, Yixiang", "Chen, Yonghao", "Yang, Zijie", "Hu, Yusong", "Chen, Xinyu", "Yi-Xiang Zhu", "Rui Meng", "Jingyu Guo", "Jiaxiang Zou", "Zi-Jie Yang", "Tao-Wen Wang", "Xin-Yu Chen"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://doi.org/10.48550/arXiv.2605.19294", "pdf_url": "https://arxiv.org/pdf/2605.19294", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C6", "C7"], "keywords": ["delay-robust", "counterfactual preference", "asynchronous VLA", "DEFLECT", "asynchronous inference", "delay robustness", "delay envelope", "preference learning", "counterfactual preference learning", "stale observation", "flow matching"], "model_class": "VLA", "techniques": ["preference post-training", "asynchronous inference", "offline preference post-training", "counterfactual preference post-training", "delay robustness", "counterfactual tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to +6.4 pp high-latency success | up to +6.4 pts success at high latency", "metric_measures": "other", "benchmarks": ["Kinetix", "LIBERO", "real robot", "simulation", "3 real tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Post-trains a VLA with automatically generated preferences (chunk from execution-time observation preferred over chunk from stale observation) so it learns to act on stale inputs without extra inference cost.", "relevance": 5, "relevance_reason": "Delay-robust async inference; enabling technique for high-latency (remote) serving.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; arXiv API: abs:\"post-training\" AND abs:\"vision-language-action\" AND abs:\"compute\"; WebSearch: simulated inference latency injection benchmark; search query; arXiv search: \"action chunk\" asynchronous inference; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.28811", "title": "DeltaWAM: Delta World Action Models for Bimanual Manipulation", "authors": ["Han Yan", "Zishang Xiang", "Haokai Jiang", "Zeyu Zhang", "Qilin Wang", "Weiyu Guo", "Yandong Guo", "Boxin Shi", "Hao Tang", "Zi-Shang Xiang", "Hao-Kai Jiang", "Ze-Yu Zhang", "Qi-Lin Wang", "Wei-Yu Guo", "Yan-Dong Guo", "Bo-Xin Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28811", "pdf_url": "https://arxiv.org/pdf/2609.28811", "code_url": "https://github.com/AIGeeksGroup/DeltaWAM", "clusters": ["C1", "C8"], "keywords": ["DeltaWAM", "efficient WAM", "bimanual"], "model_class": "world-action model", "techniques": ["delta prediction", "streaming cache", "streaming cached memory"], "base_model": "Fast-WAM", "params": "", "hardware": "", "reported_metric": "SDM cuts one-step inference latency 36.57% and FLOPs 31.55%", "metric_measures": "relative latency", "benchmarks": ["RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "World-action model predicting visual deltas plus actions and a streaming delta memory that updates cached anchor context, reducing heavy video-expert computation.", "relevance": 5, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\"); snowball round 2: cites 2106.09685, 2304.13705"}
{"id": "2608.04428", "title": "Deltoris: Enabling Real-time VLA Inference in Embodied AI via Bit-level Sparsity and Speculative Inference", "authors": ["Zheng Liu", "Zeyu Guo", "Zihan Liu", "Anbang Wu", "Han Zhao", "Fangxin Liu", "Zhezhi He", "Yinhe Han", "Jingwen Leng", "Minyi Guo", "Yiming Gan", "Yu Feng", "An-Bang Wu", "Min-Yi Guo", "Yi-Ming Gan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.04428", "pdf_url": "https://arxiv.org/pdf/2608.04428", "code_url": "", "clusters": ["C1", "C3", "C4", "C5", "C9"], "keywords": ["accelerator", "bit-level sparsity", "temporal delta", "Deltoris", "speculative inference", "diffusion VLA", "algorithm-hardware co-design", "bit-serial systolic array"], "model_class": "VLA", "techniques": ["hardware co-design", "delta computation", "speculative inference", "temporal bit sparsity", "systolic accelerator", "temporal bit-level sparsity", "custom accelerator design", "bit-level sparsity", "accelerator co-design"], "base_model": "", "params": "", "hardware": "custom accelerator (vs mobile GPU)", "reported_metric": "up to 34.2x over mobile GPUs; 6.1x over prior accelerators | up to 34.2x speedup over mobile GPUs and 6.1x over prior accelerators", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Algorithm-hardware co-design for diffusion VLAs: computes only bit-level deltas between consecutive inputs, amortizes weight loading speculatively across control steps, and implements a bit-serial systolic accelerator.", "relevance": 5, "relevance_reason": "Hardware co-design for diffusion VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search query: embodied inference energy; web: VLA accelerator FPGA ASIC design; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2604.10170", "title": "Device-Conditioned Neural Architecture Search for Efficient Robotic Manipulation", "authors": ["Yiming Wu", "Huan Wang", "Zhenghao Chen", "Ge Yuan", "Dong Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.10170", "pdf_url": "https://arxiv.org/pdf/2604.10170", "code_url": "", "clusters": ["C2", "C5", "C9"], "keywords": ["NAS", "device-conditioned QAT", "supernet", "mixed precision", "device-conditioned", "visuomotor policy"], "model_class": "diffusion/flow policy", "techniques": ["quantization-aware training", "neural architecture search", "mixed precision"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains one supernet over architectures and mixed-precision bit-widths with device lookup tables so a quantized visuomotor policy can be extracted per robot device without re-optimization.", "relevance": 5, "relevance_reason": "Hardware-conditioned quantization for robot policies.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2505.15304; venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2609.15570", "title": "DIDO: Distilling Interaction-Centric Dynamics into One-Step Denoising for World Action Models", "authors": ["Jing Lyu", "Shuanghao Bai", "Runze Xiao", "Zhenyu Liao", "Wenxing Tan", "Zihan Tang", "Ruochuan Shi", "Cheng Peng", "Yuheng Ji", "Yihao Wang", "Badong Chen", "Pengwei Wang", "Zhongyuan Wang", "Xiaoguang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.15570", "pdf_url": "https://arxiv.org/pdf/2609.15570", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["one-step video denoising", "distribution matching distillation", "interaction-centric", "one-step denoising", "DMD", "interaction-centric distillation"], "model_class": "world-action model", "techniques": ["step distillation", "distribution matching distillation", "interaction-centric representation guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Observes that one-step truncation of a video WAM keeps background but loses gripper-object dynamics, and distills converged interaction dynamics into a single denoising step.", "relevance": 5, "relevance_reason": "One-step distillation targeted at manipulation-relevant dynamics.", "foundational_technique": false, "verified": true, "found_via": "search: compression world model; snowball from 2602.15922 (citations)"}
{"id": "2026.00188", "title": "DiTPA: A DiT-Based Action Planner Accelerator Exploiting Action-Denoising-Multimodality Redundancy for Embodied Artificial Intelligence", "authors": ["Xin Zhao", "Longke Yan", "Jiancong Li", "Yongkun Wu", "Feng-Bin Tu"], "year": 2026, "venue": "ISCA 2026", "type": "paper", "url": "https://doi.org/10.1109/ISCA66397.2026.00188", "pdf_url": "https://arxiv.org/pdf/DOI:10.1109/ISCA66397.2026.00188", "code_url": "", "clusters": ["C5"], "keywords": ["DiT action planner", "accelerator", "denoising redundancy", "feature reuse"], "model_class": "diffusion/flow policy", "techniques": ["action reuse", "alternating denoising with feature reuse", "multimodality redundancy removal", "custom accelerator"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Software-hardware co-designed accelerator for diffusion-transformer action heads that exploits redundancy across consecutive actions, denoising steps and modalities.", "relevance": 5, "relevance_reason": "ISCA accelerator targeting VLA action experts.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2407.04292 (citations)"}
{"id": "2605.11750", "title": "DreamAvoid: Critical-Phase Test-Time Dreaming to Avoid Failures in VLA Policies", "authors": ["Xianzhe Fan", "Yuxiang Lu", "Shenyuan Gao", "Xiaoyang Wu", "Rui-Hua Han", "Manling Li", "Heng-Shuang Zhao", "Ruihua Han", "Hengshuang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.11750", "pdf_url": "https://arxiv.org/pdf/2605.11750", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["DreamAvoid", "test-time dreaming", "critical phase", "failure avoidance"], "model_class": "world model", "techniques": ["world-model imagination at critical phases", "dream trigger for critical phases", "candidate chunk sampling + dream evaluator"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Invokes world-model 'dreaming' only in detected critical phases to evaluate several VLA action candidates and avoid failures, limiting imagination compute to where it matters.", "relevance": 5, "relevance_reason": "Selective imagination-based safety for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2); snowball from 2602.15922 (citations)"}
{"id": "2602.06949", "title": "DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos", "authors": ["Shenyuan Gao", "William Liang", "Kaiyuan Zheng", "Ayaan Malik", "Seonghyeon Ye", "Sihyun Yu", "Wei-Cheng Tseng", "Yuzhu Dong", "Kaichun Mo", "Chen-Hsuan Lin", "Qianli Ma", "Seungjun Nah", "Loic Magne", "Jiannan Xiang", "Yuqi Xie", "Ruijie Zheng", "Dantong Niu", "You Liang Tan", "K. R. Zentner", "George Kurian", "Suneel Indupuru", "Pooya Jannaty", "Jinwei Gu", "Jun Zhang", "Jitendra Malik", "Pieter Abbeel", "Ming-Yu Liu", "Yuke Zhu", "Joel Jang", "Linxi \"Jim\" Fan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.06949", "pdf_url": "https://arxiv.org/pdf/2602.06949", "code_url": "", "clusters": ["C8"], "keywords": ["robot world model", "latent actions", "human video pretraining", "real-time distillation"], "model_class": "world model", "techniques": ["continuous latent actions", "autoregressive distillation", "few-step generation"], "base_model": "Cosmos-Predict2.5", "params": "", "hardware": "", "reported_metric": "10.81 FPS real-time rollout after distillation", "metric_measures": "throughput", "benchmarks": ["OOD robot benchmarks", "live teleoperation", "policy evaluation", "model-based planning"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "NVIDIA foundation robot world model pretrained on 44k hours of egocentric human video with continuous latent proxy actions; a distillation pipeline brings it to ~10.8 FPS for live teleoperation, policy evaluation and model-based planning.", "relevance": 5, "relevance_reason": "Seed; explicit real-time distillation of a large action-conditioned robot world model.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2604.03540", "title": "Drift-Based Policy Optimization: Native One-Step Policy Learning for Online Robot Control", "authors": ["Yuxuan Gao", "Yedong Shen", "Shi-Qi Zhang", "Wenhao Yu", "Yifan Duan", "Jia Pan", "Jia-Jia Wu", "Jia-Jun Deng", "Yanyong Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.03540", "pdf_url": "https://arxiv.org/pdf/2604.03540", "code_url": "https://github.com/YuxuanGao0822/DBPO", "clusters": ["C1", "C3"], "keywords": ["one-step policy", "NFE reduction"], "model_class": "diffusion/flow policy", "techniques": ["one-step generation", "drifting objective", "on-policy RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "100 NFEs -> 1 NFE; 9.5 ms average end-to-end latency on dual-arm UR5", "metric_measures": "other", "benchmarks": ["Diffusion Policy suite", "RoboMimic", "D4RL", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Drift-Based Policy moves iterative refinement into training so the policy produces multimodal action chunks in one network evaluation, and DBPO adds a stochastic interface for PPO-style fine-tuning while keeping 1-NFE deployment.", "relevance": 5, "relevance_reason": "direct: native one-step generative policy for high-frequency control", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2603.19199, 2608.29749"}
{"id": "2607.15065", "title": "DriftWorld: Fast World Modeling through Drifting", "authors": ["Susie Lu", "Haonan Chen", "Weirui Ye", "Yilun Du"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.15065", "pdf_url": "https://arxiv.org/pdf/2607.15065", "code_url": "https://susie-lu.github.io/driftworld/", "clusters": ["C8"], "keywords": ["drifting models", "one-pass world model", "planning by rollout search"], "model_class": "world model", "techniques": ["drifting generative model (single forward pass)", "action-conditioned drift", "rollout-based action selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "30+ fps single-pass generation; 17x faster on average than diffusion world models", "metric_measures": "throughput", "benchmarks": ["Bridge-V2", "RT-1", "Language Table", "Push-T", "Robomimic"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Action-conditioned world model built on drifting generative models that produces future frames in one forward pass, making it cheap to score many candidate action sequences for planning and policy ranking (r up to 0.99).", "relevance": 5, "relevance_reason": "Directly targets rollout speed as the planning bottleneck.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.27755", "title": "Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?", "authors": ["Guoheng Sun", "Kai Feng", "Shwai He", "Xiaochuan Gong", "Ye-Xiao He", "Ziyao Wang", "Zhe-Yu Shen", "Wanghao Ye", "R. Kompella", "Gaowen Liu", "Ang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.27755", "pdf_url": "https://arxiv.org/pdf/2606.27755", "code_url": "", "clusters": ["C2"], "keywords": ["block removal", "GateProbe", "redundancy", "language backbone"], "model_class": "VLA", "techniques": ["layer/block pruning", "fine-tuning recovery"], "base_model": "OpenVLA-OFT and others", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Block-removal study finding LLM backbones in VLAs highly redundant (two blocks can suffice) while vision and action pathways are fragile; proposes a one-shot gate sensitivity metric.", "relevance": 5, "relevance_reason": "Central redundancy analysis for VLA compression.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.07530"}
{"id": "dyna-2-world-action-model", "title": "Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models", "authors": ["Dyna Robotics"], "year": 2026, "venue": "blog (Dyna Robotics)", "type": "blog", "url": "https://www.dyna.co/dyna-2", "pdf_url": "", "code_url": "", "clusters": ["C3", "C7", "C8", "C9"], "keywords": ["world action model", "human video scaling law", "one-step video distillation", "world-action model", "video diffusion", "mixture of transformers", "flow matching", "one-step video generation"], "model_class": "world-action model", "techniques": ["mixture-of-transformers with shallow action stream", "flow matching", "one-step video distillation", "reactive policy (no future video at inference)", "mixture of transformers", "one-step generation"], "base_model": "", "params": "", "hardware": "single H100", "reported_metric": "one-step distilled video generation 10,203 ms -> 110 ms (~90x) for 3 s three-view video", "metric_measures": "relative latency", "benchmarks": ["14 internal robot tasks", "39-task zero-shot suite incl. xdof ABC", "14 robot tasks (post-training)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Joint video-action flow-matching world-action model (separate video/action DiT streams) pretrained on 1M hours of egocentric human video; shows human-to-robot transfer scaling (20%→53% normalized performance) and a one-step video generation mode ~90x faster.", "relevance": 5, "relevance_reason": "Industrial world-action model with explicit step-reduction speedup; from speaker Jason Ma's company.", "foundational_technique": false, "verified": true, "found_via": "seed; blog sweep: Dyna"}
{"id": "10.1109/lra.2026.3683324", "title": "Dynamic Maclaurin-Series-Based Vision-Language-Action Model", "authors": ["Chong Yu", "Zhongxue Gan"], "year": 2026, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://doi.org/10.1109/LRA.2026.3683324", "pdf_url": "", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["DMS-VLA", "Maclaurin series"], "model_class": "VLA", "techniques": ["weight sharing", "low-rank/series approximation", "adaptive compute"], "base_model": "", "params": "", "hardware": "H100, RTX 4090, Jetson AGX Orin", "reported_metric": "25%-55% inference speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SimplerEnv", "ALOHA"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Replaces per-layer transformer weights with shared Maclaurin-series templates and per-layer coefficients, cutting memory and adapting the number of terms per input.", "relevance": 5, "relevance_reason": "Direct VLA compression with edge-hardware speedups.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2603.07904", "title": "DyQ-VLA: Temporal-Dynamic-Aware Quantization for Embodied Vision-Language-Action Models", "authors": ["Zi-Hao Zheng", "Hangyu Cao", "Sicheng Tian", "Jia-Yu Chen", "Mao-Liang Li", "Xinhao Sun", "H. Zou", "Zhao-Bo Zhang", "Xuanzhe Liu", "Dong-Gang Cao", "Hong Mei", "Xiang Chen", "Zihao Zheng", "Jiayu Chen", "Maoliang Li", "Hailong Zou", "Zhaobo Zhang", "Donggang Cao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.07904", "pdf_url": "https://arxiv.org/pdf/2603.07904", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["dynamic quantization", "kinematic proxies", "bit switching", "bit-width switching", "kinematic proxy", "temporal sensitivity"], "model_class": "VLA", "techniques": ["dynamic mixed-precision quantization", "temporal-dynamic mixed precision"], "base_model": "", "params": "", "hardware": "", "reported_metric": "30.9% memory, 99.5% performance, 1.49x sim / up to 1.43x real-world speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Switches bit-widths at runtime based on kinematic proxies of how error-sensitive the current motion phase is, instead of a fixed precision.", "relevance": 5, "relevance_reason": "Edge-deployment-oriented quantization with measured speedups; primary cluster C1.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2602.22896", "title": "DySL-VLA: Efficient Vision-Language-Action Model Inference via Dynamic-Static Layer-Skipping for Robot Manipulation", "authors": ["Zebin Yang", "Yijiahao Qi", "Tong Xie", "Bo Yu", "Shaoshan Liu", "Meng Li"], "year": 2026, "venue": "DAC 2026", "type": "preprint", "url": "https://arxiv.org/abs/2602.22896", "pdf_url": "https://arxiv.org/pdf/2602.22896", "code_url": "https://github.com/PKU-SEC-Lab/DYSL_VLA", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["dynamic layer skipping", "action importance", "dynamic-static layer skipping", "two-stage KD", "layer skipping", "dynamic-static"], "model_class": "VLA", "techniques": ["dynamic layer skipping", "distillation", "layer skipping", "knowledge distillation", "dynamic/static layer skipping"], "base_model": "RoboFlamingo", "params": "", "hardware": "", "reported_metric": "3.75x speedup vs RoboFlamingo at iso-accuracy; 85.7x fewer trainable params", "metric_measures": "relative latency", "benchmarks": ["CALVIN"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Always runs 'informative' layers and skips 'incremental' ones depending on how critical each action is, trained via skip-aware distillation.", "relevance": 5, "relevance_reason": "Layer-skipping VLA with distillation.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search: layer skipping VLA; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2608.15502", "title": "EcoVLA: Energy-Efficient Device-Edge Co-Inference for Vision-Language-Action Models under Real-Time Constraints", "authors": ["Ao Zhou", "Bo Dai", "Le Yu", "Xingyu Liu", "Zeyu Hao", "Lingkun Long", "Chunming Hu", "Jianlei Yang", "Jian-Lei Yang"], "year": 2026, "venue": "Advanced Parallel Programming Technologies", "type": "preprint", "url": "https://arxiv.org/abs/2608.15502", "pdf_url": "https://arxiv.org/pdf/2608.15502", "code_url": "", "clusters": ["C1", "C3", "C4", "C5", "C6", "C9"], "keywords": ["device-edge co-inference", "energy", "EcoVLA", "energy efficiency", "real-time constraint", "SLO", "partitioning", "real-time constraints"], "model_class": "VLA", "techniques": ["edge-cloud split", "device-edge partitioning", "adaptive offloading", "model partitioning", "latency/energy prediction", "tensor transmission compression", "split computing", "tensor compression", "device-edge split inference", "energy-aware scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 236% higher energy efficiency at 20 Hz action output | up to 236% energy-efficiency improvement under 20 Hz constraint", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Architecture-agnostic device-edge co-inference for VLAs that picks the most energy-efficient split scheme meeting a real-time deadline and adapts to network and load changes.", "relevance": 5, "relevance_reason": "Direct device-edge VLA co-inference with energy and deadline objectives.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search query: embodied inference energy; arXiv: vision-language-action AND Jetson/edge/onboard; search query; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "eeai-corl2026-workshop-cfp", "title": "Efficient Foundation Models for Real-Time Embodied AI (CoRL 2026 Workshop) — Call for Papers and Challenge", "authors": ["Jiafan Yu", "Ouais Alsharif", "Changliu Liu", "Dhruv Shah", "Jiajun Wu", "Zhuoyang Zhang"], "year": 2026, "venue": "CoRL 2026 Workshop", "type": "talk", "url": "https://efficient-embodied-ai.github.io/", "pdf_url": "", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C6", "C7", "C8", "C9"], "keywords": ["workshop", "call for papers", "efficiency challenge", "action chunking", "mixture-of-experts", "early exit", "quantization", "edge-cloud", "safety-latency tradeoff"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Workshop site (Nov 12, 2026, Austin) with CFP covering efficient VLA/world-action architectures, compression, memory-efficient fine-tuning, composability benchmarking, hardware-aware design (GPU/TPU/Trainium/Inferentia/NPU), edge-cloud fleets, latency-safety fallbacks and efficient world models; also runs a challenge with a reference container (released Sep 25, submissions Oct 23). Papers: 4 pages, non-archival, via OpenReview, due Oct 12.", "relevance": 5, "relevance_reason": "Defines the scope of the survey itself.", "foundational_technique": false, "verified": true, "found_via": "workshop site fetch"}
{"id": "2609.12075", "title": "Efficient Vision-Language-Action Management and Serving for Robot Factories", "authors": ["Dionysios Adamopoulos", "Nattapol Chanpaisit", "Basel Fakhri", "Christina Giannoula"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.12075", "pdf_url": "https://arxiv.org/pdf/2609.12075", "code_url": "", "clusters": ["C4", "C5", "C6", "C7"], "keywords": ["VLA serving", "SLO", "multi-robot", "edge server", "SM partitioning", "GPU sharing", "Robion"], "model_class": "VLA", "techniques": ["stage disaggregation", "SM partitioning", "batching", "SLO scheduling", "intra-GPU stage disaggregation", "SM restriction", "SLO-aware scheduling", "stage disaggregation within GPU", "SLO-aware batching", "model placement"], "base_model": "", "params": "", "hardware": "multi-GPU edge server (4 GPUs)", "reported_metric": "6.7x higher robot load vs vLLM-Omni at 98% SLO; 64 robots on 4 GPUs with 8 models | up to 64 robots served on 4-GPU server within 98% SLO attainment; 6.7x higher robot load than vLLM-Omni", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Robion serves many robots' VLA requests on shared edge GPUs by co-scheduling VLM and action-diffusion stages on separate streams and placing models to meet per-request latency SLOs.", "relevance": 5, "relevance_reason": "SLO/tail-latency framing of VLA inference; throughput vs latency trade-off.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA latency; arXiv: vision-language-action AND Jetson/edge/onboard; search query; arXiv search: \"latency\" \"safety\" \"vision-language-action\""}
{"id": "2609.34608", "title": "Efficient World Action Model Inference with Adaptive Intermediate States", "authors": ["Zhinnan Liu", "Haozhi Han", "Ruge Zhang", "Teng Ma", "Tao Ma", "Zheng Liu", "Yifeng Chen", "Yunquan Zhang", "Ting Cao", "Yunxin Liu", "Kun Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34608", "pdf_url": "https://arxiv.org/pdf/2609.34608", "code_url": "", "clusters": ["C2", "C3", "C7", "C8", "C9"], "keywords": ["WAM inference acceleration", "state reuse", "training-free", "adaptive intermediate states", "denoising reuse"], "model_class": "world-action model", "techniques": ["trajectory remapping across replans", "anticipatory inference with observation rebinding", "selective layer-state reuse", "reuse of prior inference states"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.47-3.05x observation-to-action latency speedup; 2.23-3.27x GPU time per replan; 96.7-99.5% of native success retained", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboTwin 2.0"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Training-free inference engine (WAMachine) that reuses and adapts denoising state across replans, denoising steps, and transformer layers, with consistency checks deciding when stale state must be recomputed.", "relevance": 5, "relevance_reason": "Generic training-free acceleration layer for WAM deployment.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2606.10040", "title": "Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination", "authors": ["Jiajun Li", "Tiecheng Guo", "Yifan Ye", "Rongyu Zhang", "Xiaowei Chi", "Qianpu Sun", "Ying Li", "Yunfan Lou", "Jinqiao Wang", "Zhihe Lu", "Meng Guo", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.10040", "pdf_url": "https://arxiv.org/pdf/2606.10040", "code_url": "", "clusters": ["C1", "C3", "C7", "C8", "C9"], "keywords": ["compact WAM", "token-sparse video latents", "asymmetric denoising", "1B WAM", "compact video expert"], "model_class": "world-action model", "techniques": ["compact video expert transfer", "token sparsity", "fewer video than action denoising steps", "token-sparse video latents", "fewer video steps than action steps"], "base_model": "WAN-2.2-5B (transferred)", "params": "~1B", "hardware": "", "reported_metric": "98 ms per action chunk in physical deployment; >30x faster than Motus", "metric_measures": "open-loop chunk rate", "benchmarks": ["RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "A 1B WAM that treats future video as coarse guidance rather than a photoreal target, using token-sparse latents and asymmetric denoising to reach ~98 ms per chunk with success comparable to much larger WAMs.", "relevance": 5, "relevance_reason": "Direct small-WAM efficiency design with deployment latency.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "embodied-efficiency-challenge-corl2026", "title": "Embodied Efficiency Challenge / embodied-efficiency-bench (CoRL 2026 Workshop: Efficient Foundation Models for Real-Time Embodied AI)", "authors": ["Workshop organizers"], "year": 2026, "venue": "CoRL 2026 Workshop", "type": "benchmark", "url": "https://efficient-embodied-ai.github.io/", "pdf_url": "", "code_url": "", "clusters": ["C1", "C4", "C5"], "keywords": ["compress-to-a-budget", "Jetson Orin NX", "p99 latency per action", "LIBERO", "reproducibility-checked leaderboard"], "model_class": "VLA", "techniques": ["quantization", "sharding", "offload"], "base_model": "OpenVLA-7B", "params": "7B", "hardware": "Jetson Orin NX 16GB class", "reported_metric": "", "metric_measures": "other", "benchmarks": ["LIBERO-Spatial", "LIBERO-Object", "LIBERO-Goal"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Challenge compressing OpenVLA-7B to fit Jetson Orin NX 16GB with a frozen Docker harness; ranked by success within 15 points of reference, then median and p99 latency per action, then params and peak memory, with organizer re-runs of top entries.", "relevance": 5, "relevance_reason": "The target workshop's own benchmarking protocol.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: embodied-efficiency-bench"}
{"id": "2603.16952", "title": "Embodied Foundation Models at the Edge: A Survey of Deployment Constraints and Mitigation Strategies", "authors": ["Utkarsh Grover", "Ravi Ranjan", "Ming-Yang Mao", "T. Dong", "S. Praveen", "Zhen Wu", "J. M. Chang", "Tinoosh Mohsenin", "Yiying Sheng", "Agoritsa Polyzou", "Eiman Kanjo", "Xiao-Min Lin", "Grover, Utkarsh", "Ranjan, Ravi", "Mao, Mingyang", "Dong, Trung Tien", "Praveen, Satvik", "Wu, Zhenqi", "Chang, J. Morris", "Mohsenin, Tinoosh", "Sheng, Yi", "Polyzou, Agoritsa", "Kanjo, Eiman", "Lin, Xiaomin", "Mingyang Mao", "Trung Tien Dong", "Satvik Praveen", "Zhenqi Wu", "J. Morris Chang", "Yi Sheng", "Xiaomin Lin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.16952", "pdf_url": "https://arxiv.org/pdf/2603.16952", "code_url": "", "clusters": ["C2", "C4", "C5", "C7", "C9"], "keywords": ["survey", "edge deployment", "memory bandwidth", "deployment gauntlet", "edge constraints", "timing variability", "SWaP"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey organizing embodied edge deployment into eight coupled barriers (memory traffic, latency, timing variability, safety margins...), noting AR VLAs are bandwidth-bound and diffusion controllers compute-bound.", "relevance": 5, "relevance_reason": "Frames what should be measured for deployability, including jitter.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2410.21257; WebSearch: control loop jitter timing variance VLA; arXiv: vision-language-action AND Jetson/edge/onboard; arXiv search: \"latency\" \"safety\" \"vision-language-action\""}
{"id": "2607.02501", "title": "Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots", "authors": ["Ling Xu", "Borui Li", "Hao Wu", "Chuyu Han", "Xiangyu Li", "Mohan Hua", "Shiqi Jiang", "Ting Cao", "Chuanyou Li", "Sheng Zhong", "Shuai Wang", "Shenghui Zhong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.02501", "pdf_url": "https://arxiv.org/pdf/2607.02501", "code_url": "https://github.com/SEU-PAISys/Embodied.cpp", "clusters": ["C1", "C2", "C3", "C4", "C5", "C8", "C9"], "keywords": ["C++ runtime", "multi-rate execution", "heterogeneous robots", "quantization configs", "Embodied.cpp", "batch-1 latency", "heterogeneous hardware", "world-action models", "inference runtime", "heterogeneous edge", "C++", "VLA and WAM"], "model_class": "other", "techniques": ["inference runtime", "quantized inference runtime", "kernel fusion", "fused inference", "multi-rate execution", "portable inference runtime", "fused batch-1 inference", "quantization", "portable C++ inference runtime", "latency-first fused inference", "portable runtime"], "base_model": "", "params": "", "hardware": "heterogeneous edge devices", "reported_metric": "1.05-2.70x speedup, 7-77% lower VRAM vs Python | 1.05-2.70x speedup and 7-77% lower VRAM vs Python baselines", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Portable C++ runtime for VLAs and WAMs organized in five layers, targeting the embodied runtime contract of multi-rate closed-loop execution and latency-first batch-1 inference.", "relevance": 5, "relevance_reason": "Deployment runtime that operationalizes quantized embodied models.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA; search: quantization vision-language-action; search query: VLA quantization; web: processing-in-memory OR NPU accelerator embodied AI; search query; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "10.1145/3797248.3815411", "title": "Energy and Carbon Footprint of Vision-Language-Action Model Inference for Edge Robotic Systems", "authors": ["Arshia Eslami", "Mahsa Ardakani", "Amin Roostaee", "H. Zanganeh", "Ramtin Zand"], "year": 2026, "venue": "ACM (2026)", "type": "paper", "url": "https://doi.org/10.1145/3797248.3815411", "pdf_url": "https://arxiv.org/pdf/DOI:10.1145/3797248.3815411", "code_url": "", "clusters": ["C5", "C9"], "keywords": ["energy per inference", "carbon footprint", "power", "Jetson AGX Orin"], "model_class": "VLA", "techniques": ["energy characterization"], "base_model": "pi0.5, X-VLA, SmolVLA", "params": "", "hardware": "Jetson AGX Orin", "reported_metric": "all three meet 10 Hz; only pi0.5 and SmolVLA meet 20 Hz", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Measures latency, power, energy per inference and estimated carbon for pi0.5, X-VLA and SmolVLA on Jetson AGX Orin, noting bursty power draw and an accuracy-energy trade-off.", "relevance": 5, "relevance_reason": "Direct power/energy characterization of VLAs on edge hardware.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2603.02271 (citations)"}
{"id": "2604.23995", "title": "Extracting Exact Lie Derivatives Without Backpropagation: A Dual Compiler for Neural Control Barrier Functions", "authors": ["Mohammadreza Kamaldar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.23995", "pdf_url": "https://arxiv.org/pdf/2604.23995", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["neural CBF", "WCET", "embedded code generation", "forward-mode dual numbers"], "model_class": "other", "techniques": ["dual-number compiler for Lie derivatives"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Compiler emitting allocation-free C++ that computes a neural CBF and its exact Lie derivatives in one forward pass with bounded workspace, meeting WCET/memory constraints of safety coding standards.", "relevance": 5, "relevance_reason": "Directly addresses WCET/timing certification of learned safety filters.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"worst-case execution time\" neural network"}
{"id": "2609.11445", "title": "FARM: Reading Failure Signals from the Internal Predictive States of a Frozen Robotic World Model", "authors": ["Hao-Ran Pei", "Ming-Rui Luo", "Sen-Bao Wang", "Hao-Ran Lv", "Jie Guo", "Sheng Zhong", "Ruixiang Ci", "Haoran Pei", "Mingrui Luo", "Senbao Wang", "Haoran Lv", "Ruixi Ci"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.11445", "pdf_url": "https://arxiv.org/pdf/2609.11445", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["FARM", "world model readout", "failure monitoring", "VLA-JEPA", "frozen world model readout"], "model_class": "world model", "techniques": ["tiny supervised readout on frozen predictive states"], "base_model": "", "params": "33,985-parameter readout", "hardware": "", "reported_metric": "85.68/88.59 pooled AUROC/AUPRC | 0.2256 ms mean added CUDA latency; pooled AUROC 85.7", "metric_measures": "relative latency", "benchmarks": ["10-task benchmark", "real PIPER X, SO-101, Franka"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows a frozen robotic world model's internal predictive states already encode failure; a 34k-parameter readout gives step-wise failure scores for ~0.2 ms extra latency.", "relevance": 5, "relevance_reason": "Near-free safety monitoring from world-model states.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2); search query"}
{"id": "2603.16666", "title": "Fast-WAM: Do World Action Models Need Test-time Future Imagination?", "authors": ["Tianyuan Yuan", "Zibin Dong", "Yicheng Liu", "Hang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.16666", "pdf_url": "https://arxiv.org/pdf/2603.16666", "code_url": "https://yuantianyuan01.github.io/FastWAM/", "clusters": ["C1", "C7", "C8", "C9"], "keywords": ["world action model", "skip test-time imagination", "video co-training", "test-time imagination"], "model_class": "world-action model", "techniques": ["single-pass video DiT encoder", "video co-training only at train time", "skip future prediction at inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "190 ms latency; >4x faster than imagine-then-execute WAMs", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows WAMs keep most of their benefit if video prediction is used only as a training signal: the video DiT runs once as an encoder at test time, cutting latency to 190 ms (4x faster) with competitive success.", "relevance": 5, "relevance_reason": "Key efficiency finding for WAMs: test-time future imagination is largely unnecessary.", "foundational_technique": false, "verified": true, "found_via": "snowball from NVIDIA WAM blog / search; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2609.37771", "title": "Faster and Better? Benchmark Bugs and Design Limitations Distort the Evaluation of Vision-Language-Action Acceleration", "authors": ["Qiwei Chen", "Kaijun Zhou", "Nuohui Shi", "Zhiyang Li", "Yuxuan Feng", "Jinyu Gu", "Qi-Wei Chen", "Kai-Jun Zhou", "Nuo-Hui Shi", "Zhi-Yang Li", "Jin-Yu Gu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.37771", "pdf_url": "https://arxiv.org/pdf/2609.37771", "code_url": "", "clusters": ["C1", "C4", "C5", "C9"], "keywords": ["VLA acceleration evaluation", "benchmark bugs", "success checker", "motion-aware score", "acceleration evaluation", "RoboTwin", "LIBERO-Plus"], "model_class": "VLA", "techniques": ["training-free acceleration", "token pruning", "caching", "benchmark audit"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin", "LIBERO-Plus", "VLABench", "LIBERO"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Audits why training-free VLA accelerators sometimes beat the uncompressed baseline: finds 22 bugs and 4 design limitations across seven benchmarks (permissive checkers, wrong masses), and fixing them can reverse method rankings.", "relevance": 5, "relevance_reason": "Directly about the validity of efficiency-method evaluation on sim benchmarks.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark; snowball from 2604.24447 (citations)"}
{"id": "2608.02365", "title": "Faster-WAM: Do World Action Models Need Deep Action Modules?", "authors": ["Liheng Ma", "Rui Heng Yang", "Zhanguang Zhang", "Mateo Clemente", "Ziwen Hu", "Tongtong Cao", "Yingxue Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.02365", "pdf_url": "https://arxiv.org/pdf/2608.02365", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["Dock of Transformer", "shallow action head", "WAM latency"], "model_class": "world-action model", "techniques": ["single-layer action head docked to video backbone", "multi-layer KV fusion", "RoPE realignment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "66.5 ms per inference; 3.2x faster than Fast-WAM", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboTwin 2.0", "LIBERO-Plus"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Replaces deep action towers in WAMs with a one-layer action head that reads keys/values from all layers of a 30-layer video transformer, giving the lowest latency in a controlled comparison without losing accuracy.", "relevance": 5, "relevance_reason": "Architectural latency reduction for WAMs with measured end-to-end timing.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.19199", "title": "FASTER: Rethinking Real-Time Flow VLAs", "authors": ["Yuxiang Lu", "Zhe Liu", "Xianzhe Fan", "Zhenya Yang", "Jinghua Hou", "Junyi Li", "Kaixin Ding", "Hengshuang Zhao", "Lu, Yuxiang", "Liu, Zhe", "Fan, Xianzhe", "Yang, Zhenya", "Hou, Jinghua", "Li, Junyi", "Ding, Kaixin", "Zhao, Hengshuang", "Zhen-Ya Yang", "J. Li", "Kai Ding", "Heng-Shuang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.19199", "pdf_url": "https://arxiv.org/pdf/2603.19199", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C6", "C7", "C9"], "keywords": ["time to first action", "horizon-aware schedule", "reaction latency", "FASTER", "reaction time", "execution horizon", "consumer GPU", "streaming", "flow VLA", "streaming client-server", "flow schedule"], "model_class": "VLA", "techniques": ["non-uniform denoising schedule", "streaming pipeline", "horizon-aware flow schedule", "streaming client-server", "streaming client-server pipeline", "asynchronous inference", "non-uniform flow schedule", "reaction-aware async"], "base_model": "pi0.5, X-VLA", "params": "", "hardware": "consumer-grade GPUs", "reported_metric": "immediate-reaction denoising compressed ~10x into a single step", "metric_measures": "relative latency", "benchmarks": ["real robot (table tennis)", "real robot (incl. table tennis)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Analyzes reaction latency of chunked flow VLAs as governed by time-to-first-action and execution horizon, then uses a horizon-aware sampling schedule that finishes near-term actions in one step while refining later ones, cutting effective reaction latency.", "relevance": 5, "relevance_reason": "Gives a principled reaction-latency metric (TTFA + horizon) distinct from inference latency.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; WebSearch: CALVIN ABC-D efficiency latency comparison; snowball from 2509.11480 (citations); search query; arXiv API all:\"action chunking\" AND latency; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2606.20246", "title": "Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think", "authors": ["Gia-Binh Nguyen", "Trong-Bao Ho", "Thien-Loc Ha", "Khoa Vo", "Philip Lund Møller", "Quang T. Nguyen", "Long Dinh", "Tung M. Luu", "Tuan Dam", "Vu Duong", "Trung Le", "Nghi D. Q. Bui", "Minh Vu", "Tran Nguyen Le", "An Thai Le", "Ngan Le", "Daniel Sonntag", "James Zou", "Jan Peters", "Duy M. H. Nguyen", "Ngo Anh Vien"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20246", "pdf_url": "https://arxiv.org/pdf/2606.20246", "code_url": "", "clusters": ["C1", "C2", "C3"], "keywords": ["layer redundancy", "CKA", "training-free depth pruning", "layer removal", "training-free depth compression", "training-free compression", "fine-tuning cost"], "model_class": "VLA", "techniques": ["layer pruning", "training-free compression"], "base_model": "pi0, GR00T N1.5", "params": "", "hardware": "", "reported_metric": "up to 50% depth reduction; up to 30% faster inference", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboCasa", "SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Finds redundant 'twin' layers in flow-matching VLAs via a single-pass CKA analysis and removes them before fine-tuning, speeding both training and inference.", "relevance": 5, "relevance_reason": "Depth pruning of continuous-action VLAs incl. action head.", "foundational_technique": false, "verified": true, "found_via": "search query: token reduction VLA; web search: VLA pruning layers redundancy; snowball from 2502.19645 (S2 citations)"}
{"id": "2606.05254", "title": "Flash-WAM: Modality-Aware Distillation for World Action Models", "authors": ["Arman Akbari", "Ci Zhang", "Arash Akbari", "Lin Zhao", "Yixiao Chen", "Weiwei Chen", "Xuan Zhang", "Geng Yuan", "Yanzhi Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.05254", "pdf_url": "https://arxiv.org/pdf/2606.05254", "code_url": "", "clusters": ["C2", "C3", "C7", "C8", "C9"], "keywords": ["consistency distillation", "one-step WAM", "modality-aware", "modality-aware step distillation", "WAM"], "model_class": "world-action model", "techniques": ["modality-aware consistency/step distillation", "single-step video and action", "per-modality consistency parametrization"], "base_model": "LingBot-VA", "params": "", "hardware": "NVIDIA L40S; Unitree G1 humanoid", "reported_metric": "per-chunk latency 8.1 s -> 348 ms (23x) on L40S", "metric_measures": "open-loop chunk rate", "benchmarks": ["RoboTwin 2.0", "LIBERO", "real robot (Unitree G1)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Step-distills a joint video-action WAM to one step per modality by choosing different consistency parametrizations for the low-noise action stream and high-noise video stream, giving 23x speedup while naive distillation collapses.", "relevance": 5, "relevance_reason": "Distillation specifically tailored to joint video-action world models.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2609.24433", "title": "FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding", "authors": ["Hung T. Ho", "Khanh-Binh Nguyen", "Quan Nguyen", "Thanh Duong", "N. Le", "Meng Guo", "V. Ngo", "A. T. Le", "Khanh D. Nguyen", "Quang D. Nguyen", "Thanh Q. Duong", "Ngan Le", "Vien A. Ngo", "An T. Le"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24433", "pdf_url": "https://arxiv.org/pdf/2609.24433", "code_url": "https://github.com/cair-vinuni/FoldQuantVLA", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["W4A4", "TensorRT plugins", "Hadamard", "Jetson AGX Orin", "Hadamard rotation", "native integer execution", "Hadamard transform", "low-bit PTQ"], "model_class": "VLA", "techniques": ["post-training quantization", "integer kernels", "Hadamard transform", "custom TensorRT plugins", "mixed precision", "channel scaling", "block Hadamard"], "base_model": "GR00T N1.5/N1.6/N1.7; pi0.5", "params": "", "hardware": "Jetson AGX Orin; Ada GPUs", "reported_metric": "1.20-1.33x speedup over FP TensorRT on Orin; 1.25-1.52x on desktop", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "PTQ with channel scaling, block Hadamard and dynamic per-token quantization executed natively as W4A4 TensorRT plugins on Orin; mixed W8A8 for sensitive projections boosts real-robot success.", "relevance": 5, "relevance_reason": "Native low-bit execution on robot-grade hardware with real-robot trials.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2606.23085", "title": "Foresight: Failure Detection for Long-Horizon Robotic Manipulation with Action-Conditioned World Model Latents", "authors": ["Haoran Zhang", "Yifu Lu", "Boyang Wang", "Xuhui Kang", "Yen-Ling Kuo", "Zezhou Cheng", "Mengdi Wang", "Odest Chadwicke Jenkins"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.23085", "pdf_url": "https://arxiv.org/pdf/2606.23085", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["Foresight", "world model latents", "long-horizon failure detection", "failure detection", "long-horizon", "functional conformal prediction"], "model_class": "world model", "techniques": ["action-conditioned world-model embeddings", "functional conformal prediction", "causal transformer monitor", "functional conformal thresholds"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Long", "ManiSkill-Long", "BEHAVIOR-1K", "real", "real ReactorX-200/Franka"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Policy-agnostic failure detector for long-horizon manipulation built on action-conditioned world-model latents, trained only on episode-level success labels and calibrated with functional conformal prediction.", "relevance": 5, "relevance_reason": "World-model-based runtime failure monitoring for VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action; search query"}
{"id": "2511.06385", "title": "From Demonstrations to Safe Deployment: Path-Consistent Safety Filtering for Diffusion Policies", "authors": ["Ralf Römer", "Julian Balletshofer", "Jakob Thumm", "Marco Pavone", "Angela P. Schoellig", "Matthias Althoff"], "year": 2026, "venue": "ICRA 2026", "type": "paper", "url": "https://arxiv.org/abs/2511.06385", "pdf_url": "https://arxiv.org/pdf/2511.06385", "code_url": "", "clusters": ["C7"], "keywords": ["PACS", "path-consistent braking", "set-based reachability", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["path-consistent braking", "reachability verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Safety filter for diffusion policies that brakes along the generated path (staying in-distribution) and verifies safety with set-based reachability in real time for human-robot interaction.", "relevance": 5, "relevance_reason": "Real-time formally verified filter tailored to chunked generative policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" \"diffusion policy\""}
{"id": "2603.19131", "title": "From Inference Efficiency to Embodied Efficiency: Revisiting Efficiency Metrics for Vision-Language-Action Models", "authors": ["Zhuofan Li", "Hongkun Yang", "Zhenyang Chen", "Yangxuan Chen", "Yingyan", "Lin", "Chaojian Li", "Zhuo Li", "Zhen-Yang Chen", "Yangxuanxin Chen", "Y. Lin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.19131", "pdf_url": "https://arxiv.org/pdf/2603.19131", "code_url": "", "clusters": ["C1", "C2", "C4", "C5", "C9"], "keywords": ["embodied efficiency", "metrics", "task completion time", "jerk", "motion energy"], "model_class": "VLA", "techniques": ["evaluation methodology", "evaluation of compression, token sparsification, action compression", "model compression", "token sparsification", "action sequence compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "simulation"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Argues FLOPs/parameters/token throughput misrepresent VLA efficiency on robots and proposes system-level metrics such as completion time, smoothness and motion energy; compression often worsens them.", "relevance": 5, "relevance_reason": "Central to C4: redefines efficiency metrics and shows techniques' hidden embodied costs.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; snowball from 2412.01034; search query: robot policy energy; snowball from 2510.26742 (citations)"}
{"id": "2609.06036", "title": "Generator-Independent Runtime Assurance under Partial Observation", "authors": ["Guangxi Wan", "Yongbo Xie", "Yuqi Liu", "Qingwei Dong", "Qingxin Li", "Hongfei Bai", "Peng Zeng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.06036", "pdf_url": "https://arxiv.org/pdf/2609.06036", "code_url": "", "clusters": ["C7"], "keywords": ["runtime verification gate", "generator-independent safety"], "model_class": "other", "techniques": ["runtime assurance", "setwise certification", "risk ledger"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proves that certifying a whole admissible set of proposals is necessary and sufficient for runtime-assurance safety that is independent of the generator (policy, LLM), shows per-candidate certification breaks under best-of-k retry, and bounds admission under partial observation.", "relevance": 5, "relevance_reason": "Theory directly relevant to gating sampled VLA chunks behind a verifier.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2608.23927", "title": "GlanceWAM: Sparse Test-Time Imagination for World-Action Models", "authors": ["Linhan Wang", "Zijian An", "Mingyuan Zhang", "Chen Dai", "Yi Xu", "Can Cui", "Zichong Yang", "Yinlin Chen", "Lifeng Zhou", "Chang-Tien Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.23927", "pdf_url": "https://arxiv.org/pdf/2608.23927", "code_url": "https://github.com/linhanwang/GlanceWAM", "clusters": ["C6", "C7", "C8", "C9"], "keywords": ["asynchronous imagination", "sparse lookahead", "latent-space action head", "sparse imagination", "asynchronous proposer", "latent action head"], "model_class": "world-action model", "techniques": ["background single-frame lookahead on slow clock", "non-interfering attention mask", "staleness-robust training", "async lookahead on slow clock"], "base_model": "", "params": "", "hardware": "NVIDIA A100", "reported_metric": "48 ms per action chunk; 24x faster than synchronous baselines", "metric_measures": "open-loop chunk rate", "benchmarks": ["RoboCasa", "LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Imagines one future frame seconds ahead on a slow background clock while a latent action head decodes chunks every 48 ms, beating both synchronous Cosmos Policy and imagination-free co-training.", "relevance": 5, "relevance_reason": "Takes imagination off the critical path of a WAM.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2602.18813", "title": "Habilis-$β$: A Fast-Motion and Long-Lasting On-Device Vision-Language-Action Model", "authors": ["Tommoro Robotics", ":", "Jesoon Kang", "Taegeon Park", "Jisu An", "Soo Min Kimm", "Jaejoon Kim", "Jinu Pahk", "Byungju Kim", "Junseok Lee", "Namheon Baek", "Sungwan Ha", "Hojun Baek", "Eduardo Ayerve Cruz", "Wontae Kim", "Junghyeon Choi", "Yousuk Lee", "Joonmo Han", "Sunghyun Cho", "Sunghyun Kwon", "Soyoung Lee", "Jun Ki Lee", "Seung-Joon Yi", "Byoung-Tak Zhang", "Theo Taeyeong Kim"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2602.18813", "pdf_url": "https://arxiv.org/pdf/2602.18813", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5"], "keywords": ["on-device VLA", "tasks per hour", "rectified flow distillation", "Productivity-Reliability Plane", "mean time between intervention", "continuous-run protocol", "on-device", "humanoid logistics"], "model_class": "VLA", "techniques": ["rectified-flow distillation", "motion shaping", "phase-adaptive motion shaping", "classifier-free guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "real humanoid logistics: 124 TPH, 137.4 s MTBI (vs 19 TPH for pi0.5)", "metric_measures": "other", "benchmarks": ["real robot", "RoboTwin 2.0"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Tommoro Robotics' on-device VLA uses flow distillation for high-frequency edge control and proposes productivity metrics (tasks/hour, time between interventions) for continuous operation.", "relevance": 5, "relevance_reason": "On-device deployment of a VLA on a humanoid; hardware details limited in abstract.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA; search query: VLA leaderboard; arXiv: vision-language-action AND Jetson/edge/onboard"}
{"id": "2602.10377", "title": "Hardware Co-Design Scaling Laws via Roofline Modelling for On-Device LLMs", "authors": ["Luoyang Sun", "Jiwen Jiang", "Yifeng Ding", "Fengfa Li", "Yan Song", "Haifeng Zhang", "Jian Ying", "Lei Ren", "Kun Zhan", "Wei Chen", "Yan Xie", "Cheng Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.10377", "pdf_url": "https://arxiv.org/pdf/2602.10377", "code_url": "", "clusters": ["C5", "C9"], "keywords": ["hardware co-design scaling law", "roofline", "Jetson Orin", "backbone selection for VLAs", "hardware co-design scaling laws", "on-device LLM backbone selection"], "model_class": "other", "techniques": ["roofline latency modeling", "architecture search", "scaling laws", "roofline modeling"], "base_model": "", "params": "", "hardware": "Jetson Orin", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Couples a loss scaling law over architecture hyperparameters with roofline latency on Jetson Orin (1,942 candidates, 170 trained) to find accuracy-latency Pareto LLM backbones for on-device VLAs.", "relevance": 5, "relevance_reason": "Explicit hardware-aware backbone design for on-device VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv: Jetson AND (robot policy OR manipulation OR VLA); venue sweep: arXiv API VLA AND Jetson"}
{"id": "2602.13710", "title": "HBVLA: Pushing 1-Bit Post-Training Quantization for Vision-Language-Action Models", "authors": ["Xin Yan", "Zheng-Lin Wan", "Feiyang Ye", "Xing-Rui Yu", "Hang Du", "Yang You", "Ivor W. Tsang", "Zhenglin Wan", "Xingrui Yu", "Hangyu Du", "Ivor Tsang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.13710", "pdf_url": "https://arxiv.org/pdf/2602.13710", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["binarization", "1-bit PTQ", "Hessian", "Haar transform", "Haar domain"], "model_class": "VLA", "techniques": ["1-bit post-training quantization", "Hessian saliency", "binarization"], "base_model": "OpenVLA-OFT; CogACT", "params": "", "hardware": "", "reported_metric": "92.2% of FP performance (OpenVLA-OFT, LIBERO); 93.6% (CogACT, SimplerEnv)", "metric_measures": "other", "benchmarks": ["LIBERO", "SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VLA-specific binarization: policy-aware Hessian picks salient weights, sparse orthogonal transform eases the rest, then group-wise 1-bit quantization in the Haar domain.", "relevance": 5, "relevance_reason": "Targets hardware-limited robots but is primarily a quantization method (C1).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.07530; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2605.28803", "title": "HoloQ-VLA: Uniform W4A4 Quantization of Vision-Language-Action Models", "authors": ["Xin-Yu Wang", "Ming-Ze Li", "Sicheng Lyu", "Dong-Xiu Liu", "Kaicheng Yang", "Zi-Yu Zhao", "Yu-Fei Cui", "Xiaoya Chang", "Peng Lu", "Xinyu Wang", "Mingze Li", "Dongxiu Liu", "Ziyu Zhao", "Yufei Cui", "Xiao-Wen Chang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.28803", "pdf_url": "https://arxiv.org/pdf/2605.28803", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["W4A4", "Hadamard rotation", "diffusion action head", "Hadamard", "diffusion action head quantization", "uniform quantization"], "model_class": "VLA", "techniques": ["post-training quantization", "rotation", "rotation + Hadamard PTQ", "per-step scaling", "W4A4 PTQ"], "base_model": "pi0.5; GR00T-N1.5", "params": "", "hardware": "", "reported_metric": "74.2% static memory reduction at W4A4", "metric_measures": "relative memory", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Training-free PTQ that quantizes both the language backbone and the diffusion action head of VLAs uniformly to W4A4 for on-device deployment.", "relevance": 5, "relevance_reason": "Quantization (C1) aimed at integer tensor-core hardware.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; arXiv: on-device AND (robot policy OR VLA OR diffusion policy); venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2602.18397", "title": "How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-Perf", "authors": ["Wenqi Jiang", "Jason Clemons", "Karu Sankaralingam", "Christos Kozyrakis", "Jiang, Wenqi", "Clemons, Jason", "Sankaralingam, Karu", "Kozyrakis, Christos", "K. Sankaralingam", "Christoforos E. Kozyrakis"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2602.18397", "pdf_url": "https://arxiv.org/pdf/2602.18397", "code_url": "", "clusters": ["C1", "C3", "C4", "C5", "C6", "C7", "C9"], "keywords": ["performance model", "roofline", "on-device vs edge vs cloud", "dual-system", "VLA-Perf", "analytical performance model", "network latency", "asynchronous inference", "real-time inference", "inference performance modeling", "hardware"], "model_class": "VLA", "techniques": ["analytical performance modeling", "performance modeling", "asynchronous inference", "dual-system pipelines", "analytical performance modelling", "deployment placement analysis", "analytical roofline-style performance modeling", "performance analysis"], "base_model": "", "params": "", "hardware": "analytical model over edge, consumer and datacenter GPUs", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Analytical performance model for arbitrary VLA/inference-system combinations used to study how model scale, architecture, video context, async inference, dual systems and on-device/edge/cloud placement affect end-to-end latency; distills 15 takeaways.", "relevance": 5, "relevance_reason": "Hardware-aware design-space study of where and how to run VLAs (NVIDIA/Stanford authors).", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA; WebSearch: VLA inference latency energy profiling Jetson; web: Jetson Thor NVFP4 FP8 humanoid onboard; arXiv API: abs:cloud AND abs:vision-language-action; snowball: citations of 2506.07339 (S2); venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2609.10915", "title": "IMLE-VLA: Fast Single-Step Action Generation for Vision-Language-Action Policies", "authors": ["Kian Hosseinkhani", "Qinhe Peng", "George Shramko", "Mehran Aghabozorgi", "Jianing Qian", "Tristan Engst", "Alireza Moazeni", "Dinesh Jayaraman", "Ke Li", "Qin-He Peng", "Jia-Ning Qian", "Ke Li Simon Fraser University", "Universitysof Pennsylvania", "Alberta Machine Intelligence Institute", "Canada Cifar AI Chair"], "year": 2026, "venue": "IROS 2026", "type": "preprint", "url": "https://arxiv.org/abs/2609.10915", "pdf_url": "https://arxiv.org/pdf/2609.10915", "code_url": "", "clusters": ["C1"], "keywords": ["IMLE-VLA", "one-step action head", "IMLE", "inference frequency"], "model_class": "VLA", "techniques": ["single-step generation", "conditional IMLE"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "55 Hz vs 15 Hz inference (3.67x); up to 11x action throughput | 55 Hz vs 15 Hz inference frequency (3.67x); 3.9-6.6x lower per-episode VLA inference time", "metric_measures": "closed-loop control", "benchmarks": ["LIBERO", "LIBERO-plus", "real robot (Franka)", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "IMLE-VLA replaces the multi-step flow/diffusion action head of pi0.5 with a one-step generator trained via conditional implicit maximum likelihood estimation, keeping multimodality while removing iterative sampling.", "relevance": 5, "relevance_reason": "direct: removes iterative action sampling for faster VLA inference", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=0]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.25562", "title": "IndustrialVLA-Bench: A Traceable Multi-Axis Evaluation of Open Robot Policy Models", "authors": ["Yiqi Wang", "Zhifeng Rao", "Jiaqi Zhang", "Xiaoyang Li", "Zhangkai Wu", "Yiqun Duan", "Mingkai Zheng", "Fei Wang", "Shan You", "Taotao Cai"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.25562", "pdf_url": "https://arxiv.org/pdf/2609.25562", "code_url": "https://github.com/xiaoqi-7/IndustrialVLA-Bench", "clusters": ["C4"], "keywords": ["unified reporting", "evidence status", "multi-axis evaluation", "reproducibility"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus", "LIBERO-Para"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Evaluates six VLA and world-action systems under one schema with three seeds, reporting clean success, robustness, paraphrase sensitivity, observed latency and peak memory with an evidence tier; clean LIBERO differences (1.6 pts) are tiny versus robustness spread (15-31 pts).", "relevance": 5, "relevance_reason": "Exactly the kind of joint capability + deployment-cost reporting protocol C4 targets.", "foundational_technique": false, "verified": true, "found_via": "search query: inference latency robot policy"}
{"id": "2603.08546", "title": "Interactive World Simulator for Robot Policy Training and Evaluation", "authors": ["Yixuan Wang", "Rhythm Syed", "Fangyu Wu", "Mengchao Zhang", "Aykut Onol", "Jose Barreiros", "Hooshang Nayyeri", "Tony Dear", "Huan Zhang", "Yunzhu Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.08546", "pdf_url": "https://arxiv.org/pdf/2603.08546", "code_url": "", "clusters": ["C8"], "keywords": ["interactive world simulator", "consistency models", "long-horizon stability"], "model_class": "world model", "techniques": ["consistency-model image decoding", "consistency latent dynamics"], "base_model": "", "params": "", "hardware": "single RTX 4090", "reported_metric": "15 FPS, stable >10 min interaction", "metric_measures": "throughput", "benchmarks": ["real robot (rigid, deformable, piles)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Builds a fast pixel-level robot world model from moderate data using consistency models for both dynamics and decoding, stable for 10+ minutes at 15 FPS on one 4090; policies trained on its data match real-data policies.", "relevance": 5, "relevance_reason": "Real-time robot world simulator on consumer GPU.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "github-nvidia-isaac-gr00t-deployment", "title": "Isaac GR00T N1.7 deployment guide: TensorRT full-pipeline acceleration and benchmarks", "authors": ["NVIDIA"], "year": 2026, "venue": "GitHub docs (NVIDIA Isaac-GR00T)", "type": "repo", "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/main/scripts/deployment/README.md", "pdf_url": "", "code_url": "https://github.com/NVIDIA/Isaac-GR00T", "clusters": ["C2", "C4", "C5", "C9"], "keywords": ["GR00T", "TensorRT", "torch.compile", "Jetson Orin", "cross-hardware latency table", "Jetson Thor", "DGX Spark", "full-pipeline export", "Jetson AGX Thor", "end-to-end latency"], "model_class": "VLA", "techniques": ["TensorRT", "torch.compile", "TensorRT engine export", "BF16", "TensorRT engine compilation (ViT, LLM, DiT, decoders)"], "base_model": "GR00T N1.7 (Qwen3-VL backbone)", "params": "3B", "hardware": "H100, H20, RTX Pro 6000/5000, L40, L20, DGX Spark, Jetson AGX Thor, Jetson Orin", "reported_metric": "H100: 85.8 ms eager / 48.6 ms compile / 27.9 ms TensorRT; Jetson Orin: 354.0 / 227.1 / 150.9 ms (4 denoising steps, 1 camera) | TRT full pipeline (4 denoise steps, 1 camera): H100 27.9 ms (35.9 Hz); AGX Thor 80.4 ms (12.4 Hz); Jetson Orin 150.9 ms (6.6 Hz)", "metric_measures": "open-loop chunk rate", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Official end-to-end latency table for GR00T N1.7 across GPUs and Jetson under eager, torch.compile and TensorRT, with the measured scope stated (4 denoising steps, 1 camera, preprocessing+backbone+action head).", "relevance": 5, "relevance_reason": "Authoritative cross-hardware latency table for a humanoid foundation model with TensorRT.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: NVIDIA Jetson Thor GR00T inference latency; web: Jetson Thor GR00T N1 deployment TensorRT; blog sweep: NVIDIA GEAR / Isaac"}
{"id": "2607.12659", "title": "Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference", "authors": ["Zebin Yang", "Qi Wang", "Yunhe Wang", "Xiurui Guo", "Bo Yu", "Shaoshan Liu", "Jiafeng Xu", "Hao Dong", "Meng Li", "Yun-He Wang", "Jia-Feng Xu"], "year": 2026, "venue": "CoRL 2026", "type": "preprint", "url": "https://arxiv.org/abs/2607.12659", "pdf_url": "https://arxiv.org/pdf/2607.12659", "code_url": "https://github.com/PKU-SEC-Lab/Jetson-PI", "clusters": ["C1", "C2", "C3", "C4", "C5", "C6", "C7", "C9"], "keywords": ["Jetson Orin", "asynchronous inference", "future correction", "CUDA graphs", "onboard", "control frequency", "llama.cpp engine", "foresight correction", "foresight alignment"], "model_class": "VLA", "techniques": ["asynchronous inference", "future-state prediction", "CUDA graph reuse", "flow unrolling", "future correction module", "scheduling", "future-state correction", "GPU-resident buffering", "adaptive scheduling", "CUDA graphs", "foresight-aligned async inference"], "base_model": "pi0-style VLM + action expert", "params": "", "hardware": "NVIDIA Jetson Orin", "reported_metric": "8.66x (vs PyTorch) and 5.41x (vs vla.cpp) higher control frequency | 8.66x / 5.41x control-frequency improvement vs naive PyTorch / vla.cpp", "metric_measures": "closed-loop control", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Combines a learned foresight module (predicting future scene features given committed actions) with confidence-based VLM/action-expert scheduling and CUDA-level system optimizations to run VLAs asynchronously on Jetson Orin.", "relevance": 5, "relevance_reason": "Reports onboard control frequency against named runtime baselines; good composability example (algorithm + systems).", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; search query: control frequency vision-language-action; arXiv: vision-language-action AND Jetson/edge/onboard; search query; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2605.11381", "title": "Kairos: A Scalable Serving System for Physical AI", "authors": ["Yinwei Dai", "Ganesh Ananthanarayanan", "Landon P. Cox", "Xenofon Foukas", "B. Radunovic", "Ravi Netravali", "Landon Cox", "Bozidar Radunovic"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.11381", "pdf_url": "https://arxiv.org/pdf/2605.11381", "code_url": "", "clusters": ["C5", "C6"], "keywords": ["multi-robot serving", "generate-execute loop", "physical AI", "physical AI serving", "multi-robot"], "model_class": "VLA", "techniques": ["execution-aware serving", "execution-aware scheduling", "multi-robot serving"], "base_model": "", "params": "", "hardware": "", "reported_metric": "31.8-66.5% lower average end-to-end task latency vs digital-AI serving practices", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Kairos is a multi-robot serving system that schedules inference with awareness of each robot's chunk execution phase, cutting task latency as fleets grow.", "relevance": 5, "relevance_reason": "Fleet-scale serving treating the generate-execute loop as first-class (Microsoft Research).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.18397 (citations); snowball from 2603.18284"}
{"id": "2608.11521", "title": "Keep the Future, Drop the Rollout: RIFT for World Action Models", "authors": ["Chushan Zhang", "Jinguang Tong", "Xuesong Li", "Yikai Wang", "Hongdong Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.11521", "pdf_url": "https://arxiv.org/pdf/2608.11521", "code_url": "", "clusters": ["C8"], "keywords": ["rollout-free imagination", "future tokens", "K/V cache"], "model_class": "world-action model", "techniques": ["learned anticipation tokens producing future K/V cache in one pass"], "base_model": "", "params": "", "hardware": "", "reported_metric": "3.1-9.2x inference speedup; LIBERO 98.8%", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "LIBERO-Plus", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows actions depend on the final future K/V cache rather than the rollout trajectory, then learns tokens that produce that cache in one backbone pass (RIFT), eliminating iterative video generation at deployment.", "relevance": 5, "relevance_reason": "Mechanistic analysis plus a large WAM speedup.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.22335", "title": "KerColle: Unlocking Fine-Grained GPU Concurrency in Vision-Language-Action Models", "authors": ["A. Li", "Christina Giannoula", "N. Vijaykumar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22335", "pdf_url": "https://arxiv.org/pdf/2609.22335", "code_url": "", "clusters": ["C3", "C5", "C7"], "keywords": ["GPU concurrency", "CTA scheduler", "head-of-line blocking", "SM utilization", "KerColle", "VLA serving"], "model_class": "VLA", "techniques": ["online SM-aware kernel scheduling", "phase concurrency", "fine-grained kernel co-scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Characterizes GPU underutilization when overlapping VLM and action-head phases across robot requests (CTA dispatcher head-of-line blocking) and proposes fine-grained kernel concurrency to meet control rates.", "relevance": 5, "relevance_reason": "GPU microarchitecture-level characterization and scheduling for VLA inference.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.18397 (citations); snowball: citations of 2506.07339 (S2)"}
{"id": "2608.00569", "title": "Latency-Tolerant Cloud-Edge Collaborative Vision-Language-Action Models via Emergent Representational Specialization", "authors": ["Daojie Peng", "Fulong Ma", "Bingtao Wang", "Sheng Wang", "Jun Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00569", "pdf_url": "https://arxiv.org/pdf/2608.00569", "code_url": "", "clusters": ["C1", "C2", "C3", "C5", "C6", "C7", "C9"], "keywords": ["cloud-edge", "latency tolerance", "stale features", "network delay", "split inference", "cloud-edge VLA", "network jitter", "cloud-edge collaborative VLA", "representational specialization"], "model_class": "VLA", "techniques": ["cloud-edge split", "delay-robust training", "cloud-edge collaboration", "hierarchical cloud/edge policy", "delay-augmented training", "asynchronous inference", "latency-tolerant training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "63.8-78.0% success under 40-step delay window | 63.8-78.0% success with 40-step uniform delay window (VLASH <=6.4%)", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "CloudEdgeVLA trains a cloud VLA to emit slowly varying task features from delayed frames while a light edge head fuses the latest feature with current vision, keeping control robust to network delay.", "relevance": 5, "relevance_reason": "Directly tackles latency-tolerant cloud/edge split of VLA policies.", "foundational_technique": false, "verified": true, "found_via": "search query: hierarchical VLA latency; arXiv: vision-language-action AND Jetson/edge/onboard; search query; venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2605.16154", "title": "Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking", "authors": ["Vaidehi Bagaria", "Nikshep Grampurohit", "Pulkit Verma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.16154", "pdf_url": "https://arxiv.org/pdf/2605.16154", "code_url": "", "clusters": ["C3"], "keywords": ["chunk masking", "GRPO", "gradient cost"], "model_class": "VLA", "techniques": ["probabilistic chunk masking", "GRPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "gradient computation ~78% of wall-clock per step (baseline)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Finds gradient compute (~78% of step time) dominates GRPO VLA RL and masks low-signal chunks to cut it.", "relevance": 5, "relevance_reason": "Directly targets RL post-training compute.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2608.23831", "title": "Learning to Act While Waiting: RL Finetuning of Generalist Robot Policies Under Inference Latency", "authors": ["Brian Zhu", "Momen Khalil", "E Harrison", "Emanuele Poggi", "Philipp Schmitt", "Bernd Kast", "Philine Meister", "Pranav Atreya", "Qiyang Li", "Finn Ferchau", "Cesar Colmenero", "Yash Shahapurkar", "Gokul Narayanan", "Melih Erdogan", "Kai Wurm", "Georg von Wichert", "Oier Mees", "Eugen Solowjow", "Andrew Wagenmaker", "Sergey Levine"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.23831", "pdf_url": "https://arxiv.org/pdf/2608.23831", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C6", "C7"], "keywords": ["RL under latency", "asynchronous inference", "state augmentation", "inference latency", "asynchronous RL", "ARLI", "inference delay", "latency-aware finetuning", "RL under inference latency", "Markov assumption"], "model_class": "VLA", "techniques": ["asynchronous inference", "RL fine-tuning", "latency-aware RL", "state augmentation", "RL finetuning", "state augmentation with committed actions", "mid-inference observation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows inference latency breaks the Markov assumption for RL finetuning of VLAs and restores it by augmenting state with committed actions and a mid-inference observation for a low-latency RL policy.", "relevance": 5, "relevance_reason": "Shows latency changes effective dynamics, so evaluation must model delay.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv; search query: inference latency robot policy; arXiv search: \"inference delay\" robot policy"}
{"id": "2609.03715", "title": "MINERVA: How Small Can a Manipulation Policy Be and Still Solve LIBERO?", "authors": ["Kohei Sendai", "Tatsuya Matsushima", "Yusuke Iwasawa"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.03715", "pdf_url": "https://arxiv.org/pdf/2609.03715", "code_url": "", "clusters": ["C1", "C2", "C4"], "keywords": ["capacity floor", "0.54M parameters", "LIBERO", "robustness", "tiny policy", "LIBERO saturation", "CPU inference"], "model_class": "other", "techniques": ["minimal model design", "small model", "L1 regression action head", "action chunking"], "base_model": "", "params": "0.54M", "hardware": "laptop CPU", "reported_metric": "95.1% LIBERO with 0.54M params (7,700x fewer than pi0.5) | 5-9 ms per chunk (113x faster than SmolVLA, 1400x faster than pi0.5)", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "LIBERO-90", "LIBERO-Plus"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "A 0.54M-parameter visuomotor policy reaches 95.1% average on the four LIBERO suites, showing the benchmark needs little capacity and that task-ID conditioning mostly selects memorized tasks; robustness collapses on LIBERO-Plus.", "relevance": 5, "relevance_reason": "Directly questions whether LIBERO can discriminate efficient VLAs; strong methodological evidence.", "foundational_technique": false, "verified": true, "found_via": "search: distillation vision-language-action; search query: all:LIBERO AND benchmark"}
{"id": "2606.19565", "title": "Mix-QVLA: Task-Evidence-Aware Mixed-Precision Quantization of Vision-Language-Action Models", "authors": ["Navin Ranjan", "Andreas E. Savakis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.19565", "pdf_url": "https://arxiv.org/pdf/2606.19565", "code_url": "", "clusters": ["C2"], "keywords": ["mixed precision", "task evidence", "BitOps budget"], "model_class": "VLA", "techniques": ["mixed-precision post-training quantization"], "base_model": "OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "15.4 GB -> 4.1 GB memory, 96.3 vs 97.1 success, 1.52x speedup", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Scores layer sensitivity by how quantization distorts gradient-weighted task-evidence maps over the episode and allocates bits under size/BitOps budgets.", "relevance": 5, "relevance_reason": "Phase-aware mixed-precision PTQ for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization"}
{"id": "2606.09215", "title": "MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation", "authors": ["Jia Zheng", "Teli Ma", "Yudong Fan", "Zifan Wang", "Shuo Yang", "Junwei Liang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09215", "pdf_url": "https://arxiv.org/pdf/2606.09215", "code_url": "", "clusters": ["C1", "C7", "C8", "C9"], "keywords": ["humanoid loco-manipulation", "real-time WAM", "intermediate denoising features", "denoising features"], "model_class": "world-action model", "techniques": ["policy conditioned on intermediate video-denoising features", "unified whole-body motion tokens", "condition on intermediate denoising features", "unified motion latent"], "base_model": "", "params": "", "hardware": "Unitree G1", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot (9 Unitree G1 tasks)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Real-time WAM for humanoid whole-body loco-manipulation from one egocentric camera, reading intermediate video-denoising features instead of finished video and predicting unified whole-body motion tokens.", "relevance": 5, "relevance_reason": "Real-time WAM on humanoid via partial denoising.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API ti:real-time AND VLA AND humanoid"}
{"id": "2604.27792", "title": "MotuBrain: An Advanced World Action Model for Robot Control", "authors": ["Motubrain Team", "Chendong Xiang", "Fan Bao", "Hai-Tian Liu", "Heng-Kai Tan", "Hong-Zhe Bi", "James Li", "Jia-Bao Liu", "Jin-Li Pang", "Kiro Jing", "Louis Liu", "Mengchen Cai", "Rongxu Cui", "Ruowen Zhao", "Run-Qing Wang", "Shu-He Huang", "Yao Feng", "Yinze Rong", "Ze-Yuan Wang", "Jun-Yan Zhu", "Haitian Liu", "Hengkai Tan", "Hongzhe Bi", "Jiabao Liu", "Jingrui Pang", "Runqing Wang", "Shuhe Huang", "Zeyuan Wang", "Jun Zhu"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2604.27792", "pdf_url": "https://arxiv.org/pdf/2604.27792", "code_url": "", "clusters": ["C2", "C5", "C8"], "keywords": ["FP8", "step reduction", "DiT caching", "deployment stack", "unified WAM", "Motus", "deployment recipe"], "model_class": "world-action model", "techniques": ["FP8 quantization", "step reduction", "caching", "compilation", "torch compilation", "DiT caching", "action-only inference", "chunked closed-loop execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">50x speedup over naive baseline; up to 11 Hz inference", "metric_measures": "throughput", "benchmarks": ["RoboTwin 2.0", "real robot", "WorldArena", "real humanoid"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Three-stream mixture-of-transformers WAM (policy, world model, inverse dynamics in one model) with a full deployment stack (fewer steps, FP8, caching, action-only mode) giving >50x speedup to ~11 Hz.", "relevance": 5, "relevance_reason": "FP8 quantization as part of a WAM deployment recipe.", "foundational_technique": false, "verified": true, "found_via": "search: FP8 robot policy inference; search query"}
{"id": "2607.01378", "title": "Neuro-Symbolic Safety Guidance for Vision-Language-Action Models via Constrained Flow Matching", "authors": ["William English", "Hao Zheng", "Rickard Ewetz", "William H. English"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.01378", "pdf_url": "https://arxiv.org/pdf/2607.01378", "code_url": "", "clusters": ["C7"], "keywords": ["SafeLIBERO", "predictive collision avoidance", "safety guidance", "collision avoidance"], "model_class": "VLA", "techniques": ["constrained flow matching", "neuro-symbolic safety"], "base_model": "", "params": "", "hardware": "", "reported_metric": "82.8% collision avoidance, 81.6% task success on SafeLIBERO", "metric_measures": "other", "benchmarks": ["SafeLIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Enforces collision constraints on the whole predicted action trajectory of flow-matching VLAs by minimum-norm constrained corrections during denoising, anticipating collisions rather than filtering only the next action.", "relevance": 5, "relevance_reason": "Chunk-level safety enforcement during generation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=100]; snowball round 2: cites 2306.03310, 2412.04987"}
{"id": "2609.18084", "title": "Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models", "authors": ["Shahram Najam Syed", "Arthur Jakobsson", "Prayuj Sachdev", "Jeffrey Ichnowski"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18084", "pdf_url": "https://arxiv.org/pdf/2609.18084", "code_url": "", "clusters": ["C3"], "keywords": ["region-wise adaptation", "variable-rank LoRA", "diagnostic"], "model_class": "VLA", "techniques": ["LoRA", "rank allocation", "selective freezing"], "base_model": "OpenVLA-OFT, pi0, SmolVLA, DTP, Octo", "params": "", "hardware": "", "reported_metric": "matches full FT with 0.04% trainable parameters (xArm-7)", "metric_measures": "other", "benchmarks": ["LIBERO", "CALVIN", "xArm-7 real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Measures which VLA regions need adaptation for each shift type and allocates variable-rank LoRA from ten unlabeled target observations.", "relevance": 5, "relevance_reason": "Budgeted PEFT allocation for VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: LoRA fine-tuning robot policy VLA adapter arXiv 2025"}
{"id": "2608.24199", "title": "NVIDIA Cosmos-H-Dreams: Real-Time Generative Physics Simulation for Surgical Robotics", "authors": ["Javier Gamazo Tejero", "Lukas Zbinden", "Keyur Sheth", "Raghavendra K M", "Nadim Daher", "Diego Granero Maraña", "Filip Binkiewicz", "Patrick Thornycroft", "Mahdi Azizian", "Sean D. Huver"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2608.24199", "pdf_url": "https://arxiv.org/pdf/2608.24199", "code_url": "https://huggingface.co/nvidia/Cosmos-H-Dreams", "clusters": ["C8"], "keywords": ["surgical world model", "Self Forcing distillation", "streaming inference", "FlashDreams"], "model_class": "world model", "techniques": ["bidirectional teacher to causal few-step student", "Self Forcing", "streaming inference library"], "base_model": "Cosmos-H-Surgical-Simulator (Cosmos-Predict2.5)", "params": "", "hardware": "single NVIDIA RTX PRO 6000 Blackwell", "reported_metric": "~160 inference FPS streaming", "metric_measures": "throughput", "benchmarks": ["surgical robot control (Versius console, VR, policies)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Distills an action-conditioned surgical video world model into a causal few-step student that streams at ~160 FPS on one workstation GPU, driven live by consoles, VR or learned policies.", "relevance": 5, "relevance_reason": "Concrete real-time distillation of a Cosmos robot world model with hardware numbers.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2605.16241", "title": "Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation", "authors": ["Jin Shi", "Brady Zhang", "Yishun Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.16241", "pdf_url": "https://arxiv.org/pdf/2605.16241", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["distillation", "student policy", "semantic guidance", "OpenVLA teacher"], "model_class": "VLA", "techniques": ["distillation", "knowledge distillation"], "base_model": "OpenVLA-7B (teacher); pi0.5-4B (teacher)", "params": "158M student", "hardware": "RTX 4090", "reported_metric": "12.5 Hz; 3.28x speedup over OpenVLA-7B; 44x smaller", "metric_measures": "throughput", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills OpenVLA (or pi0.5) into a 158M student using teacher actions plus VLM-generated phase and direction hints used only at training time.", "relevance": 5, "relevance_reason": "VLA-to-small-policy distillation with hardware numbers.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; web search: distilling VLA into small student"}
{"id": "2603.18284", "title": "Offload or Overload: A Platform Measurement Study of Mobile Robotic Manipulation Workloads", "authors": ["Sara Pohland", "Xenofon Foukas", "Ganesh Ananthanarayanan", "A. Kolobov", "Sanjeev Mehrotra", "B. Radunovic", "Ankit Verma", "Andrey Kolobov", "Bozidar Radunovic"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.18284", "pdf_url": "https://arxiv.org/pdf/2603.18284", "code_url": "", "clusters": ["C4", "C5", "C6", "C7", "C9"], "keywords": ["measurement study", "onboard vs edge vs cloud", "battery drain", "network bandwidth", "mobile manipulation", "battery", "bandwidth", "fleet sharing", "offloading", "network latency", "mobile manipulation workloads"], "model_class": "other", "techniques": ["platform measurement", "workload characterization", "offloading analysis", "measurement study", "platform profiling"], "base_model": "", "params": "", "hardware": "onboard, edge and cloud GPUs", "reported_metric": "lossless multi-camera streaming needs ~100 Mbps uplink (per paper)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Measures mobile-manipulation foundation-model stacks on onboard, edge and cloud GPUs: small onboard GPUs cannot run the full stack, large ones drain batteries hours faster, and offloading costs accuracy and bandwidth.", "relevance": 5, "relevance_reason": "Empirical ground truth for onboard-vs-offload trade-offs (Microsoft Research).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.18397 (citations); search query; snowball: citations of 2506.07339 (S2); venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2607.10172", "title": "On the Efficiency of LoRA Fine-Tuning for Vision-Language-Action Models in Industrial Robotic Manipulation", "authors": ["Finn Ferchau", "Daniel Pommer", "Cristian Axenie"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.10172", "pdf_url": "https://arxiv.org/pdf/2607.10172", "code_url": "", "clusters": ["C3"], "keywords": ["LoRA rank sweep", "industrial assembly", "UR5e", "VRAM"], "model_class": "VLA", "techniques": ["LoRA", "component freezing"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "static peak VRAM (params+optimizer) 36.2 -> 10.8 GiB with LoRA r=32", "metric_measures": "relative memory", "benchmarks": ["UR5e precision assembly (4 tasks)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Rank/allocation/freezing sweep for LoRA on pi0 in industrial assembly: r=32 with full vision-encoder tuning matches full fine-tuning at ~30% of the static VRAM.", "relevance": 5, "relevance_reason": "Direct measured memory-efficient VLA fine-tuning study.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:LoRA"}
{"id": "10.1109/ccwc67433.2026.11393726", "title": "On-Device Small Language Model Inference Acceleration for Real-Time Robotics: A Survey", "authors": ["Anuja Dawane"], "year": 2026, "venue": "Computing and Communication Workshop and Conference", "type": "paper", "url": "https://doi.org/10.1109/CCWC67433.2026.11393726", "pdf_url": "", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["on-device SLM", "survey", "Jetson"], "model_class": "other", "techniques": ["quantization", "distillation", "KV-cache compression", "speculative decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of on-device small language model inference acceleration for real-time robotics across model, algorithm, runtime and hardware layers.", "relevance": 5, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2205.14135"}
{"id": "jetson-ai-lab-openpi-pi05-on-jetson-thor", "title": "OpenPi π₀.₅ on Jetson Thor", "authors": ["Aditya Sahu", "Anqi Liu"], "year": 2026, "venue": "Jetson AI Lab (NVIDIA) tutorial", "type": "blog", "url": "https://www.jetson-ai-lab.com/tutorials/openpi_on_thor/", "pdf_url": "", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["NVFP4", "FP8", "TensorRT", "pi0.5", "Jetson AGX Thor", "Jetson Thor"], "model_class": "VLA", "techniques": ["FP8 quantization", "NVFP4 quantization", "TensorRT", "TensorRT compilation"], "base_model": "pi0.5 (OpenPi)", "params": "", "hardware": "Jetson AGX Thor (MAXN, JetPack 7.2)", "reported_metric": "~132 ms (BF16 PyTorch) -> ~54 ms (TRT FP8) -> ~49 ms (TRT FP8+NVFP4), ~2.7x", "metric_measures": "open-loop chunk rate", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Step-by-step NVIDIA tutorial deploying pi0.5 on Jetson Thor with TensorRT FP8 and NVFP4, cutting per-inference latency ~2.7x at action horizon 10.", "relevance": 5, "relevance_reason": "Concrete edge deployment recipe and numbers for pi0.5 on Thor.", "foundational_technique": false, "verified": true, "found_via": "web search: quantized pi0 GR00T Jetson; web: Jetson Thor GR00T N1 deployment TensorRT"}
{"id": "2603.14371", "title": "OxyGen: Unified KV Cache Management for Vision-Language-Action Models under Multi-Task Parallelism", "authors": ["Xiangyu Li", "Huaizhi Tang", "Xin Ding", "Weijun Wang", "Ting Cao", "Yunxin Liu", "Yun-Xin Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.14371", "pdf_url": "https://arxiv.org/pdf/2603.14371", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C9"], "keywords": ["KV cache management", "multi-task parallelism", "mixture-of-transformers", "continuous batching", "KV cache", "multi-task", "Jetson AGX Thor", "humanoid", "mixture-of-transformers VLA", "VLA serving"], "model_class": "VLA", "techniques": ["KV cache sharing", "continuous batching", "cross-task KV sharing", "cross-frame continuous batching", "unified KV cache management", "unified KV cache sharing", "KV cache management"], "base_model": "pi0.5", "params": "", "hardware": "RTX 4090; Jetson AGX Thor", "reported_metric": "up to 3.7x speedup; >200 tok/s language + 70 Hz action simultaneously | up to 3.7x over isolated execution; >200 tokens/s language + 70 Hz action frequency simultaneously", "metric_measures": "throughput", "benchmarks": ["real humanoid", "real robot (humanoid)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Treats the KV cache as a shared resource across concurrent VLA tasks (acting, talking, memory) and across frames, avoiding redundant prefill and decoupling language decoding from fixed-rate action generation.", "relevance": 5, "relevance_reason": "On-device memory management for VLAs validated on humanoid with onboard Thor.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache; search query: Jetson VLA; snowball from 2602.18397 (citations); venue sweep: arXiv API VLA AND Jetson; snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2605.29710", "title": "PhAIL: A Real-Robot VLA Benchmark and Distributional Methodology", "authors": ["Sergey Arkhangelskiy"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2605.29710", "pdf_url": "https://arxiv.org/pdf/2605.29710", "code_url": "https://github.com/Positronic-Robotics/phail-paper", "clusters": ["C4"], "keywords": ["PhAIL", "time-to-success CDF", "human-relative throughput", "KS test", "bootstrap CI"], "model_class": "VLA", "techniques": [], "base_model": "GR00T, OpenPI, ACT", "params": "", "hardware": "Franka FR3", "reported_metric": "", "metric_measures": "", "benchmarks": ["PhAIL", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Real-robot leaderboard using the time-to-success CDF as primitive, scored by human-relative throughput with bootstrap CIs and KS significance tests, resolving close comparisons at N<=30 where binary success cannot.", "relevance": 5, "relevance_reason": "Statistically grounded real-robot protocol that folds speed into the metric.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA leaderboard"}
{"id": "2608.03682", "title": "PhyAI: Real-Time Physical AI at the Edge, Scalable Rollouts in the Cloud", "authors": ["Chenghua Wang", "Daliang Xu", "Dongqi Cai", "Duojin Sun", "Hao Zhang", "Haoze Qian", "Huaiyuan Zhang", "Jinshuo Cui", "Junbo Cui", "Kezhao Zhao", "Longxi Gao", "Mengwei Xu", "Rongjie Yi", "Ruixin Liu", "Shangguang Wang", "Tam Sikyuen", "Tianyue Zhang", "Weikai Xie", "Xuanzhe Liu", "Yingying Qin", "Yiwen Lu", "Yuan Yao", "Yuezhi Zu", "Yunhan Guo", "Yuxin Zheng", "Ziqi Guo", "Cheng-Hua Wang", "Da-Liang Xu", "Hao Qian", "Huai-Yuan Zhang", "Jing Cui", "Jun-Bo Cui", "Ke-Zhao Zhao", "Tianyue H. Zhang", "Y. Zu", "Yun-Han Guo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.03682", "pdf_url": "https://arxiv.org/pdf/2608.03682", "code_url": "https://github.com/mingti-org/phyai", "clusters": ["C4", "C5", "C6", "C7", "C8", "C9"], "keywords": ["inference engine", "unified runtime", "profiling", "FLOPs vs latency", "world-action model", "control-time roofline", "unified inference engine", "onboard/edge/cloud", "PhyAI", "edge/cloud/onboard runtime", "VLA and WAM serving", "edge-cloud", "unified inference program", "edge serving", "cloud rollouts"], "model_class": "other", "techniques": ["kernel optimization", "graph execution", "tensor parallelism", "unified inference runtime", "kernel fusion", "batching", "unified runtime with model adapters", "kernel/memory sharing", "model adapters", "shared kernels and memory management"], "base_model": "pi0, pi0.5, GR00T N1.7, MiniCPM-Robot, Cosmos3-Nano-Policy", "params": "", "hardware": "8x H20 GPUs; Hopper GPUs; edge/onboard", "reported_metric": "1.40x-4.65x speedup over official implementations; pi0.5 action expert 8.8% FLOPs but 57.2% latency at batch 1 | 1.40-4.65x speedup over official implementations; Cosmos3-Nano-Policy latency 2.46 s -> 1.18 s on 8x H20; pi0.5 ~100 samples/s at batch 32", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "One runtime for VLAs and world-action models spanning onboard, edge, and cloud, with profiling showing the action expert is latency-dominant at batch 1 and a control-time roofline separating inference-bound from environment-bound control.", "relevance": 5, "relevance_reason": "Systems-level profiling and runtime across hardware tiers; introduces roofline analysis for control.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action edge deployment benchmark latency; arXiv: vision-language-action AND Jetson/edge/onboard; snowball from 2506.07339; snowball: citations of 2506.07339 (S2); search query; venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2603.05438", "title": "Planning in 8 Tokens: A Compact Discrete Tokenizer for Latent World Model", "authors": ["Dongwon Kim", "Gawon Seo", "Jinsung Lee", "Minsu Cho", "Suha Kwak"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.05438", "pdf_url": "https://arxiv.org/pdf/2603.05438", "code_url": "", "clusters": ["C8"], "keywords": ["compact tokenizer", "8 tokens", "decision-time planning"], "model_class": "world model", "techniques": ["discrete tokenizer compressing observation to ~8 tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Compresses each observation into as few as 8 discrete tokens so an action-conditioned world model can plan at decision time orders of magnitude cheaper.", "relevance": 5, "relevance_reason": "Extreme token compression for real-time planning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2609.24840", "title": "PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control", "authors": ["Lei Ye", "Hai-Bo Gao", "Yi-Tang Li", "Peng Xu", "Zengrui Jing", "Jun Sun", "Fan-Rong Dong", "Zi-Qi Han", "Xue Wang", "Jianhua Sun", "Ce-Wu Lu", "Hao Zhao", "Liang Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24840", "pdf_url": "https://arxiv.org/pdf/2609.24840", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["onboard humanoid control", "Jetson"], "model_class": "diffusion/flow policy", "techniques": ["rolling denoising", "runtime optimization"], "base_model": "", "params": "", "hardware": "Jetson Orin NX", "reported_metric": "16.79 ms median / 19.38 ms p95 callback (< 20 ms control period)", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Predictive state-action diffusion policy for humanoids, optimized to run within the control period on an onboard Jetson and deployed on a Unitree G1.", "relevance": 5, "relevance_reason": "Onboard real-time diffusion control with measured latency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.19441", "title": "Predict Before You Deploy: Offline Prediction of Quantization-Induced Task Degradation for World Action Models", "authors": ["Jiuyi Xu", "Jinjia Guo", "Meida Chen", "Jing Du", "Yangming Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19441", "pdf_url": "https://arxiv.org/pdf/2609.19441", "code_url": "https://github.com/jiuyixu25/PreDE", "clusters": ["C1", "C2", "C3", "C4", "C8"], "keywords": ["offline prediction", "action deviation", "Franka", "PreDE", "quantization", "closed-loop degradation", "world action model", "quantization config selection", "WAM"], "model_class": "world-action model", "techniques": ["post-training quantization", "deployment validation", "offline action-deviation screening", "policy-calibrated thresholds on offline action deviation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "21/28 held-out configs decided offline, all correct", "metric_measures": "other", "benchmarks": ["real robot", "simulation", "four benchmark settings"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Predicts whether a quantization configuration will hurt closed-loop task success from offline action deviations, calibrated on a small development set; shows bit width alone does not explain task loss across five WAMs.", "relevance": 5, "relevance_reason": "Methodology to cheaply evaluate compression configs without exhaustive closed-loop runs.", "foundational_technique": false, "verified": true, "found_via": "search: quantization robot policy; search query: robot policy evaluation statistical; search query"}
{"id": "2602.02396", "title": "PRISM: Performer RS-IMLE for Single-pass Multisensory Imitation Learning", "authors": ["Amisha Bhaskar", "Pratap Tokekar", "S. di Cairano", "Alexander V. Schperberg"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.02396", "pdf_url": "https://arxiv.org/pdf/2602.02396", "code_url": "", "clusters": ["C1"], "keywords": ["multisensory", "real-time"], "model_class": "other", "techniques": ["single-pass IMLE", "Performer linear attention"], "base_model": "", "params": "", "hardware": "", "reported_metric": "30-50 Hz closed-loop control", "metric_measures": "closed-loop control", "benchmarks": ["CALVIN", "MetaWorld", "Robomimic", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "PRISM is a single-pass IMLE policy with linear attention, avoiding iterative sampling while fusing many sensors.", "relevance": 5, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2608.02545", "title": "Probabilistic Reachable-Action Verification of Visuomotor Policies via Set-Based Training", "authors": ["Yanliang Huang", "Zhuocheng Zhang", "Peng Xie", "Zhen Zhang", "Wenyuan Wu", "Majid Khadiv", "Zhuoqi Zeng", "Amr Alanwar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.02545", "pdf_url": "https://arxiv.org/pdf/2608.02545", "code_url": "", "clusters": ["C7"], "keywords": ["visuomotor verification", "reachable actions"], "model_class": "diffusion/flow policy", "techniques": ["reachability analysis", "zonotopes", "set-based training", "conformal calibration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Freezes the visual encoder, propagates zonotope sets only through the low-dimensional policy interface, trains to shrink the output enclosure, and conformally calibrates a probabilistic reachable-action radius under camera-pose perturbations.", "relevance": 5, "relevance_reason": "Directly verifies visuomotor policy action outputs with a set-based plus conformal guarantee.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:reachability AND (\"visuomotor policy\" OR \"learned policy\" OR \"robot foundation model\" OR \"vision-language-action\")"}
{"id": "2603.17850", "title": "ProbeFlow: Training-Free Adaptive Flow Matching for Vision-Language-Action Models", "authors": ["Zhou Fang", "Jiaqi Wang", "Yi Zhou", "Qiongfeng Shi", "Jia-Qi Wang", "Q. Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.17850", "pdf_url": "https://arxiv.org/pdf/2603.17850", "code_url": "", "clusters": ["C1"], "keywords": ["flow matching", "action head", "flow matching head", "adaptive inference"], "model_class": "VLA", "techniques": ["adaptive ODE steps", "training-free acceleration", "adaptive ODE step scheduling", "training-free"], "base_model": "", "params": "", "hardware": "", "reported_metric": "14.8x faster action decoding (50 -> 2.6 steps), 2.8x lower end-to-end latency on MetaWorld", "metric_measures": "relative latency", "benchmarks": ["MetaWorld", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Training-free adaptive step scheduling for flow-matching action heads based on cosine similarity of initial and lookahead velocities.", "relevance": 5, "relevance_reason": "Directly reduces action-head latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=0]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2609.33269", "title": "Q-WAM: 4-Bit Quantization of World Action Models with Action-Subspace Protection", "authors": ["Arash Akbari", "Arman Akbari", "Jingwu Luo", "Yuhao Lei", "Yi Gao", "Weiwei Chen", "Xuan Zhang", "Zhenman Fang", "Geng Yuan", "Yanzhi Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33269", "pdf_url": "https://arxiv.org/pdf/2609.33269", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["W4A4", "action observability gramian", "low-rank 16-bit branch", "MoE experts", "4-bit WAM quantization", "action-subspace protection"], "model_class": "world-action model", "techniques": ["post-training quantization", "low-rank compensation", "W4A4 PTQ", "action-sensitivity-aware channel protection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "89.6-93.0% avg success on RoboTwin 2.0, within 1.1 pp of 16-bit", "metric_measures": "other", "benchmarks": ["RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "4-bit weight-activation quantization for WAMs that measures how rounding errors in each channel combination propagate to final actions across denoising, and keeps the most action-sensitive subspace in higher precision.", "relevance": 5, "relevance_reason": "Quantization tailored to world-action models.", "foundational_technique": false, "verified": true, "found_via": "search: quantization world action model; search query"}
{"id": "2602.20309", "title": "QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models", "authors": ["Jingxuan Zhang", "Yun-Ta Hsieh", "Zhongwei Wan", "Haokun Lin", "Xin Wang", "Zi-Qi Wang", "Ying-Tie Lei", "Mi Zhang", "Yunta Hsieh", "Ziqi Wang", "Yingtie Lei"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.20309", "pdf_url": "https://arxiv.org/pdf/2602.20309", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["PTQ", "DiT action head", "calibration", "training-free PTQ", "scale calibration"], "model_class": "VLA", "techniques": ["post-training quantization", "scale-calibrated PTQ"], "base_model": "pi0 / GR00T-style VLAs with DiT heads", "params": "", "hardware": "", "reported_metric": "~70% relative memory savings on quantized components", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free PTQ that also integerizes the diffusion-transformer action head, using attention temperature matching and output-head balancing to prevent calibration drift.", "relevance": 5, "relevance_reason": "First PTQ covering DiT action heads.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2607.28405", "title": "QuantWAMs: Calibrating at the Right Granularity for World Action Models", "authors": ["Jiacheng Zhou", "Jinfan Lv", "Ruixuan Li", "Longtai Zhang", "Yan Wang", "Wenqiang Zhang", "Lizhe Qi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.28405", "pdf_url": "https://arxiv.org/pdf/2607.28405", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["calibration granularity", "Fisher saliency", "rollout auditing", "AgiBot G2", "post-training quantization", "W4A4", "closed-loop calibration"], "model_class": "world-action model", "techniques": ["post-training quantization", "mixed precision", "PTQ", "shared-basis outlier calibration", "Fisher-saliency mixed precision", "rollout-audited step protection"], "base_model": "Fast-WAM; LingBot-VA", "params": "", "hardware": "AgiBot G2 (real robot)", "reported_metric": "~29% of FP16 peak weight+activation memory for targeted blocks; sim success within 0.2-0.7 pp | peak weight+activation memory ~29% of FP16; 1.4-1.6x block-level speedup; 0.2-0.7 pp gap to FP16", "metric_measures": "relative memory", "benchmarks": ["RoboTwin 2.0", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "PTQ framework for WAMs that calibrates on closed-loop rollout states and joint video-action objectives, reaching W4A4 with sub-1-point loss and about 3.4x memory reduction on targeted blocks.", "relevance": 5, "relevance_reason": "W4A4-dominant WAM quantization with real-robot tests.", "foundational_technique": false, "verified": true, "found_via": "search: quantization world action model; search query"}
{"id": "2602.03782", "title": "QVLA: Not All Channels Are Equal in Vision-Language-Action Model's Quantization", "authors": ["Yuhao Xu", "Yantai Yang", "Z. Fan", "Yu-Fan Liu", "Yu-Ming Li", "Bing Li", "Zhi-Peng Zhang", "Zhenyang Fan", "Yufan Liu", "Yuming Li", "Zhipeng Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.03782", "pdf_url": "https://arxiv.org/pdf/2602.03782", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["channel-wise bit allocation", "action sensitivity", "0-bit pruning", "channel-wise sensitivity", "mixed precision"], "model_class": "VLA", "techniques": ["mixed-precision quantization", "pruning", "channel-aware quantization"], "base_model": "OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "29.2% of original VRAM, 98.9% of performance, 1.49x speedup", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Allocates per-channel bit-widths (including 0-bit, i.e. pruning) by measuring action-space sensitivity, outperforming LLM-style uniform quantizers like SmoothQuant on VLAs.", "relevance": 5, "relevance_reason": "Action-centric quantization framework unifying pruning and quantization.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2603.07949", "title": "RAPID: Redundancy-Aware and Compatibility-Optimal Edge-Cloud Partitioned Inference for Diverse VLA Models", "authors": ["Zihao Zheng", "Sicheng Tian", "Hangyu Cao", "Chenyue Li", "Jiayu Chen", "Maoliang Li", "Xinhao Sun", "Hailong Zou", "Guojie Luo", "Xiang Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.07949", "pdf_url": "https://arxiv.org/pdf/2603.07949", "code_url": "", "clusters": ["C1", "C3", "C5", "C6"], "keywords": ["edge-cloud partitioning", "edge-cloud partitioned inference", "step-wise redundancy", "VLA"], "model_class": "VLA", "techniques": ["edge-cloud split", "redundancy-aware partitioning", "split computing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.73x speedup with 5-7% overhead", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Edge-cloud partitioned VLA inference that avoids visual-noise-sensitive partitioning and exploits step-to-step redundancy in embodied tasks to decide what runs where.", "relevance": 5, "relevance_reason": "Edge-cloud VLA partitioning method from same group as RoboECC.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; arXiv: on-device AND (robot policy OR VLA OR diffusion policy); search query"}
{"id": "2602.01515", "title": "RAPT: Model-Predictive Out-of-Distribution Detection and Failure Diagnosis for Sim-to-Real Humanoid Deployment", "authors": ["Humphrey Munn", "Brendan Tidd", "Peter Bohm", "Marcus Gallagher", "David Howard"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.01515", "pdf_url": "https://arxiv.org/pdf/2602.01515", "code_url": "https://humphreymunn.github.io/RAPT/", "clusters": ["C7", "C8"], "keywords": ["RAPT", "50 Hz deployment monitor", "sim-to-real", "humanoid", "OOD detection"], "model_class": "other", "techniques": ["recurrent probabilistic trajectory model", "calibrated per-dimension deviation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "50 Hz monitor; 89% TPR over 78 hardware trials; +37% TPR at 0.5% FPR in sim", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Lightweight self-supervised monitor running at control rate that flags OOD execution of learned humanoid policies with very low false positives and diagnoses failure causes post hoc.", "relevance": 5, "relevance_reason": "Runtime monitor meeting high-rate control constraints.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2503.08558 (S2)"}
{"id": "2605.13778", "title": "Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs", "authors": ["Jiahui Niu", "Kefan Gu", "Yucheng Zhao", "Shengwen Liang", "Tiancai Wang", "Xing Hu", "Ying Wang", "Huawei Li", "Yu-Chen Zhao", "Tian-Cai Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.13778", "pdf_url": "https://arxiv.org/pdf/2605.13778", "code_url": "", "clusters": ["C1", "C3", "C6", "C7", "C9"], "keywords": ["speculative inference", "draft model", "diffusion VLA", "phase-aware fallback", "skip full inference"], "model_class": "VLA", "techniques": ["speculative inference", "speculative drafting + verification", "fallback to full inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "task-level avg inference latency 19.1 ms vs 58.0 ms full (3.04x)", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real conveyor sorting"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Replaces most full diffusion-VLA replanning calls with a light draft model verified in parallel by the action expert, with a phase-aware fallback to full inference.", "relevance": 5, "relevance_reason": "Direct real-time acceleration of diffusion VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; snowball: citations of 2506.07339 (S2); venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2609.19579", "title": "Recovering Aggressively Pruned Vision-Language-Action Models with Offline Hidden-State Distillation", "authors": ["Chiyoung Kim", "Sanghyuk Roy Choi", "Minhyeok Lee"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19579", "pdf_url": "https://arxiv.org/pdf/2609.19579", "code_url": "", "clusters": ["C2"], "keywords": ["width pruning", "hidden-state distillation", "offline recovery"], "model_class": "VLA", "techniques": ["structured pruning", "knowledge distillation"], "base_model": "OpenVLA-OFT; CogACT", "params": "", "hardware": "", "reported_metric": "63% backbone reduction recovered to within 3.5 pts of teacher in ~8 GPU-hours", "metric_measures": "other", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Recovers heavily width-pruned VLA backbones fully offline by matching cached teacher hidden states, replacing costly SFT+RL recovery.", "relevance": 5, "relevance_reason": "Cheap recovery of aggressively pruned VLAs.", "foundational_technique": false, "verified": true, "found_via": "search: pruning vision-language-action"}
{"id": "2608.14379", "title": "Reflex: Enabling Fast and Predictive Vision-Language-Action Models for Reaction-Critical Manipulation", "authors": ["Yuxuan Chen", "Wanruo Zhang", "Xiao Li"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2608.14379", "pdf_url": "https://arxiv.org/pdf/2608.14379", "code_url": "", "clusters": ["C1", "C2", "C4", "C5", "C7", "C9"], "keywords": ["reaction-critical benchmark", "CUDA graph", "batched encoding", "ReflexBench", "configurable latency", "reaction-critical", "simulator-control decoupling", "batched visual encoding", "benchmark", "reaction-critical manipulation", "ReflexVLA"], "model_class": "VLA", "techniques": ["CUDA graph replay", "latent future prediction", "batched visual encoding", "benchmark with configurable latency"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ReflexBench", "LIBERO", "real robot", "ReflexBench (6 dynamic tasks)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Benchmark of six dynamic tasks whose evaluator decouples simulator stepping from control and injects configurable latency under sync and async inference, plus an efficient reaction-oriented VLA.", "relevance": 5, "relevance_reason": "Latency-aware closed-loop evaluation protocol in simulation.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; WebSearch: benchmark dynamic manipulation moving objects VLA reaction latency; snowball from 2510.26742 (citations); snowball: citations of 2506.07339 (S2)"}
{"id": "2607.14695", "title": "Reflex: Real-Time VLA Control through Streaming Inference", "authors": ["Yuanchun Guo", "Bingyan Liu"], "year": 2026, "venue": "ICML 2026", "type": "preprint", "url": "https://arxiv.org/abs/2607.14695", "pdf_url": "https://arxiv.org/pdf/2607.14695", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C7"], "keywords": ["streaming inference", "KV cache", "flow matching", "operator fusion", "AdaRMSNorm", "reaction latency", "BF16 numerical stability", "Reflex"], "model_class": "VLA", "techniques": ["incremental KV cache", "asynchronous pipeline", "kernel fusion", "incremental KV caching", "async pipeline", "operator fusion", "incremental KV cache partitioning", "AdaRMSNorm"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.58x speedup; 50 Hz stable streaming; up to 54% lower reaction latency | 2.58x inference speedup; 50 Hz stable streaming; up to 54% lower reaction latency", "metric_measures": "open-loop chunk rate", "benchmarks": ["LIBERO", "Kinetix"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Streaming inference for flow-matching VLAs using timestep-invariant perception caching, a flow-phase-gated normalization for bf16 stability, and an asynchronous encode/generate pipeline.", "relevance": 5, "relevance_reason": "Reports reaction latency as distinct metric; composes caching with async and fusion.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache; search query: VLA inference performance; snowball from 2510.26742 (citations); arXiv search: \"streaming\" flow policy robot action"}
{"id": "2606.07383", "title": "RhinoVLA Technical Report", "authors": ["Huixi Technology", ":", "Chen Zhang", "Chenyang Zhou", "Guanglei Ding", "Guanghui He", "Haibin Gao", "Jiajia Chen", "Jianyong Zhang", "Lianyi Yu", "Ningyi Xu", "Ping Xu", "Qingchen Li", "Yingjun Hu", "Yijia Zhang", "Yuxi Liu", "Guang-Hui He", "Haibing Gao", "Jiajian Chen", "Pingbo Xu", "Qingcheng Li", "Ying-Jun Hu", "Yijian Zhang", "Yu-Xi Liu"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2606.07383", "pdf_url": "https://arxiv.org/pdf/2606.07383", "code_url": "https://github.com/HuixiAI/RhinoVLA", "clusters": ["C1", "C5"], "keywords": ["edge SoC", "RhinoVLA", "co-design", "real-time deployment"], "model_class": "VLA", "techniques": ["hardware co-design", "mixed precision", "token reduction", "compilation", "token-efficient backbone", "hardware-aware compilation", "parallel visual encoding", "LoRA"], "base_model": "Qwen3-VL backbone", "params": "", "hardware": "Huixi R1 edge SoC", "reported_metric": "11.69 Hz end-to-end inference", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Deployment-oriented VLA co-designed with the Huixi R1 edge SoC: token-efficient Qwen3-VL backbone, continuous action expert, hardware-aware compilation, mixed precision and parallel visual encoding.", "relevance": 5, "relevance_reason": "Onboard edge deployment with hardware co-design.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]; snowball round 2: cites 2205.14135, 2303.04137"}
{"id": "2506.17639", "title": "RLRC: Reinforcement Learning-based Recovery for Compressed Vision-Language-Action Models", "authors": ["Yuxuan Chen", "Yixin Han", "Yize Huang", "Xiao Li"], "year": 2026, "venue": "IEEE RA-L 2026", "type": "preprint", "url": "https://arxiv.org/abs/2506.17639", "pdf_url": "https://arxiv.org/pdf/2506.17639", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C9"], "keywords": ["structured pruning", "RL recovery", "compression pipeline", "compression recovery", "quantization", "composition", "compression study", "pruning"], "model_class": "VLA", "techniques": ["structured pruning", "SFT recovery", "RL fine-tuning", "quantization", "pruning", "SFT+RL recovery", "SFT", "RL recovery", "RL-based recovery"], "base_model": "OpenVLA and other VLA backbones", "params": "", "hardware": "", "reported_metric": "up to 8x memory reduction, 2.3x inference speedup", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Systematically studies VLA compression and proposes prune -> SFT+RL recovery -> quantize, using RL with critic warm-up and BC regularization to regain success rate lost to pruning.", "relevance": 5, "relevance_reason": "Direct evidence on composing pruning and quantization in VLAs and order of operations.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; snowball from 2502.19645 (S2 citations); search query: VLA quantization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2603.20711", "title": "RoboECC: Multi-Factor-Aware Edge-Cloud Collaborative Deployment for VLA Models", "authors": ["Zheng, Zihao", "Cao, Hangyu", "Chen, Jiayu", "Tian, Sicheng", "Li, Chenyue", "Li, Maoliang", "Sun, Xinhao", "Luo, Guojie", "Chen, Xiang", "Zihao Zheng", "Hangyu Cao", "Jiayu Chen", "Sicheng Tian", "Chenyue Li", "Maoliang Li", "Xinhao Sun", "Guojie Luo", "Xiang Chen"], "year": 2026, "venue": "IJCNN 2026", "type": "preprint", "url": "https://arxiv.org/abs/2603.20711", "pdf_url": "https://arxiv.org/pdf/2603.20711", "code_url": "https://github.com/zhengzihaoPKU/RoboECC", "clusters": ["C4", "C5", "C6"], "keywords": ["RoboECC", "edge-cloud", "split point", "network-aware", "edge-cloud collaborative", "network-aware adaptation"], "model_class": "VLA", "techniques": ["edge-cloud partitioning", "model-hardware co-aware segmentation", "network-adaptive deployment", "split computing", "model partitioning", "bandwidth adaptation"], "base_model": "OpenVLA, CogACT", "params": "", "hardware": "", "reported_metric": "up to 3.28x speedup with 2.55-2.62% overhead", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SimplerEnv"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Finds a model- and hardware-aware split point for VLAs between edge and cloud, then moves it online as bandwidth changes to keep latency low.", "relevance": 5, "relevance_reason": "Seed edge-cloud split-computing framework specifically for VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: VLA deployment edge devices benchmark; arXiv: on-device AND (robot policy OR VLA OR diffusion policy); search query"}
{"id": "2606.07217", "title": "Robotic Policy Adaptation via Weight-Space Meta-Learning", "authors": ["Christian Bianchi", "Siamak Yousefi", "Alessio Sampieri", "Andrea Roberti", "Luca Rigazio", "Fabio Galasso", "Luca Franco"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.07217", "pdf_url": "https://arxiv.org/pdf/2606.07217", "code_url": "", "clusters": ["C3"], "keywords": ["meta-learning", "hypernetwork", "generated LoRA", "WIZARD"], "model_class": "VLA", "techniques": ["LoRA", "hypernetwork weight generation", "meta-learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to ~14x on unseen LIBERO tasks", "metric_measures": "other", "benchmarks": ["LIBERO", "Franka real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Generates task-specific LoRA weights for a frozen VLA from a language instruction and a demo video in one forward pass, avoiding per-task fine-tuning.", "relevance": 5, "relevance_reason": "Removes fine-tuning cost at deployment via adapter generation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: LoRA fine-tuning robot policy VLA adapter arXiv 2025"}
{"id": "2609.30247", "title": "Rolling-WAM: World Action Models with Rolling Imagination", "authors": ["Yinghua Zhou", "Junjie Ye", "Yiqi Zhao", "Hao Dong", "Celina Shiyu Wang", "Ruohai Ge", "Tingyi Yang", "Basile Van Hoorick", "Gaurav Sukhatme", "Vitor Guizilini", "Yue Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.30247", "pdf_url": "https://arxiv.org/pdf/2609.30247", "code_url": "", "clusters": ["C3", "C6", "C7", "C8", "C9"], "keywords": ["rolling denoising", "staggered noise levels", "replanning latency", "humanoid"], "model_class": "world-action model", "techniques": ["rolling noise schedule across replanning cycles", "sliding window of video-action chunks"], "base_model": "", "params": "", "hardware": "", "reported_metric": "4.5x steady-state replanning speedup vs standard joint WAMs", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboTwin", "real robot (Unitree G1)", "Unitree G1"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Spreads joint video-action denoising over successive control cycles: near-term chunks are fully denoised for execution while farther chunks are partially refined and carried forward, amortizing compute.", "relevance": 5, "relevance_reason": "Diffusion-forcing style amortization for real-time WAM control.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2607.01088", "title": "ROSA: A Robotics Foundation Model Serving System for Robot Factories", "authors": ["Wenqi Jiang", "Jason Clemons", "Rowland O'Flaherty", "Hugo Hadfield", "Alperen Değirmenci", "Shuran Song", "Yashraj Narang", "Christoforos E. Kozyrakis", "Alperen Degirmenci", "Christos Kozyrakis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.01088", "pdf_url": "https://arxiv.org/pdf/2607.01088", "code_url": "", "clusters": ["C5", "C6"], "keywords": ["robot factory", "GPU pool", "serving", "battery", "Ray Serve", "robot foundation model serving", "factory", "SLO"], "model_class": "VLA", "techniques": ["shared GPU-pool serving", "SLO-driven scheduling", "multi-model pipelines", "inter-robot batching", "factory-objective scheduling"], "base_model": "", "params": "", "hardware": "server-class GPU pool", "reported_metric": "up to 12.06x factory productivity over dedicated serving", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "NVIDIA/Stanford serving system letting factory robot fleets share server GPUs over the network, scheduling for SLO-qualified productivity rather than single-request latency, with battery benefits.", "relevance": 5, "relevance_reason": "Compute placement trade-offs (onboard vs pooled) for robot foundation models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.18397 (citations); search query"}
{"id": "2609.26313", "title": "SafeLoop: Risk-Aware Rollback for Vision-Language-Action Manipulation", "authors": ["Zeyu Lou", "Tianran Zhang", "Xinquan Yue", "Ya Jing", "Chenyang Si"], "year": 2026, "venue": "IROS 2026", "type": "paper", "url": "https://arxiv.org/abs/2609.26313", "pdf_url": "https://arxiv.org/pdf/2609.26313", "code_url": "https://github.com/Loule0-0/SafeLoop/tree/release/safeloop", "clusters": ["C7", "C8"], "keywords": ["SafeLoop", "time-to-hazard", "rollback", "risk predictor"], "model_class": "VLA", "techniques": ["risk predictor (probability + time-to-hazard)", "checkpoint/rollback controller"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~70% fewer hazard cases at base control rate", "metric_measures": "other", "benchmarks": ["LIBERO (24 tasks)", "3 real tasks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "External wrapper predicting collision/drop probability and time-to-hazard, choosing to continue, checkpoint or roll back in joint space, without changing the VLA or its control rate.", "relevance": 5, "relevance_reason": "Fallback/rollback with explicit time-to-hazard.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure prediction generative robot policy"}
{"id": "2609.33218", "title": "Scope-WM: Scoped Computation for Efficient Visual World Models", "authors": ["Chunzheng Li", "Zesheng Jia", "Hongda Zhang", "Jiaying Tang", "Yuntian Wang", "Siao Liu", "Jin Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33218", "pdf_url": "https://arxiv.org/pdf/2609.33218", "code_url": "", "clusters": ["C8"], "keywords": ["scoped computation", "token selection", "MPC search reuse"], "model_class": "world model", "techniques": ["distilled action-conditioned token selector", "sparse foreground dynamics", "reuse of good action sequences in MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Push-T"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Restricts dynamics computation to prediction-relevant tokens chosen by a small distilled selector and reuses promising action sequences across MPC iterations, cutting latency and peak memory for planning.", "relevance": 5, "relevance_reason": "Latency/memory-efficient world model planning for constrained platforms.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2601.20262", "title": "Shallow-π: Knowledge Distillation for Flow-based VLAs", "authors": ["Boseong Jeon", "Yunho Choi", "Taehan Kim", "Yun-Ho Choi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.20262", "pdf_url": "https://arxiv.org/pdf/2601.20262", "code_url": "", "clusters": ["C1", "C2", "C3", "C5", "C9"], "keywords": ["layer distillation", "flow VLA", "Jetson", "layer reduction", "Jetson Thor", "humanoid", "Jetson Orin", "18→6 layers"], "model_class": "VLA", "techniques": ["distillation", "depth reduction", "knowledge distillation", "depth reduction 18->6 layers", "depth reduction of backbone and action head"], "base_model": "pi0/pi0.5-style flow VLA (18 -> 6 layers)", "params": "", "hardware": "Jetson Orin, Jetson Thor", "reported_metric": ">2x faster inference, 18->6 layers, <1% success drop", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Distills both the VLM backbone and flow action head of a pi-style VLA from 18 to 6 layers and validates onboard on Jetson Orin and Thor across robots including humanoids.", "relevance": 5, "relevance_reason": "Depth distillation of pi-style VLAs with onboard deployment.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; search: pruning vision-language-action; snowball from 2510.26742 (citations); venue sweep: arXiv API VLA AND Jetson"}
{"id": "2604.05656", "title": "SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation", "authors": ["Wuyang Luan", "Junhui Li", "Weiguang Zhao", "Wenjian Zhang", "Tieru Wu", "Rui Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.05656", "pdf_url": "https://arxiv.org/pdf/2604.05656", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["one-step flow", "self-distillation", "consistency", "one-step", "flow-matching VLA"], "model_class": "VLA", "techniques": ["one-step generation", "self-distillation", "step distillation", "consistency training"], "base_model": "pi0.5 (3B), SmolVLA (500M)", "params": "", "hardware": "", "reported_metric": "end-to-end latency 274 ms -> 83 ms (pi0.5); 9.6x denoising speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Self-distills flow-matching VLAs to one function evaluation by mixing flow targets with two-step Euler shortcut targets, with no teacher or architecture change.", "relevance": 5, "relevance_reason": "Distillation of flow VLAs' sampling.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; search: pruning vision-language-action"}
{"id": "2603.19418", "title": "Speculative Policy Orchestration: A Latency-Resilient Framework for Cloud-Robotic Manipulation", "authors": ["C. Nguyen", "Shu-Tong Jin", "Florian T. Pokorny", "Erik Elmroth", "Chanh Nguyen", "Shutong Jin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.19418", "pdf_url": "https://arxiv.org/pdf/2603.19418", "code_url": "", "clusters": ["C2", "C5", "C6", "C7", "C8", "C9"], "keywords": ["cloud robotics", "network jitter", "speculative waypoints", "edge buffer", "speculative execution", "cloud world model", "speculative policy orchestration", "epsilon-tube verifier", "speculative policy", "command starvation"], "model_class": "world model", "techniques": ["cloud world-model speculation", "edge verifier", "adaptive horizon", "speculative waypoint prefetch", "edge-side verifier", "speculative action orchestration"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">60% reduction in network-induced idle time vs blocking remote inference", "metric_measures": "other", "benchmarks": ["RLBench", "RLBench (emulated network delay)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Cloud world model pre-computes future waypoints streamed to an edge buffer; an edge epsilon-tube verifier bounds execution error and the prefetch depth adapts to tracking error, decoupling control rate from network RTT.", "relevance": 5, "relevance_reason": "Latency-resilient cloud-edge control scheme for manipulation.", "foundational_technique": false, "verified": true, "found_via": "web: TPU inference robot policy Gemini Robotics latency; search query; arXiv search: \"network latency\" robot policy learned control; venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2602.08245", "title": "STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction", "authors": ["Jin-Hao Li", "Yuxuan Cong", "Ying-Qiao Wang", "Hao Xia", "Shan Huang", "Yijia Zhang", "Ningyi Xu", "Guohao Dai"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.08245", "pdf_url": "https://arxiv.org/pdf/2602.08245", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion acceleration", "latency"], "model_class": "diffusion/flow policy", "techniques": ["warm-start", "consistency prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboMimic", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "STEP builds spatiotemporally consistent warm-start actions so diffusion policies need very few denoising steps while keeping quality.", "relevance": 5, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.00391", "title": "The Gate, Not the Cache: Gate Provenance Bounds the Closed-Loop Reliability of Training-Free VLA Token Skipping", "authors": ["Qi Luo", "Shuaijun Liu", "Hao Zhao", "Kunlin Li", "Xiaobo Wang", "Ningxing Su", "Dongsheng Wang", "Yun Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00391", "pdf_url": "https://arxiv.org/pdf/2608.00391", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["token skipping reliability", "gate provenance", "KV refresh", "token skipping", "closed-loop reliability", "error compounding"], "model_class": "VLA", "techniques": ["token skipping", "KV cache reuse", "token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "success 0.68 (reuse) / 0.31 (deletion) vs dense 1.00 at skip ratio 0.9; recovers to 0.98 with refresh", "metric_measures": "other", "benchmarks": ["LIBERO", "LIBERO-Object"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Shows training-free token skipping compounds errors across control steps when gates come from previous accelerated passes, a failure invisible to action-level detectors; a dense refresh during chunk execution restores success.", "relevance": 5, "relevance_reason": "Demonstrates open-loop/per-step accuracy metrics miss closed-loop failure of efficiency techniques.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk latency VLA; search query: VLA acceleration evaluation"}
{"id": "2609.34599", "title": "The Low-Rank Structure of VLA Reinforcement Learning", "authors": ["Minjae Oh", "Yoonah Park", "Jongwon Lim", "Yohan Jo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34599", "pdf_url": "https://arxiv.org/pdf/2609.34599", "code_url": "", "clusters": ["C3"], "keywords": ["low-rank RL updates", "timestep modules", "GR00T", "pi0.5"], "model_class": "VLA", "techniques": ["analysis", "module-targeted RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Finds RL post-training of flow VLAs produces low-rank updates concentrated in small timestep modules of the action expert.", "relevance": 5, "relevance_reason": "Suggests RL fine-tuning can be restricted to tiny parameter subsets.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.28529", "title": "The Speedup Paradox: Rethinking Inference Speed-Quality Trade-off in Embodied Tasks", "authors": ["Yujin Wang", "Junli Chen", "Yixuan Li", "Shunan Dong", "Huazhong Yang", "Yongpan Liu", "Hongyang Jia", "Yu-Jin Wang", "Hong-Yang Jia", "Wang, Yujin", "Chen, Junli", "Li, Yixuan", "Dong, Shunan", "Yang, Huazhong", "Liu, Yongpan", "Jia, Hongyang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.28529", "pdf_url": "https://arxiv.org/pdf/2606.28529", "code_url": "", "clusters": ["C1", "C2", "C4", "C5", "C7", "C9"], "keywords": ["speedup paradox", "task-level efficiency", "closed-loop effects", "lossy inference", "TISED", "task-level effect decomposition", "static vs dynamic tasks", "hardware dependence", "speed-quality trade-off", "hardware configuration"], "model_class": "VLA", "techniques": ["analysis framework", "analysis of quantization, pruning, async inference", "quantization", "pruning", "asynchronous inference", "analytical decomposition (TISED)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["static and dynamic embodied tasks"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Analytical framework decomposing how lossy inference optimizations affect task-level outcomes: on static tasks they can lengthen completion time despite lower per-step latency; on dynamic tasks moderate lossiness can raise success above baseline; sweet spots shift with hardware.", "relevance": 5, "relevance_reason": "Directly addresses how per-step efficiency translates (or not) to task-level performance.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; snowball from 2506.07530; WebSearch: action chunk size inference latency trade-off; snowball from 2602.18397 (citations)"}
{"id": "2609.32591", "title": "Think Fast, Plan Selectively: Adaptive Deliberation for Efficient Data-Driven MPC", "authors": ["Yi Xian Goh", "Sze Jue Yang", "Hao Luan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.32591", "pdf_url": "https://arxiv.org/pdf/2609.32591", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["adaptive deliberation", "System 1/System 2", "Fast-TD-MPC"], "model_class": "world model", "techniques": ["routing between policy prior and MPC planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to ~4x faster inference across 103 tasks", "metric_measures": "relative latency", "benchmarks": ["103 continuous control tasks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Routes each step either to a cheap policy or to full world-model MPC, planning only where needed, for up to 4x faster control with robustness under disturbances.", "relevance": 5, "relevance_reason": "Adaptive compute for world-model planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.24682", "title": "Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies", "authors": ["Trung Dao", "Sankalp Yamsani", "Jaden Park", "Joohyung Kim", "Yong Jae Lee"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24682", "pdf_url": "https://arxiv.org/pdf/2609.24682", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["representation distillation", "world model teacher", "cached features", "feature distillation", "world model to VLA", "cached teacher features"], "model_class": "VLA", "techniques": ["feature distillation", "feature-alignment loss to cached frozen world-model features", "projector discarded after training"], "base_model": "", "params": "0.8B", "hardware": "consumer RTX 5090", "reported_metric": "32 ms, 1.86 GB inference | deployed policy runs in 32 ms and 1.86 GB (identical cost to undistilled baseline)", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboCasa-GR1", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Transfers world-model physical grounding into a compact VLA with a single feature-alignment term against precomputed world-model features, adding zero deployment cost (32 ms, 1.86 GB on RTX 5090).", "relevance": 5, "relevance_reason": "Gets world-model benefits without runtime world-model cost.", "foundational_technique": false, "verified": true, "found_via": "web search: distilling VLA into small student; snowball from 2602.15922 (citations)"}
{"id": "10.1145/3770855.3817651", "title": "Towards Efficient Embodied Reasoning: Mixture-of-Depth Compute Allocation for Vision-Language-Action Model", "authors": ["Weiying Xie", "Qingcheng Zeng", "Zihan Meng", "Jiayu Tian", "Sibo He", "Danian Yang", "Jie Du", "Yun-Ke Wang", "Daixun Li", "Heng-Yi Wang", "Jitao Ma", "Leyuan Fang", "Yun-Song Li"], "year": 2026, "venue": "Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2", "type": "paper", "url": "https://doi.org/10.1145/3770855.3817651", "pdf_url": "https://doi.org/10.1145/3770855.3817651", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["rate-distortion", "onboard inference", "Jetson"], "model_class": "VLA", "techniques": ["mixture-of-depth", "token-wise depth allocation"], "base_model": "", "params": "", "hardware": "NVIDIA Jetson Orin", "reported_metric": "~38% lower latency, 86% fewer FLOPs", "metric_measures": "relative latency", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "MoDeVLA allocates network depth per token based on an action-aligned attribution analysis, cutting VLA compute while preserving control-relevant features.", "relevance": 5, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2608.09125", "title": "Trajectory Divergence Horizon Decision for Reliable Dual-Arm Surgical Subtask Manipulation", "authors": ["Mingwu Su", "Guankun Wang", "Jinsong Lin", "Rulin Zhou", "Ziyi Hao", "Zhiwei Fang", "Huxin Gao", "Jiewen Lai", "Jiazheng Wang", "Fan Zhang", "Hongliang Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09125", "pdf_url": "https://arxiv.org/pdf/2608.09125", "code_url": "", "clusters": ["C7"], "keywords": ["open-loop chunk truncation", "surgical robotics"], "model_class": "VLA", "techniques": ["adaptive execution horizon", "sample divergence", "replanning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "success 55%->80% (tissue), 55%->60% (needle)", "metric_measures": "other", "benchmarks": ["real dual-arm surgical benchmark"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Estimates step-wise reliability of a flow-matching VLA chunk from divergence between two noise-perturbed samples and truncates execution with a dual threshold to trigger replanning; evaluated on a dual-arm surgical benchmark.", "relevance": 5, "relevance_reason": "A concrete chunk-interruption rule driven by sample divergence.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"open-loop\" AND \"vision-language-action\" AND (execution OR monitor*)"}
{"id": "2607.27205", "title": "TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM", "authors": ["Hengyi Xie", "Chenfei Yao", "Xianjin Wu", "Yingying Zhu", "Dingkang Liang", "Xiang Bai", "Han Ding", "Heng-Yi Xie", "Chen-Fei Yao", "Xian Wu", "Xuan Xi", "Yi-Ping Tang", "Di Xu", "Ding-Kang Liang", "Xie, Hengyi", "Yao, Chenfei", "Wu, Xianjin", "Zhu, Yingying", "Liang, Dingkang", "Bai, Xiang", "Ding, Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.27205", "pdf_url": "https://arxiv.org/pdf/2607.27205", "code_url": "https://github.com/H-EmbodVis/TurboVLA", "clusters": ["C1", "C2", "C4", "C9"], "keywords": ["LLM-free VLA", "V+L->A", "compact decoder", "compact VLA", "no LLM", "0.2B", "TurboVLA", "VRAM"], "model_class": "VLA", "techniques": ["LLM-free architecture", "small backbone", "compact architecture"], "base_model": "", "params": "0.2B / 0.4B", "hardware": "RTX 4090", "reported_metric": "31.2 ms latency (~32 Hz), 0.9 GiB VRAM | 31.2 ms latency, 0.9 GiB VRAM, 97.6% LIBERO", "metric_measures": "open-loop chunk rate", "benchmarks": ["LIBERO", "RoboTwin 2.0", "RoboTwin"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Drops the LLM-centric pathway, fusing separately encoded vision and language directly into a compact action decoder; 0.2B model reaches 97.6% on LIBERO.", "relevance": 5, "relevance_reason": "Example of joint success/latency/memory reporting; C1 primary.", "foundational_technique": false, "verified": true, "found_via": "search query: tiny VLA; snowball from 2506.07530; WebSearch: embodied-efficiency-bench; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2605.08168", "title": "Understanding Asynchronous Inference Methods for Vision-Language-Action Models", "authors": ["Ayoub Agouzoul"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.08168", "pdf_url": "https://arxiv.org/pdf/2605.08168", "code_url": "https://github.com/TheAyos/async-vla-inference", "clusters": ["C1", "C2", "C4", "C6", "C7"], "keywords": ["asynchronous inference comparison", "RTC", "VLASH", "A2C2", "inference delay sweep", "asynchronous inference", "benchmark comparison", "benchmark"], "model_class": "VLA", "techniques": ["benchmarking", "real-time chunking", "training-time delay simulation", "future-state conditioning", "residual correction", "asynchronous inference", "comparative study"], "base_model": "SmolVLA; MLPMixer", "params": "", "hardware": "", "reported_metric": "A2C2 >90% solve rate up to delay d=8", "metric_measures": "other", "benchmarks": ["Kinetix", "LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Controlled head-to-head comparison of inference-time RTC, training-time RTC, VLASH and A2C2 across delays up to 20 steps, finding residual correction strongest and training-time RTC most robust among training-based methods.", "relevance": 5, "relevance_reason": "Benchmarks delay-robustness methods relevant to remote inference.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; search query: VLA benchmark latency; search query; arXiv search: \"inference delay\" robot policy"}
{"id": "2609.22293", "title": "Validating, Not Sampling: Region-Level Robustness of Vision-Language and Vision-Language-Action Models", "authors": ["Bogdan Aron", "Christopher Brix", "Benedikt Brückner", "Yanghao Zhang", "Panagiotis Kouvaros", "Alessio Lomuscio"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22293", "pdf_url": "https://arxiv.org/pdf/2609.22293", "code_url": "", "clusters": ["C7"], "keywords": ["H2V-M", "GR00T", "OpenVLA", "pi"], "model_class": "VLA", "techniques": ["robustness validation", "formal verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Sound region-level robustness validation (H2V-M) of VLMs and VLAs over continuous brightness and camera-rotation ranges; finds VLA actions can change under perturbations as small as 1 degree.", "relevance": 5, "relevance_reason": "First region-level (not sampled) validation of VLAs; directly tests certifiability of actions.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (small OR compact OR tiny) AND \"vision-language-action\""}
{"id": "2606.21386", "title": "VLA-FAIL: Efficient Task Failure Detection for Finetuned Vision-Language-Action Models", "authors": ["Florian Seligmann", "Emiliyan Gospodinov", "Enes Ulas Dincer", "Gerhard Neumann"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.21386", "pdf_url": "https://arxiv.org/pdf/2606.21386", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["VLA-FAIL", "Mahalanobis distance", "action chunk consistency", "AUCPDT"], "model_class": "VLA", "techniques": ["last-layer Mahalanobis OOD", "chunk consistency"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Low-overhead failure detection for finetuned VLAs combining last-layer Mahalanobis OOD scores with consistency of overlapping action chunks, plus a metric jointly scoring accuracy and detection latency.", "relevance": 5, "relevance_reason": "Efficient runtime monitoring suitable for on-robot use.", "foundational_technique": false, "verified": true, "found_via": "arXiv API all:\"action chunking\" AND latency"}
{"id": "2605.06175", "title": "VLA-GSE: Boosting Parameter-Efficient Fine-Tuning in VLA with Generalized and Specialized Experts", "authors": ["Yuhua Jiang", "Junjie Lu", "Xinyao Qin", "Xiaoyu Chen", "Kaixin Wang", "Feifei Gao", "Li Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.06175", "pdf_url": "https://arxiv.org/pdf/2605.06175", "code_url": "https://github.com/YuhuaJiang2002/VLA-GSE", "clusters": ["C3"], "keywords": ["PEFT", "mixture of experts", "spectral initialization", "forgetting"], "model_class": "VLA", "techniques": ["LoRA-style experts", "SVD initialization", "MoE adapters"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.51% trainable parameters; 81.2% LIBERO-Plus zero-shot", "metric_measures": "other", "benchmarks": ["LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Spectrally splits frozen backbone weights into shared and routed expert adapters, improving PEFT adaptation of VLAs while retaining VLM knowledge.", "relevance": 5, "relevance_reason": "New PEFT architecture specific to VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: parameter-efficient fine-tuning vision-language-action model arXiv"}
{"id": "2609.18663", "title": "VLA-ULAP: Interleaving Cloud VLA Calls with Ultra-Lightweight Local Action Prediction at the Edge", "authors": ["Deyu Cao", "Ryuji Oi", "Kosuke Matsushima", "Yuxuan Pan", "Ziheng Wang", "Daichi Fujiki", "Atsutake Kosuge", "De-Yu Cao", "Zi-Heng Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18663", "pdf_url": "https://arxiv.org/pdf/2609.18663", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C6", "C7", "C9"], "keywords": ["cloud-edge interleaving", "ultra-lightweight predictor", "energy", "edge-cloud split", "energy per inference", "local action predictor", "VLA call skipping", "cloud-edge", "Jetson Orin Nano", "call skipping", "cloud VLA", "call reduction", "VLA-ULAP", "ultra-lightweight local predictor", "cloud VLA + local predictor", "ultra-lightweight"], "model_class": "VLA", "techniques": ["reduced VLA invocation", "edge-cloud split", "edge-cloud offloading", "lightweight local predictor", "ultra-lightweight local predictor", "remote VLA call interleaving", "interleaved remote/local inference", "lightweight local policy", "local lightweight action predictor interleaved with remote VLA", "interleaved remote VLA calls with 7.4M local predictor"], "base_model": "GR00T, pi0.5, VLA-JEPA", "params": "~7.4M (local predictor)", "hardware": "Jetson Orin Nano (local); RTX A6000 (remote)", "reported_metric": "local predictor 20.7 ms and 0.122 J per inference; 45.3-77.3% fewer VLA calls | 20.7 ms and 0.122 J per ULAP inference vs 289.3 ms and 40.46 J for GR00T on A6000", "metric_measures": "relative latency", "benchmarks": ["simulation", "SO-101 real robot", "LIBERO", "LIBERO-Safety", "real robot (SO-101)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Interleaves remote VLA calls with a tiny local predictor, removing 45-77% of VLA calls while keeping 95-98.5% of success; reports idle-subtracted energy and energy per successful episode.", "relevance": 5, "relevance_reason": "Reports energy per inference and per successful episode with hardware — a model for energy-per-action metrics.", "foundational_technique": false, "verified": true, "found_via": "search query: Jetson VLA; WebSearch: VLA inference latency energy profiling Jetson; web: Jetson Thor NVFP4 FP8 humanoid onboard; search query; arXiv search: \"latency\" \"safety\" \"vision-language-action\"; venue sweep: arXiv API VLA AND Jetson"}
{"id": "2606.08094", "title": "vla.cpp: A Unified Inference Runtime for Vision-Language-Action Models", "authors": ["Khanh D. Nguyen", "Hung T. Ho", "Chinh T. Nguyen", "Thanh Q. Duong", "Linh D. Le", "Duy M. H. Nguyen", "Vien A. Ngo", "An T. Le", "Khanh-Binh Nguyen", "C. T. Nguyen", "Thanh Duong", "Linh Le", "D. M. H. Nguyen", "V. Ngo", "A. T. Le"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08094", "pdf_url": "https://arxiv.org/pdf/2606.08094", "code_url": "https://github.com/VinRobotics/vla.cpp", "clusters": ["C1", "C2", "C3", "C4", "C5", "C9"], "keywords": ["C++ runtime", "ternary kernels", "Jetson Orin Nano", "ternary kernel", "BitVLA", "vla.cpp", "cross-hardware profiling", "numerical checks", "llama.cpp-style", "ternary tensor-core kernel", "Apple", "Intel", "eleven VLAs"], "model_class": "VLA", "techniques": ["inference runtime", "ternary kernels", "ternary tensor-core kernels", "quantized inference runtime", "solver-step computation reuse", "unified inference runtime", "custom kernels", "computation reuse across solver steps", "unified runtime", "ternary tensor-core kernel"], "base_model": "11 VLAs incl. BitVLA, SmolVLA", "params": "", "hardware": "NVIDIA (Jetson Orin Nano 8GB, AGX Orin, RTX 3060), Apple, Intel", "reported_metric": "ternary kernel 4.0-4.6x over CUDA-core baseline for BitVLA | ternary kernel 4.0-4.6x faster than CUDA-core baseline for BitVLA", "metric_measures": "relative latency", "benchmarks": ["LIBERO-Object", "real robot", "real robot (UR10e, ALOHA)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Unified PyTorch-free runtime for 11 VLAs profiled on NVIDIA, Apple and Intel hardware; a SmolVLA case links positional-index precision to gripper errors, arguing fixed-input numerical checks must accompany rollouts.", "relevance": 5, "relevance_reason": "Deployment runtime and kernels for VLAs on heterogeneous edge hardware.", "foundational_technique": false, "verified": true, "found_via": "search query: Jetson VLA; snowball from 2506.07530; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: arXiv API VLA AND Jetson"}
{"id": "2609.24274", "title": "vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation", "authors": ["Khanh D. Nguyen", "Hoang M. Truong", "An T. Le", "Khanh Duy Nguyen", "A. T. Le"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24274", "pdf_url": "https://arxiv.org/pdf/2609.24274", "code_url": "", "clusters": ["C1", "C2", "C3", "C5"], "keywords": ["CPU inference", "SIMD", "Raspberry Pi", "Raspberry Pi 5", "int8", "SIMD micro-kernels", "thermal soak", "action supply rate"], "model_class": "other", "techniques": ["SIMD kernels", "int8 quantization", "text caching", "computation reuse", "int8", "cached text features"], "base_model": "IMPACT (ACT-based), SmolVLA and others", "params": "", "hardware": "four CPUs incl. Raspberry Pi 5", "reported_metric": "33.5 actions/s fp32, 81.2 actions/s int8 on Raspberry Pi 5; ~1.4x over compiled PyTorch | ~1.4x median over compiled PyTorch; IMPACT on Raspberry Pi 5 after 90 s thermal soak: 33.5 actions/s fp32, 81.2 actions/s int8", "metric_measures": "throughput", "benchmarks": ["LIBERO", "SO-101", "UR10e", "real robot", "real robot (SO-101, UR10e)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "CPU-only inference engine with SIMD micro-kernels that runs language-conditioned policies without a GPU, distinguishing action supply rate from feedback frequency and measuring under thermal throttling.", "relevance": 5, "relevance_reason": "GPU-free edge deployment with kernel optimization and thermally-realistic measurement.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; snowball from 2506.07530; snowball from 2602.18397 (citations)"}
{"id": "2609.25376", "title": "VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models", "authors": ["Jiu-Yi Xu", "Qing Jin", "Mei-Da Chen", "Song Wang", "Yang Sui", "Yang-Ming Shi", "Jiuyi Xu", "Meida Chen", "Yangming Shi"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.25376", "pdf_url": "https://arxiv.org/pdf/2609.25376", "code_url": "https://github.com/jiuyixu25/VLAQuantBench", "clusters": ["C1", "C2", "C4", "C5", "C9"], "keywords": ["closed-loop evaluation", "W4A4", "calibration", "layer sensitivity", "VLAQuantBench", "post-training quantization", "PTQ benchmark", "numerical formats"], "model_class": "VLA", "techniques": ["post-training quantization", "calibration", "mixed precision"], "base_model": "pi0; pi0.5; OpenVLA-OFT; X-VLA", "params": "", "hardware": "", "reported_metric": "409 runs / 94,574 simulated episodes | pi0.5 W4A4 RTN: success 7.0% to 70.5% by widening action-head subset", "metric_measures": "other", "benchmarks": ["LIBERO", "real robot", "other sim benchmarks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Controlled closed-loop PTQ study (409 runs, ~94.6k episodes) showing quantization outcomes depend on the interaction of layer scope, format and calibration rather than universal sensitivity rules; one tiny output projection governs OpenVLA-OFT recovery.", "relevance": 5, "relevance_reason": "Benchmark for closed-loop evaluation of a compression technique with statistical intervals.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.07530; search query: VLA quantization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2609.22462", "title": "VLPSA: Vision-Language-Poisson-Safe Actions for Full-Body Safety of Learned Policies", "authors": ["Meg Wilkinson", "Emily Fourney", "Joel W. Burdick", "Aaron D. Ames"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22462", "pdf_url": "https://arxiv.org/pdf/2609.22462", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["Poisson safety function", "CBF-QP", "full-body safety", "SafeLIBERO", "Poisson safety functions", "full-body collision avoidance", "safety filter"], "model_class": "VLA", "techniques": ["control barrier function safety filter", "Boolean CBF composition", "online PSF synthesis", "Poisson safe-set filter"], "base_model": "pi0.5", "params": "", "hardware": "Franka FR3", "reported_metric": "collision avoidance 23.1% -> 91.2% vs base pi0.5", "metric_measures": "other", "benchmarks": ["SafeLIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Builds Poisson safety functions online from perception to obtain a full-body CBF enforced by a QP filter on VLA actions, using dual-resolution composition for real-time rates; runs without retraining the VLA.", "relevance": 5, "relevance_reason": "Real-time CBF filter for VLAs (Ames lab).", "foundational_technique": false, "verified": true, "found_via": "WebSearch + arXiv search: \"control barrier function\" \"vision-language-action\"; venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2512.11891", "title": "VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer", "authors": ["Songqiao Hu", "Zeyi Liu", "Shuang Liu", "Jun Cen", "Zihan Meng", "Shihefeng Wang", "Xiang Li", "Xiao He"], "year": 2026, "venue": "IROS 2026", "type": "paper", "url": "https://arxiv.org/abs/2512.11891", "pdf_url": "https://arxiv.org/pdf/2512.11891", "code_url": "", "clusters": ["C7"], "keywords": ["AEGIS", "safety constraint layer", "CBF", "SafeLIBERO benchmark"], "model_class": "VLA", "techniques": ["plug-and-play CBF layer"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">50% better obstacle avoidance, ~+10% task success", "metric_measures": "other", "benchmarks": ["SafeLIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a plug-and-play CBF-based safety constraint layer to existing VLAs and introduces the SafeLIBERO benchmark for collision-critical manipulation.", "relevance": 5, "relevance_reason": "Introduces the SafeLIBERO benchmark and a CBF layer for VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"control barrier function\" \"vision-language-action\""}
{"id": "2606.13672", "title": "WEAVER, Better, Faster, Longer: An Effective World Model for Robotic Manipulation", "authors": ["Arnav Kumar Jain", "Yilin Wu", "Jesse Farebrother", "Gokul Swamy", "Andrea Bajcsy"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.13672", "pdf_url": "https://arxiv.org/pdf/2606.13672", "code_url": "", "clusters": ["C8"], "keywords": ["multi-view world model", "policy evaluation", "test-time planning", "reward prediction"], "model_class": "world model", "techniques": ["latent flow-matching prediction", "joint reward/value prediction", "memory design"], "base_model": "pi0.5 (policy improved)", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Multi-view latent world model designed for fidelity, long-horizon consistency and speed; used on hardware for policy evaluation (rho=0.87), policy improvement (+38% over pi0.5) and test-time planning.", "relevance": 5, "relevance_reason": "Efficiency is an explicit design goal alongside fidelity for planning/evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.04233", "title": "What Are We Actually Benchmarking in Robot Manipulation?", "authors": ["Tianchong Jiang", "Xiangshan Tan", "Samuel Wheeler", "Luzhe Sun", "Tewodros W. Ayalew", "Matthew Walter"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.04233", "pdf_url": "https://arxiv.org/pdf/2606.04233", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark audit", "statistical significance", "shortcut solvability", "overfitting"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "CALVIN", "SimplerEnv", "RoboCasa", "RoboTwin 2.0"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Proposes four diagnostics (shortcut solvability, significance, creeping overfitting, data-source dependence) and audits five manipulation benchmarks; LIBERO and CALVIN fail several, a 0.09B no-language probe nearly matches SOTA on LIBERO and most gains lack significance.", "relevance": 5, "relevance_reason": "Core methodology paper on benchmark validity; directly informs how efficiency claims should be evaluated.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2609.14146", "title": "When Faster VLA Deployment Changes Closed-Loop Behavior: Task Success-Latency Analysis of SmolVLA Across PyTorch and ONNX Variants", "authors": ["Rafiqul Islam"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.14146", "pdf_url": "https://arxiv.org/pdf/2609.14146", "code_url": "https://github.com/rafiqul713/smolvla-libero-onnx", "clusters": ["C2", "C4", "C5", "C9"], "keywords": ["ONNX", "SmolVLA", "deployment artifacts", "INT8", "p99 latency", "ONNX Runtime", "deployment artifact audit", "closed-loop success", "RTX 2060", "deployment changes behavior", "success-latency"], "model_class": "VLA", "techniques": ["runtime conversion", "FP16/INT8 export", "ONNX export", "FP16", "INT8 (requested)", "graph audit", "AMP"], "base_model": "SmolVLA (smolvla_libero)", "params": "", "hardware": "RTX 2060 (6 GB)", "reported_metric": "p99 latency 1181 ms -> 532-601 ms | p99 1181 ms (PyTorch+AMP) vs 601/532 ms (ONNX FP16/INT8-requested)", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "LIBERO-Spatial", "LIBERO-Object", "LIBERO Spatial", "LIBERO Object", "LIBERO Spatial/Object"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Shows ONNX export of SmolVLA halves p99 latency but drops LIBERO-Spatial success from 70% to ~40%, and that the 'INT8' artifact was actually FP32; recommends jointly reporting latency, artifact inspection and closed-loop success.", "relevance": 5, "relevance_reason": "Direct evidence that deployment optimizations silently change closed-loop behavior; p99 reporting.", "foundational_technique": false, "verified": true, "found_via": "search: quantization vision-language-action; search query: VLA latency; snowball from 2606.08094 (citations); venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2605.06222", "title": "When to Trust Imagination: Adaptive Action Execution for World Action Models", "authors": ["Rui Wang", "Yue Zhang", "Canyang Chen", "Jiehong Lin", "Zhongrui Wang", "Xiaojuan Qi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.06222", "pdf_url": "https://arxiv.org/pdf/2605.06222", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["FFDC", "replanning trigger"], "model_class": "world-action model", "techniques": ["future-reality verification", "adaptive execution horizon"], "base_model": "", "params": "", "hardware": "Unitree G1D (real)", "reported_metric": "-69.10% planning calls, -33.88% completion time (RoboTwin rand.); -74.06% calls on G1D", "metric_measures": "other", "benchmarks": ["RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight verifier compares predicted actions/visual dynamics with real observations to decide whether to keep executing a WAM plan or replan, combined with mixture-of-horizon training.", "relevance": 5, "relevance_reason": "Direct evidence for verified adaptive chunk execution reducing replanning calls.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=100]"}
{"id": "2609.34300", "title": "When World Models Lie: Adaptive Safety Analysis Under Wrong Imaginations", "authors": ["John Cao", "Somil Bansal"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34300", "pdf_url": "https://arxiv.org/pdf/2609.34300", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["adaptive conformal inference", "wrong imaginations", "latent safety filter"], "model_class": "world model", "techniques": ["online ACI on observed world-model error", "pessimistic value over uncertainty sets"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Calibrates latent safety filters online using directly observed discrepancies between predicted and observed latents, so the filter becomes pessimistic exactly when the world model is wrong.", "relevance": 5, "relevance_reason": "Addresses reliability of imagination-based safety monitors.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "semianalysis-robot-inference-on-device-vs-datacenter-2026", "title": "Where Does a Robot Think — On-Device vs Datacenter Inference", "authors": ["Ivan Chiam", "Gianluca Bencomo", "Zane Fong"], "year": 2026, "venue": "SemiAnalysis newsletter", "type": "blog", "url": "https://newsletter.semianalysis.com/p/a-brain-too-big-to-carry-on-device", "pdf_url": "", "code_url": "", "clusters": ["C5", "C6"], "keywords": ["TCO", "per-robot inference cost", "Jetson Thor", "B300", "battery", "fleet offload", "on-device vs datacenter", "robots per GPU", "jitter"], "model_class": "other", "techniques": ["cost modeling", "TCO analysis", "fleet GPU pooling"], "base_model": "", "params": "", "hardware": "Jetson Thor (40-130 W) vs B300, RTX 6000 Pro", "reported_metric": "96-robot TCO: B300 offload $18.63/h vs Thor on-device $14.97/h; offload ~46% of on-robot cost at industrial utilization; one B300 sustains 12 robots | silicon crossover ~7 robots per GPU (DRAM ~5); B300 ~$0.17/hr/PFLOP vs Jetson Thor ~$0.37/hr/PFLOP at assumed utilization; B300 sustains 12 robots at 500 ms deadline", "metric_measures": "other", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Industry analysis modelling silicon, DRAM and dollar cost of onboard Jetson Thor versus pooled datacenter GPUs for robot fleets, finding offload cheaper beyond a handful of robots per GPU at industrial utilization while flagging jitter and RF constraints.", "relevance": 5, "relevance_reason": "Rare quantitative per-unit inference cost analysis; industry blog, some numbers author-estimated. Only the first 3 of 6 authors are listed on the page.", "foundational_technique": false, "verified": true, "found_via": "web: humanoid onboard compute power budget watts; search query"}
{"id": "2602.15922", "title": "World Action Models are Zero-shot Policies", "authors": ["Seonghyeon Ye", "Yunhao Ge", "Kaiyuan Zheng", "Shenyuan Gao", "Sihyun Yu", "George Kurian", "Suneel Indupuru", "You Liang Tan", "Chuning Zhu", "Jiannan Xiang", "Ayaan Malik", "Kyungmin Lee", "William Liang", "Nadun Ranawaka", "Jiasheng Gu", "Yinzhen Xu", "Guanzhi Wang", "Fengyuan Hu", "Avnish Narayan", "Johan Bjorck", "Jing Wang", "Gwanghyun Kim", "Dantong Niu", "Ruijie Zheng", "Yuqi Xie", "Jimmy Wu", "Qi Wang", "Ryan Julian", "Danfei Xu", "Yilun Du", "Yevgen Chebotar", "Scott Reed", "Jan Kautz", "Yuke Zhu", "Linxi \"Jim\" Fan", "Joel Jang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.15922", "pdf_url": "https://arxiv.org/pdf/2602.15922", "code_url": "https://dreamzero0.github.io/", "clusters": ["C1", "C3", "C4", "C8"], "keywords": ["DreamZero", "video diffusion", "few-shot embodiment adaptation", "world action model", "WAM", "video diffusion policy", "real-time closed-loop"], "model_class": "world-action model", "techniques": ["world action model", "system optimization", "autoregressive video diffusion", "decoupled video/action denoising schedules", "parallelism and KV caching", "quantization", "CUDA kernel tuning"], "base_model": "", "params": "14B", "hardware": "H100 (training); multi-GPU inference", "reported_metric": "7 Hz closed-loop; new embodiment with 30 min play data | 7 Hz closed-loop control; 38x inference speedup from model+system optimizations", "metric_measures": "closed-loop control", "benchmarks": ["real robot", "real robot (AgiBot, Franka, others)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "14B world action model on a pretrained video diffusion backbone that jointly predicts future video and actions, doubling generalization over VLAs; a stack of algorithmic and systems optimizations gives 38x speedup to 7 Hz closed-loop control.", "relevance": 5, "relevance_reason": "Seed; flagship example of making a video-diffusion WAM run in real time.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: few-shot adaptation vision-language-action model new embodiment efficient; seed"}
{"id": "2608.01880", "title": "World Action Models in Real Time: An Empirical Study of Smooth Execution via Asynchronous Deployment", "authors": ["Motubrain Team"], "year": 2026, "venue": "arXiv (Motubrain Team)", "type": "preprint", "url": "https://arxiv.org/abs/2608.01880", "pdf_url": "https://arxiv.org/pdf/2608.01880", "code_url": "", "clusters": ["C2", "C4", "C6", "C7", "C8", "C9"], "keywords": ["world action model", "asynchronous deployment", "empirical study", "temporal alignment", "action blending", "action chunk blending", "WAM", "chunk blending"], "model_class": "world-action model", "techniques": ["asynchronous inference", "action blending", "prefix conditioning", "asynchronous execution", "prefix-conditioned generation", "blending", "async inference", "velocity guidance", "asynchronous execution strategies comparison"], "base_model": "", "params": "", "hardware": "10 Hz bimanual robot", "reported_metric": "", "metric_measures": "", "benchmarks": ["real bimanual tasks", "real robot (dynamic, precision, long-horizon)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Empirical comparison of six async deployment strategies for a world-action model; finds accurate timestamp alignment of observations/predictions/commands is essential and that prefix-conditioned generation gives the best balance of success, speed and smoothness.", "relevance": 5, "relevance_reason": "Practical guidance for running high-latency WAMs in real time.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.07339; arXiv search: \"action chunk\" asynchronous inference; search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2606.09749", "title": "Your Model Already Knows: Attention-Guided Safety Filter for Vision-Language-Action Models", "authors": ["Seongbin Park", "Fan Zhang", "Baharan Mirzasoleiman", "Shahriar Talebi", "Nader Sehatbakhsh"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09749", "pdf_url": "https://arxiv.org/pdf/2606.09749", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["attention-guided safety filter", "CBF", "moving obstacles", "training-free", "attention heads", "in-loop safety filter"], "model_class": "VLA", "techniques": ["VLA attention-head target localization", "CBF filter", "object tracking", "attention-guided localization for CBF-style filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+43% avg over init-time oracle on dynamic variant", "metric_measures": "other", "benchmarks": ["SafeLIBERO (+dynamic)"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reads the intended target from a few VLA attention heads each step, treats the rest of the scene as obstacles and runs a CBF filter plus lightweight tracker, avoiding slow VLM queries in the control loop.", "relevance": 5, "relevance_reason": "Safety filtering at control rate without slow VLM calls.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"control barrier function\" \"vision-language-action\"; venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2607.26055", "title": "πR²: Reactive Real-time Flow Policies", "authors": ["Sungjae Park", "Shubham Tulsiani"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.26055", "pdf_url": "https://arxiv.org/pdf/2607.26055", "code_url": "", "clusters": ["C2", "C3", "C7"], "keywords": ["reactive flow policy", "diffusion forcing", "latency-adaptive schedule", "fast/slow channel"], "model_class": "VLA", "techniques": ["per-position noise schedule", "fast proprio channel", "inpainting of in-flight actions"], "base_model": "GR00T-N1.7", "params": "", "hardware": "A5000 GPU", "reported_metric": "~25 Hz closed-loop replanning (~4x base), fresh observation every 40 ms", "metric_measures": "closed-loop control", "benchmarks": ["sim", "real xArm6+XHand"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Splits conditioning into a fast proprioception channel and a stale vision-language channel and uses a latency-adaptive diffusion-forcing schedule that emits one action per denoising call, letting one model adapt to varying hardware latency.", "relevance": 5, "relevance_reason": "Explicit handling of variable latency with high-rate reactive control.", "foundational_technique": false, "verified": true, "found_via": "arXiv API all:\"action chunking\" AND latency"}
{"id": "2504.16054", "title": "$π_{0.5}$: a Vision-Language-Action Model with Open-World Generalization", "authors": ["Physical Intelligence", "Kevin Black", "Noah Brown", "James Darpinian", "Karan Dhabalia", "Danny Driess", "Adnan Esmail", "Michael Equi", "Chelsea Finn", "Niccolo Fusai", "Manuel Y. Galliker", "Dibya Ghosh", "Lachy Groom", "Karol Hausman", "Brian Ichter", "Szymon Jakubczak", "Tim Jones", "Liyiming Ke", "Devin LeBlanc", "Sergey Levine", "Adrian Li-Bell", "Mohith Mothukuri", "Suraj Nair", "Karl Pertsch", "Allen Z. Ren", "Lucy Xiaoyang Shi", "Laura Smith", "Jost Tobias Springenberg", "Kyle Stachowicz", "James Tanner", "Quan Vuong", "Homer Walke", "Anna Walling", "Haohuan Wang", "Lili Yu", "Ury Zhilinsky"], "year": 2025, "venue": "CoRL 2025 / arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2504.16054", "pdf_url": "https://arxiv.org/pdf/2504.16054", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["pi0.5", "co-training", "hierarchical subtask prediction", "open-world", "hierarchical inference", "open-world generalization"], "model_class": "VLA", "techniques": ["co-training", "hierarchical inference", "action expert", "high-level subtask prediction + low-level action expert"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Physical Intelligence's pi0.5 co-trains on heterogeneous robot, web and high-level subtask data to generalize to new homes; the base for many acceleration studies.", "relevance": 5, "relevance_reason": "Common base model for efficiency work (e.g., VLASH).", "foundational_technique": false, "verified": true, "found_via": "seed; blog sweep: Physical Intelligence"}
{"id": "2511.18082", "title": "ActDistill: General Action-Guided Self-Derived Distillation for Efficient Vision-Language-Action Models", "authors": ["Wencheng Ye", "Tianshi Wang", "Lei Zhu", "Fengling Li", "Guoli Yang", "Hengtao Shen", "Wen Ye", "Tian-Shi Wang", "Guo-Li Yang", "Heng Shen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.18082", "pdf_url": "https://arxiv.org/pdf/2511.18082", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["distillation", "dynamic routing", "layer skipping", "self-derived distillation", "dynamic router", "graph-structured supervision"], "model_class": "VLA", "techniques": ["distillation", "dynamic layer routing", "knowledge distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">50% compute reduction, up to 1.67x speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SIMPLER", "SimplerEnv"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Derives a lightweight student from a VLA teacher and trains a router to execute only needed layers, supervised by hierarchical action-prediction signals.", "relevance": 5, "relevance_reason": "General VLA-to-lightweight distillation framework.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; snowball from 2503.20384"}
{"id": "2512.20276", "title": "ActionFlow: A Pipelined Action Acceleration for Vision Language Models on Edge", "authors": ["Yuntao Dai", "Hang Gu", "Teng Wang", "Qianyu Cheng", "Yifei Zheng", "Zhiyong Qiu", "Lei Gong", "Wenqi Lou", "Xuehai Zhou", "Dai, Yuntao", "Gu, Hang", "Wang, Teng", "Cheng, Qianyu", "Zheng, Yifei", "Qiu, Zhiyong", "Gong, Lei", "Lou, Wenqi", "Zhou, Xuehai", "Qian-Yu Cheng", "Xue-Hai Zhou"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.20276", "pdf_url": "https://arxiv.org/pdf/2512.20276", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C6", "C9"], "keywords": ["cross-request pipelining", "KV ring buffer", "edge", "ActionFlow", "prefill/decode batching", "OpenVLA", "edge VLA serving", "pipelined decoding", "memory-bound"], "model_class": "VLA", "techniques": ["pipelined scheduling", "operator fusion", "cross-request pipelining", "KV cache management", "prefill/decode cross-request pipelining", "pipelined action generation"], "base_model": "OpenVLA-7B", "params": "7B", "hardware": "edge platforms", "reported_metric": "2.55x improvement (reported in abstract) | motivating baseline: VLAs run at 3-5 Hz on edge devices", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Treats consecutive VLA control steps as a pipeline, batching memory-bound decode with compute-bound prefill of the next step and fusing memory ops, to raise OpenVLA frame rate on edge hardware.", "relevance": 5, "relevance_reason": "Hardware-utilization-aware scheduling exploiting memory-bound decode on edge.", "foundational_technique": false, "verified": true, "found_via": "search query: memory VLA efficient long-horizon; WebSearch: closed-loop vs open-loop control frequency VLA; arXiv: vision-language-action AND Jetson/edge/onboard; snowball from 2506.07339; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2503.24278", "title": "AutoEval: Autonomous Evaluation of Generalist Robot Manipulation Policies in the Real World", "authors": ["Zhou, Zhiyuan", "Atreya, Pranav", "Tan, You Liang", "Pertsch, Karl", "Levine, Sergey"], "year": 2025, "venue": "CoRL 2025", "type": "benchmark", "url": "https://arxiv.org/abs/2503.24278", "pdf_url": "https://arxiv.org/pdf/2503.24278", "code_url": "https://github.com/zhouzypaul/auto_eval", "clusters": ["C4"], "keywords": ["AutoEval", "autonomous evaluation", "success classifier", "reset policy", "Bridge WidowX"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "WidowX (BridgeData setup)", "reported_metric": "", "metric_measures": "", "benchmarks": ["AutoEval", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "24/7 real-world evaluation service with a learned success detector and learned reset policy; results track human-run evaluations while cutting human supervision by >99%.", "relevance": 5, "relevance_reason": "Seed real-robot evaluation infrastructure.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2506.07530", "title": "BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation", "authors": ["Hongyu Wang", "Chuyan Xiong", "Ruiping Wang", "Xilin Chen", "Rui-Ping Wang", "Xi-Lin Chen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.07530", "pdf_url": "https://arxiv.org/pdf/2506.07530", "code_url": "https://github.com/ustcwhy/BitVLA", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["1-bit", "ternary weights", "BitNet", "1.58-bit", "edge deployment", "memory-constrained edge", "1-bit VLA"], "model_class": "VLA", "techniques": ["quantization", "distillation", "ternary quantization", "quantization-aware training", "distillation (Quantize-then-Distill)", "1.58-bit quantization", "quantize-then-distill", "native 1-bit training"], "base_model": "BitNet b1.58 2B4T", "params": "~3B", "hardware": "", "reported_metric": "11.0x lower model memory, 4.4x lower end-to-end latency vs OpenVLA-OFT | 11.0x memory reduction, 4.4x end-to-end latency reduction vs OpenVLA-OFT", "metric_measures": "relative memory", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Builds a VLA natively on a ternary (1.58-bit) LLM and compresses the vision encoder to 1.58 bits via a quantize-then-distill QAT stage with a full-precision teacher; matches OpenVLA-OFT while shrinking memory ~11x.", "relevance": 5, "relevance_reason": "Extreme low-bit model targeting memory-constrained edge robot hardware.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; seed; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2506.13456", "title": "Block-wise Adaptive Caching for Accelerating Diffusion Policy", "authors": ["Kangye Ji", "Yuan Meng", "Hanyun Cui", "Ye Li", "Jianbo Zhou", "Shengjia Hua", "Lei Chen", "Zhi Wang", "Han-Yun Cui"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.13456", "pdf_url": "https://arxiv.org/pdf/2506.13456", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["BAC", "diffusion policy", "block-wise caching"], "model_class": "diffusion/flow policy", "techniques": ["feature caching", "training-free acceleration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 3x inference speedup", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free block-wise feature caching for transformer diffusion policies with an adaptive scheduler and error-truncating update order.", "relevance": 5, "relevance_reason": "Caching across denoising steps for faster diffusion policies.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]; snowball round 2: cited by 2512.15773, 2601.12894"}
{"id": "2503.08558", "title": "Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies", "authors": ["Chen Xu", "Tony Khuong Nguyen", "Emma Dixon", "Christopher Rodriguez", "Patrick Miller", "Robert Lee", "Paarth Shah", "Rares Ambrus", "Haruki Nishimura", "Masha Itkina"], "year": 2025, "venue": "RSS 2025", "type": "paper", "url": "https://arxiv.org/abs/2503.08558", "pdf_url": "https://arxiv.org/pdf/2503.08558", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["FAIL-Detect", "sequential OOD detection", "conformal prediction", "flow density estimator"], "model_class": "diffusion/flow policy", "techniques": ["scalar failure signals", "conformal prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Two-stage failure detector for IL policies trained only on successes: distills inputs/outputs into scalar uncertainty signals (best with a flow-based density estimator) and thresholds them with conformal prediction.", "relevance": 5, "relevance_reason": "Seed: failure detection without failure data.", "foundational_technique": false, "verified": true, "found_via": "seed; WebSearch"}
{"id": "2506.13725", "title": "CEED-VLA: Consistency Vision-Language-Action Model with Early-Exit Decoding", "authors": ["Wenxuan Song", "Jiayi Chen", "Pengxiang Ding", "Yuxin Huang", "Han Zhao", "Donglin Wang", "Haoang Li", "Wen-Xuan Song", "Dong-Lin Wang", "Hao-Ang Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.13725", "pdf_url": "https://arxiv.org/pdf/2506.13725", "code_url": "", "clusters": ["C1", "C2", "C9"], "keywords": ["consistency distillation", "Jacobi decoding", "early exit decoding", "early-exit decoding"], "model_class": "VLA", "techniques": ["parallel action decoding", "consistency distillation", "early exit", "consistency distillation for parallel decoding", "early-exit decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">4x inference acceleration", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Trains AR VLAs with consistency distillation so Jacobi iterations converge on several action tokens per step, and relaxes convergence to exit decoding early.", "relevance": 5, "relevance_reason": "Parallel/early-exit decoding for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA early exit; snowball from 2412.01034; venue sweep: arXiv API VLA AND (MoE OR early exit)"}
{"id": "2508.21046", "title": "CogVLA: Cognition-Aligned Vision-Language-Action Model via Instruction-Driven Routing & Sparsification", "authors": ["Wei Li", "Renshan Zhang", "Rui Shao", "Jie He", "Liqiang Nie"], "year": 2025, "venue": "NeurIPS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2508.21046", "pdf_url": "https://arxiv.org/pdf/2508.21046", "code_url": "https://github.com/JiuTian-VL/CogVLA", "clusters": ["C1", "C3"], "keywords": ["instruction-driven routing", "token sparsification", "parallel decoding", "CogVLA", "routing", "sparsification", "post-training cost"], "model_class": "VLA", "techniques": ["token pruning", "FiLM routing", "parallel action decoding", "token routing", "sparsification"], "base_model": "OpenVLA-style", "params": "", "hardware": "", "reported_metric": "2.8x lower inference latency, 2.5x lower training cost vs OpenVLA", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Uses instruction-conditioned FiLM routing to compress visual tokens in the encoder and prune instruction-irrelevant tokens in the LLM, plus bidirectional parallel action decoding.", "relevance": 5, "relevance_reason": "Instruction-aware sparsification for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; arXiv API: abs:\"vision-language-action\" AND abs:\"training cost\""}
{"id": "2501.03575", "title": "Cosmos World Foundation Model Platform for Physical AI", "authors": ["NVIDIA", ":", "Niket Agarwal", "Arslan Ali", "Maciej Bala", "Yogesh Balaji", "Erik Barker", "Tiffany Cai", "Prithvijit Chattopadhyay", "Yongxin Chen", "Yin Cui", "Yifan Ding", "Daniel Dworakowski", "Jiaojiao Fan", "Michele Fenzi", "Francesco Ferroni", "Sanja Fidler", "Dieter Fox", "Songwei Ge", "Yunhao Ge", "Jinwei Gu", "Siddharth Gururani", "Ethan He", "Jiahui Huang", "Jacob Huffman", "Pooya Jannaty", "Jingyi Jin", "Seung Wook Kim", "Gergely Klár", "Grace Lam", "Shiyi Lan", "Laura Leal-Taixe", "Anqi Li", "Zhaoshuo Li", "Chen-Hsuan Lin", "Tsung-Yi Lin", "Huan Ling", "Ming-Yu Liu", "Xian Liu", "Alice Luo", "Qianli Ma", "Hanzi Mao", "Kaichun Mo", "Arsalan Mousavian", "Seungjun Nah", "Sriharsha Niverty", "David Page", "Despoina Paschalidou", "Zeeshan Patel", "Lindsey Pavao", "Morteza Ramezanali", "Fitsum Reda", "Xiaowei Ren", "Vasanth Rao Naik Sabavat", "Ed Schmerling", "Stella Shi", "Bartosz Stefaniak", "Shitao Tang", "Lyne Tchapmi", "Przemek Tredak", "Wei-Cheng Tseng", "Jibin Varghese", "Hao Wang", "Haoxiang Wang", "Heng Wang", "Ting-Chun Wang", "Fangyin Wei", "Xinyue Wei", "Jay Zhangjie Wu", "Jiashu Xu", "Wei Yang", "Lin Yen-Chen", "Xiaohui Zeng", "Yu Zeng", "Jing Zhang", "Qinsheng Zhang", "Yuxuan Zhang", "Qingqing Zhao", "Artur Zolkowski"], "year": 2025, "venue": "arXiv (NVIDIA tech report)", "type": "tech-report", "url": "https://arxiv.org/abs/2501.03575", "pdf_url": "https://arxiv.org/pdf/2501.03575", "code_url": "https://github.com/nvidia-cosmos/cosmos-predict1", "clusters": ["C8"], "keywords": ["world foundation model", "video tokenizer", "physical AI"], "model_class": "world model", "techniques": ["diffusion and autoregressive video world models", "video tokenizers", "post-training for robotics/driving"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "NVIDIA's open platform of pretrained video world foundation models (diffusion and autoregressive), tokenizers, and data curation, intended to be post-trained into robot or driving simulators and policies.", "relevance": 5, "relevance_reason": "Seed; the most widely used open backbone for robot world models and WAMs (Cosmos Policy, DreamDojo, Cosmos-H).", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2509.11480", "title": "Cross-Platform Scaling of Vision-Language-Action Models from Edge to Cloud GPUs", "authors": ["Amir Taherin", "Juyi Lin", "Arash Akbari", "Arman Akbari", "Pu Zhao", "Weiwei Chen", "David Kaeli", "Yanzhi Wang"], "year": 2025, "venue": "Asilomar Conference on Signals, Systems, and Computers 2025", "type": "benchmark", "url": "https://arxiv.org/abs/2509.11480", "pdf_url": "https://arxiv.org/pdf/2509.11480", "code_url": "", "clusters": ["C1", "C3", "C4", "C5", "C6", "C9"], "keywords": ["cross-platform scaling", "edge vs cloud GPUs", "power budgets", "edge vs cloud", "power modes", "throughput", "memory", "cross-platform benchmarking", "edge vs datacenter GPU", "scaling", "edge vs datacenter GPUs", "VLA scaling", "power constraints", "edge to cloud scaling", "latency/throughput/memory"], "model_class": "VLA", "techniques": ["benchmarking", "hardware characterization", "power-budget analysis", "cross-platform benchmarking"], "base_model": "", "params": "", "hardware": "Jetson edge devices under power constraints; datacenter GPUs", "reported_metric": "", "metric_measures": "throughput", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Measures five VLAs on LIBERO for accuracy plus latency, throughput and peak memory across power-capped edge devices and datacenter GPUs; finds nonlinear degradation under power limits and that tokenization/backbone choice dominates throughput.", "relevance": 5, "relevance_reason": "Core hardware-aware characterization of VLAs including power budgets.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA; search query: vision-language-action latency benchmark; arXiv: vision-language-action AND Jetson/edge/onboard; arXiv API: on-device VLA; venue sweep: arXiv API VLA AND benchmark AND latency AND hardware"}
{"id": "2510.08464", "title": "Don't Run with Scissors: Pruning Breaks VLA Models but They Can Be Recovered", "authors": ["Jason Jabbour", "Dong-Ki Kim", "Max Smith", "Jay Patrikar", "Radhika Ghosal", "Youhui Wang", "Ali Agha", "Vijay Janapa Reddi", "Shayegan Omidshafiei"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://gluestick-vla.github.io/", "pdf_url": "https://arxiv.org/pdf/2510.08464", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["pruning", "safety violations", "weight interpolation", "recovery"], "model_class": "VLA", "techniques": ["unstructured/structured pruning", "training-free recovery"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["manipulation", "navigation"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows LLM-style pruning badly degrades VLAs and raises safety violations; GLUESTICK computes a one-shot dense-pruned weight-space correction applied at inference to recover success.", "relevance": 5, "relevance_reason": "Key evidence that pruning hurts VLAs, with training-free fix.", "foundational_technique": false, "verified": true, "found_via": "search: pruning vision-language-action"}
{"id": "2506.10100", "title": "EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models", "authors": ["Yantai Yang", "Yuhao Wang", "Zichen Wen", "Luo Zhongwei", "Chang Zou", "Zhipeng Zhang", "Chuan Wen", "Linfeng Zhang", "Zhongwei Luo", "Zhi-Peng Zhang", "Lin-Feng Zhang"], "year": 2025, "venue": "NeurIPS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2506.10100", "pdf_url": "https://arxiv.org/pdf/2506.10100", "code_url": "", "clusters": ["C1", "C2", "C4"], "keywords": ["training-free acceleration", "layer pruning", "token selection", "diffusion feature caching", "training-free", "visual token selection", "feature caching", "EfficientVLA", "composition"], "model_class": "VLA", "techniques": ["layer pruning", "token pruning", "feature caching", "visual token pruning", "diffusion feature caching"], "base_model": "CogACT", "params": "", "hardware": "", "reported_metric": "1.93x inference speedup, FLOPs reduced to 28.9%, -0.6% success | 1.93x inference speedup, FLOPs to 28.9%, -0.6% success on SIMPLER", "metric_measures": "relative latency", "benchmarks": ["SIMPLER", "SimplerEnv"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free framework stacking three redundancy reductions (language-layer pruning, task-aware token selection, action-head feature caching); a composition case whose speedup is well below its FLOP reduction.", "relevance": 5, "relevance_reason": "Composition of three techniques with FLOPs vs speedup gap.", "foundational_technique": false, "verified": true, "found_via": "seed; web search: VLA pruning layers redundancy; WebSearch: do VLA acceleration techniques compose"}
{"id": "2510.09459", "title": "Failure Prediction at Runtime for Generative Robot Policies", "authors": ["Ralf Römer", "Adrian Kobras", "Luca Worbis", "Angela P. Schoellig"], "year": 2025, "venue": "NeurIPS 2025", "type": "paper", "url": "https://arxiv.org/abs/2510.09459", "pdf_url": "https://arxiv.org/pdf/2510.09459", "code_url": "https://github.com/utiasDSL/fiper", "clusters": ["C7", "C8"], "keywords": ["FIPER", "failure prediction", "random network distillation", "action-chunk entropy", "conformal prediction"], "model_class": "diffusion/flow policy", "techniques": ["OOD detection (RND)", "action-chunk entropy", "conformal calibration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Predicts failures of generative IL policies without failure data by combining embedding-space OOD scores and action-chunk entropy, both conformally calibrated on successful rollouts, raising an alarm when both exceed thresholds.", "relevance": 5, "relevance_reason": "Seed: runtime failure prediction for generative policies.", "foundational_technique": false, "verified": true, "found_via": "seed; WebSearch"}
{"id": "2503.00339", "title": "Falcon: Fast Visuomotor Policies via Partial Denoising", "authors": ["Haojun Chen", "Minghao Liu", "Xiaojian Ma", "Zai-Lin Ma", "Huimin Wu", "Chengdong Ma", "Yuan-Pei Chen", "Yi-Fan Zhong", "Mingzhi Wang", "Qing Li", "Yaodong Yang"], "year": 2025, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2503.00339", "pdf_url": "https://arxiv.org/pdf/2503.00339", "code_url": "", "clusters": ["C1"], "keywords": ["training-free acceleration", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["partial denoising", "warm-start sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2-7x speedup", "metric_measures": "relative latency", "benchmarks": ["48 simulated environments", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Falcon reuses partially denoised actions from previous steps to cut diffusion-policy sampling steps, a training-free plug-in on top of other accelerators.", "relevance": 5, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2506.01953", "title": "Fast-in-Slow: A Dual-System Foundation Model Unifying Fast Manipulation within Slow Reasoning", "authors": ["Hao Chen", "Jiaming Liu", "Chenyang Gu", "Zhuoyang Liu", "Renrui Zhang", "Xiaoqi Li", "Xiao He", "Yandong Guo", "Chi-Wing Fu", "Shanghang Zhang", "Pheng-Ann Heng", "Jia-Ming Liu", "Yan-Dong Guo", "P. Heng"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "preprint", "url": "https://doi.org/10.52202/085713-3276", "pdf_url": "https://arxiv.org/pdf/2506.01953", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["dual system", "parameter sharing", "System 1 inside System 2", "Fast-in-Slow", "shared parameters", "System 1/2", "dual-system", "high-frequency control", "asynchronous"], "model_class": "VLA", "techniques": ["dual-system architecture", "asynchronous frequencies", "partially shared dual-system VLA", "dual-system", "parameter sharing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "117.7 Hz control frequency (chunk 8)", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Fast-in-Slow embeds a fast System-1 action module inside the VLM-based System-2 by partially sharing parameters, running the two at different frequencies so the robot gets high-rate control while keeping VLM reasoning.", "relevance": 5, "relevance_reason": "direct: dual-system design for high-frequency VLA execution", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA; arXiv search: \"dual-system\" \"vision-language-action\" fast slow asynchronous; snowball round 2: cites 2303.04137, 2304.13705; cited by 2510.17111, 2510.24795"}
{"id": "2501.09747", "title": "FAST: Efficient Action Tokenization for Vision-Language-Action Models", "authors": ["Karl Pertsch", "Kyle Stachowicz", "Brian Ichter", "Danny Driess", "Suraj Nair", "Quan Vuong", "Oier Mees", "Chelsea Finn", "Sergey Levine"], "year": 2025, "venue": "RSS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2501.09747", "pdf_url": "https://arxiv.org/pdf/2501.09747", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["FAST", "DCT", "frequency-space tokenization", "BPE", "action tokenization", "autoregressive VLA"], "model_class": "VLA", "techniques": ["action tokenization", "compression", "DCT-based compression tokenization"], "base_model": "pi0 (pi0-FAST)", "params": "", "hardware": "", "reported_metric": "up to 5x less training time vs diffusion VLA", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Compresses action chunks with a discrete cosine transform plus byte-pair encoding so autoregressive VLAs train effectively on high-frequency dexterous data and much faster than diffusion-based pi0.", "relevance": 5, "relevance_reason": "Token-efficient action representation widely used by efficient VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed; blog sweep: Physical Intelligence"}
{"id": "2512.04952", "title": "FASTer: Toward Efficient Autoregressive Vision Language Action Modeling via Neural Action Tokenization", "authors": ["Yicheng Liu", "Shiduo Zhang", "Zibin Dong", "Baijun Ye", "Tianyuan Yuan", "Xiaopeng Yu", "Linqi Yin", "Chenhao Lu", "Junhao Shi", "Luca Jiang-Tao Yu", "Liangtao Zheng", "Tao Jiang", "Jingjing Gong", "Xipeng Qiu", "Hang Zhao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.04952", "pdf_url": "https://arxiv.org/pdf/2512.04952", "code_url": "", "clusters": ["C1"], "keywords": ["action tokenization", "VQ tokenizer", "block-wise autoregressive"], "model_class": "VLA", "techniques": ["learned action tokenizer", "block-wise decoding", "action expert"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learned VQ tokenizer that encodes action chunks as single-channel images for high compression, paired with a block-autoregressive VLA and a small action expert for faster decoding than FAST.", "relevance": 5, "relevance_reason": "Action tokenization for efficient AR VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient"}
{"id": "2502.19645", "title": "Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success", "authors": ["Moo Jin Kim", "Chelsea Finn", "Percy Liang", "Kim, Moo Jin", "Finn, Chelsea", "Liang, Percy"], "year": 2025, "venue": "RSS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2502.19645", "pdf_url": "https://arxiv.org/pdf/2502.19645", "code_url": "https://github.com/moojink/openvla-oft", "clusters": ["C1", "C2", "C3", "C4", "C5", "C9"], "keywords": ["OpenVLA-OFT", "parallel decoding", "action chunking", "continuous actions", "L1 regression", "optimized fine-tuning", "OFT", "throughput"], "model_class": "VLA", "techniques": ["parallel action decoding", "action chunking", "continuous action head", "fine-tuning recipe", "parallel decoding", "continuous actions", "LoRA", "L1 regression", "L1 objective"], "base_model": "OpenVLA-7B", "params": "7B", "hardware": "", "reported_metric": "26x action generation throughput vs OpenVLA; LIBERO avg 97.1%", "metric_measures": "throughput", "benchmarks": ["LIBERO", "ALOHA real robot", "real robot (ALOHA)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Systematic study of how to fine-tune a VLA; the resulting OFT recipe (parallel decoding, chunking, continuous actions, L1 loss) lifts OpenVLA on LIBERO and makes high-frequency bimanual control feasible.", "relevance": 5, "relevance_reason": "Widely cited speed number whose metric type (throughput via chunking) must be distinguished from control latency.", "foundational_technique": false, "verified": true, "found_via": "seed; WebSearch: CALVIN ABC-D efficiency latency comparison; seed lookup; venue sweep: RSS 2025"}
{"id": "2509.04996", "title": "FLOWER: Democratizing Generalist Robot Policies with Efficient Vision-Language-Action Flow Policies", "authors": ["Moritz Reuss", "Hongyi Zhou", "Marcel Rühle", "Ömer Erdinç Yağmurlu", "Fabian Otto", "Rudolf Lioutikov", "Ömer Erdinç Yagmurlu"], "year": 2025, "venue": "CoRL 2025", "type": "preprint", "url": "https://arxiv.org/abs/2509.04996", "pdf_url": "https://arxiv.org/pdf/2509.04996", "code_url": "https://intuitive-robots.github.io/flower_vla/", "clusters": ["C1", "C3", "C5"], "keywords": ["flow policy", "small VLA", "layer pruning", "AdaLN", "FLOWER", "950M VLA", "Global-AdaLN", "efficient VLA", "950M", "H100 GPU hours"], "model_class": "VLA", "techniques": ["LLM layer pruning", "flow matching", "efficient conditioning", "AdaLN conditioning", "Global-AdaLN conditioning"], "base_model": "", "params": "950M", "hardware": "", "reported_metric": "pretrained in 200 H100 GPU hours", "metric_measures": "other", "benchmarks": ["CALVIN", "LIBERO", "SIMPLER", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Sub-1B flow-based VLA that drops up to half of the VLM layers and shifts capacity to the flow head, reaching state of the art on CALVIN ABC with modest pretraining compute.", "relevance": 5, "relevance_reason": "Efficient-architecture VLA with strong results.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; snowball from 2502.19645 (S2 citations); web: FPGA accelerator VLA diffusion policy"}
{"id": "2506.08822", "title": "FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency", "authors": ["Yifei Su", "Ning Liu", "Dong Chen", "Zhen Zhao", "Kun Wu", "Meng Li", "Zhiyuan Xu", "Zhengping Che", "Jian Tang"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2506.08822", "pdf_url": "https://arxiv.org/pdf/2506.08822", "code_url": "", "clusters": ["C1"], "keywords": ["flow policy", "one-step", "VLA acceleration"], "model_class": "diffusion/flow policy", "techniques": ["one-step generation", "frequency consistency"], "base_model": "", "params": "", "hardware": "", "reported_metric": "93.5 Hz inference frequency", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Imposes frequency-domain consistency on flow-based visuomotor policies to enable high-quality one-step action generation, also plugged into a VLA.", "relevance": 5, "relevance_reason": "Few-step/one-step action head for real-time control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "deepmind-gemini-robotics-on-device-2025", "title": "Gemini Robotics On-Device brings AI to local robotic devices", "authors": ["Google DeepMind", "Carolina Parada"], "year": 2025, "venue": "blog (Google DeepMind)", "type": "blog", "url": "https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/", "pdf_url": "", "code_url": "", "clusters": ["C1", "C2", "C4", "C5", "C6", "C9"], "keywords": ["on-device VLA", "offline operation", "few-shot adaptation", "ALOHA", "Apollo humanoid", "Franka FR3", "fine-tuning", "bimanual", "SDK"], "model_class": "VLA", "techniques": ["on-device model optimization", "on-device model", "few-shot fine-tuning", "on-device optimization", "fine-tuning"], "base_model": "Gemini Robotics On-Device", "params": "", "hardware": "onboard robot compute (ALOHA, Franka FR3, Apptronik Apollo)", "reported_metric": "", "metric_measures": "", "benchmarks": ["ALOHA", "Franka FR3", "Apptronik Apollo", "generalization tasks", "instruction following", "Apollo humanoid"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "DeepMind's local VLA variant that runs without network on bi-arm robots, approaching the cloud flagship's generalization and adaptable with 50-100 demos; the explicit onboard counterpart to cloud Gemini Robotics.", "relevance": 5, "relevance_reason": "Major-lab on-device VLA; hardware details undisclosed.", "foundational_technique": false, "verified": true, "found_via": "seed; web: Gemini Robotics On-Device runs locally; search query; blog sweep: Google DeepMind robotics"}
{"id": "2503.20020", "title": "Gemini Robotics: Bringing AI into the Physical World", "authors": ["Gemini Robotics Team", "Saminda Abeyruwan", "Joshua Ainslie", "Jean-Baptiste Alayrac", "Montserrat Gonzalez Arenas", "Travis Armstrong", "Ashwin Balakrishna", "Robert Baruch", "Maria Bauza", "Michiel Blokzijl", "Steven Bohez", "Konstantinos Bousmalis", "Anthony Brohan", "Thomas Buschmann", "Arunkumar Byravan", "Serkan Cabi", "Ken Caluwaerts", "Federico Casarini", "Oscar Chang", "Jose Enrique Chen", "Xi Chen", "Hao-Tien Lewis Chiang", "Krzysztof Choromanski", "David D'Ambrosio", "Sudeep Dasari", "Todor Davchev", "Coline Devin", "Norman Di Palo", "Tianli Ding", "Adil Dostmohamed", "Danny Driess", "Yilun Du", "Debidatta Dwibedi", "Michael Elabd", "Claudio Fantacci", "Cody Fong", "Erik Frey", "Chuyuan Fu", "Marissa Giustina", "Keerthana Gopalakrishnan", "Laura Graesser", "Leonard Hasenclever", "Nicolas Heess", "Brandon Hernaez", "Alexander Herzog", "R. Alex Hofer", "Jan Humplik", "Atil Iscen", "Mithun George Jacob", "Deepali Jain", "Ryan Julian", "Dmitry Kalashnikov", "M. Emre Karagozler", "Stefani Karp", "Chase Kew", "Jerad Kirkland", "Sean Kirmani", "Yuheng Kuang", "Thomas Lampe", "Antoine Laurens", "Isabel Leal", "Alex X. Lee", "Tsang-Wei Edward Lee", "Jacky Liang", "Yixin Lin", "Sharath Maddineni", "Anirudha Majumdar", "Assaf Hurwitz Michaely", "Robert Moreno", "Michael Neunert", "Francesco Nori", "Carolina Parada", "Emilio Parisotto", "Peter Pastor", "Acorn Pooley", "Kanishka Rao", "Krista Reymann", "Dorsa Sadigh", "Stefano Saliceti", "Pannag Sanketi", "Pierre Sermanet", "Dhruv Shah", "Mohit Sharma", "Kathryn Shea", "Charles Shu", "Vikas Sindhwani", "Sumeet Singh", "Radu Soricut", "Jost Tobias Springenberg", "Rachel Sterneck", "Razvan Surdulescu", "Jie Tan", "Jonathan Tompson", "Vincent Vanhoucke", "Jake Varley", "Grace Vesom", "Giulia Vezzani", "Oriol Vinyals", "Ayzaan Wahid", "Stefan Welker", "Paul Wohlhart", "Fei Xia", "Ted Xiao", "Annie Xie", "Jinyu Xie", "Peng Xu", "Sichun Xu", "Ying Xu", "Zhuo Xu", "Yuxiang Yang", "Rui Yao", "Sergey Yaroshenko", "Wenhao Yu", "Wentao Yuan", "Jingwei Zhang", "Tingnan Zhang", "Allan Zhou", "Yuxiang Zhou", "G. Team", "Montse Gonzalez Arenas", "Maria Bauzá", "M. Blokzijl", "F. Casarini", "J. Chen", "David B. D'Ambrosio", "S. Dasari", "C. Fu", "M. Giustina", "K. Gopalakrishnan", "L. Graesser", "N. Heess", "Alex Herzog", "R. Hofer", "M. Jacob", "Ryan C. Julian", "M. E. Karagozler", "T. Lee", "Yi-Xin Lin", "A. Michaely", "M. Neunert", "F. Nori", "P. Pastor", "A. Pooley", "D. Sadigh", "Pannag R. Sanketi", "P. Sermanet", "J. Springenberg", "J. Varley", "O. Vinyals", "Si-Chun Xu", "Wen-Hao Yu"], "year": 2025, "venue": "tech report (Google DeepMind)", "type": "tech-report", "url": "https://arxiv.org/abs/2503.20020", "pdf_url": "https://arxiv.org/pdf/2503.20020", "code_url": "", "clusters": ["C1", "C2", "C5", "C6", "C7", "C9"], "keywords": ["Gemini Robotics", "embodied reasoning", "generalist VLA", "cloud backbone", "local onboard decoder", "TPU/cloud", "latency hiding", "50 Hz", "local action decoder", "Gemini", "VLA", "safety considerations", "cloud backbone + local decoder", "on-robot action decoder", "action chunking", "hybrid inference"], "model_class": "VLA", "techniques": ["cloud-onboard split", "distilled backbone", "local action decoder", "hierarchical onboard/cloud architecture", "action chunking", "cloud VLM backbone with on-robot action decoder", "edge-cloud split", "latency optimization"], "base_model": "Gemini Robotics (distilled Gemini Robotics-ER backbone)", "params": "", "hardware": "cloud-hosted backbone + robot onboard computer (local decoder)", "reported_metric": "backbone query-to-response <160 ms; ~250 ms end-to-end to action chunks; effective 50 Hz control", "metric_measures": "closed-loop control", "benchmarks": ["ALOHA 2 real robot", "real robot (ALOHA 2)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Gemini-2.0-based VLA plus an embodied-reasoning model; the VLA is split into a cloud-hosted backbone and a small action decoder on the robot's computer, which masks backbone latency and yields ~50 Hz effective control from action chunks.", "relevance": 5, "relevance_reason": "Canonical industrial example of edge-cloud split with explicit latency numbers; co-authored by workshop organizer Dhruv Shah.", "foundational_technique": false, "verified": true, "found_via": "seed; web: TPU inference robot policy Gemini Robotics latency; search query; arXiv search: \"Gemini Robotics\"; author sweep: Dhruv Shah"}
{"id": "2502.00935", "title": "Generalizing Safety Beyond Collision-Avoidance via Latent-Space Reachability Analysis", "authors": ["Kensuke Nakamura", "Lasse Peters", "Andrea Bajcsy"], "year": 2025, "venue": "RSS 2025", "type": "paper", "url": "https://arxiv.org/abs/2502.00935", "pdf_url": "https://arxiv.org/pdf/2502.00935", "code_url": "https://kensukenk.github.io/latent-safety/", "clusters": ["C4", "C7", "C8"], "keywords": ["latent safety filters", "HJ reachability", "world model"], "model_class": "world model", "techniques": ["latent-space HJ reachability", "reachability analysis in generative world model latent space", "policy-agnostic safety override"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Franka manipulation (real)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Performs Hamilton-Jacobi reachability in a world model's latent space so robots can avoid failures specified only visually (e.g., spilling), overriding any base policy when needed.", "relevance": 5, "relevance_reason": "Founding latent safety filter work.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"hamilton-jacobi\" reachability learned policy safety filter; search query"}
{"id": "genie-3-deepmind-blog", "title": "Genie 3: A new frontier for world models", "authors": ["Jack Parker-Holder", "Shlomi Fruchter"], "year": 2025, "venue": "Google DeepMind blog", "type": "blog", "url": "https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/", "pdf_url": "", "code_url": "", "clusters": ["C8"], "keywords": ["real-time world model", "interactive world generation", "SIMA"], "model_class": "world model", "techniques": ["real-time autoregressive frame generation", "long-horizon visual memory", "promptable world events"], "base_model": "", "params": "", "hardware": "", "reported_metric": "24 fps at 720p, consistent for a few minutes", "metric_measures": "throughput", "benchmarks": ["SIMA agent goals"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "General-purpose world model generating text-prompted interactive environments in real time at 720p/24 fps with minutes-long consistency; tested as a training ground for the SIMA embodied agent. Limitations include a restricted action space and short interaction horizons.", "relevance": 5, "relevance_reason": "Seed (Genie 3); the reference point for real-time interactive world models.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2503.14734", "title": "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots", "authors": ["NVIDIA", ":", "Johan Bjorck", "Fernando Castañeda", "Nikita Cherniadev", "Xingye Da", "Runyu Ding", "Linxi \"Jim\" Fan", "Yu Fang", "Dieter Fox", "Fengyuan Hu", "Spencer Huang", "Joel Jang", "Zhenyu Jiang", "Jan Kautz", "Kaushil Kundalia", "Lawrence Lao", "Zhiqi Li", "Zongyu Lin", "Kevin Lin", "Guilin Liu", "Edith Llontop", "Loic Magne", "Ajay Mandlekar", "Avnish Narayan", "Soroush Nasiriany", "Scott Reed", "You Liang Tan", "Guanzhi Wang", "Zu Wang", "Jing Wang", "Qi Wang", "Jiannan Xiang", "Yuqi Xie", "Yinzhen Xu", "Zhenjia Xu", "Seonghyeon Ye", "Zhiding Yu", "Ao Zhang", "Hao Zhang", "Yizhou Zhao", "Ruijie Zheng", "Yuke Zhu", "L. Fan", "J. Jang", "Zhenyuan Jiang", "Zhi-Qi Li", "A. Mandlekar", "You Tan", "Zhen-Teng Xu", "Yu-Ke Zhu", "NVIDIA (Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yuke Zhu, et al.)"], "year": 2025, "venue": "tech report (NVIDIA)", "type": "tech-report", "url": "https://arxiv.org/abs/2503.14734", "pdf_url": "https://arxiv.org/pdf/2503.14734", "code_url": "https://github.com/NVIDIA/Isaac-GR00T", "clusters": ["C1", "C2", "C3", "C5", "C7", "C9"], "keywords": ["humanoid foundation model", "dual system", "diffusion transformer", "flow matching", "GR00T N1", "humanoid", "dual-system", "post-training", "GR00T", "Jetson", "System 1 / System 2", "System 1/System 2", "DiT action head"], "model_class": "VLA", "techniques": ["dual-system architecture", "action expert", "diffusion transformer action head", "post-training", "dual-system VLM + DiT", "dual-system VLM + diffusion transformer", "flow-matching DiT", "action chunking (H=16)", "4 denoising steps"], "base_model": "Eagle-2 VLM + DiT", "params": "2.2B (1.34B VLM)", "hardware": "NVIDIA L40 (bf16)", "reported_metric": "63.9 ms per 16-action chunk; System 2 at 10 Hz, System 1 at 120 Hz", "metric_measures": "open-loop chunk rate", "benchmarks": ["simulation", "humanoid real robot", "simulation (RoboCasa, DexMG, GR-1)", "real robot (Fourier GR-1)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "NVIDIA's open humanoid foundation model with a slow VLM System 2 and fast diffusion-transformer System 1, later deployed with TensorRT on Jetson Thor/Orin.", "relevance": 5, "relevance_reason": "Key open dual-rate VLA with explicit latency; basis for Jetson deployment.", "foundational_technique": false, "verified": true, "found_via": "seed; WebSearch: GR00T N1 post-training fine-tuning humanoid LoRA memory NVIDIA; arXiv search: GR00T N1 humanoid foundation model; blog sweep: NVIDIA GEAR / Isaac"}
{"id": "figure-helix-2025", "title": "Helix: A Vision-Language-Action Model for Generalist Humanoid Control", "authors": ["Figure AI"], "year": 2025, "venue": "blog (Figure AI)", "type": "blog", "url": "https://www.figure.ai/news/helix", "pdf_url": "", "code_url": "", "clusters": ["C1", "C2", "C5", "C6", "C9"], "keywords": ["System 1 System 2", "dual-system", "humanoid", "onboard GPUs", "dual embedded GPUs", "System 1/System 2", "200 Hz", "onboard", "onboard embedded GPUs", "dual-rate VLA", "embedded GPUs", "latent intent vector"], "model_class": "VLA", "techniques": ["dual-system architecture", "temporal offset training", "dual-system decomposition", "model-parallel deployment across two GPUs", "asynchronous latent sharing", "hierarchical slow/fast VLA", "onboard deployment", "dual-system asynchronous architecture", "latent communication between S2 and S1", "training-time temporal offset"], "base_model": "7B open-source VLM (S2) + 80M transformer (S1)", "params": "S2 7B VLM + S1 80M transformer", "hardware": "dual low-power embedded GPUs onboard Figure robot", "reported_metric": "S2 7-9 Hz; S1 200 Hz | S2 at 7-9 Hz; S1 at 200 Hz over 35-DoF action space", "metric_measures": "closed-loop control", "benchmarks": ["real robot (Figure 02)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Figure's humanoid VLA: a 7B VLM at 7-9 Hz produces a latent that conditions an 80M visuomotor transformer at 200 Hz, both running onboard; training uses a deliberate temporal offset to match deployment lag.", "relevance": 5, "relevance_reason": "Canonical industrial example of mapping a VLA onto onboard heterogeneous GPU compute.", "foundational_technique": false, "verified": true, "found_via": "seed; web: Figure Helix onboard embedded GPUs; search query; blog sweep: Figure"}
{"id": "figure-helix-logistics-blog-2025", "title": "Helix: Accelerating Real-World Logistics", "authors": ["Figure AI"], "year": 2025, "venue": "blog (Figure)", "type": "blog", "url": "https://www.figure.ai/news/helix-logistics", "pdf_url": "", "code_url": "", "clusters": ["C1", "C6", "C9"], "keywords": ["Sport Mode", "test-time action resampling", "stereo backbone", "effective throughput", "data curation"], "model_class": "VLA", "techniques": ["test-time trajectory time-compression", "stereo multiscale features", "visual proprioception"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 50% speedup via chunk resampling at 200 Hz; stereo +60% throughput; T_eff >1.0 at 50% speedup", "metric_measures": "throughput", "benchmarks": ["real robot (package triage)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Improves Helix S1 with stereo multiscale vision and self-calibration, and introduces 'Sport Mode' that resamples predicted chunks to execute faster than demonstrations without retraining.", "relevance": 5, "relevance_reason": "Test-time execution speedup and throughput metrics for an onboard VLA.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Figure"}
{"id": "2510.04898", "title": "HyperVLA: Efficient Inference in Vision-Language-Action Models via Hypernetworks", "authors": ["Zheng Xiong", "Kang Li", "Zilin Wang", "Matthew Jackson", "Jakob Foerster", "Shimon Whiteson"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.04898", "pdf_url": "https://arxiv.org/pdf/2510.04898", "code_url": "https://github.com/MasterXiong/HyperVLA", "clusters": ["C1"], "keywords": ["hypernetwork", "task-specific policy", "activated parameters"], "model_class": "VLA", "techniques": ["hypernetwork"], "base_model": "OpenVLA (comparison)", "params": "", "hardware": "", "reported_metric": "90x fewer activated parameters; 120x faster inference vs OpenVLA", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A hypernetwork generates a small task-specific policy so only that policy runs at inference, retaining multi-task capacity during training.", "relevance": 5, "relevance_reason": "Seed hypernetwork VLA.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "nvidia-blog-jetson-thor-2025", "title": "Introducing NVIDIA Jetson Thor, the Ultimate Platform for Physical AI", "authors": ["Shashank Maheshwari", "Suhas Hariharapura Sheshadri", "Leela Subramaniam Karumbunathan", "Aayush Pathak"], "year": 2025, "venue": "blog (NVIDIA Technical Blog)", "type": "blog", "url": "https://developer.nvidia.com/blog/introducing-nvidia-jetson-thor-the-ultimate-platform-for-physical-ai/", "pdf_url": "", "code_url": "", "clusters": ["C4", "C5", "C9"], "keywords": ["Jetson Thor", "Jetson Orin", "GR00T", "output token throughput", "Blackwell", "FP4", "LPDDR5X", "power envelope", "Jetson AGX Thor", "speculative decoding", "edge hardware"], "model_class": "other", "techniques": ["edge hardware platform", "FP4 inference", "speculative decoding"], "base_model": "GR00T N1.5 (benchmarked)", "params": "", "hardware": "Jetson AGX Thor T5000 (2070 sparse FP4 TFLOPS, 128 GB LPDDR5X, 273 GB/s, 40-130 W) vs Jetson AGX Orin", "reported_metric": "GR00T N1 46.7 vs 18.5 tokens/s (2.52x); N1.5 41.5 vs 15.2 (2.74x); seq len 2048, output 128, concurrency 8 | GR00T N1.5: 41.5 output tokens/s on Thor vs 15.2 on Orin (2.74x); claims 7.5x AI compute, 3.5x energy efficiency vs Orin", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Vendor launch blog reporting Thor-vs-Orin speedups for generative models including GR00T, measured as output-token throughput at concurrency 8 — not a closed-loop control rate.", "relevance": 5, "relevance_reason": "Defines the key onboard hardware target; note 273 GB/s bandwidth vs multi-PFLOP FP4 compute (memory-bound regime).", "foundational_technique": false, "verified": true, "found_via": "WebSearch: NVIDIA Jetson Thor GR00T inference latency; web: NVIDIA technical blog Jetson Thor humanoid; blog sweep: NVIDIA GEAR / Isaac"}
{"id": "2505.23705", "title": "Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better", "authors": ["Danny Driess", "Jost Tobias Springenberg", "Brian Ichter", "Lili Yu", "Adrian Li-Bell", "Karl Pertsch", "Allen Z. Ren", "Homer Walke", "Quan Vuong", "Lucy Xiaoyang Shi", "Sergey Levine"], "year": 2025, "venue": "NeurIPS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2505.23705", "pdf_url": "https://arxiv.org/pdf/2505.23705", "code_url": "", "clusters": ["C1", "C3", "C4", "C9"], "keywords": ["knowledge insulation", "stop-gradient", "action expert", "FAST tokens", "co-training", "pi0.5", "FAST tokens + action expert"], "model_class": "VLA", "techniques": ["knowledge insulation", "co-training with discrete tokens", "action expert", "gradient stopping", "discrete+continuous co-training", "co-training with VLM data", "gradient insulation", "continuous action expert"], "base_model": "PaliGemma / pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "DROID", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows naively attaching a flow action expert slows training and erodes VLM knowledge; trains the backbone on discrete FAST tokens while insulating it from action-expert gradients for faster training and fast inference.", "relevance": 5, "relevance_reason": "Training-efficiency recipe for flow-matching VLAs from Physical Intelligence.", "foundational_technique": false, "verified": true, "found_via": "search query: knowledge insulating VLA; WebSearch: knowledge insulation co-training VLA training efficiency pi0.5 arXiv; blog sweep: Physical Intelligence"}
{"id": "2502.04296", "title": "Learning Real-World Action-Video Dynamics with Heterogeneous Masked Autoregression", "authors": ["Lirui Wang", "Kevin Zhao", "Chaoqi Liu", "Xinlei Chen"], "year": 2025, "venue": "arXiv", "type": "paper", "url": "https://arxiv.org/abs/2502.04296", "pdf_url": "https://arxiv.org/pdf/2502.04296", "code_url": "https://liruiw.github.io/hma", "clusters": ["C8"], "keywords": ["HMA", "masked autoregression", "real-time video simulator"], "model_class": "world model", "techniques": ["heterogeneous pretraining", "masked autoregression over quantized/soft tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "15x faster than previous robotic video generation models", "metric_measures": "relative latency", "benchmarks": ["multi-embodiment video prediction", "policy evaluation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Masked-autoregressive action-video world model pretrained across embodiments, running ~15x faster than prior robot video generators for real-time simulation and evaluation.", "relevance": 5, "relevance_reason": "Efficiency-focused robot world model architecture.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2509.23224", "title": "Leave No Observation Behind: Real-time Correction for VLA Action Chunks", "authors": ["Kohei Sendai", "Maxime Alvarez", "Tatsuya Matsushima", "Yutaka Matsuo", "Yusuke Iwasawa"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.23224", "pdf_url": "https://arxiv.org/pdf/2509.23224", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["chunk correction", "asynchronous", "inference delay", "A2C2", "chunk correction head", "reactivity"], "model_class": "VLA", "techniques": ["residual correction head", "asynchronous inference", "lightweight correction head", "asynchronous execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+23 pts (Kinetix) and +7 pts (LIBERO Spatial) success vs RTC under delay", "metric_measures": "other", "benchmarks": ["Kinetix", "LIBERO", "LIBERO Spatial"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Adds a small per-control-step correction head that takes the latest observation plus the VLA chunk action and its index to output residual corrections, restoring closed-loop reactivity under inference delay without retraining the base VLA.", "relevance": 5, "relevance_reason": "Fast low-level correction under a slow VLA; complements RTC.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; arXiv search: real-time chunking"}
{"id": "2510.13626", "title": "LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models", "authors": ["Fei, Senyu", "Wang, Siyin", "Shi, Junhao", "Dai, Zihao", "Cai, Jikun", "Qian, Pengfang", "Ji, Li", "He, Xinzhe", "Zhang, Shiduo", "Fei, Zhaoye", "Fu, Jinlan", "Gong, Jingjing", "Qiu, Xipeng"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2510.13626", "pdf_url": "https://arxiv.org/pdf/2510.13626", "code_url": "https://github.com/sylvestf/LIBERO-plus", "clusters": ["C4"], "keywords": ["LIBERO-Plus", "robustness", "camera viewpoint", "initial state", "language insensitivity"], "model_class": "VLA", "techniques": [], "base_model": "OpenVLA, OpenVLA-OFT, pi0, pi0-FAST, Nora, WorldVLA, UniVLA, RIPT-VLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Seven-dimension controlled perturbation suite for LIBERO; VLAs fall from ~95% to below 30% under camera/initial-state changes and largely ignore language.", "relevance": 5, "relevance_reason": "Seed robustness variant, increasingly used to test compressed VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2510.03827", "title": "LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization", "authors": ["Xueyang Zhou", "Yangming Xu", "Guiyao Tie", "Yongchao Chen", "Guowen Zhang", "Duanfeng Chu", "Pan Zhou", "Lichao Sun"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2510.03827", "pdf_url": "https://arxiv.org/pdf/2510.03827", "code_url": "https://github.com/Zxy-MLlab/LIBERO-PRO", "clusters": ["C4"], "keywords": ["LIBERO-PRO", "memorization", "evaluation flaws", "perturbation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-PRO"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Shows that VLAs scoring >90% on standard LIBERO drop to ~0% when objects, initial states, instructions or environments are perturbed, arguing LIBERO rewards rote memorization.", "relevance": 5, "relevance_reason": "Seed robustness variant; implies LIBERO success retention is a weak signal for judging compressed/accelerated VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2506.18897", "title": "MinD: Learning A Dual-System World Model for Real-Time Planning and Implicit Risk Analysis", "authors": ["Xiaowei Chi", "Kuangzhi Ge", "Jiaming Liu", "Siyuan Zhou", "Peidong Jia", "Zichen He", "Yuzhen Liu", "Tingguang Li", "Lei Han", "Sirui Han", "Shanghang Zhang", "Yike Guo", "Pei-Dong Jia", "Si-Rui Han", "Yi-Ke Guo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.18897", "pdf_url": "https://arxiv.org/pdf/2506.18897", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["dual-system world model", "single-step low-res latents", "implicit risk analysis", "real-time planning", "risk analysis", "video generation"], "model_class": "world-action model", "techniques": ["asynchronous low-frequency video diffusion + high-frequency diffusion policy", "single-step latent prediction", "video-action alignment", "low-frequency video generation + high-frequency action", "implicit risk prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RL-Bench", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Dual-rate world model where a slow video diffuser produces one-step low-resolution latents that guide a fast diffusion policy, and whose predictions also serve as implicit risk signals.", "relevance": 5, "relevance_reason": "Combines real-time planning and risk analysis with a world model.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2503.20384", "title": "MoLe-VLA: Dynamic Layer-skipping Vision Language Action Model via Mixture-of-Layers for Efficient Robot Manipulation", "authors": ["Rongyu Zhang", "Menghang Dong", "Yuan Zhang", "Liang Heng", "Xiaowei Chi", "Gaole Dai", "Li Du", "Yuan Du", "Shanghang Zhang", "Rong-Yu Zhang", "Dan Wang", "Shang-Hang Zhang"], "year": 2025, "venue": "AAAI 2026", "type": "preprint", "url": "https://arxiv.org/abs/2503.20384", "pdf_url": "https://arxiv.org/pdf/2503.20384", "code_url": "", "clusters": ["C1", "C2", "C9"], "keywords": ["mixture of layers", "layer skipping", "router", "mixture-of-layers", "self-distillation", "routing"], "model_class": "VLA", "techniques": ["dynamic layer skipping", "self-distillation", "self-knowledge distillation", "router distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 5.6x lower compute vs standard LLM", "metric_measures": "other", "benchmarks": ["RLBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Treats LLM layers as experts and uses a spatial-temporal router to activate a subset per state, with cognition self-distillation to compensate for skipped layers.", "relevance": 5, "relevance_reason": "MoE-style layer skipping for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; search: layer skipping VLA; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2507.10543", "title": "MP1: MeanFlow Tames Policy Learning in 1-step for Robotic Manipulation", "authors": ["Ju-Yi Sheng", "Zi-Yi Wang", "Pei-Ming Li", "Meng-Yuan Liu"], "year": 2025, "venue": "AAAI Conference on Artificial Intelligence", "type": "paper", "url": "https://arxiv.org/abs/2507.10543", "pdf_url": "https://arxiv.org/pdf/2507.10543", "code_url": "", "clusters": ["C1"], "keywords": ["one-step policy", "point cloud"], "model_class": "diffusion/flow policy", "techniques": ["MeanFlow", "1-NFE generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "6.8 ms average inference; 19x faster than DP3", "metric_measures": "relative latency", "benchmarks": ["Adroit", "Meta-World", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "MP1 applies MeanFlow to 3D point-cloud policies to generate trajectories in a single network evaluation without consistency losses.", "relevance": 5, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2405.07503"}
{"id": "2512.01809", "title": "Much Ado About Noising: Dispelling the Myths of Generative Robotic Control", "authors": ["Chaoyi Pan", "Giri Anantharaman", "Nai-Chieh Huang", "Claire Jin", "Daniel Pfrommer", "Chenyang Yuan", "Frank Permenter", "Guannan Qu", "Nicholas Boffi", "Guanya Shi", "Max Simchowitz"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.01809", "pdf_url": "https://arxiv.org/pdf/2512.01809", "code_url": "", "clusters": ["C1"], "keywords": ["generative control policies", "iterative computation", "minimal iterative policy"], "model_class": "diffusion/flow policy", "techniques": ["analysis", "two-step regression policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Finds generative policies' advantage comes from supervised iterative computation, not multimodality; a two-step regression policy matches flow policies and beats distilled shortcuts.", "relevance": 5, "relevance_reason": "Challenges need for many-step generation.", "foundational_technique": false, "verified": true, "found_via": "search query: shortcut model policy robot"}
{"id": "2510.25122", "title": "NanoVLA: Routing Decoupled Vision-Language Understanding for Nano-sized Generalist Robotic Policies", "authors": ["Jiahong Chen", "Jing Wang", "Long Chen", "Chuwei Cai", "Jinghui Lu", "Jia-Hong Chen", "C. Cai"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.25122", "pdf_url": "https://arxiv.org/pdf/2510.25122", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C9"], "keywords": ["nano VLA", "late fusion", "dynamic routing", "Jetson Orin Nano", "NanoVLA", "edge", "routing", "late fusion caching"], "model_class": "VLA", "techniques": ["late vision-language fusion", "caching", "dynamic backbone routing", "long-short chunking", "compact architecture", "vision-language decoupling", "dynamic routing", "long-short action chunking", "decoupled vision-language routing", "small models"], "base_model": "", "params": "", "hardware": "Jetson Orin Nano", "reported_metric": "up to 52x faster inference on edge, 98% fewer parameters | up to 52x faster inference on edge vs prior SOTA VLAs with 98% fewer parameters", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Lightweight VLA family that fuses vision and language late so features can be cached and routes between small and large backbones, targeting Jetson Orin Nano.", "relevance": 5, "relevance_reason": "Combines several efficiency techniques on edge hardware; example of composition.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; snowball from 2506.07530; search query: vision-language-action edge deployment benchmark latency; web: Qualcomm Snapdragon NPU on-device robot policy VLA; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2508.00697", "title": "On-Device Diffusion Transformer Policy for Efficient Robot Manipulation", "authors": ["Yiming Wu", "Huan Wang", "Zhenghao Chen", "Jianxin Pang", "Dong Xu", "Yi-Ming Wu", "Zheng-Hao Chen"], "year": 2025, "venue": "ICCV 2025", "type": "paper", "url": "https://arxiv.org/abs/2508.00697", "pdf_url": "https://arxiv.org/pdf/2508.00697", "code_url": "", "clusters": ["C1", "C2", "C3", "C5", "C9"], "keywords": ["on-device", "pruning", "consistency distillation", "mobile", "LightDP", "mobile devices", "denoiser pruning", "step reduction", "diffusion transformer policy", "on-device diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["structured pruning", "consistency distillation", "retraining", "pruning with retraining", "network compression/pruning", "step reduction via consistency distillation"], "base_model": "", "params": "", "hardware": "mobile devices", "reported_metric": "", "metric_measures": "", "benchmarks": ["PushT", "Robomimic", "CALVIN", "LIBERO", "real robot", "simulation"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "LightDP prunes the diffusion-transformer denoiser with a unified prune-and-retrain pipeline and combines it with fewer sampling steps for real-time diffusion policies on mobile hardware.", "relevance": 5, "relevance_reason": "Pruning+distillation of diffusion policies for on-device use.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy distillation; search: pruning world model robot; web: Jetson diffusion policy real-time onboard; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "openpi-remote-inference", "title": "openpi: Remote inference (policy server over WebSocket)", "authors": ["Physical Intelligence"], "year": 2025, "venue": "GitHub docs", "type": "repo", "url": "https://github.com/Physical-Intelligence/openpi/blob/main/docs/remote_inference.md", "pdf_url": "", "code_url": "https://github.com/Physical-Intelligence/openpi", "clusters": ["C6"], "keywords": ["policy server", "websocket", "remote inference"], "model_class": "VLA", "techniques": ["remote policy server", "client-side image resizing", "open-loop chunk execution"], "base_model": "pi0 / pi0-FAST / pi0.5", "params": "", "hardware": "off-robot GPU server", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "openpi's standard deployment path serves pi0-family policies from a GPU server over WebSocket; robots send resized images/state and receive action chunks, querying only every N steps.", "relevance": 5, "relevance_reason": "Seed: de facto remote-inference pattern for open VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2503.02310", "title": "PD-VLA: Accelerating Vision-Language-Action Model Integrated with Action Chunking via Parallel Decoding", "authors": ["Wenxuan Song", "Jiayi Chen", "Pengxiang Ding", "Han Zhao", "Wei Zhao", "Zhide Zhong", "Zongyuan Ge", "Zhijun Li", "Donglin Wang", "Jun Ma", "Lujia Wang", "Haoang Li"], "year": 2025, "venue": "IROS 2025", "type": "paper", "url": "https://arxiv.org/abs/2503.02310", "pdf_url": "https://arxiv.org/pdf/2503.02310", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["parallel decoding", "Jacobi decoding", "action chunking", "Jacobi fixed-point iteration", "training-free acceleration"], "model_class": "VLA", "techniques": ["parallel action decoding", "fixed-point iteration", "parallel fixed-point decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.52x execution frequency vs base VLA", "metric_measures": "closed-loop control", "benchmarks": ["LIBERO", "real robot", "simulation"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Turns autoregressive decoding of chunked action tokens into parallel fixed-point (Jacobi) iterations, speeding up decoding without retraining.", "relevance": 5, "relevance_reason": "Training-free VLA decoding acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; venue sweep: IROS 2025"}
{"id": "2506.07339", "title": "Real-Time Execution of Action Chunking Flow Policies", "authors": ["Kevin Black", "Manuel Y. Galliker", "Sergey Levine"], "year": 2025, "venue": "NeurIPS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2506.07339", "pdf_url": "https://arxiv.org/pdf/2506.07339", "code_url": "https://github.com/Physical-Intelligence/real-time-chunking-kinetix", "clusters": ["C1", "C2", "C3", "C4", "C6", "C7", "C9"], "keywords": ["real-time chunking", "asynchronous execution", "inpainting", "inference delay", "Kinetix", "action chunking", "RTC"], "model_class": "diffusion/flow policy", "techniques": ["asynchronous inference", "action inpainting", "real-time chunking", "inpainting", "inpainting-based chunk continuation", "guided inpainting", "action chunking", "asynchronous chunking", "guided inpainting of action chunk"], "base_model": "pi0 / pi0.5", "params": "", "hardware": "", "reported_metric": "robust to inference delays >300 ms on precise tasks (match striking, Ethernet plugging)", "metric_measures": "other", "benchmarks": ["Kinetix", "real bimanual robot", "Kinetix (12 dynamic tasks)", "real robot bimanual", "Kinetix dynamic tasks", "real bimanual manipulation", "real robot bimanual (6 tasks)", "Kinetix (sim)", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Proposes real-time chunking (RTC): while the robot executes the current chunk, the next chunk is generated with actions that will already have executed frozen and the rest inpainted via guidance, so any diffusion/flow VLA can run asynchronously without retraining and without pauses or jerks at boundaries.", "relevance": 5, "relevance_reason": "Seed work defining asynchronous execution robust to inference delay for large VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; WebSearch: closed-loop vs open-loop control frequency VLA; search query; seed; arXiv API all:\"action chunking\" AND latency; blog sweep: Physical Intelligence"}
{"id": "2505.10547", "title": "Real-Time Out-of-Distribution Failure Prevention via Multi-Modal Reasoning", "authors": ["Milan Ganai", "Rohan Sinha", "Christopher Agia", "Daniel Morton", "Luigi Di Lillo", "Marco Pavone"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.10547", "pdf_url": "https://arxiv.org/pdf/2505.10547", "code_url": "", "clusters": ["C7"], "keywords": ["FORTRESS", "fallback strategies", "semantic safety", "runtime monitor", "foundation model reasoning"], "model_class": "other", "techniques": ["slow-rate foundation-model anticipation", "fast fallback planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses multimodal foundation models at low rate to anticipate failure modes and precompute safe fallback sets, then rapidly plans to them when a runtime monitor triggers, avoiding slow on-the-fly reasoning.", "relevance": 5, "relevance_reason": "Explicitly handles slow foundation-model latency for real-time fallback.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2410.04640 (S2)"}
{"id": "2506.18123", "title": "RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies", "authors": ["Atreya, Pranav", "Pertsch, Karl", "Lee, Tony", "Kim, Moo Jin", "Jain, Arhan", "Kuramshin, Artur", "Eppner, Clemens", "Neary, Cyrus", "Hu, Edward", "Ramos, Fabio", "Tremblay, Jonathan", "Arora, Kanav", "Ellis, Kirsty", "Macesanu, Luca", "Villasevil, Marcel Torne", "Leonard, Matthew", "Cho, Meedeum", "Aslan, Ozgur", "Dass, Shivin", "Wang, Jie", "Reger, William", "Yuan, Xingfang", "Yang, Xuning", "Gupta, Abhishek", "Jayaraman, Dinesh", "Berseth, Glen", "Daniilidis, Kostas", "Martin-Martin, Roberto", "Lee, Youngwoon", "Liang, Percy", "Finn, Chelsea", "Levine, Sergey"], "year": 2025, "venue": "CoRL 2025", "type": "benchmark", "url": "https://arxiv.org/abs/2506.18123", "pdf_url": "https://arxiv.org/pdf/2506.18123", "code_url": "", "clusters": ["C4"], "keywords": ["RoboArena", "crowd-sourced evaluation", "pairwise double-blind", "DROID", "ranking"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "DROID Franka platform", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboArena", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distributed real-world evaluation where evaluators at seven institutions freely pick tasks but run double-blind pairwise comparisons; aggregated preferences rank seven generalist policies more accurately than centralized evaluation.", "relevance": 5, "relevance_reason": "Seed real-robot evaluation protocol.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2510.17950", "title": "RoboChallenge: Large-scale Real-robot Evaluation of Embodied Policies", "authors": ["Adina Yakefu", "Bin Xie", "Chongyang Xu", "Enwen Zhang", "Erjin Zhou", "Fan Jia", "Haitao Yang", "Haoqiang Fan", "Haowei Zhang", "Hongyang Peng", "Jing Tan", "Junwen Huang", "Kai Liu", "Kaixin Liu", "Kefan Gu", "Qinglun Zhang", "Ruitao Zhang", "Saike Huang", "Shen Cheng", "Shuaicheng Liu", "Tiancai Wang", "Tiezhen Wang", "Wei Sun", "Wenbin Tang", "Yajun Wei"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2510.17950", "pdf_url": "https://arxiv.org/pdf/2510.17950", "code_url": "", "clusters": ["C4"], "keywords": ["RoboChallenge", "Table30", "online real-robot evaluation", "hosted robots"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "UR5, Franka Panda, Cobot Magic ALOHA, ARX-5", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboChallenge Table30", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Online hosted fleet of real robots for large-scale, reproducible VLA evaluation; initial Table30 benchmark of 30 tabletop tasks across four embodiments with specialist and generalist tracks.", "relevance": 5, "relevance_reason": "Major real-robot evaluation infrastructure.", "foundational_technique": false, "verified": true, "found_via": "search query: RoboChallenge real robot online evaluation"}
{"id": "2510.26742", "title": "Running VLAs at Real-time Speed", "authors": ["Yunchao Ma", "Yizhuang Zhou", "Yunhuan Yang", "Tiancai Wang", "Haoqiang Fan", "Yi-Zhuang Zhou", "Tian-Cai Wang", "Hao-Qiang Fan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.26742", "pdf_url": "https://arxiv.org/pdf/2510.26742", "code_url": "https://github.com/Dexmal/realtime-vla", "clusters": ["C1", "C2", "C3", "C5", "C7"], "keywords": ["real-time pi0", "kernel optimization", "streaming inference", "CUDA graphs", "Triton", "kernel fusion", "split-k", "RTX 4090", "pi0", "real-time VLA", "consumer GPU"], "model_class": "VLA", "techniques": ["inference engineering", "streaming inference", "CUDA graph capture", "graph simplification / QKV fusion", "Triton tile tuning", "kernel fusion", "partial split-k", "inference overhead elimination"], "base_model": "pi0", "params": "3.3B", "hardware": "single consumer GPU", "reported_metric": "30 Hz frame rate; up to 480 Hz trajectory frequency | pi0 latency: 1 view 20.0 ms, 2 views 27.3 ms, 3 views 36.8 ms (vs 105-114 ms naive PyTorch, 43.8-67.6 ms OpenPI JAX); 30 Hz frame rate, up to 480 Hz trajectory frequency", "metric_measures": "open-loop chunk rate", "benchmarks": ["real robot (falling pen grasp)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Bag of systems optimizations that run a multi-view pi0-class VLA at 30 Hz with high-rate trajectories on one consumer GPU, enabling tasks like catching a falling pen, plus a streaming inference framework.", "relevance": 5, "relevance_reason": "Canonical kernel-level optimization study for a flow VLA with per-optimization breakdown.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; web: TensorRT-LLM OpenVLA acceleration kernel fusion; snowball: citations of 2506.07339 (S2)"}
{"id": "2506.09937", "title": "SAFE: Multitask Failure Detection for Vision-Language-Action Models", "authors": ["Qiao Gu", "Yuanliang Ju", "Shengxiang Sun", "Igor Gilitschenski", "Haruki Nishimura", "Masha Itkina", "Florian Shkurti"], "year": 2025, "venue": "NeurIPS 2025", "type": "paper", "url": "https://arxiv.org/abs/2506.09937", "pdf_url": "https://arxiv.org/pdf/2506.09937", "code_url": "https://github.com/vla-safe/SAFE", "clusters": ["C7", "C8"], "keywords": ["SAFE", "multitask failure detection", "VLA internal features", "conformal prediction"], "model_class": "VLA", "techniques": ["feature-based failure probe", "functional conformal prediction"], "base_model": "OpenVLA, pi0, pi0-FAST", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Trains a lightweight probe on VLA internal features to output a per-step failure score that generalizes to unseen tasks, calibrated with conformal prediction to trade detection accuracy vs time.", "relevance": 5, "relevance_reason": "Seed: runtime failure detection for generalist VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed; WebSearch"}
{"id": "2505.15304", "title": "Saliency-Aware Quantized Imitation Learning for Efficient Robotic Control", "authors": ["Seongmin Park", "Hyungmin Kim", "Sangwoo Kim", "Wonseok Jeon", "Juyoung Yang", "Byeongwook Jeon", "Yoonseon Oh", "Jungwook Choi"], "year": 2025, "venue": "ICCV 2025", "type": "preprint", "url": "https://arxiv.org/abs/2505.15304", "pdf_url": "https://arxiv.org/pdf/2505.15304", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["saliency", "QAT", "mission-critical states", "saliency-aware QAT", "edge GPU", "energy", "imitation learning"], "model_class": "VLA", "techniques": ["quantization-aware training", "saliency-weighted loss", "quantization-aware training with saliency weighting", "saliency-aware quantization-aware training"], "base_model": "OpenVLA and other IL policies", "params": "", "hardware": "edge GPU", "reported_metric": "up to 2.5x speedup and 2.5x energy savings (4-bit weights) | 4-bit VLA: up to 2.5x speedup and 2.5x energy savings on an edge GPU", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot", "self-driving", "physics simulation"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Adds saliency-based loss weighting to quantization-aware training so low-bit VLAs keep fidelity at mission-critical states, measured on an edge GPU.", "relevance": 5, "relevance_reason": "Direct follow-up of QAIL for quantized VLAs.", "foundational_technique": false, "verified": true, "found_via": "search: robot policy quantization; arXiv: on-device AND (robot policy OR VLA OR diffusion policy); venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2506.01844", "title": "SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics", "authors": ["Mustafa Shukor", "Dana Aubakirova", "Francesco Capuano", "Pepijn Kooijmans", "Steven Palma", "Adil Zouitine", "Michel Aractingi", "Caroline Pascal", "Martino Russi", "Andres Marafioti", "Simon Alibert", "Matthieu Cord", "Thomas Wolf", "Remi Cadene", "Andrés Marafioti", "Rémi Cadène", "Shukor, Mustafa", "Aubakirova, Dana", "Capuano, Francesco", "Kooijmans, Pepijn", "Palma, Steven", "Zouitine, Adil", "Aractingi, Michel", "Pascal, Caroline", "Russi, Martino", "Marafioti, Andres", "Alibert, Simon", "Cord, Matthieu", "Wolf, Thomas", "Cadene, Remi"], "year": 2025, "venue": "arXiv (Hugging Face tech report)", "type": "tech-report", "url": "https://arxiv.org/abs/2506.01844", "pdf_url": "https://arxiv.org/pdf/2506.01844", "code_url": "https://github.com/huggingface/lerobot", "clusters": ["C1", "C2", "C3", "C4", "C5", "C6", "C7", "C9"], "keywords": ["small VLA", "asynchronous inference", "community data", "action chunking", "layer skipping", "async inference", "single-GPU training", "LeRobot", "SmolVLA", "consumer hardware", "PolicyServer/RobotClient", "asynchronous inference stack", "flow matching action expert"], "model_class": "VLA", "techniques": ["small backbone", "layer skipping", "asynchronous inference", "flow matching action expert", "compact architecture", "layer truncation", "visual token reduction", "compact VLM backbone", "flow matching", "small VLM backbone", "action chunking", "remote policy server", "layer skipping (first 16 LLM layers)", "reduced visual tokens"], "base_model": "SmolVLM-2", "params": "~450M (~100M action expert)", "hardware": "consumer GPUs / CPU; remote GPU server for async stack", "reported_metric": "async inference ~30% faster task completion vs synchronous (per Hugging Face) | LIBERO 87.3% vs pi0 86.0%; async ~30% faster task completion (9.7 s vs 13.75 s pick-place)", "metric_measures": "other", "benchmarks": ["LIBERO", "Meta-World", "SO-100 real robot", "real robot", "real robot (SO-100)", "SO-100/SO-101 real robot", "real robot (SO-100/101)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Compact community VLA trained on LeRobot data that introduces an open asynchronous inference stack decoupling action prediction from execution for higher effective control rates.", "relevance": 5, "relevance_reason": "Seed work: open reference implementation of asynchronous remote VLA inference.", "foundational_technique": false, "verified": true, "found_via": "seed; web search: TinyVLA SmolVLA compact VLA; WebSearch: closed-loop vs open-loop control frequency VLA; search query; arXiv search: SmolVLA asynchronous inference; blog sweep: Hugging Face LeRobot"}
{"id": "2506.01392", "title": "Sparse Imagination for Efficient Visual World Model Planning", "authors": ["Junha Chun", "Youngjoon Jeong", "Taesup Kim"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.01392", "pdf_url": "https://arxiv.org/pdf/2506.01392", "code_url": "", "clusters": ["C8"], "keywords": ["sparse imagination", "token dropping", "world model planning"], "model_class": "world model", "techniques": ["randomized grouped attention", "adaptive token count at rollout"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["visual planning tasks", "real-world with VLAs"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains a transformer world model to tolerate processing only a subset of tokens, so planning rollouts can be run sparsely according to the available compute budget.", "relevance": 5, "relevance_reason": "Compute-adaptive latent rollouts for robot planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2507.22424", "title": "Spec-VLA: Speculative Decoding for Vision-Language-Action Models with Relaxed Acceptance", "authors": ["Songsheng Wang", "Rucheng Yu", "Zhihang Yuan", "Chao Yu", "Feng Gao", "Yu Wang", "Derek F. Wong"], "year": 2025, "venue": "EMNLP 2025", "type": "paper", "url": "https://arxiv.org/abs/2507.22424", "pdf_url": "https://arxiv.org/pdf/2507.22424", "code_url": "", "clusters": ["C1"], "keywords": ["speculative decoding", "relaxed acceptance", "action tokens"], "model_class": "VLA", "techniques": ["speculative decoding"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "1.42x speedup vs OpenVLA; +44% acceptance length", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "First speculative decoding framework for VLAs, relaxing acceptance by tolerating nearby action-token bins.", "relevance": 5, "relevance_reason": "Speculative decoding for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: speculative decoding VLA"}
{"id": "2509.05614", "title": "SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning", "authors": ["Hanzhen Wang", "Jiaming Xu", "Yushun Xiang", "Jiayi Pan", "Yongkang Zhou", "Yong-Lu Li", "Guohao Dai", "Jia-Ming Xu", "Yu-Shun Xiang", "Guo-Hao Dai"], "year": 2025, "venue": "ICML 2026", "type": "preprint", "url": "https://arxiv.org/abs/2509.05614", "pdf_url": "https://arxiv.org/pdf/2509.05614", "code_url": "", "clusters": ["C1", "C2", "C4"], "keywords": ["token pruning", "spatial-temporal consistency", "roofline", "FlashAttention", "SpecPrune-VLA", "visual token pruning", "action-aware"], "model_class": "VLA", "techniques": ["static action-level token pruning", "layer-level dynamic pruning", "action-aware controller", "token pruning", "training-free acceleration", "speculative pruning"], "base_model": "OpenVLA-OFT, DB-OFT, CogACT", "params": "7B", "hardware": "A800-80GB; RTX 3090; RTX 4090", "reported_metric": "up to 1.57x (LIBERO) and 1.70x (real) speedup; 1.32x over FlashAttention baseline (98.5 ms vs 109.0 ms A800); 1.42x combined with action-expert caching; 1.48x on RTX 4090", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SimplerEnv", "real robot (Flexiv Rizon4)", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Training-free two-level visual token pruning that uses history across control steps plus per-layer importance, with pruning aggressiveness tied to end-effector speed; includes roofline analysis and composes with FlashAttention and caching.", "relevance": 5, "relevance_reason": "Explicit composition ablations (pruning + FlashAttention, + caching) and cross-GPU speedups.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q1: arXiv abs:VLA AND abs:\"combined with\" AND abs:acceleration; recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]; snowball round 2: cites 2106.09685, 2304.13705"}
{"id": "2509.09090", "title": "SQAP-VLA: A Synergistic Quantization-Aware Pruning Framework for High-Performance Vision-Language-Action Models", "authors": ["Hengyu Fang", "Yijiang Liu", "Yuan Du", "Li Du", "Huanrui Yang", "Heng-Yu Fang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.09090", "pdf_url": "https://arxiv.org/pdf/2509.09090", "code_url": "", "clusters": ["C1", "C2", "C4", "C9"], "keywords": ["quantization-aware pruning", "training-free", "token pruning", "quantization", "quantization-aware token pruning", "composability", "incompatibility"], "model_class": "VLA", "techniques": ["quantization", "token pruning", "post-training quantization", "joint quantization and token pruning"], "base_model": "OpenVLA / standard VLAs", "params": "", "hardware": "", "reported_metric": "1.93x speedup, up to +4.5% avg success", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Co-designs quantization with token pruning, introducing quantization-aware pruning criteria so both can be applied together without the incompatibility seen when stacking them naively.", "relevance": 5, "relevance_reason": "Key evidence that efficiency techniques do not compose naively.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; search: vision-language-action quantization; search query: VLA quantization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2506.15799", "title": "Steering Your Diffusion Policy with Latent Space Reinforcement Learning", "authors": ["Andrew Wagenmaker", "Mitsuhiko Nakamoto", "Yunchu Zhang", "Seohong Park", "Waleed Yagoub", "Anusha Nagabandi", "Abhishek Gupta", "Sergey Levine"], "year": 2025, "venue": "CoRL 2025", "type": "preprint", "url": "https://arxiv.org/abs/2506.15799", "pdf_url": "https://arxiv.org/pdf/2506.15799", "code_url": "", "clusters": ["C3"], "keywords": ["DSRL", "latent noise RL", "black-box steering", "pi0"], "model_class": "diffusion/flow policy", "techniques": ["noise-space RL", "SAC"], "base_model": "pi0; diffusion policies", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Adapts a frozen diffusion/flow policy by training a small RL actor over its input noise, needing only black-box access and few samples; applied to pi0.", "relevance": 5, "relevance_reason": "Weight-free, sample-efficient adaptation of VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: diffusion steering latent space RL DSRL fine-tune pi0 noise policy"}
{"id": "2512.05964", "title": "Training-Time Action Conditioning for Efficient Real-Time Chunking", "authors": ["Kevin Black", "Allen Z. Ren", "Michael Equi", "Sergey Levine"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.05964", "pdf_url": "https://arxiv.org/pdf/2512.05964", "code_url": "", "clusters": ["C1", "C2", "C3", "C6", "C7", "C9"], "keywords": ["training-time RTC", "action prefix conditioning", "inference delay", "inference delay simulation", "simulated delay"], "model_class": "VLA", "techniques": ["asynchronous inference", "training-time delay simulation", "simulated delay at training", "action prefix conditioning", "simulated inference delay in training", "action-prefix conditioning"], "base_model": "pi0.6", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real box building, espresso making", "Kinetix", "real robot (box building, espresso)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Replaces inference-time inpainting in RTC with training-time conditioning on committed action prefixes under simulated delays, removing guidance overhead; matches RTC on pi0.6 real tasks and beats it at high delay in simulation.", "relevance": 5, "relevance_reason": "Direct PI follow-up that makes delay-robust async execution cheaper.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search query; arXiv search: real-time chunking; blog sweep: Physical Intelligence"}
{"id": "2505.00779", "title": "Uncertainty-aware Latent Safety Filters for Avoiding Out-of-Distribution Failures", "authors": ["Junwon Seo", "Kensuke Nakamura", "Andrea Bajcsy"], "year": 2025, "venue": "CoRL 2025", "type": "paper", "url": "https://arxiv.org/abs/2505.00779", "pdf_url": "https://arxiv.org/pdf/2505.00779", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["latent safety filter", "epistemic uncertainty", "HJ reachability", "OOD", "UNISafe", "conformal prediction"], "model_class": "world model", "techniques": ["uncertainty-augmented latent reachability", "conformal calibration", "conformal OOD threshold", "safety filter"], "base_model": "", "params": "", "hardware": "Franka manipulator", "reported_metric": "", "metric_measures": "", "benchmarks": ["vision-based manipulation (sim + real)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Augments latent-space HJ reachability with world-model epistemic uncertainty (conformally calibrated) so the safety filter also steers away from unseen OOD hazards.", "relevance": 5, "relevance_reason": "Core imagination-based safety monitoring method.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: conformal prediction robot policy failure; search query"}
{"id": "2506.09985", "title": "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning", "authors": ["Mido Assran", "Adrien Bardes", "David Fan", "Quentin Garrido", "Russell Howes", "Mojtaba", "Komeili", "Matthew Muckley", "Ammar Rizvi", "Claire Roberts", "Koustuv Sinha", "Artem Zholus", "Sergio Arnaud", "Abha Gejji", "Ada Martin", "Francois Robert Hogan", "Daniel Dugas", "Piotr Bojanowski", "Vasil Khalidov", "Patrick Labatut", "Francisco Massa", "Marc Szafraniec", "Kapil Krishnakumar", "Yong Li", "Xiaodong Ma", "Sarath Chandar", "Franziska Meier", "Yann LeCun", "Michael Rabbat", "Nicolas Ballas"], "year": 2025, "venue": "arXiv (Meta FAIR tech report)", "type": "tech-report", "url": "https://arxiv.org/abs/2506.09985", "pdf_url": "https://arxiv.org/pdf/2506.09985", "code_url": "https://github.com/facebookresearch/vjepa2", "clusters": ["C8"], "keywords": ["JEPA", "latent planning", "action-conditioned world model", "zero-shot planning"], "model_class": "world model", "techniques": ["self-supervised video pretraining", "latent action-conditioned prediction", "CEM planning in embedding space"], "base_model": "V-JEPA 2 ViT-g", "params": "~300M action-conditioned predictor (V-JEPA 2-AC)", "hardware": "", "reported_metric": "~16 s per action step for V-JEPA 2-AC planning vs ~4 min for a Cosmos video-generation baseline (per third-party summaries of the paper)", "metric_measures": "relative latency", "benchmarks": ["Something-Something v2", "Epic-Kitchens-100", "real robot (Franka pick-and-place)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Pretrains a joint-embedding predictive video model on 1M+ hours of internet video, then post-trains a small action-conditioned latent predictor on <62h of DROID robot video to plan pick-and-place on Franka arms zero-shot via image-goal planning in representation space.", "relevance": 5, "relevance_reason": "Seed; shows latent (non-pixel) world models plan an order of magnitude faster than video-generation world models.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2509.09372", "title": "VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model", "authors": ["Yihao Wang", "Pengxiang Ding", "Lingxiao Li", "Can Cui", "Zirui Ge", "Xinyang Tong", "Wenxuan Song", "Han Zhao", "Wei Zhao", "Pengxu Hou", "Siteng Huang", "Yifan Tang", "Wenhui Wang", "Ru Zhang", "Jianyi Liu", "Donglin Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.09372", "pdf_url": "https://arxiv.org/pdf/2509.09372", "code_url": "https://github.com/OpenHelix-Team/VLA-Adapter", "clusters": ["C1", "C3"], "keywords": ["tiny VLA", "bridge attention", "0.5B backbone", "no robot pretraining"], "model_class": "VLA", "techniques": ["small backbone", "lightweight policy module", "lightweight policy head"], "base_model": "", "params": "0.5B backbone", "hardware": "single consumer-grade GPU", "reported_metric": "trains in 8 hours on one consumer GPU", "metric_measures": "other", "benchmarks": ["LIBERO", "CALVIN", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Studies which VL features best condition actions and proposes a lightweight bridge-attention policy on a 0.5B backbone, trainable in 8 hours on one consumer GPU with fast inference.", "relevance": 5, "relevance_reason": "Drastically lowers VLA training/adaptation cost.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA-Adapter; WebSearch: LoRA fine-tuning robot policy VLA adapter arXiv 2025"}
{"id": "2502.02175", "title": "VLA-Cache: Towards Efficient Vision-Language-Action Model via Adaptive Token Caching in Robotic Manipulation", "authors": ["Siyu Xu", "Yunke Wang", "Chenghao Xia", "Dihao Zhu", "Tao Huang", "Chang Xu", "Yun-Ke Wang", "Di Zhu"], "year": 2025, "venue": "NeurIPS 2025", "type": "preprint", "url": "https://doi.org/10.48550/arXiv.2502.02175", "pdf_url": "https://arxiv.org/pdf/2502.02175", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["token caching", "KV reuse", "training-free", "VLA-Cache", "control frequency"], "model_class": "VLA", "techniques": ["KV cache reuse", "token caching", "KV token caching"], "base_model": "OpenVLA, CogACT", "params": "", "hardware": "", "reported_metric": "up to 1.7x CUDA latency speedup; +15% control frequency", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SIMPLER", "real robot", "SimplerEnv"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Training-free method that reuses KV entries of visual tokens that barely change between consecutive frames, while recomputing task-relevant tokens with a layer-adaptive reuse ratio.", "relevance": 5, "relevance_reason": "Canonical accelerator; distinguishes latency speedup from control-rate gain.", "foundational_technique": false, "verified": true, "found_via": "seed; search query: token caching VLA; snowball round 2: cited by 2506.10100, 2510.24795"}
{"id": "2511.16449", "title": "VLA-Pruner: Temporal-Aware Dual-Level Visual Token Pruning for Efficient Vision-Language-Action Inference", "authors": ["Ziyan Liu", "Yeqiu Chen", "Hongyi Cai", "Tao Lin", "Shuo Yang", "Zheng Liu", "Bo Zhao", "Ye-Qiu Chen", "Hong-Yi Cai"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.16449", "pdf_url": "https://arxiv.org/pdf/2511.16449", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["visual token pruning", "action-decode attention", "temporal-aware"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.99x speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SIMPLER"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA-Pruner scores tokens using both prefill semantic salience and temporally smoothed action-decode relevance, fixing VLM-style pruning that drops action-critical tokens.", "relevance": 5, "relevance_reason": "Title unmistakably about VLA acceleration; unverified", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; snowball round 2: cites 2205.14135, 2303.04137"}
{"id": "2512.01031", "title": "VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference", "authors": ["Jiaming Tang", "Yufei Sun", "Yilong Zhao", "Shang Yang", "Yujun Lin", "Zhuoyang Zhang", "James Hou", "Yao Lu", "Zhijian Liu", "Song Han"], "year": 2025, "venue": "IROS 2026", "type": "preprint", "url": "https://arxiv.org/abs/2512.01031", "pdf_url": "https://arxiv.org/pdf/2512.01031", "code_url": "https://github.com/mit-han-lab/vlash", "clusters": ["C1", "C2", "C3", "C4", "C5", "C6", "C7", "C9"], "keywords": ["asynchronous inference", "future-state conditioning", "reaction latency", "VLASH", "future-state-aware", "future-state projection", "action quantization"], "model_class": "VLA", "techniques": ["asynchronous inference", "state roll-forward", "action quantization", "future-state conditioning", "state roll-forward conditioning", "state forward projection"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "up to 11.8x lower reaction latency vs synchronous; 1.5-2.0x task completion speedup with action quantization", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot", "real robot (ping-pong, whack-a-mole)", "real robot ping-pong, whack-a-mole"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Makes asynchronous VLA inference stable by conditioning on the robot state projected forward through the still-executing action chunk, with no architecture change or runtime overhead, enabling fast-reaction tasks.", "relevance": 5, "relevance_reason": "Key asynchronous-inference method, the main latency-hiding tool for remote/offboard VLA serving.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search query: VLA quantization; search query; arXiv search: \"action chunk\" asynchronous inference; author sweep: Zhuoyang Zhang"}
{"id": "2410.24164", "title": "$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control", "authors": ["Kevin Black", "Noah Brown", "Danny Driess", "Adnan Esmail", "Michael Equi", "Chelsea Finn", "Niccolo Fusai", "Lachy Groom", "Karol Hausman", "Brian Ichter", "Szymon Jakubczak", "Tim Jones", "Liyiming Ke", "Sergey Levine", "Adrian Li-Bell", "Mohith Mothukuri", "Suraj Nair", "Karl Pertsch", "Lucy Xiaoyang Shi", "James Tanner", "Quan Vuong", "Anna Walling", "Haohuan Wang", "Ury Zhilinsky", "Black, Kevin", "Brown, Noah", "Driess, Danny", "Esmail, Adnan", "Equi, Michael", "Finn, Chelsea", "Fusai, Niccolo", "Groom, Lachy", "Hausman, Karol", "Ichter, Brian", "Jakubczak, Szymon", "Jones, Tim", "Ke, Liyiming", "Levine, Sergey", "Li-Bell, Adrian", "Mothukuri, Mohith", "Nair, Suraj", "Pertsch, Karl", "Shi, Lucy Xiaoyang", "Tanner, James", "Vuong, Quan", "Walling, Anna", "Wang, Haohuan", "Zhilinsky, Ury", "A. Esmail", "M. Equi", "S. Jakubczak", "L. Shi", "Hao-Huan Wang", "et al. (Physical Intelligence)"], "year": 2024, "venue": "RSS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2410.24164", "pdf_url": "https://arxiv.org/pdf/2410.24164", "code_url": "https://github.com/Physical-Intelligence/openpi", "clusters": ["C1", "C2", "C3", "C4", "C5", "C9"], "keywords": ["flow matching", "action expert", "VLA", "action chunking", "pi0", "fine-tuning", "flow matching VLA", "inference time", "latency breakdown", "RTX 4090", "off-board inference"], "model_class": "VLA", "techniques": ["flow matching action expert", "action chunking", "flow matching", "post-training on curated data", "flow-matching action expert", "KV caching of prefix", "action chunking (H=50)", "10 integration steps"], "base_model": "PaliGemma 3B", "params": "~3.3B", "hardware": "NVIDIA RTX 4090", "reported_metric": "73 ms onboard (14 ms image encoders, 32 ms observation pass, 27 ms for 10 flow steps); 86 ms off-board incl. 13 ms network", "metric_measures": "open-loop chunk rate", "benchmarks": ["real robot dexterous tasks", "real robot", "real robot (many platforms)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Flow-matching VLA with a VLM backbone and action expert trained on diverse multi-robot data; its reported consumer-GPU inference time is a common latency reference point.", "relevance": 5, "relevance_reason": "Seminal base model used as the latency reference in many efficiency comparisons (consumer-GPU latency figure not confirmed from fetched abstract, so omitted).", "foundational_technique": false, "verified": true, "found_via": "seed; WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025; WebSearch: Physical Intelligence openpi inference speed; blog sweep: Physical Intelligence"}
{"id": "2408.17355", "title": "Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling", "authors": ["Yuejiang Liu", "Jubayer Ibn Hamid", "Annie Xie", "Yoonho Lee", "Maximilian Du", "Chelsea Finn", "Yue-Jiang Liu"], "year": 2024, "venue": "ICLR 2025", "type": "preprint", "url": "https://arxiv.org/abs/2408.17355", "pdf_url": "https://arxiv.org/pdf/2408.17355", "code_url": "https://bid-robot.github.io/", "clusters": ["C1", "C2", "C7"], "keywords": ["action chunking", "test-time sampling", "reactivity", "bidirectional decoding", "closed-loop resampling"], "model_class": "diffusion/flow policy", "techniques": ["action chunking", "test-time sampling", "test-time sampling and selection", "backward coherence", "forward contrast", "closed-loop resampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["7 sim benchmarks", "real robot", "7 simulation benchmarks", "2 real-world tasks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Analyzes the consistency vs reactivity tradeoff of action chunking and proposes BID, which samples many candidate chunks each step and picks one coherent with prior decisions and contrasted against a weaker policy, enabling closed-loop chunked policies.", "relevance": 5, "relevance_reason": "Seed work on reactive decoding of chunked policies; notes compute cost limits high-frequency use.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; seed; WebSearch; snowball round 2: cited by 2502.19645, 2503.02310"}
{"id": "2405.07503", "title": "Consistency Policy: Accelerated Visuomotor Policies via Consistency Distillation", "authors": ["Aaditya Prasad", "Kevin Lin", "Jimmy Wu", "Linqi Zhou", "Jeannette Bohg"], "year": 2024, "venue": "RSS 2024", "type": "preprint", "url": "https://consistency-policy.github.io/", "pdf_url": "https://arxiv.org/pdf/2405.07503", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["consistency distillation", "diffusion policy", "laptop GPU"], "model_class": "diffusion/flow policy", "techniques": ["consistency distillation", "few-step sampling", "step distillation"], "base_model": "Diffusion Policy", "params": "", "hardware": "laptop GPU", "reported_metric": "order-of-magnitude faster inference than fastest alternative", "metric_measures": "relative latency", "benchmarks": ["Robomimic", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills a Diffusion Policy teacher into a consistency model that generates action sequences in one or few steps, enabling low-latency control on resource-limited robots.", "relevance": 5, "relevance_reason": "Seminal diffusion-policy distillation work.", "foundational_technique": false, "verified": true, "found_via": "seed; web search: consistency policy distillation"}
{"id": "2407.04292", "title": "DaDu-Corki: Algorithm-Architecture Co-Design for Embodied AI-powered Robotic Manipulation", "authors": ["Yiyang Huang", "Yuhui Hao", "Bo Yu", "Feng Yan", "Yuxin Yang", "Feng Min", "Yinhe Han", "Lin Ma", "Shaoshan Liu", "Qiang Liu", "Yiming Gan", "Yu-Hsiang Hao", "Shao-Shan Liu", "Yi-Ming Gan"], "year": 2024, "venue": "ISCA 2025", "type": "preprint", "url": "https://arxiv.org/abs/2407.04292", "pdf_url": "https://arxiv.org/pdf/2407.04292", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["Corki", "algorithm-architecture co-design", "trajectory prediction", "accelerator", "torque control accelerator", "communication overlap"], "model_class": "VLA", "techniques": ["trajectory prediction", "hardware accelerator", "trajectory-level prediction to reduce LLM calls", "custom dataflow accelerator for computed-torque control", "compute-communication pipelining"], "base_model": "", "params": "", "hardware": "Xilinx Zynq-7000 ZC706 FPGA (accelerator prototype)", "reported_metric": "up to 5.1x fewer LLM inferences, up to 5.9x speedup, up to 13.9% success improvement; control accelerated up to 29.0x", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Predicts near-future trajectories to reduce LLM inference frequency and co-designs a transformer accelerator, decoupling inference, control and communication to cut latency and energy.", "relevance": 5, "relevance_reason": "Seminal computer-architecture paper on embodied-AI co-design (ISCA).", "foundational_technique": false, "verified": true, "found_via": "search query: embodied inference energy; web: algorithm-hardware co-design embodied AI accelerator"}
{"id": "2411.02359", "title": "DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Execution", "authors": ["Yang Yue", "Yulin Wang", "Bingyi Kang", "Yizeng Han", "Shenzhi Wang", "Shiji Song", "Jiashi Feng", "Gao Huang", "Bing-Yi Kang", "Shen-Zhi Wang", "Jia-Shi Feng"], "year": 2024, "venue": "NeurIPS 2024", "type": "preprint", "url": "https://arxiv.org/abs/2411.02359", "pdf_url": "https://arxiv.org/pdf/2411.02359", "code_url": "https://github.com/yueyang130/DeeR-VLA", "clusters": ["C1", "C2", "C4", "C5", "C9"], "keywords": ["dynamic early exit", "multi-exit MLLM", "resource constraints", "compute budget", "DeeR-VLA", "early exit", "dynamic inference", "budgeted compute", "GPU memory budget", "power budget", "CALVIN", "robot execution"], "model_class": "VLA", "techniques": ["early exit", "dynamic inference", "dynamic early exit", "resource-constrained termination criteria", "multi-exit architecture"], "base_model": "OpenFlamingo-based RoboFlamingo", "params": "", "hardware": "", "reported_metric": "5.2-6.5x lower LLM compute; 2-6x lower LLM GPU memory", "metric_measures": "other", "benchmarks": ["CALVIN"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds multiple exits to a robotic MLLM and learns termination criteria under average/peak compute or memory budgets, activating only as much of the model as each situation needs.", "relevance": 5, "relevance_reason": "Explicitly budget-conditioned inference for robot hardware limits.", "foundational_technique": false, "verified": true, "found_via": "seed; web search: VLA pruning layers redundancy; search query: VLA power consumption; snowball (DeeR-VLA seed lookup); venue sweep: arXiv API VLA AND (MoE OR early exit)"}
{"id": "2411.04983", "title": "DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning", "authors": ["Gaoyue Zhou", "Hengkai Pan", "Yann LeCun", "Lerrel Pinto"], "year": 2024, "venue": "ICML 2025", "type": "paper", "url": "https://arxiv.org/abs/2411.04983", "pdf_url": "https://arxiv.org/pdf/2411.04983", "code_url": "https://dino-wm.github.io/", "clusters": ["C8"], "keywords": ["DINO-WM", "latent planning", "pretrained visual features", "zero-shot planning"], "model_class": "world model", "techniques": ["frozen DINOv2 patch features", "ViT latent predictor", "CEM/MPC in feature space"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["PointMaze", "Push-T", "Wall", "Rope/Granular deformables"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Predicts future DINOv2 patch features instead of pixels and plans action sequences toward goal features, enabling task-agnostic zero-shot planning without reconstruction cost.", "relevance": 5, "relevance_reason": "Reference latent world model for efficient planning; baseline for most sparse/efficient WM work.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2412.12953", "title": "Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning", "authors": ["Moritz Reuss", "Jyothish Pari", "Pulkit Agrawal", "Rudolf Lioutikov"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.12953", "pdf_url": "https://arxiv.org/pdf/2412.12953", "code_url": "", "clusters": ["C1"], "keywords": ["mixture of denoising experts", "noise-conditioned routing", "expert caching"], "model_class": "diffusion/flow policy", "techniques": ["mixture of experts", "expert caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "40% fewer active parameters; 90% lower inference cost (FLOPs)", "metric_measures": "other", "benchmarks": ["CALVIN", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MoDE: diffusion transformer policy with sparse experts routed by noise level, allowing experts to be precomputed/cached for 90% fewer FLOPs.", "relevance": 5, "relevance_reason": "Efficient MoE diffusion policy.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion transformer policy efficient"}
{"id": "2405.05941", "title": "Evaluating Real-World Robot Manipulation Policies in Simulation", "authors": ["Li, Xuanlin", "Hsu, Kyle", "Gu, Jiayuan", "Pertsch, Karl", "Mees, Oier", "Walke, Homer Rich", "Fu, Chuyuan", "Lunawat, Ishikaa", "Sieh, Isabel", "Kirmani, Sean", "Levine, Sergey", "Wu, Jiajun", "Finn, Chelsea", "Su, Hao", "Vuong, Quan", "Xiao, Ted"], "year": 2024, "venue": "CoRL 2024", "type": "benchmark", "url": "https://arxiv.org/abs/2405.05941", "pdf_url": "https://arxiv.org/pdf/2405.05941", "code_url": "https://github.com/simpler-env/SimplerEnv", "clusters": ["C4"], "keywords": ["SimplerEnv", "SIMPLER", "real-to-sim", "sim-real correlation", "Google Robot", "WidowX Bridge"], "model_class": "VLA", "techniques": [], "base_model": "RT-1, RT-1-X, Octo", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Simulated counterparts of Google Robot and WidowX/Bridge setups that reduce control and visual gaps without full digital twins; >1500 paired sim/real evaluations show strong performance correlation.", "relevance": 5, "relevance_reason": "Seed benchmark; standard real-to-sim proxy used to report VLA accuracy after compression.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2412.05408", "title": "FogROS2-FT: Fault Tolerant Cloud Robotics", "authors": ["Kaiyuan Chen", "Kush Hari", "Trinity Chung", "Michael Wang", "Nan Tian", "Christian Juette", "Jeffrey Ichnowski", "Liu Ren", "John Kubiatowicz", "Ion Stoica", "Ken Goldberg"], "year": 2024, "venue": "IROS 2024 (Best Paper Finalist)", "type": "paper", "url": "https://arxiv.org/abs/2412.05408", "pdf_url": "https://arxiv.org/pdf/2412.05408", "code_url": "https://github.com/BerkeleyAutomation/FogROS2", "clusters": ["C6", "C7"], "keywords": ["fault tolerance", "multi-cloud", "spot instances", "tail latency"], "model_class": "other", "techniques": ["request replication", "multi-cloud routing", "spot instances"], "base_model": "", "params": "", "hardware": "multi-cloud (AWS/GCP etc.)", "reported_metric": "up to 2.2x cost reduction and 5.53x lower P99 latency for motion planning", "metric_measures": "relative latency", "benchmarks": ["object detection", "semantic segmentation", "motion planning", "scan-pick-and-place real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Replicates stateless robot services across several clouds and takes the first reply, so robots keep working through provider outages or bad QoS and can safely use cheap spot instances.", "relevance": 5, "relevance_reason": "Addresses cloud-robot reliability, tail latency and cost, central to cloud-hosted foundation-model policies.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2410.05273", "title": "HiRT: Enhancing Robotic Control with Hierarchical Robot Transformers", "authors": ["Jianke Zhang", "Yanjiang Guo", "Xiaoyu Chen", "Yen-Jen Wang", "Yucheng Hu", "Chengming Shi", "Jianyu Chen"], "year": 2024, "venue": "CoRL 2024", "type": "paper", "url": "https://arxiv.org/abs/2410.05273", "pdf_url": "https://arxiv.org/pdf/2410.05273", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["hierarchical robot transformer", "low-frequency VLM", "high-frequency policy", "HiRT"], "model_class": "VLA", "techniques": ["dual-system architecture", "feature caching", "slow VLM features + fast vision policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2x control frequency on static tasks | 2x control frequency in static tasks; dynamic task success 48% -> 75%", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Runs the VLM backbone at low frequency to produce slowly updated features that condition a fast lightweight vision policy for real-time interaction in dynamic tasks.", "relevance": 5, "relevance_reason": "Classic fast servo under slow VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: hierarchical VLA latency; arXiv search: Hierarchical Robot Transformers HiRT"}
{"id": "stanford-minivla-2024", "title": "MiniVLA: A Better VLA with a Smaller Footprint", "authors": ["Suneel Belkhale", "Dorsa Sadigh"], "year": 2024, "venue": "blog (Stanford AI Lab)", "type": "blog", "url": "https://ai.stanford.edu/blog/minivla/", "pdf_url": "", "code_url": "https://github.com/Stanford-ILIAD/openvla-mini", "clusters": ["C1"], "keywords": ["small VLA", "Qwen2.5-0.5B", "VQ action chunking", "multi-image"], "model_class": "VLA", "techniques": ["small backbone", "residual VQ action chunking"], "base_model": "OpenVLA recipe with Qwen 2.5 0.5B", "params": "~1B", "hardware": "NVIDIA L40s", "reported_metric": "12.5 Hz vs 5 Hz for OpenVLA (2.5x)", "metric_measures": "throughput", "benchmarks": ["LIBERO-90"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shrinks OpenVLA from 7B to ~1B with a Qwen2.5-0.5B backbone and adds residual-VQ action chunking and history/wrist images, raising LIBERO-90 from 62% to 82%.", "relevance": 5, "relevance_reason": "Seed small-VLA work.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2410.21257", "title": "One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation", "authors": ["Zhendong Wang", "Zhaoshuo Li", "Ajay Mandlekar", "Zhenjia Xu", "Jiaojiao Fan", "Yashraj Narang", "Linxi Fan", "Yuke Zhu", "Yogesh Balaji", "Mingyuan Zhou", "Ming-Yu Liu", "Yu Zeng", "Zhao-Shuo Li", "A. Mandlekar", "Jiao-Jiao Fan", "L. Fan", "Yu-Ke Zhu", "Yuan Zeng"], "year": 2024, "venue": "ICML 2025", "type": "preprint", "url": "https://arxiv.org/abs/2410.21257", "pdf_url": "https://arxiv.org/pdf/2410.21257", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["one-step diffusion", "distillation", "KL", "one-step", "KL distillation", "Franka"], "model_class": "diffusion/flow policy", "techniques": ["diffusion distillation", "one-step generation"], "base_model": "Diffusion Policy", "params": "", "hardware": "", "reported_metric": "action prediction 1.5 Hz -> 62 Hz | action prediction frequency 1.5 Hz -> 62 Hz", "metric_measures": "throughput", "benchmarks": ["Robomimic", "real Franka", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "OneDP distills a pretrained diffusion policy into a single-step generator by minimizing KL along the diffusion chain at a small extra training cost.", "relevance": 5, "relevance_reason": "Key one-step policy distillation baseline.", "foundational_technique": false, "verified": true, "found_via": "search query: one-step diffusion policy; web search: consistency policy distillation"}
{"id": "jetson-ai-lab-openvla-tutorial", "title": "OpenVLA - Vision/Language Action Models for Embodied Robotics (Fine Tuning and Deployment Guide)", "authors": ["NVIDIA Jetson AI Lab (Dustin Franklin)"], "year": 2024, "venue": "blog / NVIDIA Jetson AI Lab tutorial", "type": "blog", "url": "https://www.jetson-ai-lab.com/openvla.html", "pdf_url": "", "code_url": "https://github.com/dusty-nv/NanoLLM", "clusters": ["C2", "C4", "C5"], "keywords": ["OpenVLA", "Jetson AGX Orin", "MLC", "INT4", "FP8", "NanoLLM", "MimicGen"], "model_class": "VLA", "techniques": ["INT4/FP8 weight quantization (MLC)", "TensorRT FP16 vision encoders", "on-device LoRA"], "base_model": "OpenVLA-7B", "params": "7B", "hardware": "Jetson AGX Orin 64GB (JetPack 6); lists AGX Orin 32GB and Orin NX 16GB as supported", "reported_metric": "Bridge val: FP16 95.3% action acc / 840 ms (1.19 FPS); FP8 95.2% / 471 ms (2.12); INT4 90.1% / 336 ms (2.97). MimicGen block stacking success: FP16 86%, FP8 85%, INT4 84% (~335 ms INT4)", "metric_measures": "closed-loop control", "benchmarks": ["BridgeData V2 subset (action accuracy)", "MimicGen block stacking (sim)"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "NVIDIA tutorial quantizing OpenVLA-7B to FP8/INT4 with MLC on Jetson AGX Orin; INT4 is 2.5x faster than FP16 with ~2 points lower sim task success after fine-tuning.", "relevance": 5, "relevance_reason": "Most direct public measurement of OpenVLA-7B at INT4/FP8 on a Jetson with closed-loop (sim) success; not LIBERO and not on the 16 GB NX.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: web archive of jetson-ai-lab.com/openvla.html (linked from openvla GitHub discussions)"}
{"id": "2406.09246", "title": "OpenVLA: An Open-Source Vision-Language-Action Model", "authors": ["Moo Jin Kim", "Karl Pertsch", "Siddharth Karamcheti", "Ted Xiao", "Ashwin Balakrishna", "Suraj Nair", "Rafael Rafailov", "Ethan Foster", "Grace Lam", "Pannag Sanketi", "Quan Vuong", "Thomas Kollar", "Benjamin Burchfiel", "Russ Tedrake", "Dorsa Sadigh", "Sergey Levine", "Percy Liang", "Chelsea Finn", "A. Balakrishna", "E. Foster", "G. Lam", "Pannag R. Sanketi", "et al."], "year": 2024, "venue": "CoRL 2024", "type": "preprint", "url": "https://arxiv.org/abs/2406.09246", "pdf_url": "https://arxiv.org/pdf/2406.09246", "code_url": "https://github.com/openvla/openvla", "clusters": ["C1", "C2", "C3", "C4", "C5", "C9"], "keywords": ["open-source VLA", "Llama 2", "DINOv2+SigLIP", "4-bit inference", "LoRA", "LoRA fine-tuning", "quantized serving", "OpenVLA", "quantization", "inference speed", "RTX 4090", "open VLA", "consumer GPU"], "model_class": "VLA", "techniques": ["action tokenization", "LoRA fine-tuning", "4-bit quantization (inference)", "LoRA", "quantization", "int8/int4 quantization", "4/8-bit quantization"], "base_model": "Llama 2 + DINOv2/SigLIP (Prismatic)", "params": "7B", "hardware": "RTX 4090, H100, A5000", "reported_metric": "4-bit: 71.9% BridgeData V2 success (matching bf16), GPU memory 16.8 GB -> 7.0 GB | ~6 Hz on RTX 4090 in bf16; A5000: 1.2 Hz (int8), 3 Hz (int4); memory 16.8 GB bf16 / 10.2 GB int8 / 7.0 GB int4", "metric_measures": "closed-loop control", "benchmarks": ["BridgeData V2", "Google robot", "LIBERO", "real robot", "BridgeData V2 WidowX", "Franka real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Open 7B VLA whose paper reports GPU memory and control rates across precisions and GPUs, establishing the standard profiling baseline that later edge work attacks.", "relevance": 5, "relevance_reason": "Base model for most VLA PEFT work; includes first systematic LoRA/quantization study for VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed; venue sweep: CoRL 2024"}
{"id": "2412.01034", "title": "Quantization-Aware Imitation-Learning for Resource-Efficient Robotic Control", "authors": ["Seongmin Park", "Hyungmin Kim", "Wonseok Jeon", "Juyoung Yang", "Byeongwook Jeon", "Yoonseon Oh", "Jungwook Choi"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.01034", "pdf_url": "https://arxiv.org/pdf/2412.01034", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["QAT", "4-bit weights", "edge GPU", "autonomous driving", "quantization-aware imitation learning", "energy savings", "low-bit policy"], "model_class": "VLA", "techniques": ["quantization-aware training", "low-bit quantization", "quantization-aware fine-tuning", "4-bit weights", "quantization-aware training for IL"], "base_model": "OpenVLA; IL self-driving policies", "params": "", "hardware": "edge GPU (Jetson-class); low-end GPU for driving", "reported_metric": "up to 2.5x speedup, 2.5x energy saving (W4 manipulation); 3.7x speedup, 3.1x energy (W4A4 driving)", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot", "CARLA-style driving"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "QAIL fine-tunes imitation-learning policies (incl. OpenVLA) with a quantization-robust objective so 4-bit weight/activation models keep accuracy; measured speed and energy savings on edge GPUs.", "relevance": 5, "relevance_reason": "Seed work; first QAT framework for IL/VLA policies.", "foundational_technique": false, "verified": true, "found_via": "seed; arXiv: on-device AND (robot policy OR VLA OR diffusion policy); venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2406.04339", "title": "RoboMamba: Efficient Vision-Language-Action Model for Robotic Reasoning and Manipulation", "authors": ["Jiaming Liu", "Mengzhen Liu", "Zhenyu Wang", "Pengju An", "Xiaoqi Li", "Kaichen Zhou", "Senqiao Yang", "Renrui Zhang", "Yandong Guo", "Shanghang Zhang", "Jia-Ming Liu", "Meng-Zhen Liu", "Lily Lee", "Yan-Dong Guo"], "year": 2024, "venue": "NeurIPS 2024", "type": "preprint", "url": "https://arxiv.org/abs/2406.04339", "pdf_url": "https://arxiv.org/pdf/2406.04339", "code_url": "", "clusters": ["C1"], "keywords": ["Mamba", "state space model", "VLA", "linear complexity", "RoboMamba"], "model_class": "VLA", "techniques": ["state space model backbone", "parameter-efficient fine-tuning", "state space model", "Mamba"], "base_model": "Mamba LLM", "params": "", "hardware": "", "reported_metric": "3x faster inference than existing VLAs", "metric_measures": "relative latency", "benchmarks": ["RLBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Builds a VLA on a Mamba state-space language backbone for linear-time inference, adding a small SE(3) pose head fine-tuned on ~0.1% of parameters.", "relevance": 5, "relevance_reason": "Efficient SSM VLA per title; unverified.", "foundational_technique": false, "verified": true, "found_via": "seed; snowball round 2: cited by 2502.02175, 2503.02310"}
{"id": "2409.09491", "title": "Robot Learning as an Empirical Science: Best Practices for Policy Evaluation", "authors": ["Hadas Kress-Gazit", "Kunimatsu Hashimoto", "Naveen Kuppuswamy", "Paarth Shah", "Phoebe Horgan", "Gordon Richardson", "Siyuan Feng", "Benjamin Burchfiel"], "year": 2024, "venue": "arXiv", "type": "paper", "url": "https://arxiv.org/abs/2409.09491", "pdf_url": "https://arxiv.org/pdf/2409.09491", "code_url": "", "clusters": ["C4"], "keywords": ["best practices", "policy evaluation", "success rate", "failure modes"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Position/best-practice paper urging reporting of conditions, complementary metrics, statistical analysis and qualitative failure descriptions for physical robot evaluation.", "relevance": 5, "relevance_reason": "Foundational methodology reference for evaluation protocols.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation statistical"}
{"id": "2309.05837", "title": "The Safety Filter: A Unified View of Safety-Critical Control in Autonomous Systems", "authors": ["Kai-Chieh Hsu", "Haimin Hu", "Jaime F. Fisac"], "year": 2024, "venue": "Annual Review of Control, Robotics, and Autonomous Systems 7 (2024)", "type": "paper", "url": "https://doi.org/10.1146/annurev-control-071723-102940", "pdf_url": "https://arxiv.org/pdf/2309.05837", "code_url": "", "clusters": ["C7"], "keywords": ["safety filter", "survey", "HJ reachability", "CBF", "monitoring and intervention"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Review that unifies safety filter families (HJ reachability, CBFs, predictive filters) into a shared monitor-plus-intervention structure and outlines scalable synthesis and robust monitoring directions.", "relevance": 5, "relevance_reason": "Seed survey framing runtime safety filters for learned policies.", "foundational_technique": true, "verified": true, "found_via": "seed; WebSearch"}
{"id": "2409.12514", "title": "TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation", "authors": ["Junjie Wen", "Yichen Zhu", "Jinming Li", "Minjie Zhu", "Kun Wu", "Zhiyuan Xu", "Ning Liu", "Ran Cheng", "Chaomin Shen", "Yaxin Peng", "Feifei Feng", "Jian Tang", "Yi-Chen Zhu", "Zhibin Tang", "Ya-Xin Peng", "Fei-Fei Feng"], "year": 2024, "venue": "IEEE RA-L", "type": "preprint", "url": "https://arxiv.org/abs/2409.12514", "pdf_url": "https://arxiv.org/pdf/2409.12514", "code_url": "", "clusters": ["C1", "C2", "C5"], "keywords": ["compact VLA", "diffusion head", "data efficiency", "small VLM backbone"], "model_class": "VLA", "techniques": ["small backbone", "diffusion policy head", "compact architecture", "small multimodal backbone", "diffusion policy decoder"], "base_model": "", "params": "<1.4B backbones", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MetaWorld", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "A family of compact VLAs initialized from fast small multimodal models and paired with a diffusion policy decoder, skipping large-scale robot pretraining while beating OpenVLA on speed.", "relevance": 5, "relevance_reason": "Canonical compact-VLA baseline used in compression comparisons.", "foundational_technique": false, "verified": true, "found_via": "seed; web search: TinyVLA compact VLA"}
{"id": "2410.08001", "title": "Towards Synergistic, Generalized, and Efficient Dual-System for Robotic Manipulation", "authors": ["Qingwen Bu", "Hongyang Li", "Li Chen", "Jisong Cai", "Jia Zeng", "Heming Cui", "Maoqing Yao", "Yu Qiao"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.08001", "pdf_url": "https://arxiv.org/pdf/2410.08001", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["generalist-specialist", "dual system", "diffusion specialist", "RoboDual", "control frequency"], "model_class": "VLA", "techniques": ["dual-system architecture", "VLA generalist + small diffusion specialist"], "base_model": "OpenVLA", "params": "20M trainable specialist", "hardware": "", "reported_metric": "3.8x higher control frequency in real-world | 3.8x higher control frequency in real deployment; +26.7% real-world vs OpenVLA", "metric_measures": "closed-loop control", "benchmarks": ["CALVIN", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Pairs a slow VLA generalist with a lightweight diffusion-transformer specialist that rolls out multi-step actions conditioned on generalist outputs, raising control rate and precision.", "relevance": 5, "relevance_reason": "Canonical fast low-level policy under slow VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA; arXiv search: RoboDual generalist specialist"}
{"id": "2410.04640", "title": "Unpacking Failure Modes of Generative Policies: Runtime Monitoring of Consistency and Progress", "authors": ["Christopher Agia", "Rohan Sinha", "Jingyun Yang", "Zi-ang Cao", "Rika Antonova", "Marco Pavone", "Jeannette Bohg"], "year": 2024, "venue": "CoRL 2024", "type": "paper", "url": "https://arxiv.org/abs/2410.04640", "pdf_url": "https://arxiv.org/pdf/2410.04640", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["Sentinel", "runtime monitoring", "temporal action consistency", "VLM progress monitoring"], "model_class": "diffusion/flow policy", "techniques": ["statistical temporal consistency", "VLM monitoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "18% more failures detected than either detector alone", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Splits runtime monitoring into fast statistical detection of erratic behavior via temporal action consistency of generative policies and slower VLM checks for task-progress failures.", "relevance": 5, "relevance_reason": "Seed: runtime monitors for generative policies; fast vs slow detector split.", "foundational_technique": false, "verified": true, "found_via": "seed; arXiv search: runtime monitoring robot policy"}
{"id": "2303.04137", "title": "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion", "authors": ["Cheng Chi", "Zhenjia Xu", "Siyuan Feng", "Eric Cousineau", "Yilun Du", "Benjamin Burchfiel", "Russ Tedrake", "Shuran Song", "Zhen-Jia Xu", "Si-Yuan Feng", "B. Burchfiel"], "year": 2023, "venue": "RSS 2023 (extended journal version)", "type": "paper", "url": "https://arxiv.org/abs/2303.04137", "pdf_url": "https://arxiv.org/pdf/2303.04137", "code_url": "https://diffusion-policy.cs.columbia.edu", "clusters": ["C1", "C2", "C7"], "keywords": ["diffusion policy", "receding horizon", "action sequences"], "model_class": "diffusion/flow policy", "techniques": ["diffusion action generation", "action chunking", "action diffusion", "receding-horizon execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Robomimic", "Push-T", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Introduces Diffusion Policy, generating action sequences by conditional denoising and executing them with receding-horizon control, the base for many chunked-execution methods.", "relevance": 5, "relevance_reason": "Seminal chunked generative policy whose latency/open-loop tradeoffs C7 methods address.", "foundational_technique": true, "verified": true, "found_via": "seed; snowball: references of 2506.07339 (S2)"}
{"id": "2304.13705", "title": "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware", "authors": ["Tony Z. Zhao", "Vikash Kumar", "Sergey Levine", "Chelsea Finn"], "year": 2023, "venue": "RSS 2023", "type": "preprint", "url": "https://arxiv.org/abs/2304.13705", "pdf_url": "https://arxiv.org/pdf/2304.13705", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["ACT", "action chunking", "temporal ensembling", "ALOHA"], "model_class": "other", "techniques": ["action chunking", "temporal ensembling", "CVAE", "CVAE transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot bimanual (6 tasks)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Introduces ALOHA low-cost bimanual teleoperation and Action Chunking with Transformers (ACT), predicting action sequences and smoothing them via temporal ensembling of overlapping chunks.", "relevance": 5, "relevance_reason": "Seminal origin of action chunking and temporal ensembling, the open-loop/smoothing baseline for C7 methods.", "foundational_technique": false, "verified": true, "found_via": "seed; seed; arXiv search title"}
{"id": "2306.03310", "title": "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning", "authors": ["Bo Liu", "Yifeng Zhu", "Chongkai Gao", "Yihao Feng", "Qiang Liu", "Yuke Zhu", "Peter Stone"], "year": 2023, "venue": "NeurIPS 2023 Datasets & Benchmarks", "type": "benchmark", "url": "https://arxiv.org/abs/2306.03310", "pdf_url": "https://arxiv.org/pdf/2306.03310", "code_url": "https://github.com/Lifelong-Robot-Learning/LIBERO", "clusters": ["C4"], "keywords": ["LIBERO", "lifelong learning", "simulation benchmark", "LIBERO-Spatial", "LIBERO-Object", "LIBERO-Goal", "LIBERO-Long"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Procedurally generated manipulation benchmark with four suites (130 tasks) plus teleoperated demos, originally for lifelong learning; its Spatial/Object/Goal/Long suites became the de facto VLA success-rate benchmark on which most VLA acceleration papers report.", "relevance": 5, "relevance_reason": "Seed benchmark; the standard testbed where efficiency methods report accuracy retention.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2301.04104", "title": "Mastering Diverse Domains through World Models", "authors": ["Danijar Hafner", "Jurgis Pasukonis", "Jimmy Ba", "Timothy Lillicrap"], "year": 2023, "venue": "Nature 2025 (arXiv 2023)", "type": "paper", "url": "https://www.nature.com/articles/s41586-025-08744-2", "pdf_url": "https://arxiv.org/pdf/2301.04104", "code_url": "https://github.com/danijar/dreamerv3", "clusters": ["C8"], "keywords": ["model-based RL", "latent imagination", "RSSM"], "model_class": "world model", "techniques": ["latent world model", "imagination-based policy learning", "symlog/normalization robustness"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Atari", "DMC", "Minecraft", "Crafter", "150+ tasks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "General model-based RL agent that learns a recurrent latent world model and trains actor-critic purely in imagined rollouts; fixed hyperparameters work across 150+ tasks and it was first to get Minecraft diamonds from scratch.", "relevance": 5, "relevance_reason": "Seed; canonical compact latent world model whose imagination-based learning underpins most later robot world-model RL.", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2310.17552", "title": "Model-Based Runtime Monitoring with Interactive Imitation Learning", "authors": ["Huihan Liu", "Shivin Dass", "Roberto Martín-Martín", "Yuke Zhu"], "year": 2023, "venue": "ICRA 2024", "type": "paper", "url": "https://arxiv.org/abs/2310.17552", "pdf_url": "https://arxiv.org/pdf/2310.17552", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["runtime monitoring", "latent dynamics", "failure classifier", "interactive imitation learning"], "model_class": "world model", "techniques": ["cVAE latent dynamics", "failure classifier on imagined futures"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Learns a latent dynamics model plus failure classifier from deployment data to simulate outcomes of upcoming actions and flag risky states before failures occur.", "relevance": 5, "relevance_reason": "Foundational world-model-based runtime monitor.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2312.01990", "title": "SARA-RT: Scaling up Robotics Transformers with Self-Adaptive Robust Attention", "authors": ["Isabel Leal", "Krzysztof Choromanski", "Deepali Jain", "Kumar Avinava Dubey", "Jake Varley", "Michael S. Ryoo", "Yao Lu", "Frederick Liu"], "year": 2023, "venue": "ICRA 2024", "type": "paper", "url": "https://arxiv.org/abs/2312.01990", "pdf_url": "https://arxiv.org/pdf/2312.01990", "code_url": "", "clusters": ["C1", "C2", "C9"], "keywords": ["linear attention", "up-training", "RT-2", "point cloud transformer", "on-robot deployment"], "model_class": "VLA", "techniques": ["linear attention conversion (up-training)"], "base_model": "RT-2", "params": "", "hardware": "", "reported_metric": "best SARA-RT-2 +10.6% accuracy and 14% faster than RT-2 (per DeepMind blog)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Up-training converts pretrained quadratic-attention robot transformers, including billion-parameter VLAs, into linear-attention versions for on-robot deployment while preserving quality.", "relevance": 5, "relevance_reason": "Early Google DeepMind work explicitly targeting on-robot efficiency of VLAs.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Google DeepMind robotics"}
{"id": "2310.05905", "title": "TAIL: Task-specific Adapters for Imitation Learning with Large Pretrained Models", "authors": ["Zuxin Liu", "Jesse Zhang", "Kavosh Asadi", "Yao Liu", "Ding Zhao", "Shoham Sabach", "Rasool Fakoor"], "year": 2023, "venue": "ICLR 2024", "type": "preprint", "url": "https://arxiv.org/abs/2310.05905", "pdf_url": "https://arxiv.org/pdf/2310.05905", "code_url": "", "clusters": ["C3"], "keywords": ["TAIL", "task-specific adapters", "continual imitation learning"], "model_class": "other", "techniques": ["LoRA", "bottleneck adapters", "prefix/P-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "LoRA uses ~1% of full-FT trainable parameters (abstract)", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Early systematic comparison of adapters, prompt tuning and LoRA for adapting large pretrained policies to new control tasks in continual imitation learning; LoRA best with ~1% params and no forgetting.", "relevance": 5, "relevance_reason": "Foundational PEFT-for-robot-policy study.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: TAIL task-specific adapters imitation learning LoRA robot continual"}
{"id": "2310.16828", "title": "TD-MPC2: Scalable, Robust World Models for Continuous Control", "authors": ["Nicklas Hansen", "Hao Su", "Xiaolong Wang"], "year": 2023, "venue": "ICLR 2024", "type": "paper", "url": "https://arxiv.org/abs/2310.16828", "pdf_url": "https://arxiv.org/pdf/2310.16828", "code_url": "https://github.com/nicklashansen/tdmpc2", "clusters": ["C8"], "keywords": ["TD-MPC", "latent MPC", "decoder-free world model"], "model_class": "world model", "techniques": ["implicit latent world model", "MPPI planning in latent space", "TD learning"], "base_model": "", "params": "up to 317M", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DMControl", "Meta-World", "ManiSkill2", "MyoSuite"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Improves TD-MPC into a scalable decoder-free latent world model that plans with sampling-based MPC in latent space; a single 317M agent handles 80 tasks across embodiments with one hyperparameter set.", "relevance": 5, "relevance_reason": "Seed; reference design for cheap latent-space online planning with a learned world model.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2205.09778", "title": "FogROS2: An Adaptive Platform for Cloud and Fog Robotics Using ROS 2", "authors": ["Jeffrey Ichnowski", "Kaiyuan Chen", "Karthik Dharmarajan", "Simeon Adebola", "Michael Danielczuk", "Vıctor Mayoral-Vilches", "Nikhil Jha", "Hugo Zhan", "Edith LLontop", "Derek Xu", "Camilo Buscaron", "John Kubiatowicz", "Ion Stoica", "Joseph Gonzalez", "Ken Goldberg"], "year": 2022, "venue": "ICRA 2023", "type": "paper", "url": "https://arxiv.org/abs/2205.09778", "pdf_url": "https://arxiv.org/pdf/2205.09778", "code_url": "https://github.com/BerkeleyAutomation/FogROS2", "clusters": ["C6"], "keywords": ["fog robotics", "cloud offloading", "ROS 2"], "model_class": "other", "techniques": ["cloud offloading", "video compression", "automated cloud provisioning"], "base_model": "", "params": "", "hardware": "AWS cloud instances", "reported_metric": "grasp planning 14 s -> 1.2 s; motion planning 45x faster; 97% lower image RTT vs FogROS1", "metric_measures": "relative latency", "benchmarks": ["SLAM", "grasp planning", "motion planning"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Open-source ROS 2 framework that launches selected unmodified ROS 2 nodes on cloud machines with secure networking and compressed image transport, letting low-compute robots offload SLAM, grasp and motion planning.", "relevance": 5, "relevance_reason": "Seed system for cloud/fog robotics offloading; basis of many later remote-inference robot works.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "1x-world-model-self-learning-blog-2026", "title": "1X World Model (1XWM) as a robot policy", "authors": ["1X Technologies"], "year": 2026, "venue": "blog (1X)", "type": "blog", "url": "https://www.1x.tech/discover/world-model-self-learning", "pdf_url": "", "code_url": "", "clusters": ["C6", "C7", "C9"], "keywords": ["video world model policy", "inverse dynamics model", "offboard inference"], "model_class": "world model", "techniques": ["text-to-video planning + IDM", "sliding window frames"], "base_model": "", "params": "14B video backbone", "hardware": "offboard multi-GPU server", "reported_metric": "11 s per WM inference (multi-GPU) for 5 s of video; IDM 1 s", "metric_measures": "other", "benchmarks": ["real robot (NEO)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Uses a 14B text-conditioned video diffusion model to imagine task execution and an IDM to extract actions, run offboard; latency (~11 s) highlights the gap for world-model policies.", "relevance": 4, "relevance_reason": "Concrete latency of a world-model-as-policy system; motivates efficient world models.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: 1X"}
{"id": "2604.09244", "title": "2D or 3D: Who Governs Salience in VLA Models? -- Tri-Stage Token Pruning Framework with Modality Salience Awareness", "authors": ["Zihao Zheng", "Sicheng Tian", "Zhihao Mao", "Lingyue Zhang", "Chenyue Li", "Ziyun Zhang", "Hong Gao", "Yuchen Huang", "Yutong Xu", "Guojie Luo", "Xiang Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.09244", "pdf_url": "https://arxiv.org/pdf/2604.09244", "code_url": "", "clusters": ["C2"], "keywords": ["MVLA"], "model_class": "VLA", "techniques": ["token pruning", "modality salience"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 2.55x inference speedup, 5.8% overhead", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Tri-stage analysis of 2D vs 3D token salience in multi-visual-modal VLAs and a matching pruning framework that selects tokens per modality.", "relevance": 4, "relevance_reason": "Token pruning for 2D+3D VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2601.14628", "title": "A Brain-inspired Embodied Intelligence for Fluid and Fast Reflexive Robotics Control", "authors": ["Wei-Yu Guo", "He Zhang", "Peng-Teng Li", "Tiefu Cai", "Ziyang Chen", "Yan-Dong Guo", "Xiao He", "Yong-Kui Yang", "Ying Sun", "Hui Xiong"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2601.14628", "pdf_url": "https://arxiv.org/pdf/2601.14628", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["NeuroVLA", "reflexes"], "model_class": "VLA", "techniques": ["neuromorphic computing"], "base_model": "", "params": "", "hardware": "neuromorphic processor", "reported_metric": "0.4 W; safety reflexes <20 ms", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Bio-inspired VLA with cortex/cerebellum/spinal layers deployed on neuromorphic hardware, reporting very low power and fast safety reflexes.", "relevance": 4, "relevance_reason": "Energy-efficient neuromorphic deployment.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310; cited by 2607.10172"}
{"id": "2604.04913", "title": "A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens", "authors": ["Tommie Kerssies", "G. Berton", "Ju He", "Qi-Hang Yu", "Wufei Ma", "Daan de Geus", "G. Dubbelman", "Liang-Chieh Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.04913", "pdf_url": "https://arxiv.org/pdf/2604.04913", "code_url": "", "clusters": ["C8"], "keywords": ["efficient generative world model"], "model_class": "world model", "techniques": ["delta tokens", "feature-space prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "35x fewer parameters and 2,000x fewer FLOPs than existing generative world models; 1,024x token reduction at 512x512", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Encodes frame-to-frame feature changes as one token per frame, enabling a far smaller and cheaper generative world model.", "relevance": 4, "relevance_reason": "Directly about world-model efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2606.00267, 2609.33177"}
{"id": "2606.10366", "title": "A Practical Recipe Towards Improving Sim-and-Real Correlation for VLA Evaluation", "authors": ["Shuo Wang", "Hanyuan Xu", "Yingdong Hu", "Fanqi Lin", "Yang Gao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.10366", "pdf_url": "https://arxiv.org/pdf/2606.10366", "code_url": "", "clusters": ["C4"], "keywords": ["sim-real correlation", "ranking consistency", "VLA evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["multiple sim platforms", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Systematic study across simulators, VLAs and perturbations of whether sim evaluation preserves real-world rankings and failure patterns, with recipes for improving alignment.", "relevance": 4, "relevance_reason": "Methodology for trusting sim results.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA evaluation simulation real correlation"}
{"id": "vla-edge-processor-token-filtering-pnm-2026", "title": "A Scalable Vision-Language-Action Edge Processor with Token Filtering and Hybrid Processing-Near-Memory for 6.7-ms Perception-to-Action Latency", "authors": ["Jin Wang", "Zhengke Yang", "Yulong Chen", "Yufu Zhang", "Haolong Li", "Ruijie Peng"], "year": 2026, "venue": "2026 IEEE/JSAP Symposium on VLSI Technology and Circuits", "type": "paper", "url": "", "pdf_url": "", "code_url": "", "clusters": ["C5", "C9"], "keywords": ["VLA accelerator", "processing-near-memory", "token filtering", "RRAM-SRAM"], "model_class": "VLA", "techniques": ["SimHash temporal token filtering", "hybrid RRAM-SRAM PNM", "zero external memory access"], "base_model": "", "params": "", "hardware": "16-chip 55 nm module", "reported_metric": "6.7 ms perception-to-action (150 Hz), 12.74 mJ", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Custom VLA edge chip that filters temporally redundant tokens and keeps weights in near-memory compute, reaching 6.7 ms perception-to-action latency.", "relevance": 4, "relevance_reason": "Dedicated hardware for onboard VLA inference.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2604.05672", "title": "A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model", "authors": ["Kai-Dong Zhang", "Jian Zhang", "Rongtao Xu", "Yu Sun", "Shuoshuo Xue", "You-Peng Wen", "Xiaoyu Guo", "Minghao Guo", "Weijia Liufu", "Zi-Hou Liu", "Kang-Yu Ji", "Yang-Song Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.05672", "pdf_url": "https://arxiv.org/pdf/2604.05672", "code_url": "", "clusters": ["C1", "C2", "C3", "C9"], "keywords": ["truncated VLM", "adaptive", "open-source", "low-cost"], "model_class": "VLA", "techniques": ["backbone truncation", "efficient action head"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A1 is an open VLA that truncates the VLM backbone and streamlines the action head for high-throughput inference on commodity hardware.", "relevance": 4, "relevance_reason": "Efficient open VLA.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2601.19634", "title": "AC^2-VLA: Action-Context-Aware Adaptive Computation in Vision-Language-Action Models for Efficient Robotic Manipulation", "authors": ["Wenda Yu", "Tianshi Wang", "Fengling Li", "Jingjing Li", "Lei Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.19634", "pdf_url": "https://arxiv.org/pdf/2601.19634", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["adaptive computation", "cognition reuse", "token pruning", "layer skipping", "selective execution"], "model_class": "VLA", "techniques": ["feature reuse", "token pruning", "dynamic layer skipping", "self-distillation", "caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.79x speedup; FLOPs 29.4%", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Conditions how much of the VLA to run on observations, instruction and previous actions, unifying reuse across timesteps, token pruning, and selective component execution.", "relevance": 4, "relevance_reason": "Dynamic depth/compute reduction in VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; search: pruning vision-language-action"}
{"id": "2603.18464", "title": "AcceRL: A Distributed Asynchronous Reinforcement Learning and World Model Framework for Vision-Language-Action Models", "authors": ["Chengxuan Lu", "Shukuan Wang", "Yanjie Li", "Yingying Fang", "Huoyan Wang", "Tian Zhang", "Wei Liu", "Shiji Jin", "Fuyuan Qian", "Peiming Li", "Chao Xu", "Baigui Sun", "Yang Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.18464", "pdf_url": "https://arxiv.org/pdf/2603.18464", "code_url": "", "clusters": ["C3", "C6", "C8"], "keywords": ["asynchronous RL", "distributed", "world model", "asynchronous distributed RL", "world model integration", "VLA"], "model_class": "other", "techniques": ["asynchronous distributed RL", "world-model rollouts", "decoupled rollout/inference/update", "plug-in world models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Distributed asynchronous RL framework for VLAs that removes synchronization bubbles and can plug in world models as rollout environments.", "relevance": 4, "relevance_reason": "Training-system efficiency for VLA RL.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations); search query"}
{"id": "2606.15285", "title": "Acting While Understanding: Asynchronous Semantic-Action Decoupling for Real-Time Vision-Language-Action Models", "authors": ["Yan, Shenhao", "Wang, Ge", "Liu, Qi", "Meng, Weilin", "Yang, Jiahao", "Yao, Chengsi", "Feng, Fan", "Ma, Xiaoguang", "Zhao, Yiming", "Han, Yatong", "Shenhao Yan", "Ge Wang", "Qi Liu", "Weilin Meng", "Jiahao Yang", "Chengsi Yao", "Fan Feng", "Xiaoguang Ma", "Yiming Zhao", "Yatong Han", "Shen-Hao Yan", "Jia-Hao Yang", "Cheng-Si Yao", "Xiao-Guang Ma", "Yi-Ming Zhao", "Ya-Tong Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.15285", "pdf_url": "https://arxiv.org/pdf/2606.15285", "code_url": "", "clusters": ["C1", "C2", "C4", "C6", "C7", "C9"], "keywords": ["semantic-action decoupling", "server-side throughput", "closed-loop", "low/high frequency modules", "asynchronous", "stale semantics", "dual-rate"], "model_class": "VLA", "techniques": ["asynchronous semantic updates", "time-misalignment training", "asynchronous semantic/action decoupling", "historical action conditioning", "low-frequency semantic module", "high-frequency action module", "asynchronous low-frequency understanding module"], "base_model": "pi0.5, UniVLA", "params": "", "hardware": "", "reported_metric": "up to 35.6 Hz server-side action-module inference throughput", "metric_measures": "throughput", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Splits an existing VLA at its semantic-action interface so a slow module asynchronously refreshes semantic conditions and a fast module outputs actions continuously, compensating stale semantics.", "relevance": 4, "relevance_reason": "Slow/fast split that maps naturally onto cloud/edge partitioning.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: closed-loop vs open-loop control frequency VLA; snowball from 2506.01844; snowball: citations of 2506.07339 (S2); venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2606.25985", "title": "Action ControlNet: A Lightweight Delay-Aware Adapter for Smooth Asynchronous Control in Vision-Language-Action Models", "authors": ["Tiecheng Guo", "Meng Guo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.25985", "pdf_url": "https://arxiv.org/pdf/2606.25985", "code_url": "", "clusters": ["C1", "C2", "C6", "C7"], "keywords": ["delay-aware adapter", "asynchronous control", "chunk handoff", "Action ControlNet", "chunk stitching"], "model_class": "VLA", "techniques": ["asynchronous inference", "residual adapter", "residual adapter on action head", "residual adapter conditioned on executed motion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Kinetix", "Meta-World", "SO-ARM101", "Meta-World MT50", "SO-ARM101 real"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Lightweight ControlNet-style adapter that conditions a mostly frozen diffusion/flow action head on the executed motion suffix, smoothing asynchronous handoffs under inference delay.", "relevance": 4, "relevance_reason": "Delay-robust async control with little training.", "foundational_technique": false, "verified": true, "found_via": "search query: lightweight VLA; snowball from 2506.07339; arXiv search: \"action chunk\" asynchronous inference"}
{"id": "2605.25537", "title": "Action-Prior Denoising for Smooth Real-Time Chunking", "authors": ["Dongyang Liu", "Zhaowen Zheng", "Yu Sun", "Longxu Zhang", "Yixuan Liu", "Hao Wan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.25537", "pdf_url": "https://arxiv.org/pdf/2605.25537", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["soft RTC", "training-time RTC", "action prior", "action-prior denoising", "real-time chunking"], "model_class": "diffusion/flow policy", "techniques": ["asynchronous inference", "training-time delay simulation", "training-time RTC", "soft prefix masking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "solve rate 0.809 vs 0.815 (hard RTC); -9.1% action delta, -9.6% jerk at high delay", "metric_measures": "other", "benchmarks": ["Kinetix", "real robot", "real robot sorting"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Generalizes training-time RTC from a binary prefix mask to soft priors built from partially denoised previous-chunk actions, keeping later overlap actions close to the old plan and reducing jerk at high delay with near-naive runtime.", "relevance": 4, "relevance_reason": "Smoothness under inference delay for chunked policies.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk inference delay; arXiv search: real-time chunking"}
{"id": "2607.14439", "title": "Active Real-World Factor-Based Evaluation for Generalist Robot Policies", "authors": ["Liao, Andrew", "Cui, Hanchen", "Desingh, Karthik", "Deshwal, Aryan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.14439", "pdf_url": "https://arxiv.org/pdf/2607.14439", "code_url": "", "clusters": ["C4"], "keywords": ["active evaluation", "sequential experimental design", "surrogate model", "task factors"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Treats real-world policy evaluation as active experimental design over task factors, choosing configurations that maximize information about the performance distribution and failure regions.", "relevance": 4, "relevance_reason": "Sample-efficient real-robot evaluation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RobotArena infinity"}
{"id": "2606.08508", "title": "ActProbe: Action-Space Probe for Early Failure Detection of Generative Robot Policies", "authors": ["Bingjia Huang", "Xiangyu Li", "Xiang Wang", "Liang Mi", "Zixu Hao", "Weijun Wang", "Hao Wu", "Kun Li", "Yunxin Liu", "Ting Cao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08508", "pdf_url": "https://arxiv.org/pdf/2606.08508", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["ActProbe", "action-space probe", "temporal consistency error", "early failure detection"], "model_class": "diffusion/flow policy", "techniques": ["action-only LSTM-MLP probe"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+12.7% hypervolume on F1-timeliness frontier; 2.9x fewer PPO interactions", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Black-box failure detector using only emitted action chunks (consecutive-chunk consistency error and chunk magnitude) from a single forward pass to raise early alerts.", "relevance": 4, "relevance_reason": "Zero-overhead black-box runtime monitor.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure prediction generative robot policy"}
{"id": "2603.11984", "title": "Ada3Drift: Adaptive Training-Time Drifting for One-Step 3D Visuomotor Robotic Manipulation", "authors": ["Chong-Yang Xu", "Yi Zou", "Zi-Li Feng", "Fan-Man Meng", "Shuai-Cheng Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.11984", "pdf_url": "https://arxiv.org/pdf/2603.11984", "code_url": "", "clusters": ["C1"], "keywords": ["3D policy", "point cloud", "one-step"], "model_class": "diffusion/flow policy", "techniques": ["one-step generation", "drifting models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1 NFE vs 10 NFE for diffusion baselines", "metric_measures": "other", "benchmarks": ["Adroit", "Meta-World", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Point-cloud policy built on drifting models that generates action chunks in a single step with adaptive training-time drifting.", "relevance": 4, "relevance_reason": "One-step policy vs 10-step diffusion", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.11363", "title": "Adaptation of Generalist Robot Policies with Minimal Data", "authors": ["Shreyas Kowshik", "Sreyas Venkataraman", "Leo Wang", "Niharika Pant", "Max Simchowitz", "Aviral Kumar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.11363", "pdf_url": "https://arxiv.org/pdf/2608.11363", "code_url": "", "clusters": ["C3"], "keywords": ["MiDAS", "minimal-data adaptation", "offline-to-online RL"], "model_class": "VLA", "techniques": ["BC anchoring", "offline-to-online RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Offline-to-online recipe that adapts a VLA from one or few demos followed by autonomous online interaction.", "relevance": 4, "relevance_reason": "Minimal-data adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2604.04161", "title": "Adaptive Action Chunking at Inference-time for Vision-Language-Action Models", "authors": ["Yuanchang Liang", "Xiaobo Wang", "Kai Wang", "Shuo Wang", "Xiaojiang Peng", "Haoyu Chen", "David Kim Huat Chua", "Prahlad Vadakkepat"], "year": 2026, "venue": "CVPR 2026", "type": "paper", "url": "https://arxiv.org/abs/2604.04161", "pdf_url": "https://arxiv.org/pdf/2604.04161", "code_url": "", "clusters": ["C1"], "keywords": ["adaptive chunk size", "action entropy"], "model_class": "VLA", "techniques": ["adaptive action chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Chooses chunk length at inference based on predicted-action entropy, balancing reactivity and consistency.", "relevance": 4, "relevance_reason": "Adaptive chunking.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking"}
{"id": "2608.06008", "title": "Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features", "authors": ["Sining Ang", "Yuguang Yang", "Yan Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.06008", "pdf_url": "https://arxiv.org/pdf/2608.06008", "code_url": "", "clusters": ["C8"], "keywords": ["Adaptive-WAM", "video diffusion"], "model_class": "world-action model", "techniques": ["early exit", "intermediate features"], "base_model": "Wan2.2-5B", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Attaches trajectory heads to intermediate DiT blocks of a video diffusion world model and exits early once a quality scorer is satisfied, skipping future-video synthesis.", "relevance": 4, "relevance_reason": "Early-exit WAM planning.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\")"}
{"id": "2608.29208", "title": "AdaVLA: Adaptive Step Flow Matching for Training-free Acceleration of Vision-Language-Action Models", "authors": ["Sunghwan Han", "Youngtae Han", "Youngmin Yi"], "year": 2026, "venue": "IROS 2026", "type": "preprint", "url": "https://arxiv.org/abs/2608.29208", "pdf_url": "https://arxiv.org/pdf/2608.29208", "code_url": "", "clusters": ["C1", "C3", "C9"], "keywords": ["adaptive steps", "trajectory curvature", "MLP pruning", "training-free", "adaptive step flow matching"], "model_class": "VLA", "techniques": ["adaptive sampling steps", "dynamic pruning", "adaptive number of flow steps"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses flow-trajectory curvature as a confidence signal to cut ODE steps and adjust MLP pruning online without training data.", "relevance": 4, "relevance_reason": "Training-free flow VLA acceleration on Jetson.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA; venue sweep: arXiv API VLA AND Jetson"}
{"id": "2606.06660", "title": "AEGIS: A Backup Reflex for Physical AI", "authors": ["J. Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.06660", "pdf_url": "https://arxiv.org/pdf/2606.06660", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["AEGIS backup reflex", "selective escalation", "activation probe", "policy switching"], "model_class": "VLA", "techniques": ["activation probe early warning", "gated switch to stronger policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "recovers 10.1% of lost trajectories vs 4.6% blind escalation; stronger policy used on 38% of steps", "metric_measures": "other", "benchmarks": ["LIBERO-Spatial"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Probe on a weak policy’s activations flags high-risk steps and hands control to a stronger policy only for those steps, framing fallback as a timing decision.", "relevance": 4, "relevance_reason": "Fallback-policy switching triggered by runtime monitor.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2609.20761", "title": "Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control", "authors": ["Hanchu Zhou", "Brendan Lynch", "Raman Goyal", "Dechen Gao", "Begum Kasap", "Boqi Zhao", "Junshan Zhang", "Han-Chu Zhou", "De-Chen Gao", "Bo Zhao", "Jun-Shan Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20761", "pdf_url": "https://arxiv.org/pdf/2609.20761", "code_url": "", "clusters": ["C8"], "keywords": ["Agile-WAM", "tactile"], "model_class": "world-action model", "techniques": ["shared latent flow matching", "multi-horizon prediction", "compact tactile WAM"], "base_model": "", "params": "", "hardware": "", "reported_metric": "11.9 ms inference latency", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Tactile WAM that jointly generates action-chunk latents and future visual/tactile latents with one flow-matching process from a shared latent, predicting vision at a longer offset than touch.", "relevance": 4, "relevance_reason": "Small tactile WAM at ~12 ms latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2410.24164, 2412.04987"}
{"id": "2608.29537", "title": "AGM: Achievement-Grounded Memory for Closed-Loop Agents with Frozen VLA Policies", "authors": ["Hongbo Gao", "Zeyu Ni", "Xin Wen", "Siyu Xu", "Ruifeng Li", "Rui Xu", "Yunke Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29537", "pdf_url": "https://arxiv.org/pdf/2608.29537", "code_url": "", "clusters": ["C7"], "keywords": ["open-loop chunks", "progress tracking", "frozen VLA"], "model_class": "VLA", "techniques": ["achievement verification", "external memory", "lightweight verifier"], "base_model": "", "params": "2.43M (verifier head)", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboMME", "real robot"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a 2.43M-parameter verification head plus point tracking and gripper-load cues so a frozen VLA only advances its task pointer after physically verifying each subgoal, deciding when to continue, retry, or stop.", "relevance": 4, "relevance_reason": "Directly addresses the continue/retry/terminate decision for open-loop chunk execution with a cheap verifier.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*)"}
{"id": "2601.12277", "title": "An Efficient and Multi-Modal Navigation System with One-Step World Model", "authors": ["Wangtian Shen", "Ziyang Meng", "Jinming Ma", "Mingliang Zhou", "Diyun Xiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.12277", "pdf_url": "https://arxiv.org/pdf/2601.12277", "code_url": "", "clusters": ["C8"], "keywords": ["one-step navigation world model", "real-time planning"], "model_class": "world model", "techniques": ["lightweight one-step generation instead of multi-step AR diffusion", "iterative optimization planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Lightweight navigation world model with one-step prediction that makes world-model-based trajectory optimization fast enough for real-time robot navigation.", "relevance": 4, "relevance_reason": "One-step world model for real-time planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.15046", "title": "AnoleVLA: Lightweight Vision-Language-Action Model with Deep State Space Models for Mobile Manipulation", "authors": ["Yusuke Takagi", "Motonari Kambara", "Daichi Yashima", "Koki Seno", "Kento Tokura", "Komei Sugiura"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.15046", "pdf_url": "https://arxiv.org/pdf/2603.15046", "code_url": "", "clusters": ["C1"], "keywords": ["deep state space model", "lightweight VLA", "mobile manipulation"], "model_class": "VLA", "techniques": ["state space model backbone"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~3x faster inference than a large VLA", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight VLA built on a deep SSM for multimodal sequences, faster and more successful than a large VLA on a mobile manipulator.", "relevance": 4, "relevance_reason": "SSM-based lightweight VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: lightweight VLA"}
{"id": "2609.33299", "title": "AquaWAM: A Dynamics-aware World Action Model for Underwater Embodied Agents", "authors": ["Cunhao Zhu", "Yifeng Wang", "Dongliang Xu", "Yunzhong Hou", "Yue Yao", "Chi Harold Liu", "Cun-Hao Zhu", "Yi-Feng Wang", "Dong-Liang Xu", "Yun-Zhong Hou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33299", "pdf_url": "https://arxiv.org/pdf/2609.33299", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["underwater", "AquaWAM", "world action model"], "model_class": "world-action model", "techniques": ["state-space world model", "compact state prediction"], "base_model": "", "params": "", "hardware": "NVIDIA Jetson AGX Orin", "reported_metric": "2.7x faster action decisions than U0; 72.6% success on USIM", "metric_measures": "relative latency", "benchmarks": ["USIM"], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Underwater world-action model predicting compact navigation states (DVL/IMU/pressure) including passive dynamics rather than images, cutting model size and compute.", "relevance": 4, "relevance_reason": "Onboard Jetson Orin speed comparison for a WAM.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2310.16828, 2402.15391"}
{"id": "2603.10126", "title": "AR-VLA: True Autoregressive Action Expert for Vision-Language-Action Models", "authors": ["Yutong Hu", "Jan-Nico Zaech", "Nikolay Nikolov", "Yuanqi Yao", "Sombit Dey", "Giuliano Albanese", "Renaud Detry", "Luc Van Gool", "Danda Paudel"], "year": 2026, "venue": "RSS 2026", "type": "paper", "url": "https://arxiv.org/abs/2603.10126", "pdf_url": "https://arxiv.org/pdf/2603.10126", "code_url": "", "clusters": ["C1"], "keywords": ["autoregressive action expert", "long-lived memory", "asynchronous prefixes"], "model_class": "VLA", "techniques": ["autoregressive action expert", "asynchronous conditioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Standalone autoregressive action expert with its own history that runs continuously while refreshing vision-language prefixes at a slower rate, re-anchoring for staleness.", "relevance": 4, "relevance_reason": "Frequency-decoupled action expert.", "foundational_technique": false, "verified": true, "found_via": "search query: autoregressive action expert VLA"}
{"id": "2607.17257", "title": "Asynchronous Multimodal Diffusion Policy Composition via Latency-Aware Guidance Fusion", "authors": ["Zihao He", "Hongjie Fang", "Shirun Tang", "Cewu Lu", "Haoshu Fang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.17257", "pdf_url": "https://arxiv.org/pdf/2607.17257", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["LAG-Fusion", "latency-aware guidance fusion", "asynchronous multimodal composition"], "model_class": "diffusion/flow policy", "techniques": ["asynchronous guidance fusion", "reference-frame rebasing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lets modality-specific diffusion policies run at their native rates and inject denoising guidance whenever available, with a rebasing rule to keep asynchronous composition consistent.", "relevance": 4, "relevance_reason": "Explicitly latency-aware multi-rate policy composition.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"asynchronous\" \"diffusion policy\" delay"}
{"id": "2609.23896", "title": "BarrierFormer: Transformer-Guided Predictive Barrier Enforcement for Safe Robot Control", "authors": ["Anandsingh Chauhan", "Kunal Garg"], "year": 2026, "venue": "CoRL 2026", "type": "paper", "url": "https://arxiv.org/abs/2609.23896", "pdf_url": "https://arxiv.org/pdf/2609.23896", "code_url": "", "clusters": ["C7"], "keywords": ["CBF", "receding horizon", "safe control"], "model_class": "other", "techniques": ["control barrier functions", "transformer", "predictive safety"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A causal transformer trained with rollout-level CBF supervision predicts future rollouts and enforces barrier conditions over a horizon, avoiding MPC-style online optimization.", "relevance": 4, "relevance_reason": "Predictive (horizon-level) barrier enforcement matches the chunk-certification problem better than myopic CBF filters.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\")"}
{"id": "2605.00078", "title": "Being-H0.7: A Latent World-Action Model from Egocentric Videos", "authors": ["Hao Luo", "Wanpeng Zhang", "Yicheng Feng", "Sipeng Zheng", "Haiweng Xu", "Chaoyi Xu", "Ziheng Xi", "Yuhui Fu", "Zongqing Lu"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2605.00078", "pdf_url": "https://arxiv.org/pdf/2605.00078", "code_url": "", "clusters": ["C8"], "keywords": ["latent WAM", "latent queries", "egocentric video"], "model_class": "world-action model", "techniques": ["learnable latent reasoning queries", "future-informed dual branch (train only)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Brings future-aware reasoning into a VLA-style policy through learnable latent queries trained against future information, with no future frame generation at deployment.", "relevance": 4, "relevance_reason": "Latent WAM avoiding pixel prediction cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from NVIDIA WAM blog"}
{"id": "2603.24941", "title": "Beyond Attention Magnitude: Leveraging Inter-layer Rank Consistency for Efficient Vision-Language-Action Models", "authors": ["Peiju Liu", "Jinming Liu", "Xipeng Qiu", "Xuanjing Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.24941", "pdf_url": "https://arxiv.org/pdf/2603.24941", "code_url": "", "clusters": ["C1"], "keywords": ["visual token reduction", "rank consistency", "training-free"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "CogACT", "params": "", "hardware": "", "reported_metric": "78% fewer visual tokens, +6% success", "metric_measures": "other", "benchmarks": ["SIMPLER"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Selects visual tokens using consistency of their ranking across layers rather than raw attention magnitude, cutting tokens substantially without training.", "relevance": 4, "relevance_reason": "VLA token pruning.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient"}
{"id": "2603.13616", "title": "Beyond Binary Success: Sample-Efficient and Statistically Rigorous Robot Policy Comparison", "authors": ["David Snyder", "Apurva Badithela", "Nikolai Matni", "George Pappas", "Anirudha Majumdar", "Masha Itkina", "Haruki Nishimura"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.13616", "pdf_url": "https://arxiv.org/pdf/2603.13616", "code_url": "", "clusters": ["C4"], "keywords": ["safe anytime-valid inference", "sequential testing", "policy comparison", "partial credit"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot", "simulation"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Sequential, anytime-valid statistical tests for comparing robot policies on binary and continuous metrics (progress, reward, smoothness), allowing early stopping at a set confidence.", "relevance": 4, "relevance_reason": "Statistically rigorous comparison beyond binary success; applies to smoothness metrics affected by efficiency methods.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation statistical"}
{"id": "2602.20566", "title": "BFA++: Hierarchical Best-Feature-Aware Token Prune for Multi-View Vision Language Action Model", "authors": ["Haosheng Li", "Weixin Mao", "Zihan Lan", "Hongwei Xiong", "Hongan Wang", "Chenyang Si", "Ziwei Liu", "Xiaoming Deng", "Hua Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.20566", "pdf_url": "https://arxiv.org/pdf/2602.20566", "code_url": "", "clusters": ["C1"], "keywords": ["multi-view token pruning"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "pi0, RDT", "params": "", "hardware": "", "reported_metric": "1.8x (pi0) / 1.5x (RDT) speedup", "metric_measures": "relative latency", "benchmarks": ["RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Hierarchical pruning for multi-camera VLAs: intra-view predictors keep task regions, inter-view predictors drop cameras irrelevant to the current phase.", "relevance": 4, "relevance_reason": "Token pruning for multi-view VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning"}
{"id": "2604.25416", "title": "Biased Dreams: Limitations to Epistemic Uncertainty Quantification in Latent Dynamics Models", "authors": ["Julia Berger", "Bernd Frauenknecht", "Sebastian Trimpe", "Bastian Leibe"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.25416", "pdf_url": "https://arxiv.org/pdf/2604.25416", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["epistemic uncertainty", "RSSM ensembles", "compounding error"], "model_class": "world model", "techniques": ["analysis of ensemble disagreement in latent rollouts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Finds ensemble disagreement in Dreamer-style latent models tracks local but not accumulated rollout error because of an attractor effect, questioning its use as a safety signal.", "relevance": 4, "relevance_reason": "Limits of uncertainty-based world-model safety monitors.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2606.05271", "title": "BIDENT: Heterogeneous Operator-level Mapping for Efficient Edge Inference", "authors": ["Hoseok Kim", "Arghadip Das", "Soumendu Ghosh", "Arnab Raha", "Vijay Raghunathan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.05271", "pdf_url": "https://arxiv.org/pdf/2606.05271", "code_url": "", "clusters": ["C5"], "keywords": ["CPU/GPU/NPU", "operator-level mapping", "edge SoC", "VLA pipelines"], "model_class": "other", "techniques": ["operator-to-processing-unit assignment via shortest path", "intra-model parallelism"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Maps individual operators of models (including multi-stage VLA pipelines) to the best CPU/GPU/NPU on an edge SoC using profiled costs, for latency or energy minimization.", "relevance": 4, "relevance_reason": "Heterogeneous accelerator scheduling applicable to VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv: energy AND vision-language-action AND (edge OR battery OR power)"}
{"id": "2609.36305", "title": "Bilinear World Models: Learning Representations with Structured Dynamics for Efficient Control", "authors": ["Antonio Pariente", "Ignacio Boero", "Nikolai Matni", "Alejandro Ribeiro"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36305", "pdf_url": "https://arxiv.org/pdf/2609.36305", "code_url": "", "clusters": ["C8"], "keywords": ["bilinear dynamics", "structured latent world model", "planning time"], "model_class": "world model", "techniques": ["bilinear-parameterized latent dynamics", "action recoverability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "planning time reduced by nearly three orders of magnitude", "metric_measures": "relative latency", "benchmarks": ["2D/3D control tasks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Constrains JEPA latent dynamics to a bilinear form so planning becomes nearly closed-form, cutting planning time ~1000x while keeping control accuracy.", "relevance": 4, "relevance_reason": "Structural route to very fast latent planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.15475", "title": "Bit-Flip Attacks on Vision-Language-Action Models: Action-Decoding Architecture Shapes the Vulnerability", "authors": ["Yu-Dong Gao", "Ling-Han Chen", "Wenhan Wu", "Mia Zhou", "Jiyao Wang", "Kaiyang Ji", "Ming-Yu Guo", "Honglong Chen", "Yudong Gao", "Linghan Chen", "Kaiyan Ji", "Mingyu Guo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15475", "pdf_url": "https://arxiv.org/pdf/2608.15475", "code_url": "", "clusters": ["C2", "C7", "C8", "C9"], "keywords": ["bit-flip attack", "Rowhammer", "INT8", "security", "robustness of quantized VLAs"], "model_class": "VLA", "techniques": ["quantization (INT8) robustness analysis", "fault injection analysis"], "base_model": "pi0 and other VLAs", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Shows a handful of gradient-chosen bit flips in INT8 VLA weights can zero closed-loop success, with flow-matching heads needing far more flips than regression/token heads; protecting a few percent of weights helps.", "relevance": 4, "relevance_reason": "Safety/security of quantized deployed VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.07530; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2605.13382", "title": "BlockVLA: Accelerating Autoregressive VLA via Block Diffusion Finetuning", "authors": ["Ruiheng Wang", "Shuanghao Bai", "Haoran Zhang", "Badong Chen", "Xiangyu Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.13382", "pdf_url": "https://arxiv.org/pdf/2605.13382", "code_url": "", "clusters": ["C1"], "keywords": ["block diffusion", "discrete diffusion", "KV cache reuse"], "model_class": "VLA", "techniques": ["block diffusion decoding", "KV cache reuse"], "base_model": "", "params": "", "hardware": "", "reported_metric": "3.3x faster than discrete diffusion baselines", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SIMPLER"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fine-tunes AR VLA backbones into block diffusion policies: autoregressive across blocks with prefix KV reuse, parallel denoising within blocks.", "relevance": 4, "relevance_reason": "Parallel decoding VLA design.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "2605.30660", "title": "BOKBO (Best of K Bad Options): Calibrated Abstention for VLA Policies", "authors": ["Anya Singh", "Cabrel Happi", "Jai Relan", "Varun Nair", "Vidyut Baradwaj"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30660", "pdf_url": "https://arxiv.org/pdf/2605.30660", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["BOKBO", "conformal abstention", "K-sample inference", "executed-violation rate"], "model_class": "VLA", "techniques": ["conformal abstention layer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds conformal abstention to best-of-K VLA sampling so the system refuses to execute when all candidates are likely unsafe, with finite-sample guarantees on executed-violation rate.", "relevance": 4, "relevance_reason": "Guaranteed fallback/abstention for test-time scaled VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety\" \"action chunking\""}
{"id": "10.1109/ur69537.2026.11626770", "title": "Breaking the Resource Barrier: Parameter-Efficient Hierarchical VLA Fine-Tuning via Single-View Semantic Reasoning", "authors": ["Woo-Kyoung Jeong", "Yongwoo Gu", "June-sup Yi", "Hyoung-Rock Kim"], "year": 2026, "venue": "2026 23rd International Conference on Ubiquitous Robots (UR)", "type": "paper", "url": "https://doi.org/10.1109/UR69537.2026.11626770", "pdf_url": "", "code_url": "", "clusters": ["C3"], "keywords": ["parameter-efficient fine-tuning", "hierarchical VLA"], "model_class": "VLA", "techniques": ["LoRA", "PEFT", "critic-based action selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "98.1% average success on LIBERO", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Parameter-efficient hierarchical VLA fine-tuning with single-view semantic reasoning and critic-selected actions, matching full fine-tuning under limited compute.", "relevance": 4, "relevance_reason": "PEFT for VLAs under resource constraints", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2306.03310"}
{"id": "2605.26577", "title": "Bridging Control with Neural Network Verifier alpha-beta-CROWN: A Tutorial", "authors": ["Haoyu Li", "Xiangru Zhong", "Hao Cheng", "Bin Hu", "Huan Zhang"], "year": 2026, "venue": "ACC 2026 tutorial", "type": "talk", "url": "https://arxiv.org/abs/2605.26577", "pdf_url": "https://arxiv.org/pdf/2605.26577", "code_url": "", "clusters": ["C7"], "keywords": ["alpha-beta-CROWN", "control verification"], "model_class": "other", "techniques": ["bound propagation", "neural certificates"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Tutorial showing how to use the general-purpose alpha-beta-CROWN bounding engine to verify stability and safety certificates of learned controllers beyond specific structural assumptions.", "relevance": 4, "relevance_reason": "Practical bridge from NN verifiers to controller certification.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:Reluplex OR Marabou OR NNV OR beta-CROWN OR VNN-COMP"}
{"id": "2607.09045", "title": "Can the Cloud Drive? Infrastructure Feasibility of Offloading Autonomous Driving Across 5G and 6G", "authors": ["Pouya Parsa", "Kawon Han", "Seongjin Choi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.09045", "pdf_url": "https://arxiv.org/pdf/2607.09045", "code_url": "", "clusters": ["C5", "C6", "C9"], "keywords": ["cloud offload", "roofline GPU service model", "5G/6G", "cost", "autonomous driving VLA", "5G/6G offloading", "TCO", "roofline", "cloud offloading", "driving VLA", "~60 TFLOPs"], "model_class": "VLA", "techniques": ["analytical feasibility modeling", "analytical latency/cost model", "feature-level offloading", "feasibility analysis of offloading"], "base_model": "", "params": "", "hardware": "B300-class cloud GPUs (modelled)", "reported_metric": "driving VLA forward pass ~60 TFLOPs | autoregressive FP16 VLA decode ~114 ms on 2025 hardware; onboard VLA baseline up to $8,500 per vehicle-year", "metric_measures": "other", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Analytical study coupling cellular uplink limits, GPU roofline service and utilization-aware cost to determine when cloud offload of driving VLAs meets reactive/deliberative latency budgets and becomes cheaper than onboard compute.", "relevance": 4, "relevance_reason": "Rare explicit TCO + network feasibility analysis for offloaded VLA inference (driving domain).", "foundational_technique": false, "verified": true, "found_via": "arXiv: offloading AND (VLA OR robot foundation model); search query; venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2601.21998", "title": "Causal World Modeling for Robot Control", "authors": ["Lin Li", "Qihang Zhang", "Yiming Luo", "Shuai Yang", "Ruilin Wang", "Fei Han", "Mingrui Yu", "Zelin Gao", "Nan Xue", "Xing Zhu", "Yujun Shen", "Yinghao Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.21998", "pdf_url": "https://arxiv.org/pdf/2601.21998", "code_url": "https://github.com/robbyant/lingbot-va", "clusters": ["C8"], "keywords": ["LingBot-VA", "autoregressive video-action diffusion", "MoT"], "model_class": "world-action model", "techniques": ["shared vision-action latent MoT", "closed-loop rollout with real observations", "asynchronous inference pipeline"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Autoregressive diffusion WAM that predicts frames and actions together and pipelines action prediction with motor execution; widely used as a base for later WAM acceleration work.", "relevance": 4, "relevance_reason": "Base model for several efficiency papers (Flash-WAM, QuantWAMs).", "foundational_technique": false, "verified": true, "found_via": "snowball from NVIDIA WAM blog"}
{"id": "2606.25473", "title": "Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models", "authors": ["Kaiwen Zheng", "Guande He", "Min Zhao", "Jintao Zhang", "Hua-Yu Chen", "Jianfei Chen", "Chen-Hsuan Lin", "Mingying Liu", "Jun Zhu", "Qianli Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.25473", "pdf_url": "https://arxiv.org/pdf/2606.25473", "code_url": "", "clusters": ["C8"], "keywords": ["streaming video", "interactive world model"], "model_class": "world model", "techniques": ["diffusion distillation", "consistency models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1-2 sampling steps (Wan2.1-1.3B)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Autoregressive diffusion distillation recipe for streaming video and interactive world models, applied to Cosmos.", "relevance": 4, "relevance_reason": "Few-step distillation of world models.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2506.08009, 2511.00062"}
{"id": "2608.20791", "title": "CertVLA: Certified Defense against Physical Visual Attacks for Vision-Language-Action Models", "authors": ["Hui Lu", "Zhijie Peng", "Yuqi Lin", "Zaijia Yang", "Jiaming He", "Shuhan Ye", "Yi Yu", "Hanwei Zhu", "Bingquan Shen", "Alex Kot", "Xudong Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.20791", "pdf_url": "https://arxiv.org/pdf/2608.20791", "code_url": "", "clusters": ["C7"], "keywords": ["certification", "adversarial patch", "VLA"], "model_class": "VLA", "techniques": ["certified robustness", "conformal calibration", "masking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Provides a certified defense for continuous VLA actions under bounded patch/texture attacks by checking consistency across covering masks and conformally calibrating an episode-level score, extending query certificates to whole rollouts.", "relevance": 4, "relevance_reason": "One of the few formal certificates stated for closed-loop VLA action outputs.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*)"}
{"id": "2607.26789", "title": "CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation", "authors": ["Yu-Shan Liu", "Pei-Bo Sun", "Xin-Tao Chao", "Zhen-Ya Yang", "Yi-Fan Xie", "Lingfeng Zhang", "Shoujie Li", "Chenyu Tang", "Fangda Chen", "Xiao-Ping Zhang", "Wen-Bo Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.26789", "pdf_url": "https://arxiv.org/pdf/2607.26789", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["CheckVLA", "execution-time verification", "action-conditioned world model", "conformal risk"], "model_class": "world model", "techniques": ["action-conditioned world-model verification", "conformal calibration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Verifies open-loop chunk execution against a frozen action-conditioned world model, with conformally calibrated thresholds bounding unnecessary interventions.", "relevance": 4, "relevance_reason": "Runtime verification of committed open-loop chunks.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2607.12992", "title": "ChunkFlow: Towards Continuity-Consistent Chunked Policy Learning", "authors": ["Zhao Yang", "Yinan Shi", "Mingyuan Yao", "Wenyao Xue", "Yawei Jueluo", "Longjun Liu", "Yi-Nan Shi"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.12992", "pdf_url": "https://arxiv.org/pdf/2607.12992", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["boundary jitter", "chunked policy", "ChunkFlow"], "model_class": "VLA", "techniques": ["chunk overlap blending", "continuity losses", "AWAC fine-tuning", "chunk continuity"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Splits each chunk into frozen, editable and future zones, blends overlaps deterministically at execution and trains with seam and higher-order continuity losses plus history corruption, then AWAC fine-tunes.", "relevance": 4, "relevance_reason": "Addresses chunk-boundary jitter that arises when chunks are replaced mid-execution.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:timing AND (\"vision-language-action\" OR \"robot policy\") AND (deadline* OR jitter); snowball round 2: cites 2402.15391, 2405.12213"}
{"id": "2605.06247", "title": "CKT-WAM: Parameter-Efficient Context Knowledge Transfer Between World Action Models", "authors": ["Yuhua Jiang", "Yijun Guo", "Hongbing Yang", "Guojun Lei", "Nuo Chen", "Yinuo Zhang", "Shaoqiang Yan", "Bo Lin", "Feifei Gao", "Biqing Qi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.06247", "pdf_url": "https://arxiv.org/pdf/2605.06247", "code_url": "https://github.com/YuhuaJiang2002/CKT-WAM", "clusters": ["C2", "C3", "C4", "C8"], "keywords": ["knowledge transfer", "world action model", "adapters", "PEFT", "knowledge transfer between WAMs", "compact context", "parameter-efficient"], "model_class": "world-action model", "techniques": ["adapters", "context distillation", "mixture of adapters", "teacher hidden-state compression into text-embedding context"], "base_model": "", "params": "", "hardware": "", "reported_metric": "86.1% LIBERO-Plus with 1.17% trainable parameters", "metric_measures": "other", "benchmarks": ["LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Transfers knowledge between world-action models through a compact learned context in the text embedding space using routed adapters, training ~1% of parameters.", "relevance": 4, "relevance_reason": "Parameter-efficient WAM distillation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: parameter-efficient fine-tuning vision-language-action model arXiv; snowball from 2602.15922 (citations)"}
{"id": "2601.09512", "title": "CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion", "authors": ["Ralf Römer", "Yi Zhang", "Yuming Li", "Angela P. Schoellig"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.09512", "pdf_url": "https://arxiv.org/pdf/2601.09512", "code_url": "", "clusters": ["C3"], "keywords": ["CLARE", "adapter routing", "exemplar-free continual learning"], "model_class": "VLA", "techniques": ["adapters", "dynamic expansion", "routing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Exemplar-free continual VLA learning by adding lightweight adapters only where layer-wise feature similarity says they are needed.", "relevance": 4, "relevance_reason": "Parameter-efficient continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"continual learning\""}
{"id": "2603.01953", "title": "Closed-Loop Action Chunks with Dynamic Corrections for Training-Free Diffusion Policy", "authors": ["Peng-Yuan Wu", "Pingrui Zhang", "Zhi-Gang Wang", "Dong Wang", "Bin Zhao", "Xuelong Li"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.01953", "pdf_url": "https://arxiv.org/pdf/2603.01953", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["dynamic scenes", "training-free"], "model_class": "diffusion/flow policy", "techniques": ["closed-loop chunk correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "5% additional computation", "metric_measures": "other", "benchmarks": ["PushT", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "DCDP injects real-time dynamics features to correct action chunks during execution, improving responsiveness with little extra compute.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2601.21712", "title": "CoFreeVLA: Short-Horizon Collision-Free Dual-Arm Manipulation via Vision-Language-Action Model and Risk Estimation", "authors": ["Yaohua Liu", "Binkai Ou", "Hengjun Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.21712", "pdf_url": "https://arxiv.org/pdf/2601.21712", "code_url": "", "clusters": ["C7"], "keywords": ["CoFreeVLA", "self-collision risk estimator", "dual-arm", "action gating"], "model_class": "VLA", "techniques": ["short-horizon collision risk estimator", "risk-guided recovery"], "base_model": "", "params": "", "hardware": "", "reported_metric": "collision rate 0.54 -> 0.23; success 0.45 -> 0.61", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight short-horizon self-collision estimator gates risky VLA commands on dual-arm robots and synthesizes recovery trajectories.", "relevance": 4, "relevance_reason": "Runtime safety gating of VLA actions.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: safe robot foundation models"}
{"id": "2609.21908", "title": "CommitFlow: Semantic Commitment Verification and Local Correction for Long-Horizon Robot Manipulation VLA Execution", "authors": ["Zixiang Zhao", "Yansong Feng", "Yang Yang", "Chaoyu Wang", "Haoran Xiao", "Hui Zhang", "Chuang Cheng", "Jianjun Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21908", "pdf_url": "https://arxiv.org/pdf/2609.21908", "code_url": "", "clusters": ["C7"], "keywords": ["CommitFlow"], "model_class": "VLA", "techniques": ["commitment monitoring", "local correction"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "75.9% mean success on RoboTwin 2.0 (+22.7 over pi0.5)", "metric_measures": "other", "benchmarks": ["RoboTwin 2.0"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Monitors semantic stage commitments against state evidence, holds back dependent actions when unmet, and applies minimal local corrections to a frozen VLA.", "relevance": 4, "relevance_reason": "Commitment-level gating of chunk execution.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=0]"}
{"id": "2606.31497", "title": "Communication-Aware Robot Execution for Cloud Inference under Spatially Heterogeneous Connectivity", "authors": ["Fengkai Liu", "Yuichi Ohsita", "Masayuki Murata", "Hideyuki Shimonishi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.31497", "pdf_url": "https://arxiv.org/pdf/2606.31497", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["cloud inference", "wireless connectivity", "communication map", "request-response window"], "model_class": "other", "techniques": ["communication-aware planning", "cloud offloading"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["indoor wireless simulation from measurements"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Treats where the robot submits its next cloud request as a motion decision, planning toward spots with good wireless coverage so the next cloud result arrives before the current primitive runs out.", "relevance": 4, "relevance_reason": "Network-aware execution for cloud-hosted foundation-model robots.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.25371", "title": "Conformal Recovery-Deadline Certificates for Runtime Assurance of Adapting Controllers", "authors": ["Alireza Shojaei"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.25371", "pdf_url": "https://arxiv.org/pdf/2606.25371", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["simplex", "recovery deadline"], "model_class": "other", "techniques": ["runtime assurance", "split conformal prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces latching runtime-assurance trips with a split-conformal upper bound on an adapting controller's recovery time, delaying fallback within a certified deadline and backed by a verified hard-limit monitor.", "relevance": 4, "relevance_reason": "Gives a principled rule for when to interrupt versus let a learned controller recover.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2603.09298", "title": "CORAL: Scalable Multi-Task Robot Learning via LoRA Experts", "authors": ["Yuankai Luo", "Woping Chen", "Tong Liang", "Zhenguo Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.09298", "pdf_url": "https://arxiv.org/pdf/2603.09298", "code_url": "", "clusters": ["C3"], "keywords": ["LoRA experts", "multi-task", "storage"], "model_class": "VLA", "techniques": ["LoRA", "expert routing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Freezes one VLA backbone and attaches one LoRA expert per task, avoiding per-task full checkpoints and interference.", "relevance": 4, "relevance_reason": "Storage/deployment-efficient multi-task adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2604.20246", "title": "Cortex 2.0: Grounding World Models in Real-World Industrial Deployment", "authors": ["Adriana Aida", "Walid Amer", "Katarina Bankovic", "Dhruv Behl", "Fabian Busch", "Annie Bhalla", "Minh Duong", "Florian Gienger", "Rohan Godse", "Denis Grachev", "Ralf Gulde", "Elisa Hagensieker", "Junpeng Hu", "Shivam Joshi", "Tobias Knobloch", "Likith Kumar", "Damien LaRocque", "Keerthana Lokesh", "Omar Moured", "Khiem Nguyen", "Christian Preyss", "Ranjith Sriganesan", "Vikram Singh", "Carsten Sponner", "Anh Tong", "Dominik Tuscher", "Marc Tuscher", "Pavan Upputuri"], "year": 2026, "venue": "arXiv (Sereact tech report)", "type": "tech-report", "url": "https://arxiv.org/abs/2604.20246", "pdf_url": "https://arxiv.org/pdf/2604.20246", "code_url": "", "clusters": ["C8"], "keywords": ["plan-and-act", "latent trajectory scoring", "industrial deployment"], "model_class": "world-action model", "techniques": ["candidate future generation in visual latent space", "success/efficiency scoring", "commit best candidate"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot (industrial pick/sort/unpack)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Industrial system that rolls out candidate futures in a visual latent space, scores them for success and efficiency, and executes the best one, outperforming reactive VLAs in cluttered, contact-rich settings.", "relevance": 4, "relevance_reason": "Latent-space world-model planning in real industrial deployment.", "foundational_technique": false, "verified": true, "found_via": "snowball from NVIDIA WAM blog"}
{"id": "2601.16163", "title": "Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning", "authors": ["Moo Jin Kim", "Yihuai Gao", "Tsung-Yi Lin", "Yen-Chen Lin", "Yunhao Ge", "Grace Lam", "Percy Liang", "Shuran Song", "Ming-Yu Liu", "Chelsea Finn", "Jinwei Gu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.16163", "pdf_url": "https://arxiv.org/pdf/2601.16163", "code_url": "https://nvidia-cosmos.github.io/cosmos-cookbook/", "clusters": ["C3", "C4", "C8"], "keywords": ["Cosmos Policy", "video model fine-tuning", "single-stage post-training", "latent frame injection", "value prediction", "test-time planning"], "model_class": "world-action model", "techniques": ["single-stage post-training", "single-stage post-training of video model", "actions/values as latent frames", "best-of-N planning with predicted value"], "base_model": "Cosmos-Predict2", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboCasa", "real bimanual"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Post-trains a pretrained video diffusion model to output action chunks, future states and values as extra latent frames, making one model a policy, world model and value function usable for test-time planning.", "relevance": 4, "relevance_reason": "Unifies world model and policy; planning variant highlights compute cost of imagination.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations); snowball from NVIDIA WAM blog"}
{"id": "2607.25487", "title": "CoTinyVLA: Chain-of-Thought Distillation for a Sub-Billion-Parameter Vision-Language-Action Model", "authors": ["Minhyeok Lee", "Chiyoung Kim", "Chanhoe Gu", "Seongrok Kim", "Sanghyuk Roy Choi", "Donghwan Hwang", "Donghun Ryu", "Seokhyun Kim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.25487", "pdf_url": "https://arxiv.org/pdf/2607.25487", "code_url": "", "clusters": ["C2"], "keywords": ["chain-of-thought distillation", "sub-billion VLA", "LIBERO-Plus robustness"], "model_class": "VLA", "techniques": ["CoT distillation"], "base_model": "Qwen3.5-0.8B (35B teacher)", "params": "0.9B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills hierarchical plan/think reasoning from a 35B teacher into a 0.9B VLA, beating 7B baselines on perturbation robustness.", "relevance": 4, "relevance_reason": "Shows compact distilled VLAs can retain robustness.", "foundational_technique": false, "verified": true, "found_via": "web search: TinyVLA compact VLA distillation"}
{"id": "2601.14140", "title": "CREATE: Cross-Layer Resilience Characterization and Optimization for Efficient yet Reliable Embodied AI Systems", "authors": ["Tong Xie", "Yijiahao Qi", "Jinqi Wen", "Zishen Wan", "Yanchi Dong", "Zihao Wang", "Shaofei Cai", "Yitao Liang", "Tian-Yu Jia", "Yuan Wang", "Run-Sheng Wang", "Meng Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.14140", "pdf_url": "https://arxiv.org/pdf/2601.14140", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["undervolting", "bit errors", "resilience", "energy-reliability", "LLM planner + RL controller"], "model_class": "other", "techniques": ["error injection study", "circuit-level anomaly clearance", "weight rotation for fault tolerance", "voltage scaling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Characterizes fault tolerance of LLM-planner/RL-controller embodied systems under low-voltage bit errors and co-optimizes energy and reliability across circuit, model and application layers.", "relevance": 4, "relevance_reason": "Energy-reliability hardware co-design for battery-powered embodied agents.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2407.04292 (citations)"}
{"id": "2609.28984", "title": "CrossSafe: Towards Cross-Embodiment Latent Safety Filters", "authors": ["Ihab Tabbara", "Yuxuan Yang", "Hussein Sibai"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28984", "pdf_url": "https://arxiv.org/pdf/2609.28984", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["latent safety filter", "Hamilton-Jacobi reachability", "cross-embodiment", "cross-embodiment safety filter", "latent HJ reachability", "latent"], "model_class": "world model", "techniques": ["latent-space HJ reachability", "embodiment-conditioned filter", "shared latent safety reasoning with embodiment-specific realization", "latent safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Learns an embodiment-conditioned HJ reachability value function and safety policy in latent space shared across robots, generalizing whole-body collision avoidance zero-shot to held-out embodiments.", "relevance": 4, "relevance_reason": "Scaling latent safety filters across embodiments.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" \"vision-language-action\"; snowball from 2502.00935 (citations); venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2605.21854", "title": "CrossVLA: Cross-Paradigm Post-Training and Inference Optimization for Vision-Language-Action Models", "authors": ["Zhi Liu"], "year": 2026, "venue": "workshop draft", "type": "preprint", "url": "https://arxiv.org/abs/2605.21854", "pdf_url": "https://arxiv.org/pdf/2605.21854", "code_url": "https://github.com/lz-googlefycy/vla", "clusters": ["C1", "C3", "C4", "C5", "C9"], "keywords": ["latency anatomy", "DPO", "prefix caching", "LoRA vs DoRA", "flow-matching DPO", "post-training", "prefix KV caching", "denoise loop", "KV cache ceiling"], "model_class": "VLA", "techniques": ["profiling", "KV caching", "DPO", "LoRA", "DoRA", "inference profiling"], "base_model": "OpenVLA, pi0.5", "params": "", "hardware": "", "reported_metric": "denoising loop is 78.6% of sample_actions latency; prefix KV caching capped at 21% speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Empirical study that, among other things, profiles flow-VLA inference and finds the denoise loop dominates latency while prefix caching gives limited gains and can hurt success.", "relevance": 4, "relevance_reason": "Useful profiling data point; main contribution is post-training.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; snowball from 2502.19645 (S2 citations); search query: VLA benchmark latency; arXiv: vision-language-action AND (accelerator OR hardware OR GPU OR energy)"}
{"id": "2605.13276", "title": "D-VLA: A High-Concurrency Distributed Asynchronous Reinforcement Learning Framework for Vision-Language-Action Models", "authors": ["Yucheng Guo", "Yongjian Guo", "Zhong Guan", "Wen Huang", "Haoran Sun", "Haodong Yue", "Xiaolong Xiang", "Shuai Di", "Zhen Sun", "Luqiao Wang", "Junwu Xiong", "Yicheng Gong", "Yu-Cheng Guo", "Yong-Jian Guo"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.13276", "pdf_url": "https://arxiv.org/pdf/2605.13276", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["D-VLA", "RL training throughput"], "model_class": "VLA", "techniques": ["distributed RL", "asynchronous pipeline", "VRAM management"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distributed asynchronous RL framework for billion-parameter VLAs that decouples training data and weight planes and overlaps sampling, inference, gradients and parameter sync.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]; snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2609.33177", "title": "DeltaWAM: Change-Centric Visual Foresight via Delta Tokens for an Efficient World-Action Model", "authors": ["Tianyun Jiang", "Wenrui Bao", "Bingxin Xu", "Yu Tian", "Yuzhang Shang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33177", "pdf_url": "https://arxiv.org/pdf/2609.33177", "code_url": "", "clusters": ["C8"], "keywords": ["delta tokens", "change-centric foresight", "DINO features"], "model_class": "world-action model", "techniques": ["one delta vector per future frame", "flow-matching action expert on predicted transitions"], "base_model": "", "params": "0.725B", "hardware": "2x H100 (training, 256 GPU-hours)", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Predicts a single compact delta token per future frame encoding change in DINO features instead of reconstructing whole scenes, giving a sub-1B WAM trained on just 256 GPU-hours.", "relevance": 4, "relevance_reason": "Very compact future representation for efficient WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2609.16503", "title": "Dense to MoE Adaptation for Compact Vision Language Action Policies", "authors": ["Muchun Niu", "Shuang Chen", "Yuzhou Wu", "Xiaobing Tu", "Yinggui Wang", "Jinkui Ren", "Xiantao Zhang", "Linfeng Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.16503", "pdf_url": "https://arxiv.org/pdf/2609.16503", "code_url": "", "clusters": ["C1", "C2", "C3", "C9"], "keywords": ["dense-to-MoE", "expert deactivation", "compact VLA", "AdaDE", "Dense2MoE", "expert masking during fine-tuning"], "model_class": "VLA", "techniques": ["MoE conversion", "structured pruning", "fine-tuning-time pruning", "Dense2MoE conversion", "router-statistics expert masking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "40% LLM params deactivated, 95.7% LIBERO success | 40% LLM params deactivated retaining 95.7% LIBERO success", "metric_measures": "relative memory", "benchmarks": ["LIBERO", "RoboTwin 2.0"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "AdaDE converts dense FFNs to MoE during fine-tuning and prunes unused experts, deactivating 40% of LLM parameters with small success loss.", "relevance": 4, "relevance_reason": "Compression integrated into fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA mixture-of-experts; arXiv API: abs:\"vision-language-action\" AND abs:adapter AND abs:\"parameter\"; venue sweep: arXiv API VLA AND (MoE OR early exit)"}
{"id": "2603.10469", "title": "DepthCache: Depth-Guided Training-Free Visual Token Merging for Vision-Language-Action Model Inference", "authors": ["Yuquan Li", "Lianjie Ma", "Han Ding", "Lijun Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.10469", "pdf_url": "https://arxiv.org/pdf/2603.10469", "code_url": "", "clusters": ["C1"], "keywords": ["depth-guided token merging", "training-free"], "model_class": "VLA", "techniques": ["token merging"], "base_model": "pi0.5, OpenVLA, GR00T", "params": "", "hardware": "", "reported_metric": "up to 1.28x speedup, <1% success drop", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses depth to merge distant background tokens aggressively while preserving near-field workspace tokens, spreading merging over frames.", "relevance": 4, "relevance_reason": "Token merging for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: token merging robot"}
{"id": "2603.23149", "title": "Describe-Then-Act: Proactive Agent Steering via Distilled Language-Action World Models", "authors": ["Massimiliano Pappa", "Luca Romani", "Valentino Sacco", "Alessio Palma", "Stéphane Lathuilière", "Fabio Galasso", "Xavier Alameda-Pineda", "I. Spinelli"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.23149", "pdf_url": "https://arxiv.org/pdf/2603.23149", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["failure prevention", "distillation", "fast steering"], "model_class": "other", "techniques": ["cross-modal distillation", "language world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "14x speedup over baselines", "metric_measures": "relative latency", "benchmarks": ["MetaWorld", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DILLO distills a VLM teacher into a latent-conditioned LLM that predicts semantic outcomes of planned actions as text, enabling fast proactive policy steering without visual simulation.", "relevance": 4, "relevance_reason": "direct: low-latency safety steering replacing visual world models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2301.04104"}
{"id": "2604.16484", "title": "DexWorldModel: Causal Latent World Modeling towards Automated Learning of Embodied Tasks", "authors": ["Yueci Deng", "Guiliang Liu", "Kui Jia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.16484", "pdf_url": "https://arxiv.org/pdf/2604.16484", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["causal latent world model", "DINOv3 targets", "O(T) memory"], "model_class": "world-action model", "techniques": ["latent generative targets", "causal inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "CLWM predicts DINOv3 features instead of pixels to avoid redundant reconstruction, memory growth and sequential latency in WAMs.", "relevance": 4, "relevance_reason": "Efficient latent WAM.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2606.12402", "title": "DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners?", "authors": ["Jadelynn Dao", "Milan Ganai", "Yasmina Abukhadra", "Ajay Sridhar", "Mozhgan Nasr Azadani", "Katie Luo", "Clark Barrett", "Jiajun Wu", "Chelsea Finn", "Marco Pavone"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.12402", "pdf_url": "https://arxiv.org/pdf/2606.12402", "code_url": "", "clusters": ["C1", "C5", "C9"], "keywords": ["test-time compute", "routing", "VLM planner", "latency-success Pareto"], "model_class": "other", "techniques": ["compute routing", "adaptive model selection"], "base_model": "", "params": "", "hardware": "Franka arm (real); simulated benchmarks", "reported_metric": "matches or exceeds stronger model's success at up to 65% lower average latency", "metric_measures": "relative latency", "benchmarks": ["VLABench", "RoboMME", "real robot (Franka)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Routes each planning query to an appropriate amount of test-time compute (CoT depth, model size, memory) using scene context, improving the success–cost Pareto frontier of VLM planners.", "relevance": 4, "relevance_reason": "Adaptive compute allocation for embodied planners with measured latency savings; co-authored by speaker Jiajun Wu.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "2604.25050", "title": "DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors", "authors": ["Pengcheng Wang", "Kaiwen Hong", "Chensheng Peng", "K. Driggs-Campbell", "Masayoshi Tomizuka", "Chenfeng Xu", "Chen Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.25050", "pdf_url": "https://arxiv.org/pdf/2604.25050", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["DiscreteRTC", "discrete diffusion policy", "native inpainting"], "model_class": "diffusion/flow policy", "techniques": ["discrete diffusion unmasking as inpainting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Argues discrete (masked) diffusion policies natively support real-time chunking since committed actions are just unmasked tokens, avoiding RTC guidance overhead.", "relevance": 4, "relevance_reason": "Lower-overhead async execution.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: real-time chunking"}
{"id": "2602.00440", "title": "DISK: Dynamic Inference SKipping for World Models", "authors": ["Anugunj Naman", "Gaibo Zhang", "Ayushman Singh", "Yaguang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.00440", "pdf_url": "https://arxiv.org/pdf/2602.00440", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["dynamic inference skipping", "training-free", "driving world model"], "model_class": "world model", "techniques": ["dual-branch skip controllers", "latent-difference skip tests in AR rollouts"], "base_model": "", "params": "", "hardware": "NVIDIA L40S", "reported_metric": "2x speedup on trajectory diffusion, 1.6x on video diffusion at equal planning error", "metric_measures": "relative latency", "benchmarks": ["nuPlan", "nuScenes", "NAVSIM"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Training-free adaptive step skipping for coupled video/trajectory diffusion world models that keeps closed-loop driving metrics while speeding rollouts 1.6-2x.", "relevance": 4, "relevance_reason": "Adaptive inference for world-model rollouts.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2607.00666", "title": "Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts", "authors": ["Taewook Kang", "Taeheon Kim", "Donghyun Shin", "Jonghyun Choi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.00666", "pdf_url": "https://arxiv.org/pdf/2607.00666", "code_url": "", "clusters": ["C3"], "keywords": ["weight arithmetic", "one-shot adaptation", "domain shift"], "model_class": "VLA", "techniques": ["task/domain vector arithmetic", "SVD subspace alignment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Adapts VLAs to camera/robot shifts via weight-vector arithmetic needing just one demonstration.", "relevance": 4, "relevance_reason": "One-shot, near training-free adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.18589", "title": "DREAM-Chunk: Reactive Action Chunking with Latent World Model", "authors": ["Wenxi Chen", "Kai-Di Zhang", "Chi Lin", "Zhi-Yuan Zhang", "Yuanya She", "Yue-Jiang Liu", "Raymond A. Yeh", "Shao-Shuai Mou", "Yan Gu", "Kaidi Zhang", "Zhiyuan Zhang", "Yu She", "Yuejiang Liu", "Shaoshuai Mou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18589", "pdf_url": "https://arxiv.org/pdf/2606.18589", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["DREAM-Chunk", "latent world model", "test-time selection", "reactive chunking", "test-time scaling"], "model_class": "world model", "techniques": ["candidate chunk sampling + latent rollout matching", "lightweight latent world model", "candidate chunk rollout and matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Samples multiple candidate chunks, predicts latent futures with a small world model and switches to the chunk whose prediction best matches the observed rollout, adding reactivity without fine-tuning.", "relevance": 4, "relevance_reason": "Cheap latent imagination for closed-loop robustness.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2); search query"}
{"id": "2606.08737", "title": "Dream-Tac: A Unified Tactile World Action Model for Contact-Rich Robot Manipulation", "authors": ["Yunfan Lou", "Yifan Ye", "Yankai Fu", "Jun Cen", "Xiaowei Chi", "Yaoxu Lyu", "Peidong Jia", "Sirui Han", "Zhihe Lu", "Shanghang Zhang", "Yi-Fan Ye", "Yan-Kai Fu", "Pei-Dong Jia", "Si-Rui Han"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.08737", "pdf_url": "https://arxiv.org/pdf/2606.08737", "code_url": "https://github.com/LYFCLOUDFAN/Dream-Tac", "clusters": ["C1", "C8"], "keywords": ["Dream-Tac", "tactile world-action model", "real-time"], "model_class": "world-action model", "techniques": ["diffusion caching", "visuotactile fusion", "cache-based diffusion acceleration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.8x faster inference, 2.9x faster training", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Visuotactile world-action model with contact-gated fusion plus training and cache-based inference acceleration for real-time deployment.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\"); snowball round 2: cites 2205.14135, 2301.04104"}
{"id": "2603.24587", "title": "DreamerAD: Efficient Reinforcement Learning via Latent World Model for Autonomous Driving", "authors": ["Pengxuan Yang", "Yupeng Zheng", "Deheng Qian", "Zebin Xing", "Qichao Zhang", "Linbo Wang", "Yichen Zhang", "Shaoyu Guo", "Zhongpu Xia", "Qiang Chen", "Junyu Han", "Lingyun Xu", "Yifeng Pan", "Dongbin Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.24587", "pdf_url": "https://arxiv.org/pdf/2603.24587", "code_url": "", "clusters": ["C8"], "keywords": ["latent world model RL", "shortcut forcing", "1-step diffusion"], "model_class": "world model", "techniques": ["shortcut forcing (100 -> 1 step)", "latent reward model", "GRPO in latent imagination"], "base_model": "", "params": "", "hardware": "", "reported_metric": "diffusion sampling 100 -> 1 step, 80x speedup (from ~2 s/frame)", "metric_measures": "relative latency", "benchmarks": ["NavSim v2"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Makes RL in a driving video world model practical by compressing sampling from 100 steps to 1 (80x) with shortcut forcing and doing credit assignment with a latent reward model.", "relevance": 4, "relevance_reason": "Step compression of world models for fast imagination RL.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.29749", "title": "DriftingVLA: Native One-Step Vision-Language-Action Generation via Per-Dimension Temporal Drifting", "authors": ["Yuxuan Gao", "Shiqi Zhang", "Yedong Shen", "Yifan Duan", "Wenhao Yu", "Xin Zhang", "Siyuan Cao", "Jiajun Deng", "Yanyong Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29749", "pdf_url": "https://arxiv.org/pdf/2608.29749", "code_url": "", "clusters": ["C1"], "keywords": ["one-step VLA", "distribution drifting", "per-dimension"], "model_class": "VLA", "techniques": ["one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboTwin 2.0"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Native one-step VLA trained with a distribution-drifting objective applied per action dimension over time.", "relevance": 4, "relevance_reason": "One-step VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk latency VLA"}
{"id": "2609.33148", "title": "DroneWAM: Efficient World Action Model for Drone Visual Navigation", "authors": ["Liang Yao", "Fan Liu", "Hongbo Lu", "Wei Xu", "Jianyu Jiang", "Yijun Shen", "Chuanyi Zhang", "Pai Peng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33148", "pdf_url": "https://arxiv.org/pdf/2609.33148", "code_url": "", "clusters": ["C8"], "keywords": ["drone navigation", "JEPA WAM", "adaptive rollout depth"], "model_class": "world-action model", "techniques": ["representation-space prediction", "token resampler compression", "gate-controlled adaptive rollout depth"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DroneNav-6D (sim)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "JEPA-style WAM for drones that compresses features into few tokens and adaptively chooses imagination depth per scene to meet onboard efficiency limits.", "relevance": 4, "relevance_reason": "Compute-adaptive latent WAM for a constrained platform.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2609.24868", "title": "DualWAM: Dual-System World Action Models for Asynchronous Global Planning and Local Refinement", "authors": ["Yixin Zheng", "Jiangran Lyu", "Yuntian Deng", "Kai Liu", "Yizhou Zhou", "Yizhou Wang", "Xiaoguang Zhao", "He Wang", "Zhizheng Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24868", "pdf_url": "https://arxiv.org/pdf/2609.24868", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["dual-system WAM", "global plan + local refinement", "wrist-camera refinement"], "model_class": "world-action model", "techniques": ["periodic high-noise global denoising", "fast low-noise local refinement from intermediate state", "asynchronous systems"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Runs a slow global world-action denoiser periodically and a fast wrist-camera refiner that completes low-noise denoising from its intermediate state, giving high-frequency closed-loop updates.", "relevance": 4, "relevance_reason": "Dual-rate WAM for responsiveness.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2606.32028", "title": "DVG-WM: Disentangled Video Generation Enables Efficient Embodied World Model for Robotic Manipulation", "authors": ["Ziyu Shan", "Zhenyu Wu", "Xiaofeng Wang", "Zheng Zhu", "Ziwei Wang", "Zhen-Yu Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.32028", "pdf_url": "https://arxiv.org/pdf/2606.32028", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["disentangled video generation", "dynamics then synthesis", "cascade", "efficient world model"], "model_class": "world model", "techniques": ["coarse dynamics preview then refinement", "flow matching from dynamics to video latents", "latent degradation for details", "disentangled dynamics vs rendering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 3.97x acceleration", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Separates low-level dynamics prediction from high-resolution rendering in a robot video world model, giving up to ~4x faster generation with better contact detail.", "relevance": 4, "relevance_reason": "Efficient embodied video world model.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "dyna-2.1-physical-agent-blog-2026", "title": "Dyna-2.1: A Physical Agent for End-to-End Workflows", "authors": ["Dyna Robotics"], "year": 2026, "venue": "blog (Dyna Robotics)", "type": "blog", "url": "https://www.dyna.co/dyna-2.1", "pdf_url": "", "code_url": "", "clusters": ["C1", "C6", "C9"], "keywords": ["hierarchical control", "VLM orchestrator", "whole-body controller", "unified robot representation"], "model_class": "world-action model", "techniques": ["multi-rate hierarchy", "RL whole-body controller"], "base_model": "", "params": "", "hardware": "Taku semi-humanoid (21-DoF upper body, wheeled base)", "reported_metric": "DYNA-2 policy ~5 Hz; whole-body controller 100 Hz", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Three-rate stack: VLM workflow orchestrator, ~5 Hz DYNA-2 policy producing whole-body task-space targets, and a 100 Hz RL whole-body controller; task-space representation lets human video train the policy.", "relevance": 4, "relevance_reason": "Concrete multi-rate deployment architecture with control rates from a speaker's company.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Dyna"}
{"id": "2605.11567", "title": "Dynamic Execution Commitment of Vision-Language-Action Models", "authors": ["Feng Chen", "Xianghui Wang", "Yuxuan Chen", "Boying Li", "Yefei He", "Zeyu Zhang", "Yicheng Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.11567", "pdf_url": "https://arxiv.org/pdf/2605.11567", "code_url": "https://inceptionwang.github.io/A3/", "clusters": ["C2", "C7"], "keywords": ["A3", "adaptive action acceptance", "execution horizon", "self-speculative verification"], "model_class": "VLA", "techniques": ["group sampling consensus", "prefix verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Decides how many chunk actions to commit by self-speculative verification: accepts only the longest prefix whose actions remain consistent when re-decoded, adapting commitment to state-dependent reliability.", "relevance": 4, "relevance_reason": "State-dependent open-loop commitment trading robustness vs throughput.", "foundational_technique": false, "verified": true, "found_via": "arXiv API all:\"action chunking\" AND latency"}
{"id": "2605.24924", "title": "Dynamic Neural Koopman Distillation for Real-Time Robot Control Using Diffusion Models", "authors": ["Lei Zheng", "Peiqi Yu", "Zengqi Peng", "Changliu Liu", "Armin Lederer"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24924", "pdf_url": "https://arxiv.org/pdf/2605.24924", "code_url": "", "clusters": ["C1", "C2", "C3", "C9"], "keywords": ["Koopman", "one-step distillation", "Kinova", "diffusion policy", "Koopman layer"], "model_class": "diffusion/flow policy", "techniques": ["diffusion distillation", "distillation to single step", "Koopman operator"], "base_model": "", "params": "", "hardware": "Kinova manipulator (real)", "reported_metric": "millisecond-regime inference latency (single forward pass)", "metric_measures": "other", "benchmarks": ["D4RL", "real robot", "D4RL MuJoCo", "real robot (Kinova)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills multi-step diffusion policies into a single pass using a factorized dynamic Koopman layer with state-dependent modal gains, keeping multimodality.", "relevance": 4, "relevance_reason": "Diffusion step reduction for real-time control by organizer Changliu Liu.", "foundational_technique": false, "verified": true, "found_via": "search: one-step policy distillation robot; author sweep: Changliu Liu"}
{"id": "2608.01452", "title": "DynamicManip: Enabling Dynamic Manipulation from a Single Static Demonstration", "authors": ["Hao-Ran Liao", "Pengyu Wang", "Shuoyu Chen", "Kehan Cheng", "Xuhang Chen", "Yuhao Lin", "Mu Lin", "Zhizhao Liang", "Xiaoyi Fan", "Chengyi Xing", "Dan Niu", "Yi-Lin Wei", "Wei-Shi Zheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.01452", "pdf_url": "https://arxiv.org/pdf/2608.01452", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["dynamic manipulation", "low latency"], "model_class": "other", "techniques": ["adaptive inference frequency", "data augmentation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "32.9% lower policy-query latency", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Synthesizes dynamic-manipulation demos from one static demo and uses a policy that adapts its inference frequency to task dynamics.", "relevance": 4, "relevance_reason": "Latency-aware control for dynamic manipulation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.00793", "title": "DynamicWAM: Dual-Path Motion Conditioning for World-Action Models in Dynamic Manipulation", "authors": ["Yunfan Lou", "Hewen Gao", "Xiyu Zhu", "Zhuoran Qiao", "Xuan Han", "Yifan Yang", "Yifan Ye", "Boxian Yao", "Zhibo Pang", "He-Wen Gao", "Yi-Fan Yang", "Yi-Fan Ye", "Bo-Xian Yao", "Zhi-Bo Pang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00793", "pdf_url": "https://arxiv.org/pdf/2608.00793", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["DynamicWAM", "dynamic manipulation"], "model_class": "world-action model", "techniques": ["distillation", "real-time chunking", "asynchronous execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DOMINO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Compact world-action model for moving objects with optical-flow and kinematic conditioning, using a distilled backbone and RTC-style asynchronous execution for responsiveness.", "relevance": 4, "relevance_reason": "Distilled WAM with asynchronous real-time execution.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2410.24164, 2502.19645"}
{"id": "2603.16188", "title": "ECHO: Edge-Cloud Humanoid Orchestration for Language-to-Motion Control", "authors": ["Haozhe Jia", "Jianfei Song", "Yuan Zhang", "Honglei Jin", "Youcheng Fan", "Wenshuo Chen", "Wei Zhang", "Yutao Yue"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.16188", "pdf_url": "https://arxiv.org/pdf/2603.16188", "code_url": "", "clusters": ["C6", "C9"], "keywords": ["humanoid", "cloud text-to-motion", "edge RL tracker", "edge-cloud humanoid", "text-to-motion", "RL tracker"], "model_class": "diffusion/flow policy", "techniques": ["cloud generation + onboard tracking", "compact motion representation", "cloud diffusion generator + edge RL tracker"], "base_model": "", "params": "", "hardware": "cloud GPU + onboard humanoid compute", "reported_metric": "cloud motion generation ~1 s (10 DDIM steps)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Cloud diffusion model turns language into a compact 38-D robot-native motion reference that an onboard RL tracker executes closed-loop on a humanoid.", "relevance": 4, "relevance_reason": "Clear hierarchical cloud-generation/onboard-control split for humanoids.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:edge-cloud AND abs:robot; venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2602.15836", "title": "EdgeNav-QE: QLoRA Quantization and Dynamic Early Exit for LAM-based Navigation on Edge Devices", "authors": ["Mengyun Liu", "Shanshan Huang", "Jian Jiang"], "year": 2026, "venue": "ACM TOMM (Trans. Multimedia Computing, Communications & Applications)", "type": "paper", "url": "https://arxiv.org/abs/2602.15836", "pdf_url": "https://arxiv.org/pdf/2602.15836", "code_url": "", "clusters": ["C2", "C3"], "keywords": ["large action model", "navigation", "4-bit", "early exit", "edge"], "model_class": "other", "techniques": ["4-bit quantization", "QLoRA", "dynamic early exit"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Quantizes a large action model backbone to 4-bit with QLoRA adaptation and adds early-exit branches so simple navigation steps terminate early on edge devices.", "relevance": 4, "relevance_reason": "Combined quantization+early exit for embodied navigation on edge.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246"}
{"id": "2609.35570", "title": "EdgeVLN: Runtime-Aware Deployment Ready Quantized Vision Language Navigation Model", "authors": ["Rithvik Jonna", "Man Namgung", "Aakash Gurram", "Tinoosh Mohsenin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.35570", "pdf_url": "https://arxiv.org/pdf/2609.35570", "code_url": "", "clusters": ["C1", "C2", "C4", "C5", "C9"], "keywords": ["VLN", "llama.cpp", "IQ4", "stop head", "EdgeVLN", "quantization", "Jetson Orin NX", "energy", "resident memory", "quantized VLN", "edge deployment", "runtime-aware"], "model_class": "other", "techniques": ["weight quantization (2-8 bit)", "weight quantization", "memory token pruning", "weight quantization 8-2 bit", "llama.cpp driver", "lightweight stop head", "quantization", "latent trajectory termination"], "base_model": "StreamVLN", "params": "", "hardware": "NVIDIA Jetson Orin NX 16GB", "reported_metric": "58.02% SR at 4-bit, +0.013 s per step | 58.02% SR on 4-bit model with +0.013 s latency per step", "metric_measures": "other", "benchmarks": ["R2R VLN-CE"], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Characterizes a navigation VLM across 8-to-2-bit quantization and runtimes on Jetson Orin NX, measuring latency, energy and resident memory, and adds a cheap stop head reusing backbone states.", "relevance": 4, "relevance_reason": "Runtime-aware edge deployment on Orin NX with energy/memory measurements.", "foundational_technique": false, "verified": true, "found_via": "search: quantization vision-language-action; search query: vision-language-action compression combined quantization pruning; snowball from 2603.02271 (citations); venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2608.14586", "title": "Efficient Block-Layer Parallel Inference for Vision-Language-Action on Hybrid Architectures", "authors": ["Haibo HU", "Lianming Huang", "Qiao Li", "Nan Guan", "Chun Jason Xue"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.14586", "pdf_url": "https://arxiv.org/pdf/2608.14586", "code_url": "", "clusters": ["C5"], "keywords": ["autonomous driving VLA", "CPU-GPU hybrid", "layer partitioning", "GPU memory"], "model_class": "VLA", "techniques": ["heterogeneous CPU-GPU partitioning", "cross-frame asynchronous pipeline"], "base_model": "Orion, MindDrive", "params": "", "hardware": "vehicle onboard CPU+GPU", "reported_metric": "Orion 521 ms -> 408 ms; MindDrive 443 ms -> 306.2 ms; Orion peak GPU memory 45 GB -> 29 GB", "metric_measures": "relative latency", "benchmarks": ["Bench2Drive", "real vehicle"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Offloads the tail of a driving VLA's LLM to underused vehicle CPUs through an asynchronous cross-frame pipeline, freeing GPU memory so the VLA fits beside the full Autoware stack.", "relevance": 4, "relevance_reason": "Heterogeneous-processor scheduling for onboard VLA under memory limits.", "foundational_technique": false, "verified": true, "found_via": "arXiv: vision-language-action AND Jetson/edge/onboard"}
{"id": "2602.03983", "title": "Efficient Long-Horizon Vision-Language-Action Models via Static-Dynamic Disentanglement", "authors": ["Weikang Qiu", "Huashuo Lei", "Tinglin Huang", "Rex Ying"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.03983", "pdf_url": "https://arxiv.org/pdf/2602.03983", "code_url": "", "clusters": ["C1"], "keywords": ["static-dynamic tokens", "KV cache reuse", "multi-frame"], "model_class": "VLA", "techniques": ["KV cache reuse", "token disentanglement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.0x (sim) / 2.2x (real) inference speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SIMPLER", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Splits visual input into static and dynamic tokens, keeping one copy of static tokens and reusing their KV cache via a recache gate, enabling cheap multi-frame context.", "relevance": 4, "relevance_reason": "KV reuse for long-horizon VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient"}
{"id": "2606.14375", "title": "Elastic Queries Reinforcement Learning: Self-Aware Policy Execution for VLA Models", "authors": ["Ge Wang", "Xinyu Tan", "Xiang Li", "Man Luo", "Chengsi Yao", "Shenhao Yan", "Jiahao Yang", "Fan Feng", "Honghao Cai", "Xiangyuan Wang", "Zhixin Mai", "Yiming Zhao", "Yatong Han", "Zhen Li", "Gennan Wang", "Xin-Yu Tan", "Cheng Yao", "Shen-Hao Yan", "Xiang-Yuan Wang", "Yimin Zhao", "Ya-Tong Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.14375", "pdf_url": "https://arxiv.org/pdf/2606.14375", "code_url": "", "clusters": ["C1", "C3", "C7"], "keywords": ["elastic inference", "compute allocation", "denoising budget", "difficulty-aware"], "model_class": "VLA", "techniques": ["adaptive chunk length", "adaptive denoising steps", "RL scheduling", "adaptive compute", "variable chunk length"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "A lightweight adaptor chooses latent input, denoising budget and chunk length per VLA query, trained with RL on critic-ensemble disagreement, reducing amortized inference cost.", "relevance": 4, "relevance_reason": "Learns when to execute longer open-loop chunks versus re-query.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"open-loop\" AND \"vision-language-action\" AND (execution OR monitor*); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.31132", "title": "ELASTIC: Efficiently Learning to Adaptively Scale Test-Time Compute for Generative Control Policies", "authors": ["Andrew Zou Li", "Gokul Swamy", "Yonatan Bisk", "Andrea Bajcsy"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.31132", "pdf_url": "https://arxiv.org/pdf/2606.31132", "code_url": "", "clusters": ["C1"], "keywords": ["test-time compute", "adaptive denoising", "meta-policy"], "model_class": "VLA", "techniques": ["adaptive compute allocation"], "base_model": "pi0.5, Diffusion Policy", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns a meta-policy that chooses denoising steps and parallel samples per state for generative policies, matching best-of-N success at lower latency.", "relevance": 4, "relevance_reason": "Adaptive compute for generative policies.", "foundational_technique": false, "verified": true, "found_via": "search query: adaptive computation robot policy"}
{"id": "2605.29438", "title": "ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models", "authors": ["Ye Li", "Huanan Liu", "Kangye Ji", "Yuan Meng", "Jiajun Fan", "Yuansong Wang", "Shiyu Qin", "Chenglei Wu", "Shu-Tao Xia", "Zhi Wang", "Hua-Yang Liu", "Jia-Jun Fan", "Shi-Yu Qin", "Shuhui Xia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.29438", "pdf_url": "https://arxiv.org/pdf/2605.29438", "code_url": "", "clusters": ["C1"], "keywords": ["adaptive inference", "training-free acceleration", "GR00T", "CogACT"], "model_class": "VLA", "techniques": ["dynamic compute scheduling", "feature reuse", "denoising reuse"], "base_model": "GR00T, CogACT", "params": "", "hardware": "", "reported_metric": "up to 2.55x (GR00T) / 3.77x (CogACT) speedup; 13.8 Hz -> 26.3 Hz on real GR00T tasks | 13.8 Hz -> 26.3 Hz control frequency on real GR00T tasks; up to 2.55x (GR00T) / 3.77x (CogACT) speedup", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Plug-in scheduler that decides per step how much of the vision encoder, LLM and action head to recompute versus reuse, based on representation stability, motion cues and progress.", "relevance": 4, "relevance_reason": "Directly about VLA inference acceleration with measured control-rate gains.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]; snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2607.26657", "title": "Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control", "authors": ["Weili Zeng", "Yitong Xing", "Fulong Liu", "Chengqun Yang", "Antao Xiang", "Feng Tian", "Jingnan Gao", "Jisong Cai", "Xin Wang", "Xiaomin Wu", "Yao Mu", "Xiaokang Yang", "Yichao Yan", "Wei-Li Zeng", "Yi-Tong Xing", "Fu-Long Liu", "Cheng-Qun Yang", "An-Tao Xiang", "Jing-Nan Gao", "Ji-Song Cai", "Xiao-Min Wu", "Yi-Chao Yan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.26657", "pdf_url": "https://arxiv.org/pdf/2607.26657", "code_url": "", "clusters": ["C8"], "keywords": ["folding imagination", "predictive representations", "ultra-efficient control", "world generator", "efficient control"], "model_class": "world-action model", "techniques": ["supervise current-only encoder with generator intermediate states", "no generation at deployment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Internalizes a world generator's intermediate future-construction computation into a representation predicted from the present alone, so actions are produced without running the generator.", "relevance": 4, "relevance_reason": "Title in scope (efficient world-model-based control); unverified", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations); snowball round 2: cites 2306.03310, 2402.15391"}
{"id": "2602.00780", "title": "Environment-Aware Adaptive Pruning with Interleaved Inference Orchestration for Vision-Language-Action Models", "authors": ["Yuting Huang", "Leilei Ding", "Zhipeng Tang", "Zenghuan Zhu", "Jiajun Deng", "Xinrui Lin", "Shuo Liu", "Haojie Ren", "Jianmin Ji", "Yanyong Zhang", "Yu-Ting Huang", "Lei-Lei Ding", "Zhi-Peng Tang", "Zeng-Hua Zhu", "Jia-Jun Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.00780", "pdf_url": "https://arxiv.org/pdf/2602.00780", "code_url": "", "clusters": ["C1", "C2", "C9"], "keywords": ["adaptive channel pruning", "inference orchestration", "training-free", "environment-aware", "adaptive pruning", "dynamic sparsity", "layer pruning"], "model_class": "VLA", "techniques": ["structured pruning", "scheduling", "dynamic channel pruning", "environment-aware adaptive pruning", "interleaved inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.60x speedup (2.18x with token pruning)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "EcoVLA adapts channel sparsity to the changing environment using temporal consistency and interleaves pruned/unpruned inference, combinable with other accelerations.", "relevance": 4, "relevance_reason": "Training-free parameter sparsification for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; snowball from 2506.17639; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2605.25477", "title": "EXPO-FT: Sample-Efficient Reinforcement Learning Finetuning for Vision-Language-Action Models", "authors": ["Perry Dong", "Kuo-Han Hung", "Tian Gao", "Dorsa Sadigh", "Chelsea Finn"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.25477", "pdf_url": "https://arxiv.org/pdf/2605.25477", "code_url": "", "clusters": ["C3"], "keywords": ["sample-efficient RL", "real-world", "human corrections"], "model_class": "VLA", "techniques": ["online RL", "human-in-the-loop"], "base_model": "", "params": "", "hardware": "", "reported_metric": "30/30 success after avg 19.1 min online robot data", "metric_measures": "other", "benchmarks": ["real robot precision/dynamic tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "System for stable sample-efficient real-world RL fine-tuning of pretrained VLAs reaching perfect success in ~20 minutes of robot data.", "relevance": 4, "relevance_reason": "Highly sample-efficient post-training.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv"}
{"id": "2608.10232", "title": "FACT: Failure-Aware Causal Training for World-Action Models", "authors": ["Quanquan Peng", "Yutong Liang", "Rui Yan", "Nicklas Hansen", "Xiaolong Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.10232", "pdf_url": "https://arxiv.org/pdf/2608.10232", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["failure-aware training", "action-conditioned progress", "WAM"], "model_class": "world-action model", "techniques": ["action-conditioned causal WAM trained on failure rollouts", "progress prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains a causal WAM to predict video and task progress conditioned on executed actions, including failures, so it can anticipate consequences of bad actions.", "relevance": 4, "relevance_reason": "Makes WAM imagination usable for failure anticipation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2609.03611", "title": "FailBench: How Reliable are VLMs at Judging Robot Task Success?", "authors": ["Navasardyan, Zaruhi", "Danielyan, Tatul", "Davtyan, Hrant"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.03611", "pdf_url": "https://arxiv.org/pdf/2609.03611", "code_url": "", "clusters": ["C4"], "keywords": ["FailBench", "VLM judge reliability", "failure detection", "success detection"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["FailBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "2,197 manipulation attempts from 14 sources to test VLM success/failure judges; the best reaches only 0.77 balanced accuracy and judges are biased toward predicting success.", "relevance": 4, "relevance_reason": "Validates (and warns about) automated success detection used by AutoEval-style and world-model evaluators.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: OpenReview CoRL 2025 evaluation protocol VLA benchmark flaws"}
{"id": "2607.27511", "title": "Failure Detection for Surgical Robot Imitation Policies via Flow-Matching World Modeling", "authors": ["Zhefeng Huang", "Yi-Lin Cai", "Ankit B. Patel", "Mohammad Hajiha", "Brendan Browne", "Yue Chen", "Yilin Cai", "Ankit Patel"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.27511", "pdf_url": "https://arxiv.org/pdf/2607.27511", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["surgical robot", "failure detection", "flow-matching world model", "surgical failure detection", "conformal"], "model_class": "world model", "techniques": ["action-conditioned flow-matching dynamics", "inverse-transport nonconformity score", "conformal calibration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["surgical tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Detects surgical-robot policy failures without failure data by scoring how inconsistent observed latents are with a nominal action-conditioned flow-matching world model, calibrated conformally.", "relevance": 4, "relevance_reason": "Runtime failure detection in safety-critical domain.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2); search query"}
{"id": "2609.04277", "title": "FailureSpot: Label-Efficient Timestamp-Level Failure Detection for Vision-Language-Action Models", "authors": ["Jie Ma", "Zongxi Liu", "Yi Zhu", "Zong-Xi Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.04277", "pdf_url": "https://arxiv.org/pdf/2609.04277", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "action chunk", "runtime monitoring", "timestamp-level"], "model_class": "VLA", "techniques": ["failure detection", "probing internal representations", "weak supervision", "active learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains a lightweight proactive failure detector on VLA internal features with timestamp-level labels built cheaply from unlabeled action chunks, so failures are localized to the chunk where they begin rather than whole trajectories.", "relevance": 4, "relevance_reason": "Per-chunk failure localization is a direct trigger for interrupting open-loop chunk execution.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunk*\" AND (safety OR safe); snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2608.06434", "title": "Fast and Accurate: An Adaptive VLA Inference Framework through Environment-aware Model Selection", "authors": ["Yuewei Sun", "Lang Qin", "Zechuan Tian", "Jingwen Li", "Guiqin Wang", "Shengzeng Huo", "Wenxin Ren", "Tao Fang", "Xiaochen Zhang", "Guanqing Deng", "Xiang Wang", "Xiaowen Dong", "Qinghai Guo", "Yuxin Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.06434", "pdf_url": "https://arxiv.org/pdf/2608.06434", "code_url": "", "clusters": ["C1"], "keywords": ["model selection", "RL switching", "fast-slow"], "model_class": "VLA", "techniques": ["model switching", "dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "effective action frequency 93.4 Hz on LIBERO", "metric_measures": "closed-loop control", "benchmarks": ["LIBERO", "real dual-arm robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "An RL switching policy chooses between a decoupled large deliberative VLA and a small reactive one, invoking the big model sparsely.", "relevance": 4, "relevance_reason": "Adaptive dual-system inference.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "2602.01801", "title": "Fast Autoregressive Video Diffusion and World Models with Temporal Cache Compression and Sparse Attention", "authors": ["Dvir Samuel", "Issar Tzachor", "Matan Levy", "Michael Green", "Gal Chechik", "Rami Ben-Ari"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.01801", "pdf_url": "https://arxiv.org/pdf/2602.01801", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["FAST-AR", "KV cache compression", "sparse attention"], "model_class": "world model", "techniques": ["temporal KV cache compression", "ANN-based sparse cross/self attention", "training-free"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Training-free attention framework for autoregressive video diffusion/world models that bounds KV growth via temporal correspondence and sparsifies attention with approximate nearest neighbors.", "relevance": 4, "relevance_reason": "General acceleration for AR world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.29350", "title": "Fast Enough to Act: Spatio-Temporal Visual Token Merging for Low-Latency Robotic VLMs and VLAs", "authors": ["Junzhou Chen", "Jindong Wang", "Gang Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.29350", "pdf_url": "https://arxiv.org/pdf/2606.29350", "code_url": "", "clusters": ["C1"], "keywords": ["token merging", "spatio-temporal", "training-free"], "model_class": "VLA", "techniques": ["token merging"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "8.3x speedup at 1024x1024 input", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Merges redundant visual tokens across frames using 3D spatio-temporal coordinates during encoding, with positional correction for merged tokens.", "relevance": 4, "relevance_reason": "Token merging for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: token merging robot"}
{"id": "2606.26217", "title": "Fast LeWorldModel", "authors": ["Yuntian Gao", "Xiangyu Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.26217", "pdf_url": "https://arxiv.org/pdf/2606.26217", "code_url": "", "clusters": ["C8"], "keywords": ["LeWorldModel", "action-prefix prediction", "parallel latent rollout"], "model_class": "world model", "techniques": ["parallel prediction of latents for all action prefixes"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Replaces step-by-step latent rollout in a JEPA world model with parallel prediction of the latent reached after each action prefix, speeding planning and reducing error accumulation.", "relevance": 4, "relevance_reason": "Faster JEPA planning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2609.30833", "title": "Fast Plans, Faithful Actions: Closing the Planning-Execution Gap in Hierarchical Vision-Language-Action Models", "authors": ["Chuanliang Xie", "Boyu Ma", "Gen Li", "Yizhou Liu", "Houwang Chen", "Xinyu Zhou", "Jianfei Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.30833", "pdf_url": "https://arxiv.org/pdf/2609.30833", "code_url": "", "clusters": ["C1"], "keywords": ["hierarchical VLA", "block-autoregressive planning", "waypoints"], "model_class": "VLA", "techniques": ["block-wise decoding", "goal modulation"], "base_model": "pi0.5-derived", "params": "", "hardware": "", "reported_metric": "VLM forward passes 57 -> 8; 8.7x lower planning latency", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Finds hierarchical waypoint planners are slow and underused; uses waypoint-aligned block-AR decoding and goal modulation to fix both.", "relevance": 4, "relevance_reason": "Hierarchical VLA latency.", "foundational_technique": false, "verified": true, "found_via": "search query: autoregressive action expert VLA"}
{"id": "2601.09708", "title": "Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning", "authors": ["Chi-Pin Huang", "Yunze Man", "Zhiding Yu", "Min-Hung Chen", "Jan Kautz", "Yu-Chiang Frank Wang", "Fu-En Yang"], "year": 2026, "venue": "CVPR 2026", "type": "paper", "url": "https://arxiv.org/abs/2601.09708", "pdf_url": "https://arxiv.org/pdf/2601.09708", "code_url": "", "clusters": ["C1"], "keywords": ["latent chain-of-thought", "reasoning VLA", "distillation"], "model_class": "VLA", "techniques": ["latent reasoning", "distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 89.3% lower inference latency vs reasoning VLAs", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Compresses explicit chain-of-thought in reasoning VLAs into short verbalizable latent tokens distilled from a teacher, sharply cutting reasoning latency.", "relevance": 4, "relevance_reason": "Efficient reasoning VLA design.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient"}
{"id": "2608.04404", "title": "Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models", "authors": ["Weiheng Zhao", "Haoyi Jiang", "Xin Shi", "Liu Liu", "Fan Huang", "Zhizhong Su", "Wei Sui", "Xinggang Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.04404", "pdf_url": "https://arxiv.org/pdf/2608.04404", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["sparse future conditioning", "SparseMoT", "Interval KV-Fusion"], "model_class": "world-action model", "techniques": ["compute-once future representations", "selective video-action interaction", "KV fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.21x faster than Joint-WAM; LIBERO-Plus success 49.1% -> 73.6% vs Fast-WAM", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "LIBERO-Plus", "RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Argues inference-time future features matter for robustness, and keeps them cheaply by computing future representations once and fusing them only at a few network stages, beating Fast-WAM on OOD while staying 2.2x faster than joint WAMs.", "relevance": 4, "relevance_reason": "Efficiency/robustness trade-off analysis for WAM test-time imagination.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.23648", "title": "FAVLA: A Force-Adaptive Fast-Slow VLA model for Contact-Rich Robotic Manipulation", "authors": ["Yao Li", "Peiyuan Tang", "Wu-Yang Zhang", "Cheng Zhu", "Yifan Duan", "Wei-Kai Shi", "Xiaodong Zhang", "Z. Yang", "Jianmin Ji", "Yan-Yong Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.23648", "pdf_url": "https://arxiv.org/pdf/2602.23648", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["fast-slow", "force", "reactivity"], "model_class": "VLA", "techniques": ["fast-slow", "force adapter", "adaptive execution frequency"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "FAVLA decouples a slow VLM planner from a fast force-conditioned action expert whose execution frequency is adaptively scheduled from predicted force variation for reactive contact-rich control.", "relevance": 4, "relevance_reason": "direct: fast-slow VLA with adaptive control frequency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2304.13705"}
{"id": "2609.21216", "title": "Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies", "authors": ["Zhipeng Tang", "Xinda Chen", "Weining Rao", "Xiao Li", "Wenting Tan", "Yuning Wang", "Xiao Shi", "Xiaofang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21216", "pdf_url": "https://arxiv.org/pdf/2609.21216", "code_url": "", "clusters": ["C1"], "keywords": ["flow matching", "few-step", "endpoint correction"], "model_class": "VLA", "techniques": ["few-step sampling", "residual corrector"], "base_model": "pi0-style flow VLA, SmolVLA", "params": "", "hardware": "", "reported_metric": "2-step config 2.12x speedup; 5-step 30.2% lower forward latency vs 10-step", "metric_measures": "relative latency", "benchmarks": ["RoboTwin"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Spends part of the integration budget on a learned one-shot residual correction of a few-step flow sample, improving the quality-latency frontier of frozen flow VLAs.", "relevance": 4, "relevance_reason": "Few-step flow action sampling.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference"}
{"id": "2609.18623", "title": "FIVE-VLA: Fast and EffectIVE Autonomous Driving with Recurrent Action Memory", "authors": ["Kemal Oksuz", "Alexandru Buburuzan", "Yuhan Yao", "Puneet K. Dokania"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18623", "pdf_url": "https://arxiv.org/pdf/2609.18623", "code_url": "", "clusters": ["C1"], "keywords": ["driving VLA", "efficient encoder"], "model_class": "VLA", "techniques": ["token reduction", "recurrent action memory", "skip text generation"], "base_model": "", "params": "641M", "hardware": "A100; T4 (edge proxy)", "reported_metric": "~30 fps on A100, ~4 fps on T4 (8-30x faster than prior methods)", "metric_measures": "throughput", "benchmarks": ["Bench2Drive", "NVIDIA Physical AI AV dataset"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA with an efficient encoder emitting 98 tokens for 448x896 images, single-pass trajectory output (no text), and a recurrent action memory; 641M parameters.", "relevance": 4, "relevance_reason": "Token-efficient compact VLA with explicit edge-proxy speed numbers.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2605.15492", "title": "FLASH: Efficient Visuomotor Policy via Sparse Sampling", "authors": ["Jiaqi Bai", "Jindou Jia", "Yuxuan Hu", "Gen Li", "Xiangyu Chen", "Tuo An", "Kuangji Zuo", "Jianfei Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.15492", "pdf_url": "https://arxiv.org/pdf/2605.15492", "code_url": "", "clusters": ["C1"], "keywords": ["Legendre polynomial actions", "history-anchored flow", "single step"], "model_class": "diffusion/flow policy", "techniques": ["continuous action parameterization", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "31.40 ms per-episode inference; up to 175x faster than diffusion policies", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Represents action chunks as Legendre polynomials and starts flow from previous coefficients, enabling accurate single-step long-horizon generation.", "relevance": 4, "relevance_reason": "Action representation for fast policies.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching robot policy efficient"}
{"id": "2608.27384", "title": "FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference", "authors": ["Zekai Li", "Jiaming Tang", "Zhijian Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.27384", "pdf_url": "https://arxiv.org/pdf/2608.27384", "code_url": "", "clusters": ["C1", "C2", "C3", "C6", "C7"], "keywords": ["streaming action decoding", "multi-noise-level buffer", "chunk-wise causal attention", "asynchronous inference"], "model_class": "VLA", "techniques": ["rolling diffusion", "asynchronous inference", "streaming multi-noise-level action buffer", "chunk-wise causal attention", "streaming action buffer", "multi-noise-level chunks"], "base_model": "", "params": "", "hardware": "single GPU", "reported_metric": ">=30 Hz control frequency", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Keeps a streaming buffer of chunks at different noise levels decoded with chunk-wise causal attention so each inference step yields one executable chunk, giving smooth asynchronous execution without extra future-state conditioning.", "relevance": 4, "relevance_reason": "Latency + async consistency for flow VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; search query; arXiv API all:\"action chunking\" AND latency"}
{"id": "2608.10860", "title": "Flex-$π$: A Multi-Stream World-Action Model with Compute Flexibility", "authors": ["Ge Yan", "Jinghao Liu", "Yuzhi Fan", "Lei Cai", "Minwen Liao", "Jesse Zhang", "Dieter Fox"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.10860", "pdf_url": "https://arxiv.org/pdf/2608.10860", "code_url": "", "clusters": ["C1", "C7", "C8", "C9"], "keywords": ["multi-stream WAM", "compute flexibility", "pointmap supervision"], "model_class": "world-action model", "techniques": ["per-stream dropout", "cross-modality forcing", "action-only fast mode", "configurable compute streams"], "base_model": "", "params": "6B", "hardware": "", "reported_metric": "runs faster than pi0.5 in action-only mode (per abstract)", "metric_measures": "relative latency", "benchmarks": ["real bimanual manipulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "6B WAM supervised on RGB, 3D pointmaps and DINO semantics in a shared VAE latent; stream dropout lets one checkpoint run anything from fast action-only inference to full joint generation.", "relevance": 4, "relevance_reason": "Compute-adjustable WAM inference.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2608.00635", "title": "FlowPilot: Real-Time World-Action Modeling for Agile UAV Navigation", "authors": ["Runqing Wang", "Ding Yu", "Pengyuan Min", "Xinhong Zhang", "Wei Xiao", "Yu Hu", "Jie Chen", "Fu Zhang", "Gang Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00635", "pdf_url": "https://arxiv.org/pdf/2608.00635", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["UAV WAM", "onboard navigation", "Bernstein trajectories"], "model_class": "world-action model", "techniques": ["compact dual-stream MoT", "action-centric deployment (no depth generation)", "polynomial trajectory parameterization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["sim + real UAV flight"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Compact WAM for agile drone flight that co-trains on future depth prediction but outputs only smooth trajectories at deployment, running onboard in real time.", "relevance": 4, "relevance_reason": "Onboard real-time WAM.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2603.24806", "title": "FODMP: Fast One-Step Diffusion of Movement Primitives Generation for Time-Dependent Robot Actions", "authors": ["Xirui Shi", "Arya Ebrahimi", "Yi Hu", "Jun Jin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.24806", "pdf_url": "https://arxiv.org/pdf/2603.24806", "code_url": "", "clusters": ["C1"], "keywords": ["fast inference", "time-dependent actions"], "model_class": "diffusion/flow policy", "techniques": ["one-step diffusion", "distillation", "movement primitives"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 10x faster than MPD, 7x faster than action-chunking diffusion policies", "metric_measures": "relative latency", "benchmarks": ["MetaWorld", "ManiSkill", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distils diffusion over movement primitives into one-step generation for fast, time-dependent closed-loop control.", "relevance": 4, "relevance_reason": "Large speedups enabling dynamic catching", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2406.01586"}
{"id": "2602.12322", "title": "ForeAct: Steering Your VLA with Efficient Visual Foresight Planning", "authors": ["Zhuoyang Zhang", "Shang Yang", "Qinghao Hu", "Luke J. Huang", "James Hou", "Yufei Sun", "Yao Lu", "Song Han"], "year": 2026, "venue": "CVPR 2026 (Highlight)", "type": "preprint", "url": "https://arxiv.org/abs/2602.12322", "pdf_url": "https://arxiv.org/pdf/2602.12322", "code_url": "", "clusters": ["C1", "C7", "C9"], "keywords": ["visual foresight", "subgoal image generation", "planner"], "model_class": "world model", "techniques": ["efficient foresight image generation", "visual subgoal conditioning"], "base_model": "pi0", "params": "", "hardware": "H100", "reported_metric": "0.33 s per 640x480 foresight image on H100; 87.4% success vs 46.5% for pi0", "metric_measures": "other", "benchmarks": ["11 real-world multi-step tasks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "An efficient foresight-image generator (pretrained on >1M cross-embodiment episodes) imagines the next subgoal observation so an unchanged VLA can focus on visuomotor control.", "relevance": 4, "relevance_reason": "Efficient world-model planner for VLAs by organizer Zhuoyang Zhang.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Zhuoyang Zhang"}
{"id": "2608.11605", "title": "Foresight Without Seeing: Latent Futures for World Action Models", "authors": ["Jiakai Huang", "Zhongbo Wu", "Zheng Zhang", "Zihan Wang", "Shan You", "Tao Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.11605", "pdf_url": "https://arxiv.org/pdf/2608.11605", "code_url": "", "clusters": ["C8"], "keywords": ["Future-KV", "dynamics registers", "direct-policy WAM"], "model_class": "world-action model", "techniques": ["single video-DiT prefill producing future K/V", "latent action teacher supervision (train only)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Gives a direct-policy WAM access to predicted dynamics via one prefill pass over current latents plus stochastic future slots, with no future video decoding at deployment.", "relevance": 4, "relevance_reason": "Cheap implicit foresight for WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.20735", "title": "ForeTime-VLA: Causal Future-Token Distillation from a World Action Model for Conveyor-Belt Manipulation", "authors": ["Siyuan Ma", "Yutian Zhang", "Boshi Zhang", "Qinglian Wu", "Jiaqi Zhai", "Dong Wei", "Xiaojin Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.20735", "pdf_url": "https://arxiv.org/pdf/2608.20735", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["future-token distillation", "WAM teacher", "dynamic objects"], "model_class": "VLA", "techniques": ["distill WAM future representation into VLA", "causal history encoder"], "base_model": "pi0.5; Fast-WAM teacher", "params": "", "hardware": "", "reported_metric": "2.46-2.93% latency overhead vs base pi0.5", "metric_measures": "relative latency", "benchmarks": ["conveyor-belt manipulation (real)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Distills a compact 64-D future-aware target from a frozen WAM teacher into a pi0.5 policy, so the VLA anticipates moving objects with ~3% latency overhead and no teacher at runtime.", "relevance": 4, "relevance_reason": "World-model knowledge transferred to a fast policy.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.06987", "title": "Foundational World Models Accurately Detect Bimanual Manipulator Failures", "authors": ["Isaac R. Ward", "Michelle Ho", "Houjun Liu", "Aaron Feldman", "Joseph Vincent", "Liam Kruse", "Sean Cheong", "Duncan Eddy", "Mykel J. Kochenderfer", "Mac Schwager"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.06987", "pdf_url": "https://arxiv.org/pdf/2603.06987", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["bimanual failure detection", "Cosmos tokenizer latent", "conformal runtime monitor"], "model_class": "world model", "techniques": ["probabilistic latent world model in Cosmos tokenizer space", "uncertainty as nonconformity score"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Push-T", "Bimanual Cable Manipulation dataset"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains a probabilistic world model in a pretrained video tokenizer's latent space and uses its predictive uncertainty as a conformal runtime monitor for bimanual failures; releases a new bimanual dataset.", "relevance": 4, "relevance_reason": "World-model uncertainty for safety monitoring.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.30965", "title": "FRAM: Trajectory-Guided Visual Feature Selection for Compact Language-Conditioned Robot Manipulation", "authors": ["Hiroshi Ito", "Hyogo Hiruma", "Yoshiki Kanai", "Takahiro Yoshida", "Akira Kanazawa", "Hiroki Yamada"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.30965", "pdf_url": "https://arxiv.org/pdf/2609.30965", "code_url": "", "clusters": ["C1"], "keywords": ["FRAM", "small policy", "feature selection"], "model_class": "other", "techniques": ["trajectory-guided feature selection", "small policy"], "base_model": "", "params": "138.7M", "hardware": "", "reported_metric": "92.2% LIBERO vs 94.2% for pi0 (3.3B)", "metric_measures": "other", "benchmarks": ["LIBERO", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "FRAM is a small policy that predicts future end-effector trajectories and reads local visual features at those points, approaching pi0 performance with far fewer parameters.", "relevance": 4, "relevance_reason": "Compact policy near 3.3B-model performance.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (small OR compact OR tiny) AND \"vision-language-action\"; snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2604.24391", "title": "FreqCache: Accelerating Embodied VLN Models with Adaptive Frequency-Guided Token Caching", "authors": ["Zihao Zheng", "Xingyu Zhou", "Z. Mao", "Songyu Sun", "Lingyu Zhang", "Yu-Long Ao", "Yu Feng", "Qiong Zhang", "Yong-Hua Lin", "Xiang Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.24391", "pdf_url": "https://arxiv.org/pdf/2604.24391", "code_url": "", "clusters": ["C1"], "keywords": ["VLN", "frequency domain"], "model_class": "VLA", "techniques": ["token caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.59x speedup", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Frequency-guided token caching that accelerates VLN models training-free.", "relevance": 4, "relevance_reason": "Token caching acceleration.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2502.02175, 2603.07949"}
{"id": "2607.24008", "title": "FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking", "authors": ["Hai Jiang", "Yixian Zou", "Binbin Liang", "Boqian Liu", "Fanman Meng", "Shuaicheng Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.24008", "pdf_url": "https://arxiv.org/pdf/2607.24008", "code_url": "", "clusters": ["C1", "C2", "C6", "C7"], "keywords": ["asynchronous execution", "observation prediction", "RTC", "anticipatory conditioning", "execution-time observation prediction", "FutureRTC", "asynchronous VLA"], "model_class": "VLA", "techniques": ["asynchronous inference", "state/observation forecasting", "future observation prediction", "state correction", "state roll-forward correction", "observation forecasting", "policy consistency loss"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Plug-in adapter that forecasts the proprioceptive state and visual features the robot will have at execution time and feeds them to an unchanged VLA, reducing prediction-execution misalignment under delay.", "relevance": 4, "relevance_reason": "Delay compensation for async VLA execution.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; search query; arXiv search: \"action chunk\" asynchronous inference"}
{"id": "2609.05588", "title": "GE-Act 2.0: Pretraining and Scaling a World-Action Model for Robotic Manipulation", "authors": ["AgiBot Research Team", "Renhang Liu", "Wenzhi Zhao", "Zhuo Yang", "Liliang Chen", "Pengfei Zhou", "Shengcong Chen", "Guanghui Ren", "Youlun Peng", "Rongjun Jin", "Nan Wang", "Sukai Wang", "Xindong He", "Jinyuan Feng", "Ziyu Xiong", "Linqing Zhong", "Yifei Wei", "Feng Han", "Long Zhang", "Da Huang", "Nanshu Zhao", "Chenghao Yin", "Mo Wu", "Zhaodong Yan", "Kongtao Hu", "Yuxiang Yan", "Aogelijiang Niyazi", "Yu Fang", "Jia Zeng", "Lizhu Meng", "Daizhen Lv", "Haoyu Cao", "Zhiwen Hou", "Lianjin Ye", "Yuehan Niu", "Zhikai Cai", "Xuan Hu", "Hui Min", "Xiongfeng Cai", "Yue Liao", "Jing Wu", "Soujanya Poria", "Ye Li", "Sanping Zhou", "Maoqing Yao"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2609.05588", "pdf_url": "https://arxiv.org/pdf/2609.05588", "code_url": "", "clusters": ["C8"], "keywords": ["GE-Act 2.0", "single-step visual planner", "control-oriented autoencoder"], "model_class": "world-action model", "techniques": ["aggressive control-oriented compression", "one-pass future prediction", "separate IDM pretraining", "selective optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["100 tasks / 20 skill groups (AgiBot G1/G2)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "AgiBot WAM trained from scratch with a compressive control-oriented autoencoder and a single-step visual planner that produces the full future state in one pass, scaled to 30k hours of co-training data.", "relevance": 4, "relevance_reason": "One-pass (non-iterative) visual planner design for efficient WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2605.27491", "title": "GE-Sim 2.0: A Roadmap Towards Comprehensive Closed-loop Video World Simulators for Robotic Manipulation", "authors": ["Boxiang Qiu", "Liliang Chen", "Yue Liao", "Nan Wang", "Lintao Wang", "Jiayi Luo", "Wenzhi Zhao", "Shengcong Chen", "Di Chen", "Ye Li", "Chen Gao", "Shuicheng Yan", "Si Liu", "Maoqing Yao", "Guanghui Ren"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2605.27491", "pdf_url": "https://arxiv.org/pdf/2605.27491", "code_url": "", "clusters": ["C3", "C4", "C8"], "keywords": ["GE-Sim 2.0", "video world simulator", "world judge", "closed-loop evaluation", "closed-loop video simulator", "frame skipping"], "model_class": "world model", "techniques": ["random-stride training for up to 4x frame skipping", "state expert", "world judge"], "base_model": "", "params": "", "hardware": "H100", "reported_metric": "25-frame rollout in 2.3 s on one H100", "metric_measures": "other", "benchmarks": ["closed-loop policy evaluation (AgiBot)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "AgiBot's second-generation closed-loop video world simulator with a state decoder and judge; random-stride training allows up to 4x frame skipping at inference to extend horizon cheaply.", "relevance": 4, "relevance_reason": "Efficient closed-loop neural simulator (Genie Envisioner line).", "foundational_technique": false, "verified": true, "found_via": "search query: VLA leaderboard; search query"}
{"id": "gemini-robotics-2-blog-2026-c6", "title": "Gemini Robotics 2 brings whole body intelligence to robots", "authors": ["Carolina Parada", "Google DeepMind"], "year": 2026, "venue": "blog (Google DeepMind)", "type": "blog", "url": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/", "pdf_url": "", "code_url": "", "clusters": ["C2", "C6", "C8", "C9"], "keywords": ["cloud ER model", "on-device VLA", "model family split", "Gemini Robotics 2", "On-Device 2", "whole-body humanoid", "ASIMOV-Agentic safety"], "model_class": "VLA", "techniques": ["hierarchical cloud reasoning / on-device VLA", "on-device model", "embodiment adaptation"], "base_model": "", "params": "", "hardware": "Apptronik Apollo 2, Franka Duo, Dexmate, SO101, Trossen", "reported_metric": "", "metric_measures": "", "benchmarks": ["Apptronik Apollo 2", "Franka Duo", "SO101", "Trossen"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "July 2026 release of Gemini Robotics 2 (VLA), ER 2 (reasoning VLM) and On-Device 2, which adapts to new embodiments in hours from <200 examples; includes a safety report and an agentic safety benchmark. No latency/control-rate figures given.", "relevance": 4, "relevance_reason": "Industry reference architecture splitting cloud reasoning and on-device control.", "foundational_technique": false, "verified": true, "found_via": "search query; blog sweep: Google DeepMind robotics"}
{"id": "2603.17834", "title": "Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control", "authors": ["Zunzhe Zhang", "Runhan Huang", "Yicheng Liu", "Shaoting Zhu", "Linzhan Mou", "Hang Zhao", "Run-Han Huang", "Shao-Ting Zhu", "Lin-Zhan Mou"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.17834", "pdf_url": "https://arxiv.org/pdf/2603.17834", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["GeCO", "time-unconditional flow"], "model_class": "diffusion/flow policy", "techniques": ["time-unconditional flow matching", "adaptive inference", "adaptive-step flow matching", "early exit", "OOD detection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reformulates flow-matching control as optimization so inference exits early on easy states and refines on hard ones; the field norm doubles as a training-free OOD safety signal.", "relevance": 4, "relevance_reason": "Adaptive inference compute plus a safety monitor; scales to pi0 VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.13960", "title": "GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch", "authors": ["GigaWorld Team", "Angen Ye", "Angyuan Ma", "Boyuan Wang", "Chaojun Ni", "Fangzheng Ye", "Guan Huang", "Guo Li", "Guosheng Zhao", "Haodong Yan", "Hengtao Li", "Jiwen Lu", "Kai Wang", "Mingming Yu", "Qitang Hu", "Qiuping Deng", "Songling Liu", "Xiaoyu Tian", "Xiaofeng Wang", "Xinyu Zhou", "Xiuwei Xu", "Xinze Chen", "Yang Wang", "Yejun Zeng", "Yifan Chang", "Yun Ye", "Zhenyu Wu", "Zhanqian Wu", "Zheng Zhu"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2607.13960", "pdf_url": "https://arxiv.org/pdf/2607.13960", "code_url": "", "clusters": ["C1", "C7", "C8", "C9"], "keywords": ["action-centered WAM", "MoT", "AutoResearch", "no video at inference"], "model_class": "world-action model", "techniques": ["future video only in training", "mixture-of-transformers expert separation", "action-only decoding", "training-only video dynamics", "action-only decoding at inference"], "base_model": "", "params": "", "hardware": "RTX 4090", "reported_metric": "85 ms inference latency", "metric_measures": "relative latency", "benchmarks": ["robot manipulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Action-centered WAM that trains on future visual dynamics but decodes only actions at test time through a separated MoT expert, reaching 85 ms latency on a local RTX 4090.", "relevance": 4, "relevance_reason": "WAM with local consumer-GPU latency number.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2603.17240", "title": "GigaWorld-Policy: An Efficient Action-Centered World--Action Model", "authors": ["Angen Ye", "Boyuan Wang", "Chaojun Ni", "Guan Huang", "Guosheng Zhao", "Hao Li", "Hengtao Li", "Jie Li", "Jindi Lv", "Jingyu Liu", "Min Cao", "Peng Li", "Qiuping Deng", "Wenjun Mei", "Xiaofeng Wang", "Xinze Chen", "Xinyu Zhou", "Yang Wang", "Yifan Chang", "Yifan Li", "Yukun Zhou", "Yun Ye", "Zhichao Liu", "Zheng Zhu", "An-Gen Ye", "Heng-Tao Li", "Jin-Di Lv", "Mingyu Cao", "Yi-Fan Chang", "Yi-Fan Li", "Yu-Kun Zhou", "Zhi-Chao Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.17240", "pdf_url": "https://arxiv.org/pdf/2603.17240", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["GigaWorld-Policy", "world-action model", "efficient action decoding", "video backbone"], "model_class": "world-action model", "techniques": ["causal masking", "optional video generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "9x faster than Motus WAM baseline", "metric_measures": "relative latency", "benchmarks": ["RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Action-centred world-action model whose causal design lets it skip future-video generation at inference, decoding actions much faster than prior WAMs.", "relevance": 4, "relevance_reason": "Explicitly targets WAM inference overhead", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=100]; snowball round 2: cited by 2604.26694, 2606.09215"}
{"id": "2603.17573", "title": "HeiSD: Hybrid Speculative Decoding for Embodied Vision-Language-Action Models with Kinematic Awareness", "authors": ["Zihao Zheng", "Zhihao Mao", "Sicheng Tian", "Maoliang Li", "Jiayu Chen", "Xinhao Sun", "Zhaobo Zhang", "Xuanzhe Liu", "Donggang Cao", "Hong Mei", "Xiang Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.17573", "pdf_url": "https://arxiv.org/pdf/2603.17573", "code_url": "", "clusters": ["C1"], "keywords": ["hybrid speculative decoding", "retrieval drafts", "kinematics"], "model_class": "VLA", "techniques": ["speculative decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 2.45x (sim), 2.06-2.41x (real) speedup", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Combines drafter-based and retrieval-based speculative decoding for VLAs, choosing between them with a kinematics-based metric and relaxing sequence acceptance.", "relevance": 4, "relevance_reason": "Speculative decoding for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: speculative decoding VLA"}
{"id": "2605.30834", "title": "Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring", "authors": ["Seongheon Park", "Wendi Li", "Changdae Oh", "Samuel Yeh", "Zsolt Kira", "Michael Hagenow", "Sharon Li"], "year": 2026, "venue": "NeurIPS 2026", "type": "paper", "url": "https://arxiv.org/abs/2605.30834", "pdf_url": "https://arxiv.org/pdf/2605.30834", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["Hide-and-Seek", "weakly supervised failure detection", "contrastive"], "model_class": "VLA", "techniques": ["inter/intra-trajectory contrastive learning", "conformal prediction"], "base_model": "OpenVLA, pi0, pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "VLABench", "real"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns step-level failure signals for VLA runtime monitoring from only trajectory-level labels via contrastive objectives, avoiding action resampling or external models.", "relevance": 4, "relevance_reason": "Cheap VLA runtime monitor.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: FAIL-Detect"}
{"id": "2604.03208", "title": "Hierarchical Planning with Latent World Models", "authors": ["Wancong Zhang", "Basile Terver", "Artem Zholus", "Soham Chitnis", "Harsh Sutaria", "Mido Assran", "Randall Balestriero", "Amir Bar", "Adrien Bardes", "Yann LeCun", "Nicolas Ballas"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.03208", "pdf_url": "https://arxiv.org/pdf/2604.03208", "code_url": "", "clusters": ["C8"], "keywords": ["hierarchical latent MPC", "macro-actions", "multi-timescale world model"], "model_class": "world model", "techniques": ["multi-scale latent world models", "latent macro-action encoder", "subgoal latent matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real Franka manipulation", "navigation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Hierarchical MPC over latent world models at several temporal scales with compressed macro-actions, keeping long-horizon search tractable and reducing planning cost on real Franka tasks.", "relevance": 4, "relevance_reason": "Reduces exponential planning cost of latent world-model MPC.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.22860", "title": "HiL-ResRL: A Model-Agnostic Finetuning Adapter via Human-in-the-loop Residual Reinforcement Learning", "authors": ["Jingyi Liu", "Zhaohong Mai", "ShunSen He", "Hang Ren", "Chao Wang", "Shunbo Zhou", "XiaoDong Wu", "Heng Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.22860", "pdf_url": "https://arxiv.org/pdf/2606.22860", "code_url": "", "clusters": ["C3"], "keywords": ["residual RL", "model-agnostic adapter", "human-in-the-loop"], "model_class": "VLA", "techniques": ["residual policy", "online RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">95% success within 1.5 h of real-world RL", "metric_measures": "other", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains a small residual RL policy on top of frozen VLA actions with human guidance, making real-world fine-tuning architecture-agnostic and fast.", "relevance": 4, "relevance_reason": "Lightweight residual adaptation leaves the VLA frozen.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv"}
{"id": "2605.28726", "title": "How VLAs Fail Differently: Black-Box Action Monitoring Reveals Architecture-Specific Failure Signatures", "authors": ["Krishna Gupta"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.28726", "pdf_url": "https://arxiv.org/pdf/2605.28726", "code_url": "https://github.com/krishnam94/vla-edge", "clusters": ["C7", "C8"], "keywords": ["black-box action monitoring", "SafeContract", "direction reversal", "jerk", "velocity checks"], "model_class": "other", "techniques": ["training-free action monitors", "conformal calibration"], "base_model": "VQ-BeT, Diffusion Policy, ACT", "params": "", "hardware": "", "reported_metric": "direction-reversal AUROC 0.93/0.79/0.91; velocity-violation AUROC 0.41-0.69", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Finds motor-command failure signatures differ by architecture: direction reversal predicts failure universally while common velocity-limit checks carry almost no signal for continuous policies.", "relevance": 4, "relevance_reason": "Cheap black-box runtime monitors; questions standard velocity checks.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2602.13718", "title": "HybridFlow: A 2-NFE Generative Policy for Real-Time Robotic Manipulation", "authors": ["Zhenchen Dong", "Fulin Chen", "Jinna Fu", "Jiaming Wu", "Qingran Wu", "Shengyuan Yu", "Hongyu Yu", "Yide Liu", "Zhen Dong", "Jin-Na Fu", "Jia-Min Wu", "Sheng-Yuan Yu", "Fu-Lin Chen", "Yi-De Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.13718", "pdf_url": "https://arxiv.org/pdf/2602.13718", "code_url": "", "clusters": ["C1", "C3", "C5", "C9"], "keywords": ["MeanFlow", "2-NFE", "Jetson Thor", "Jetson AGX Thor", "real-time generative policy", "two-step generative policy"], "model_class": "diffusion/flow policy", "techniques": ["mean flow", "few-step sampling", "few-step flow policy", "MeanFlow global jump + local refine", "few-step generation"], "base_model": "", "params": "", "hardware": "Jetson AGX Thor", "reported_metric": "~8x lower action-generation latency than 16-step Diffusion Policy", "metric_measures": "relative latency", "benchmarks": ["RoboMimic", "real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Two-evaluation policy: a MeanFlow global jump, parameter-free re-noising, and one local refinement step with the same network, no distillation.", "relevance": 4, "relevance_reason": "Primarily C3, but provides Thor-measured latency comparison.", "foundational_technique": false, "verified": true, "found_via": "search query: one-step diffusion policy; arXiv: Jetson AND (robot policy OR manipulation OR VLA); venue sweep: arXiv API VLA AND Jetson; snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2605.01581", "title": "Hyper-DP3: Frequency-Aware Right-Sizing of 3D Diffusion Policies for Visuomotor Control", "authors": ["Jin-Hao Zhang", "Zhexuan Zhou", "Hui-Zhe Li", "Yi-Chen Lai", "Wenlong Xia", "Haoming Song", "You-Min Gong", "Jie Mei"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.01581", "pdf_url": "https://arxiv.org/pdf/2605.01581", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["pocket-scale diffusion policy", "frequency analysis"], "model_class": "diffusion/flow policy", "techniques": ["two-step DDIM", "lightweight decoder", "right-sizing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "<1% of parameters of prior 3D diffusion policies; two-step inference", "metric_measures": "other", "benchmarks": ["RoboTwin2.0", "Adroit", "MetaWorld"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Shows action denoising saturates after few steps and builds a tiny 3D diffusion policy with two-step DDIM inference.", "relevance": 4, "relevance_reason": "Model right-sizing and few-step sampling for diffusion policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2604.16405", "title": "ICAT: Incident-Case-Grounded Adaptive Testing for Physical-Risk Prediction in Embodied World Models", "authors": ["Zhenglin Lai", "Sirui Huang", "Yuteng Li", "Changxin Huang", "Jianqiang Li", "Bingzhe Wu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2604.16405", "pdf_url": "https://arxiv.org/pdf/2604.16405", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["physical-risk prediction", "incident reports", "world model testing"], "model_class": "world model", "techniques": ["incident-grounded risk-case generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Benchmark grounded in real incident reports showing mainstream video world models downplay danger cues and miscalibrate severity for hazardous actions.", "relevance": 4, "relevance_reason": "Tests whether world models can serve as safety predictors.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.19531", "title": "ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?", "authors": ["Yuyang Zhang", "Wenyao Zhang", "Zekun Qi", "He Zhang", "Haitao Lin", "Jingbo Zhang", "Yao Mu", "Xiaokang Yang", "Wenjun Zeng", "Xin Jin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.19531", "pdf_url": "https://arxiv.org/pdf/2606.19531", "code_url": "", "clusters": ["C1", "C7", "C8", "C9"], "keywords": ["image editing prior", "KV-cache conditioning", "WAM without video", "image editing vs video generation", "ImageWAM"], "model_class": "world-action model", "techniques": ["pretrained image-editing model as world prior", "action expert on editing KV caches", "no target-frame decoding", "single-frame future editing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "FLOPs 1/6 and latency 1/4 of video-based WAMs", "metric_measures": "relative latency", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Replaces video generation in WAMs with a single-target-frame image-editing prior whose KV caches condition a flow-matching action expert, cutting compute 6x and latency 4x.", "relevance": 4, "relevance_reason": "Cheaper world prior for WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2606.01098", "title": "Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry", "authors": ["Zemin Yang", "Yao He", "Yiming Zhong", "Yuhao Zhang", "Xinge Zhu", "Yao Mu", "Qingqiu Huang", "Yue-Xin Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.01098", "pdf_url": "https://arxiv.org/pdf/2606.01098", "code_url": "", "clusters": ["C1"], "keywords": ["one-step policy"], "model_class": "diffusion/flow policy", "techniques": ["one-step generation", "implicit drifting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Implicit Drifting Policy is a one-step imitation policy that injects training-time correction via conditional expert geometry, avoiding iterative diffusion sampling for high-frequency control.", "relevance": 4, "relevance_reason": "direct: one-step action generation", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2606.02486", "title": "Intercepting the Future: Latent-Space Predictive World Model for Dynamic VLA Manipulation", "authors": ["Shahram Najam Syed", "Arthur Jakobsson", "Haoran Hao", "Jeffrey Ichnowski"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.02486", "pdf_url": "https://arxiv.org/pdf/2606.02486", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["AHEAD", "dynamic manipulation", "dynamic objects", "predict-then-act"], "model_class": "world model", "techniques": ["latent world model", "latency compensation", "adaptive horizon"], "base_model": "OpenVLA 7B", "params": "4.9M added", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["20 dynamic sim scenarios", "real robot (xArm 7)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Wraps a frozen OpenVLA with a 4.9M-parameter motion-aware latent world model that forecasts future patch tokens over an adaptive horizon, compensating inference latency for moving objects.", "relevance": 4, "relevance_reason": "Directly addresses the observation-to-execution latency gap with a lightweight add-on world model.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (small OR compact OR tiny) AND \"vision-language-action\"; snowball round 2: cites 2106.09685; cited by 2609.09250"}
{"id": "2609.27656", "title": "InternW0: A Foundational Physical World Model for Efficient Real-World Interactions", "authors": ["Jisong Cai", "Yao Mu", "Ganlin Yang", "Zhe Cao", "Zhangzheng Tu", "Xing Gao", "Kailin Li", "Xinyu Zhan", "Lixin Yang", "Yangkun Zhu", "Haoxiang Ma", "Ming Zhou", "Qiaojun Yu", "Yufei Xue", "Liqun He", "Yifei Yao", "Yifan Zhu", "Long Ling", "Bingqi Jiang", "Haoyu Guo", "Xueyue Zhu", "Bowen Zhou", "Bin Zhao", "Tianfan Xue", "Chunhua Shen", "Weinan Zhang"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2609.27656", "pdf_url": "https://arxiv.org/pdf/2609.27656", "code_url": "", "clusters": ["C8"], "keywords": ["InternW0", "asymmetric video-action", "K/V reuse"], "model_class": "world-action model", "techniques": ["high-capacity video expert + light fast action expert", "layerwise K/V reuse with observation adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shanghai AI Lab physical world model that runs a light action expert at a faster timescale than the video expert and reuses adapted K/V instead of regenerating the future each step.", "relevance": 4, "relevance_reason": "Multi-frequency WAM design.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "figure-helix-02-blog-2026", "title": "Introducing Helix 02: Full-Body Autonomy", "authors": ["Figure AI"], "year": 2026, "venue": "blog (Figure)", "type": "blog", "url": "https://www.figure.ai/news/helix-02", "pdf_url": "", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["System 0", "whole-body controller", "1 kHz", "three-rate hierarchy", "tactile"], "model_class": "VLA", "techniques": ["three-level hierarchy (S2/S1/S0)", "sim-to-real RL whole-body controller"], "base_model": "", "params": "S0 10M", "hardware": "", "reported_metric": "S0 at 1 kHz; S1 at 200 Hz", "metric_measures": "closed-loop control", "benchmarks": ["real robot (4-min dishwasher task, 61 actions)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Adds a 10M-parameter learned whole-body controller (S0, 1 kHz) under the 200 Hz S1 and semantic S2, enabling full-body loco-manipulation from one visuomotor stack.", "relevance": 4, "relevance_reason": "Multi-rate onboard architecture with tiny high-rate controller.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Figure"}
{"id": "2608.09381", "title": "JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling", "authors": ["Yihan Lin", "Jiawei He", "Shifeng Bao", "Chen Zhao", "Yang Li", "Xiaobo Wang", "Yan Wang", "Cheng Chi", "Jing Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09381", "pdf_url": "https://arxiv.org/pdf/2608.09381", "code_url": "", "clusters": ["C8"], "keywords": ["latent WAM", "V-JEPA space", "shared predictor"], "model_class": "world-action model", "techniques": ["latent transition prediction in V-JEPA space", "no future video generation"], "base_model": "V-JEPA; pi0.5 instantiation", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "WAM built in V-JEPA latent space that couples latent transition prediction and action generation through one predictor, avoiding video-generation deployment cost.", "relevance": 4, "relevance_reason": "JEPA route to cheap WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.06049", "title": "L-SDPPO: Policy Optimization of Spiking Diffusion Policy for Intra-vehicular Robotic Manipulation", "authors": ["Li-Wen Zhang", "Dongruo Zhou", "Guang-Hui Sun", "Yifei Zheng", "Yuhui Hu", "Kaihong Ouyang", "Zuo-Quan Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.06049", "pdf_url": "https://arxiv.org/pdf/2606.06049", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["spiking diffusion policy", "low energy"], "model_class": "diffusion/flow policy", "techniques": ["spiking neural network", "RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "L-SDPPO uses an RL-optimized spiking diffusion policy for low-energy manipulation on power-constrained spacecraft robots, with a latency-injection mechanism for dynamic perception.", "relevance": 4, "relevance_reason": "direct: energy-efficient diffusion policy", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2601.05248", "title": "LaST$_{0}$: Latent Spatio-Temporal Chain-of-Thought for Robotic Vision-Language-Action Model", "authors": ["Zhuoyang Liu", "Jiaming Liu", "Hao Chen", "Jiale Yu", "Ziyu Guo", "Chengkai Hou", "Chenyang Gu", "Xiangju Mi", "Renrui Zhang", "Kun Wu", "Zhengping Che", "Jian Tang", "Pheng-Ann Heng", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.05248", "pdf_url": "https://arxiv.org/pdf/2601.05248", "code_url": "", "clusters": ["C1"], "keywords": ["latent spatio-temporal CoT", "mixture of transformers", "dual system"], "model_class": "VLA", "techniques": ["latent reasoning", "dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Latent CoT over future visual dynamics, 3D structure and proprioception in a mixture-of-transformers with a low-frequency reasoning expert and a high-frequency acting expert.", "relevance": 4, "relevance_reason": "Dual-rate latent reasoning VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "2608.24882", "title": "Latent Action as Intention Enables Efficient Future Imagination for World Action Models", "authors": ["Xiang Li", "Yupeng Zheng", "Songen Gu", "Huailiang Ma", "Feng Yu", "Yuhang Zheng", "Xian Nie", "Shanshuai Yuan", "Yujie Zang", "Weize Li", "Shuai Tian", "Moyang Liu", "Ya-Qin Zhang", "Wenchao Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.24882", "pdf_url": "https://arxiv.org/pdf/2608.24882", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["latent action", "LAWA", "future intention", "latent action intention", "efficient imagination"], "model_class": "world-action model", "techniques": ["discrete latent-action tokenizer", "joint denoising of latent state and actions", "no video branch at inference", "latent-action future imagination"], "base_model": "", "params": "", "hardware": "", "reported_metric": "42.9% lower inference latency than matched Joint-WAM at similar performance", "metric_measures": "relative latency", "benchmarks": ["RoboCasa", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Keeps test-time future imagination but in a compact latent-action space instead of pixels, recovering Joint-WAM generalization that Fast-WAM loses while cutting latency by ~43%.", "relevance": 4, "relevance_reason": "Middle ground between full imagination and none for WAM efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2605.02739", "title": "Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference", "authors": ["Yudong Liu", "Yuan Li", "Zijia Tang", "Yuxi Zheng", "Yueqian Lin", "Qinsi Wang", "Yi Li", "Shuangjun Liu", "Shuai Zhang", "Taotao Jing", "Dashan Gao", "Ning Bi", "Jingwei Sun", "Yiran Chen", "Hai Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.02739", "pdf_url": "https://arxiv.org/pdf/2605.02739", "code_url": "", "clusters": ["C1"], "keywords": ["dual-system VLA", "feature delta prediction", "VLM call reduction"], "model_class": "VLA", "techniques": ["feature forecasting", "reduced backbone invocation"], "base_model": "GR00T N1.6, pi0.5", "params": "", "hardware": "", "reported_metric": "50-75% fewer VLM calls; 1.65-1.73x per-episode speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboCasa", "ALOHA sim"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns to predict how VLM outputs change between steps so the action head can run on predicted features while the backbone runs only periodically.", "relevance": 4, "relevance_reason": "Fast-slow dual-system efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "2605.08732", "title": "Latent Geometry Beyond Search: Amortizing Planning in World Models", "authors": ["Hoang-Quan Nguyen", "Xiao-Hao Xu", "Xiao-Nan Huang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.08732", "pdf_url": "https://arxiv.org/pdf/2605.08732", "code_url": "", "clusters": ["C8"], "keywords": ["GC-IDM", "latent planning"], "model_class": "world model", "techniques": ["amortized planning", "inverse dynamics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "100-130x lower per-decision cost vs CEM", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces iterative latent-space planning (CEM) with a learned goal-conditioned inverse dynamics model, matching performance at a fraction of per-decision cost.", "relevance": 4, "relevance_reason": "Large per-decision compute reduction for world-model planning.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104; cited by 2609.32591"}
{"id": "2602.01166", "title": "Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models", "authors": ["Shuanghao Bai", "Jing Lyu", "Wanqi Zhou", "Zhe Li", "Dakai Wang", "Lei Xing", "Xiaoguang Zhao", "Pengwei Wang", "Zhongyuan Wang", "Cheng Chi", "Badong Chen", "Shanghang Zhang"], "year": 2026, "venue": "ICML 2026", "type": "paper", "url": "https://arxiv.org/abs/2602.01166", "pdf_url": "https://arxiv.org/pdf/2602.01166", "code_url": "", "clusters": ["C1"], "keywords": ["latent reasoning", "curriculum"], "model_class": "VLA", "techniques": ["latent reasoning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 90% lower inference latency vs explicit CoT", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Curriculum moves VLAs from explicit textual/visual CoT to continuous latent reasoning, eliminating CoT generation at inference.", "relevance": 4, "relevance_reason": "Efficient reasoning VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: reasoning VLA latency"}
{"id": "2606.15768", "title": "LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies", "authors": ["Jialei Chen", "Kai Wang", "Kang Chen", "Shuaihang Chen", "Feng Gao", "Wenhao Tang", "Zhiyuan Li", "Weilin Liu", "Zhuyu Yao", "Boxun Li", "Yuanbo Xu", "Chao Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.15768", "pdf_url": "https://arxiv.org/pdf/2606.15768", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["latent WAM", "latent visual subgoals", "vision foundation model features"], "model_class": "world-action model", "techniques": ["latent-action-conditioned latent world model", "feature-space subgoal prediction", "latent world model on foundation features"], "base_model": "", "params": "", "hardware": "", "reported_metric": "187 ms per action-chunk; up to 24x lower wall-clock latency than pixel-space WAMs", "metric_measures": "open-loop chunk rate", "benchmarks": ["LIBERO", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Conditions the policy on predicted future features in a frozen vision-foundation-model latent space instead of generated video, keeping foresight while being up to 24x faster than pixel WAMs.", "relevance": 4, "relevance_reason": "Latent-space foresight for fast WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2606.16253", "title": "Learned Image Compression for Vision-Language-Action Models", "authors": ["Hyeonjun Kim", "Jegwang Ryu", "Sangbeom Ha", "Junhyeok Lee", "Jun-Hyuk Kim", "Hyemin Ahn", "Jaeho Lee", "S. Ha"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.16253", "pdf_url": "https://arxiv.org/pdf/2606.16253", "code_url": "", "clusters": ["C2", "C6"], "keywords": ["SPARC", "bandwidth", "remote inference"], "model_class": "VLA", "techniques": ["learned compression", "rate allocation", "learned image compression", "spatially adaptive rate control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa365", "VLABench", "LIBERO", "real remote control"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learned image codec that allocates bitrate by task relevance so VLAs keep control performance under tight bandwidth in distributed/remote deployment.", "relevance": 4, "relevance_reason": "Addresses camera-stream bandwidth for split/remote VLA deployment.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]; snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2602.12978", "title": "Learning Native Continuation for Action Chunking Flow Policies", "authors": ["Yufeng Liu", "Hang Yu", "Juntu Zhao", "Bocheng Li", "Di Zhang", "Mingzhu Li", "Wenxuan Wu", "Yingdong Hu", "Junyuan Xie", "Junliang Guo", "Dequan Wang", "Yang Gao", "Jun-Tu Zhao", "Ming-Zhe Li", "Ying-Dong Hu"], "year": 2026, "venue": "RSS 2026", "type": "preprint", "url": "https://arxiv.org/abs/2602.12978", "pdf_url": "https://arxiv.org/pdf/2602.12978", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["action chunking", "flow policy", "continuation", "smoothness", "Legato", "native continuation", "training-time continuation"], "model_class": "VLA", "techniques": ["training-time continuation", "asynchronous inference", "schedule-shaped mixture initialization", "flow reshaping"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~10% better smoothness and completion time vs RTC", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Trains flow VLAs to natively continue from known committed actions by starting denoising from a schedule-shaped mix of actions and noise, avoiding RTC-induced mode switching.", "relevance": 4, "relevance_reason": "Smooth async execution built into policy.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; arXiv search: real-time chunking"}
{"id": "2603.22703", "title": "Learning Safe-Stoppability Monitors for Humanoid Robots", "authors": ["Yifan Sun", "Yiyuan Pan", "Shangtao Li", "Caiwu Ding", "Tao Cui", "Lingyun Wang", "Changliu Liu"], "year": 2026, "venue": "IROS 2026", "type": "preprint", "url": "https://arxiv.org/abs/2603.22703", "pdf_url": "https://arxiv.org/pdf/2603.22703", "code_url": "https://intelligent-control-lab.github.io/humanoid_stoppability/", "clusters": ["C7", "C8", "C9"], "keywords": ["safe-stoppability", "emergency stop", "fallback controller", "humanoid", "PRISM", "safe-stoppable set"], "model_class": "other", "techniques": ["learned stoppability monitor", "importance sampling", "fallback policy", "learned safety monitor"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Formalizes humanoid E-stop as executing a balance-preserving fallback controller and learns a neural monitor of states from which such a stop is safe, refined via importance sampling and transferred to hardware.", "relevance": 4, "relevance_reason": "Fallback-policy safety for learned humanoid controllers (Changliu Liu lab).", "foundational_technique": false, "verified": true, "found_via": "WebSearch: Changliu Liu safe control; arXiv search \"emergency stop\"; author sweep: Changliu Liu"}
{"id": "2602.00686", "title": "Learning to Accelerate Vision-Language-Action Models through Adaptive Visual Token Caching", "authors": ["Yujie Wei", "Jiahan Fan", "Jiyu Guo", "Ruichen Zhen", "Rui Shao", "Xiu Su", "Zeke Xie", "Hongxun Yao", "Shuo Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.00686", "pdf_url": "https://arxiv.org/pdf/2602.00686", "code_url": "", "clusters": ["C1"], "keywords": ["learned token caching", "cache ratio predictor"], "model_class": "VLA", "techniques": ["token caching", "learned policy for compute allocation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.76x wall-clock speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SIMPLER", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Replaces heuristic caching with trainable modules deciding which and how many visual tokens to reuse, optimized end-to-end with a differentiable relaxation.", "relevance": 4, "relevance_reason": "Learned KV/token reuse.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference"}
{"id": "2608.05084", "title": "Learning When to Stop: Prefix-Optimal Dynamic Diffusion Policies for Continuous Control", "authors": ["Rohit Kumar Salla", "M. Saravanan", "Simon Stepputtis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.05084", "pdf_url": "https://arxiv.org/pdf/2608.05084", "code_url": "", "clusters": ["C1"], "keywords": ["dynamic denoising"], "model_class": "diffusion/flow policy", "techniques": ["adaptive early stopping", "prefix value function"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~2.7x fewer denoising iterations", "metric_measures": "other", "benchmarks": ["MuJoCo"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "POGP learns a value at each denoising step and stops sampling once more steps would not help.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2409.00588"}
{"id": "2602.22818", "title": "LeRobot: An Open-Source Library for End-to-End Robot Learning", "authors": ["Remi Cadene", "Simon Aliberts", "Francesco Capuano", "Michel Aractingi", "Adil Zouitine", "Pepijn Kooijmans", "Jade Choghari", "Martino Russi", "Caroline Pascal", "Steven Palma", "Mustafa Shukor", "Jess Moss", "Alexander Soare", "Dana Aubakirova", "Quentin Lhoest", "Quentin Gallouédec", "Thomas Wolf"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.22818", "pdf_url": "https://arxiv.org/pdf/2602.22818", "code_url": "https://github.com/huggingface/lerobot", "clusters": ["C1", "C3", "C6"], "keywords": ["LeRobot", "open-source library", "asynchronous inference stack", "training stack", "robot learning library", "asynchronous inference", "open source"], "model_class": "other", "techniques": ["asynchronous inference", "tooling", "library", "asynchronous inference stack", "dataset streaming"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Hugging Face's end-to-end robot learning library covering middleware, datasets, policy implementations and a generalized asynchronous inference stack.", "relevance": 4, "relevance_reason": "Standard tool providing robot-client/policy-server async inference.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; arXiv API: abs:LeRobot; search query"}
{"id": "2606.05737", "title": "Let It Be Simple: One-Step Action Generation for Vision-Language-Action Models", "authors": ["Yitong Chen", "Shiduo Zhang", "Jingjing Gong", "Xipeng Qiu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.05737", "pdf_url": "https://arxiv.org/pdf/2606.05737", "code_url": "", "clusters": ["C1"], "keywords": ["one-step VLA", "high-noise training", "condition-target structure"], "model_class": "VLA", "techniques": ["one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus", "LIBERO-Pro", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Argues VLAs resemble image-to-text generation, so one-step flow decoding works with high-noise training and no special distillation.", "relevance": 4, "relevance_reason": "Insight on one-step VLA decoding.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching VLA one-step"}
{"id": "2609.36518", "title": "LIBERO-MAX: Do Robot Policies Adapt When the World Changes?", "authors": ["Yunbei Zhang", "Zijian Jin", "Yuanzhe Liu", "Janet Wang", "Xilun Zhang", "Yuyou Zhang", "Zhenyu Zhang", "Daoan Zhang", "Shuaicheng Niu", "Gen Li", "Jianfei Yang", "Jihun Hamm", "Ismini Lourentzou", "Weirui Ye", "Bo Liu", "Peter Stone", "Marco Pavone"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.36518", "pdf_url": "https://arxiv.org/pdf/2609.36518", "code_url": "", "clusters": ["C4"], "keywords": ["LIBERO-MAX", "mid-task changes", "paired evaluation", "query cadence"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-MAX"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "8,000 paired episodes with and without a mid-execution event (object moves, viewpoint shift, obstacle); 14 policies lose 11-26 points, and changing re-query cadence does not close the gap.", "relevance": 4, "relevance_reason": "Measures reactivity to dynamic change, the property most affected by latency and chunk length; paired protocol is methodologically useful.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2609.24350", "title": "LIBERO-VPro: Benchmarking Closed-Loop Visual Robustness of Robotic Foundation Models", "authors": ["Huiqiong Li", "Zhiting Mei", "Anirudha Majumdar", "Jingjing Chen", "Yu-Gang Jiang", "Bin Zhu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.24350", "pdf_url": "https://arxiv.org/pdf/2609.24350", "code_url": "", "clusters": ["C4"], "keywords": ["LIBERO-VPro", "visual robustness", "camera staleness", "closed-loop evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "Franka Research 3", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-VPro", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Closed-loop visual-robustness benchmark on LIBERO with degradation, stale-camera, source-consistency and scene-change perturbations; ~196k sim episodes plus 200 real rollouts show VLAs and world-action models are very sensitive to stale or missing frames.", "relevance": 4, "relevance_reason": "Camera staleness perturbation directly probes latency/asynchrony effects that efficiency methods introduce.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2606.08242", "title": "Light-WAM: Efficient World Action Models with State-Fusion Action Decoding", "authors": ["Ziang Li", "Dongzhou Cheng", "Yibin Wang", "Shiyue Wang", "Xiaoyang Xu", "Lingxuan Weng", "Juan Wang", "Jiaqi Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08242", "pdf_url": "https://arxiv.org/pdf/2606.08242", "code_url": "", "clusters": ["C1", "C7", "C8", "C9"], "keywords": ["lightweight WAM", "downsampled latent supervision", "single-pass action head", "single-pass action decoding"], "model_class": "world-action model", "techniques": ["compact video backbone", "multi-layer state fusion action expert", "downsampled future-video supervision", "state-fusion action expert"], "base_model": "", "params": "0.44B trainable", "hardware": "", "reported_metric": "72.03 ms inference latency, 4.1 GiB peak GPU memory", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboTwin 2.0"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Small WAM that uses future video only as downsampled training supervision and decodes action chunks in one forward pass from fused backbone states, giving ~72 ms latency and 4.1 GiB memory.", "relevance": 4, "relevance_reason": "Compact WAM with reported latency and memory.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2608.03701", "title": "LiLa-WAM: Lightweight Latent Reasoning World-Action Model for Robotic Manipulation", "authors": ["Fan Yang", "Yuting Su", "Xiaobo Wang", "Yuncheng You", "Fugui Fan", "Yuting Wu", "Minghui Wu", "Chenxu Zhao", "JiaHong Ning", "Peiguang Jing"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.03701", "pdf_url": "https://arxiv.org/pdf/2608.03701", "code_url": "", "clusters": ["C8"], "keywords": ["lightweight latent WAM", "single 24GB GPU training"], "model_class": "world-action model", "techniques": ["compact latent reasoning space", "end-to-end single-stage training"], "base_model": "", "params": "", "hardware": "single 24GB GPU (training)", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Compact latent-space WAM trainable end-to-end on one 24 GB GPU, avoiding pixel-level prediction and multi-stage latent construction.", "relevance": 4, "relevance_reason": "Low-cost WAM training and inference.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2603.16666 (citations)"}
{"id": "2609.28431", "title": "LiMA: Bridging Long-term Imagination to Real-time Dexterous Manipulation via Asynchronous Diffusion", "authors": ["Ning Chen", "Yankai Fu", "Junkai Zhao", "Qianpu Sun", "Guocai Yao", "Pengwei Wang", "Zhongyuan Wang", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28431", "pdf_url": "https://arxiv.org/pdf/2609.28431", "code_url": "", "clusters": ["C2", "C4", "C7", "C8"], "keywords": ["LiMA", "asynchronous dual-system diffusion", "dexterous", "asynchronous diffusion", "dexterous manipulation", "slow intent / fast reaction"], "model_class": "world-action model", "techniques": ["slow intent planner + fast refinement", "latent Schrödinger bridge coupling", "multi-scale slow-fast generative hierarchy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Asynchronous dual-system framework where a slow world-action model generates sparse long-horizon intent and a fast system refines dense high-frequency motion for dexterous manipulation.", "relevance": 4, "relevance_reason": "Slow-fast decoupling of WAM imagination.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"dual-system\" \"vision-language-action\" fast slow asynchronous; snowball from 2603.16666 (citations)"}
{"id": "2603.03380", "title": "LiteVLA-Edge: Quantized On-Device Multimodal Control for Embedded Robotics", "authors": ["Justin Williams", "Kishor Datta Gupta", "Roy George", "Mrinmoy Sarkar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.03380", "pdf_url": "https://arxiv.org/pdf/2603.03380", "code_url": "", "clusters": ["C1", "C2", "C4", "C5", "C9"], "keywords": ["GGUF", "llama.cpp", "ROS 2", "on-device", "LiteVLA-Edge", "GGUF 4-bit", "Jetson Orin", "4-bit", "embedded VLA", "quantized on-device control"], "model_class": "VLA", "techniques": ["4-bit post-training quantization", "post-training 4-bit quantization", "llama.cpp", "4-bit GGUF quantization", "llama.cpp GPU inference", "quantization", "deployment pipeline"], "base_model": "", "params": "", "hardware": "Jetson Orin-class", "reported_metric": "150.5 ms mean end-to-end latency (~6.6 Hz)", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Systems recipe: fine-tune a compact VLA, quantize to 4-bit GGUF and run it through llama.cpp on a Jetson inside a ROS 2 loop, establishing timing feasibility rather than task success.", "relevance": 4, "relevance_reason": "Onboard quantized VLA deployment baseline; no task-level eval.", "foundational_technique": false, "verified": true, "found_via": "search: quantization robot policy; search query: VLA quantization; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2605.00884", "title": "LiteVLA-H: Dual-Rate Vision-Language-Action Inference for Onboard Aerial Guidance and Semantic Perception", "authors": ["Justin williams", "Kishor Datta Gupta", "Roy George", "Mrinmoy Sarkar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.00884", "pdf_url": "https://arxiv.org/pdf/2605.00884", "code_url": "", "clusters": ["C1", "C3", "C5", "C6", "C9"], "keywords": ["aerial VLA", "dual-rate", "prefill-dominated latency", "UAV", "Jetson AGX Orin", "onboard aerial VLA", "dual-rate inference", "dual-rate VLA", "aerial", "onboard"], "model_class": "VLA", "techniques": ["dual-rate scheduling", "small backbone", "compact VLA", "knowledge-preserving fine-tuning", "dual-rate inference"], "base_model": "", "params": "256M", "hardware": "NVIDIA Jetson AGX Orin", "reported_metric": "reactive action tokens at 50.65 ms (19.74 Hz); semantic outputs at 6.08-6.67 Hz | action tokens at 50.65 ms (19.74 Hz); semantic outputs 149.90-164.57 ms (6.08-6.67 Hz)", "metric_measures": "open-loop chunk rate", "benchmarks": ["aerial guidance"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "A 256M-parameter aerial VLA on Jetson AGX Orin that alternates a fast short-token guidance mode with a slower semantic narration mode, motivated by the finding that multimodal prefill dominates edge latency.", "relevance": 4, "relevance_reason": "Onboard alternative where comms are constrained; notes prefill dominates latency.", "foundational_technique": false, "verified": true, "found_via": "search query: Jetson VLA; arXiv: vision-language-action AND Jetson/edge/onboard; snowball from 2602.18397; venue sweep: arXiv API VLA AND Jetson"}
{"id": "2605.09948", "title": "LoopVLA: Learning Sufficiency in Recurrent Refinement for Vision-Language-Action Models", "authors": ["Boyang Shen", "Kaixiang Yang", "Hao Wang", "Qiuyu Yu", "Qiang Xie", "Qiang Li", "Zhiwei Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.09948", "pdf_url": "https://arxiv.org/pdf/2605.09948", "code_url": "", "clusters": ["C1", "C2", "C9"], "keywords": ["recurrent refinement", "weight sharing", "sufficiency estimation", "shared block", "adaptive depth", "sufficiency"], "model_class": "VLA", "techniques": ["recurrent depth", "adaptive computation", "weight sharing", "adaptive depth", "learned halting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "45% fewer parameters; up to 1.7x throughput", "metric_measures": "throughput", "benchmarks": ["LIBERO", "LIBERO-Plus", "VLA-Arena"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Applies a shared transformer block recurrently and learns when the representation is sufficient to stop, trading depth for adaptive iterations.", "relevance": 4, "relevance_reason": "Weight sharing + learned early exit for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA early exit; search: early exit robot policy; venue sweep: arXiv API VLA AND (MoE OR early exit)"}
{"id": "2609.34375", "title": "LRC-JEPA: Disentangling Dynamics and Residual Context for Efficient World Models", "authors": ["Luzhe Huang", "Lei Chu", "Jingyi Liang", "Yuhuan Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34375", "pdf_url": "https://arxiv.org/pdf/2609.34375", "code_url": "", "clusters": ["C8"], "keywords": ["compact JEPA", "residual context", "disentanglement"], "model_class": "world model", "techniques": ["compact predictive latent + residual context embeddings", "only compact latent propagated for planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["4 simulated control envs", "Bridge-v2"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Lightweight JEPA world model that routes appearance into a separate residual context so only a small dynamic latent is rolled out for planning, matching much larger pretrained models.", "relevance": 4, "relevance_reason": "Small latent world model for efficient planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.03593", "title": "Making Embodied AI Reliable: A Community Agenda from Testing to Formal Verification", "authors": ["Xi Zheng", "Dulanga Weerakoon", "Yintong Huo", "Teresa Yeo", "Guy Van Den Broeck", "Vijay Ganesh", "Daniel Neider", "Biplav Srivastava", "Ivan Ruchkin", "Archan Misra", "Corina Pasareanu"], "year": 2026, "venue": "arXiv", "type": "paper", "url": "https://arxiv.org/abs/2606.03593", "pdf_url": "https://arxiv.org/pdf/2606.03593", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["embodied AI reliability", "position"], "model_class": "other", "techniques": ["testing", "formal verification", "runtime assurance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Community agenda from the AAAI-26 bridge program arguing embodied-AI reliability needs integrated scenario testing, compositional verification and adaptive runtime assurance.", "relevance": 4, "relevance_reason": "Maps open problems in verifying embodied foundation-model systems.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2608.21407", "title": "Mamba-based Selective State Space Modeling Improves the Accuracy-Complexity Tradeoff of SmolVLA Vision-Language-Action Experts", "authors": ["Farida Mohsen", "Thowayba Elkaffash", "Mohammad Reza Chalak Qazani", "Mohamed Mabrok", "Nader Meskin", "Ali Safa"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.21407", "pdf_url": "https://arxiv.org/pdf/2608.21407", "code_url": "", "clusters": ["C1"], "keywords": ["Mamba action expert", "SmolVLA", "execution horizon"], "model_class": "VLA", "techniques": ["state space model"], "base_model": "SmolVLA", "params": "", "hardware": "", "reported_metric": "24% fewer parameters at matched success (N=1)", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces SmolVLA's attention action expert with Mamba, showing better success retention with long open-loop horizons and fewer parameters.", "relevance": 4, "relevance_reason": "SSM action expert.", "foundational_technique": false, "verified": true, "found_via": "search query: Mamba robot policy"}
{"id": "10.1109/ccnc65079.2026.11366568", "title": "MambaVLA: A Scalable and Efficient Vision-Language-Action Model with State Space Architecture", "authors": ["Sai Navaneet Peddapalli", "Manisha Lingala", "Sangmoon Lee", "Ju H. Park"], "year": 2026, "venue": "Consumer Communications and Networking Conference", "type": "paper", "url": "https://doi.org/10.1109/CCNC65079.2026.11366568", "pdf_url": "", "code_url": "", "clusters": ["C1"], "keywords": ["Mamba", "efficient VLA"], "model_class": "VLA", "techniques": ["state space model", "linear-time sequence modeling", "INT4 LLM"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA built on a Mamba state-space backbone with flow-matching actions, matching transformer VLAs at lower compute and faster inference.", "relevance": 4, "relevance_reason": "Efficient backbone architecture for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2605.29766", "title": "MARS Policy: Multimodality Only When It Matters", "authors": ["Jin-Dou Jia", "Tuo An", "Yuxuan Hu", "Gen Li", "Jing-Liang Li", "Bo-Han Hou", "Xiangyu Chen", "Jiaqiang Bai", "Bofan Lyu", "Jian-Fei Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.29766", "pdf_url": "https://arxiv.org/pdf/2605.29766", "code_url": "", "clusters": ["C1"], "keywords": ["multimodality", "inference latency"], "model_class": "diffusion/flow policy", "techniques": ["adaptive stochasticity", "deterministic fallback"], "base_model": "", "params": "", "hardware": "", "reported_metric": "83.20% inference latency reduction (real-world)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Invokes stochastic multimodal generation only in phases that need it and uses deterministic prediction elsewhere, cutting inference cost.", "relevance": 4, "relevance_reason": "Directly reduces generative-policy inference latency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.01469", "title": "Mean-Flow based One-Step Vision-Language-Action", "authors": ["Yang Chen", "Xiaoguang Ma", "Bin Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.01469", "pdf_url": "https://arxiv.org/pdf/2603.01469", "code_url": "", "clusters": ["C1"], "keywords": ["MeanFlow", "one-step VLA"], "model_class": "VLA", "techniques": ["mean flow", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "8.7x faster than SmolVLA, 83.9x faster than Diffusion Policy (generation speed)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Swaps the flow-matching action head for a MeanFlow average-velocity formulation enabling one-step generation in a VLA.", "relevance": 4, "relevance_reason": "One-step VLA generation.", "foundational_technique": false, "verified": true, "found_via": "search query: one-step diffusion policy"}
{"id": "2609.27167", "title": "Median Temporal Ensembling: Training-Free Robust Aggregation for Action-Chunked Visuomotor Policies", "authors": ["Yuhang Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.27167", "pdf_url": "https://arxiv.org/pdf/2609.27167", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["temporal ensembling", "median aggregation", "robustness", "median temporal ensembling", "robust aggregation", "adversarial corruption", "blank frames"], "model_class": "other", "techniques": ["temporal ensembling", "coordinate-wise median ensembling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows exponential-mean temporal ensembling has breakdown point zero and replaces it with a coordinate-wise median over overlapping chunk predictions, bounding damage from corrupted predictions or blank camera frames with no retraining.", "relevance": 4, "relevance_reason": "Robustness of the chunk aggregation layer at runtime.", "foundational_technique": false, "verified": true, "found_via": "search query: temporal ensembling; arXiv search: temporal ensembling action chunking"}
{"id": "2605.30571", "title": "Memory-Bound but Not Bandwidth-Limited: The Physical AI Inference Gap in Batch-1 LLM Decode", "authors": ["Josef Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30571", "pdf_url": "https://arxiv.org/pdf/2605.30571", "code_url": "", "clusters": ["C5"], "keywords": ["batch-1 decode", "HBM bandwidth utilization", "CUDA Graphs", "physical AI"], "model_class": "other", "techniques": ["bandwidth-utilization characterization", "CUDA graphs"], "base_model": "", "params": "", "hardware": "H100 SXM5, A100-80GB, L40S, L4", "reported_metric": "Qwen-2.5-7B ctx 2048: L4 ~81% of analytic memory floor vs H100 ~27%; CUDA Graphs 1.259x on H100", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows batch-1 decode typical of physical AI is memory-dominated but achieved bandwidth fraction falls on faster GPUs because of launch/overhead terms, which CUDA graphs partially recover.", "relevance": 4, "relevance_reason": "Refines the 'memory-bandwidth bottleneck' account for robot-style batch-1 inference.", "foundational_technique": true, "verified": true, "found_via": "arXiv: memory-bound AND (robot OR VLA OR embodied)"}
{"id": "2606.20562", "title": "MemoryWAM: Efficient World Action Modeling with Persistent Memory", "authors": ["Sizhe Yang", "Juncheng Mu", "Tianming Wei", "Chenhao Lu", "Xiaofan Li", "Linning Xu", "Zhengrong Xue", "Zhecheng Yuan", "Dahua Lin", "Jiangmiao Pang", "Huazhe Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20562", "pdf_url": "https://arxiv.org/pdf/2606.20562", "code_url": "", "clusters": ["C1", "C7", "C8", "C9"], "keywords": ["persistent memory", "gist tokens", "long-horizon WAM", "WAM"], "model_class": "world-action model", "techniques": ["hybrid memory (recent + anchor frames + gist tokens)", "compressed long-range attention", "persistent memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["memory-dependent manipulation (sim + real)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Adds bounded-cost persistent memory to a WAM by mixing recent frames, event-boundary anchors and compact summary tokens, so non-Markovian tasks are handled without the latency/memory growth of full history.", "relevance": 4, "relevance_reason": "Memory-efficiency for long-context world-action models.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2605.29360", "title": "MiraBench: Evaluating Action-Conditioned Reliability in Robotic World Models", "authors": ["Tianzhuo Yang", "Zihan Shen", "Zirui Mi", "Zhaoyi Zhang", "Jiayi Zhou", "Jiaming Ji", "Juntao Dai", "Jiawei Chen", "Boyuan Chen", "Yaodong Yang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2605.29360", "pdf_url": "https://arxiv.org/pdf/2605.29360", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["action-conditioned reliability", "optimism bias", "world model benchmark"], "model_class": "world model", "techniques": ["hierarchical evaluation: physics, action fidelity, optimism bias"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MiraBench (16k human judgments)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Benchmark testing whether robot world models follow commanded actions and correctly predict failure under failure-inducing actions, rather than just looking realistic.", "relevance": 4, "relevance_reason": "Measures optimism bias that undermines world-model safety monitoring.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2605.02881", "title": "MolmoAct2: Action Reasoning Models for Real-world Deployment", "authors": ["Haoquan Fang", "Jiafei Duan", "Donovan Clay", "Sam Wang", "Shuo Liu", "Weikai Huang", "Xiang Fan", "Wei-Chuan Tsai", "Shirui Chen", "Yi Ru Wang", "Shanli Xing", "Jaemin Cho", "Jae Sung Park", "Ainaz Eftekhar", "Peter Sushko", "Karen Farley", "Angad Wadhwa", "Cole Harrison", "Winson Han", "Ying-Chun Lee", "Eli VanderBilt", "Rose Hendrix", "Suveen Ellawela", "Lucas Ngoo", "Joyce Chai", "Zhongzheng Ren", "Ali Farhadi", "Dieter Fox", "Ranjay Krishna"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2605.02881", "pdf_url": "https://arxiv.org/pdf/2605.02881", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["open VLA", "action tokenizer", "flow expert", "adaptive-depth reasoning", "MolmoAct2", "open action reasoning", "deployment"], "model_class": "VLA", "techniques": ["action tokenization", "action expert", "selective re-prediction", "reasoning latency reduction", "open weights on cheap hardware"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "AI2's open action-reasoning VLA with an open FAST-style tokenizer, a flow-matching expert grafted via per-layer KV conditioning, and a reasoning variant that recomputes depth tokens only for changed regions.", "relevance": 4, "relevance_reason": "Deployment-oriented open VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache; venue sweep: arXiv API VLA AND benchmark AND latency AND hardware"}
{"id": "2609.20709", "title": "MoWAM: Explicit Future Motion Prediction for Efficient World Action Models", "authors": ["Jiayu Wang", "Bin Zhu", "Yue Yu", "Jingjing Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20709", "pdf_url": "https://arxiv.org/pdf/2609.20709", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["future motion prediction", "verifier-based test-time scaling", "efficient WAM"], "model_class": "world-action model", "techniques": ["explicit compact motion prediction replaces video", "candidate sampling with progress verifier", "explicit motion instead of video"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Swaps future video generation for prediction of compact future robot motion, which is cheap enough to sample several motion-action candidates and pick one with a task-progress verifier.", "relevance": 4, "relevance_reason": "Compact future representation enabling cheap test-time planning.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2607.06216", "title": "MoWorld: A Flash World Model", "authors": ["Team Moxin", "De-Yi Ji", "Tianrun Chen", "Xin Zhang", "Jia-Le Yang", "Qi Zhu", "An Zhao", "Zi-Hao Xie", "Han Wang", "Xuan-Yi Liu", "Yi-Xiang Zhou", "Pei-Wang Liu", "Yi Tan", "Cheng Chen", "Da-Yi Zhu", "Ming-Yu Wei", "Han-Jie Xu", "Jun Liao", "Si-Qi Li", "Ling-Yun Lu", "Hong Fang", "Hong-Ming Tan", "You-Jiang Zhu", "Tai-Yu Zhang", "Zejian Li", "Chao Ding", "Lan-Yun Zhu", "Yun-He Pan", "Lingyun Sun"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.06216", "pdf_url": "https://arxiv.org/pdf/2607.06216", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["flash world model", "real-time"], "model_class": "world model", "techniques": ["mixed-precision", "parallel inference", "NPU deployment"], "base_model": "", "params": "", "hardware": "NPU", "reported_metric": "up to 50 FPS on NPU; 30-50% of inference cost of existing world models", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cost-effective world model with a mixed-precision parallel inference stack, the first real-time interactive world model on an NPU.", "relevance": 4, "relevance_reason": "Efficient onboard world model", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2501.03575, 2506.08009"}
{"id": "2603.07799", "title": "MWM: Mobile World Models for Action-Conditioned Consistent Prediction", "authors": ["Han Yan", "Zi-Shang Xiang", "Ze-Yu Zhang", "Hao Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.07799", "pdf_url": "https://arxiv.org/pdf/2603.07799", "code_url": "", "clusters": ["C8"], "keywords": ["navigation world model", "action-conditioned consistency"], "model_class": "world model", "techniques": ["few-step diffusion distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["benchmark", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Mobile navigation world model with consistency post-training and inference-consistent few-step distillation.", "relevance": 4, "relevance_reason": "Efficient few-step world model for deployment", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2301.04104"}
{"id": "2607.08639", "title": "Native Video-Action Pretraining for Generalizable Robot Control", "authors": ["Qihang Zhang", "Lin Li", "Luyao Zhang", "Shuai Yang", "Yiming Luo", "Shuaiting Li", "Ruilin Wang", "Junke Wang", "Jiahao Shao", "Gangwei Xu", "Jiaming Zhou", "Yishu Shen", "Yudong Jin", "Fangyi Xu", "Shuailei Ma", "Jiaqi Liao", "Guanxing Lu", "Zifan Shi", "Yongkun Wen", "Yujie Zhao", "Weixuan Tang", "Xinyang Wang", "Chaojian Li", "Jiapeng Zhu", "Ka Leong Cheng", "Nan Xue", "Xing Zhu", "Yujun Shen", "Yinghao Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.08639", "pdf_url": "https://arxiv.org/pdf/2607.08639", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["LingBot-VA 2.0", "sparse MoE", "causal pretraining", "asynchronous inference"], "model_class": "world-action model", "techniques": ["semantic visual-action tokenizer", "sparse MoE backbone", "async latent prediction with re-grounding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Video-action foundation model built from scratch for embodiment, using a sparse MoE backbone for capacity without latency and asynchronous latent prediction re-grounded on the latest observation.", "relevance": 4, "relevance_reason": "Designed for high-frequency WAM inference.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2607.06678", "title": "NativeMEM: Native Memory Compression for Long-Horizon Robotic Manipulation", "authors": ["Ziye Wang", "Modi Shi", "Chaojun Ni", "Jiazhi Yang", "Mengdi Li", "Zhizhong Su", "Tianwei Lin", "Hongyang Li", "Mo-Di Shi", "Jia-Zhi Yang", "Meng-Di Li", "Zhi-Zhong Su", "Tian-Wei Lin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.06678", "pdf_url": "https://arxiv.org/pdf/2607.06678", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["long-horizon memory", "memory", "long-horizon"], "model_class": "VLA", "techniques": ["memory token compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Reuses the VLA's own vision encoder to compress each past frame per view into one memory token appended to the input, giving long history with negligible latency.", "relevance": 4, "relevance_reason": "Memory compression for long-horizon VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]; snowball round 2: cites 2410.24164, 2502.19645"}
{"id": "2608.16503", "title": "NebulaVLA: A Dual-Frequency Vision-Language-Action Model With Guide Action for Robotic Manipulation", "authors": ["Cong Zhao", "Shuai Tian", "Xu Zhang", "Baocheng Ni", "Xinguo Song", "Xueying Sun", "Shu Jiang", "Shouchang Yang", "Bo Tang", "Jin Deng", "Ge Zhu", "YongCheng Wang", "Jin Xu", "Ri Yang", "Congyu Zhao", "Bao-Cheng Ni", "Xin-Guo Song", "Shou-Chang Yang", "Bosch Tang", "Yong-Chen Wang", "Richa Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.16503", "pdf_url": "https://arxiv.org/pdf/2608.16503", "code_url": "", "clusters": ["C1"], "keywords": ["NebulaVLA", "dual-system", "asynchronous inference"], "model_class": "VLA", "techniques": ["dual-frequency architecture", "asynchronous inference", "dual-frequency asynchronous"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~2.7x faster action generation; 85.5% avg success LIBERO-Plus", "metric_measures": "relative latency", "benchmarks": ["LIBERO-Plus"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Asynchronous dual-frequency VLA separating slow semantic reasoning from fast action control, with a unified action representation and mask-based smoothness guidance.", "relevance": 4, "relevance_reason": "Dual-rate architecture with reported generation speedup.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]; snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2606.11187", "title": "Next Forcing: Causal World Modeling with Multi-Chunk Prediction", "authors": ["Gangwei Xu", "Qihang Zhang", "Jiaming Zhou", "Xing Zhu", "Yujun Shen", "Xin Yang", "Yinghao Xu", "Jia-Ming Zhou", "Xing-Hui Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.11187", "pdf_url": "https://arxiv.org/pdf/2606.11187", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["Next Forcing", "world action model", "autoregressive video", "multi-token prediction"], "model_class": "world-action model", "techniques": ["multi-chunk prediction", "parallel chunk decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2x inference acceleration; 2.3x faster convergence", "metric_measures": "relative latency", "benchmarks": ["RoboTwin", "PhyWorld"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds lightweight multi-chunk prediction heads to autoregressive world-action models, speeding convergence and allowing the next video chunk to be predicted in parallel at inference.", "relevance": 4, "relevance_reason": "Directly addresses slow WAM inference/training.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2609.09250", "title": "No Free Checker: A Survey of Verifiers for Robot Policies", "authors": ["Yang Wan", "Xihang Yue", "Zhirui Liu", "Ziyuan Chu", "Shuxun Wang", "Yuhan Chen", "Xiaonan Jiang", "Xukun Zhu", "Yubo Dong", "Linchao Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.09250", "pdf_url": "https://arxiv.org/pdf/2609.09250", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["verifiers survey", "availability vs credibility", "runtime monitors", "reward models"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of ~150 robot-policy verifiers (success detectors, reward models, runtime monitors, safety filters, temporal logic) compared by verdict cost/latency/density versus credibility.", "relevance": 4, "relevance_reason": "Organizes monitors by timeliness/cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" \"vision-language-action\"; venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2602.04215", "title": "OAT: Ordered Action Tokenization", "authors": ["Chaoqi Liu", "Xiaoshen Han", "Jiawei Gao", "Yue Zhao", "Haonan Chen", "Yilun Du"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.04215", "pdf_url": "https://arxiv.org/pdf/2602.04215", "code_url": "", "clusters": ["C1"], "keywords": ["ordered tokens", "FSQ", "anytime decoding"], "model_class": "VLA", "techniques": ["learned action tokenizer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learned tokenizer producing causally ordered action tokens, so prefixes decode to coarse actions and inference cost can be traded for fidelity.", "relevance": 4, "relevance_reason": "Action tokenization enabling anytime inference.", "foundational_technique": false, "verified": true, "found_via": "search query: action tokenization"}
{"id": "msr-offloaded-inference-blog-2026", "title": "Offloaded inference for real-world physical AI robotics", "authors": ["Ganesh Ananthanarayanan", "Matthew Balkwill", "Xenofon Foukas", "Sanjeev Mehrotra", "Bozidar Radunovic"], "year": 2026, "venue": "Microsoft Research blog", "type": "blog", "url": "https://www.microsoft.com/en-us/research/blog/offloaded-inference-for-real-world-physical-ai-robotics/", "pdf_url": "", "code_url": "https://github.com/microsoft/physical-ai-toolchain", "clusters": ["C6"], "keywords": ["Physical AI Toolchain", "Kubernetes", "offloaded inference", "battery life"], "model_class": "VLA", "techniques": ["containerized offload", "robot-edge-cloud orchestration"], "base_model": "", "params": "", "hardware": "Jetson Nano/Orin AGX/Thor, Raspberry Pi 5 onboard; Jetson Thor, A100 offload", "reported_metric": "VLA accuracy drop ~50% on smaller onboard GPUs; larger onboard GPU drains battery up to 160% faster", "metric_measures": "other", "benchmarks": ["Mobile ALOHA", "Stretch-3", "SO-101", "UR10e"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Announces open-source offloaded-inference support in Microsoft's Physical AI Toolchain, summarizing measurements that onboard GPUs either lack capacity or drain batteries while offloading restores accuracy.", "relevance": 4, "relevance_reason": "Industry tooling and measurements for robot inference offloading (companion to 2603.18284).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.36413", "title": "One from Infinity: Actualizing Futures from Pretrained World Models into Robot Actions", "authors": ["Bang Du", "Yichen Xie", "Shuqi Zhao", "Yuxin Chen", "Menglin Wu", "Masayoshi Tomizuka", "Yi-Chen Xie", "Shu-Qi Zhao", "Yu-Xin Chen", "Meng-Ling Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36413", "pdf_url": "https://arxiv.org/pdf/2609.36413", "code_url": "", "clusters": ["C1", "C3", "C8"], "keywords": ["RoboActualizer", "frozen world model", "parameter-efficient", "low latency"], "model_class": "world-action model", "techniques": ["frozen world model", "tiny actualizer", "flow matching"], "base_model": "", "params": "~60M trainable", "hardware": "", "reported_metric": "39 ms latency", "metric_measures": "other", "benchmarks": ["LIBERO", "LIBERO-Plus", "RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RoboActualizer keeps a pretrained video world model frozen and trains only a tiny (~60M) DiT-based actualizer to select and realize task-conditioned futures, needing up to 100x fewer trainable parameters and a single GPU.", "relevance": 4, "relevance_reason": "direct: parameter-efficient adaptation of world models with low latency", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight); snowball round 2: cites 2306.03310, 2406.09246"}
{"id": "2605.07931", "title": "One Token Per Frame: Reconsidering Visual Bandwidth in World Models for VLA Policy", "authors": ["Zuojin Tang", "Shengchao Yuan", "Xiaoxin Bai", "Zhiyuan Jing", "De Ma", "Gang Pan", "Bin Liu", "Zhi Jing", "Bing Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.07931", "pdf_url": "https://arxiv.org/pdf/2605.07931", "code_url": "", "clusters": ["C1", "C2", "C8"], "keywords": ["OneWM-VLA", "visual bandwidth", "compact latent"], "model_class": "VLA", "techniques": ["attention pooling", "compact latent prediction", "token compression", "latent future prediction"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MetaWorld MT50", "LIBERO", "real robot (Piper)", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "OneWM-VLA adds visual future prediction to a VLA using a single pooled token per view per future step, keeping the generative sequence compact while improving control.", "relevance": 4, "relevance_reason": "direct: compact world-model tokens for efficient VLA", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=0]; snowball round 2: cites 2303.04137, 2304.13705; cited by 2608.09771"}
{"id": "2603.12480", "title": "One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies", "authors": ["Shaolong Li", "Lichao Sun", "Yong-Chao Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.12480", "pdf_url": "https://arxiv.org/pdf/2603.12480", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["one-step flow", "self-consistency", "warm start"], "model_class": "diffusion/flow policy", "techniques": ["self-distillation", "one-step generation"], "base_model": "flow policies; pi0.5", "params": "", "hardware": "", "reported_metric": ">100x faster action generation than 100-step baselines", "metric_measures": "relative latency", "benchmarks": ["RoboTwin 2.0", "56 simulated tasks"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains single-step flow policies from scratch via self-consistency and self-guided regularization, even improving on 10-step pi0.5.", "relevance": 4, "relevance_reason": "Teacher-free step distillation.", "foundational_technique": false, "verified": true, "found_via": "search: one-step policy distillation robot"}
{"id": "s2-online-failure-detection-time-aware-metrics-rtcsa-2026", "title": "Online Failure Detection for Robot Manipulation with Time-Aware Metrics and Efficient Models", "authors": ["Tian-Rui Ma", "Min-Hao Fan", "Wei-Chen Liu"], "year": 2026, "venue": "IEEE RTCSA 2026", "type": "paper", "url": "https://www.semanticscholar.org/paper/7c1ecaa5b9f132c53ac795154a86142f2037844c", "pdf_url": "", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["AUTC", "time-aware metric", "online failure detection", "efficient detector"], "model_class": "VLA", "techniques": ["transformer failure detector", "time-aware evaluation metric"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proposes AUTC, a scalar metric integrating failure-detection accuracy over timeliness, and an efficient transformer sequential failure detector for VLA manipulation.", "relevance": 4, "relevance_reason": "Timeliness-aware evaluation of runtime monitors at an embedded real-time venue.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2605.11678", "title": "OOM-Free Alpamayo via CPU-GPU Memory Swapping for Vision-Language-Action Models", "authors": ["Seungwoo Roh", "Huiyeong Kim", "Jong-Chan Kim", "Seung-Wook. Roh", "H. Kim"], "year": 2026, "venue": "IEEE RTCSA 2026", "type": "paper", "url": "https://arxiv.org/abs/2605.11678", "pdf_url": "https://arxiv.org/pdf/2605.11678", "code_url": "", "clusters": ["C1", "C3", "C5", "C8"], "keywords": ["memory swapping", "layer offloading", "driving VLA", "CPU-GPU memory swapping", "VRAM-constrained", "Alpamayo"], "model_class": "VLA", "techniques": ["CPU-GPU offloading", "pipelining", "demand layering", "transfer-compute overlap", "GPU-resident layer selection", "performance prediction model"], "base_model": "Alpamayo-R1-10B", "params": "10B", "hardware": "RTX 5070 Ti (16 GB)", "reported_metric": "up to 3.55x speedup over baseline offloading | up to 3.55x speedup over conventional memory offloading at BF16", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Runs a 21.5 GB driving VLA on a 16 GB GPU by streaming layers from CPU memory with pipelined transfers and a profiled choice of which layers stay GPU-resident.", "relevance": 4, "relevance_reason": "Memory-capacity-aware system design for VLA on commodity hardware.", "foundational_technique": false, "verified": true, "found_via": "search query: efficient VLA; web: profiling OpenVLA inference latency edge GPU memory-bound"}
{"id": "github-openvla-compat-vla-lite", "title": "openvla-compat (vla-lite): OpenVLA-7B on modern transformers and on an 8 GB GPU", "authors": ["jashshah999"], "year": 2026, "venue": "GitHub", "type": "repo", "url": "https://github.com/jashshah999/vla-lite", "pdf_url": "", "code_url": "https://github.com/jashshah999/vla-lite", "clusters": ["C2", "C4"], "keywords": ["OpenVLA", "bitsandbytes", "nf4", "int8", "silent vision bypass", "fidelity"], "model_class": "VLA", "techniques": ["weight-only quantization (bitsandbytes int8/nf4)"], "base_model": "openvla-7b-finetuned-libero-spatial", "params": "7B", "hardware": "RTX 5000 Ada Laptop 16 GB", "reported_metric": "peak VRAM bf16 15.29 GB / int8 8.05 GB / nf4 4.60 GB; p50 latency 260 / 422 / 245 ms; nf4 flips gripper bin on 2/16 frames, int8 1/16", "metric_measures": "relative memory", "benchmarks": ["16 LIBERO frames, open-loop action fidelity (no closed-loop SR)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Measures bitsandbytes int8/nf4 OpenVLA on a 16 GB laptop GPU: 3.3x memory cut but no speedup (int8 slower), and 4-bit flips discrete gripper decisions; also documents a transformers>=4.50 bug that silently ignores the image.", "relevance": 4, "relevance_reason": "Concrete memory/latency/fidelity numbers for fitting OpenVLA-7B under 16 GB; explicitly lacks LIBERO closed-loop SR, highlighting the gap.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub issues repo:openvla/openvla quantization (#311)"}
{"id": "2607.06262", "title": "Optimal Transport Q-Learning for Flow Policy Steering and Acceleration", "authors": ["Andreas Sochopoulos", "Esmeralda S. Whitammer", "Nikolaos Tsagkas", "João Moura", "Michael Gienger", "Sethu Vijayakumar", "S. Vijayakumar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.06262", "pdf_url": "https://arxiv.org/pdf/2607.06262", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["OTQL", "flow policy acceleration", "Q-learning"], "model_class": "diffusion/flow policy", "techniques": ["optimal transport flow matching", "advantage weighting", "RL post-training", "step reduction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "VLA success 38% -> 76%; 70% fewer inference steps per action", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Advantage-weighted conditional optimal-transport flow matching that RL-fine-tunes and simultaneously shortens flow policies within 50-60 episodes without separate distillation.", "relevance": 4, "relevance_reason": "Accelerates flow-policy inference via fine-tuning with a small interaction budget.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=0]; snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2609.10550", "title": "Optimizing AI Inference Across the Deployment Stack", "authors": ["Tejinder Singh", "John Pflueger", "Jeebak Mitra", "Robert Lincourt", "Mitchell Markow", "Bhavesh A. Patel"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.10550", "pdf_url": "https://arxiv.org/pdf/2609.10550", "code_url": "", "clusters": ["C2", "C4", "C5"], "keywords": ["cross-layer interactions", "evidence protocol", "roofline", "queuing", "Jetson AGX Orin"], "model_class": "other", "techniques": ["quantization", "pruning", "distillation", "graph fusion", "dynamic batching"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin; A100; H100", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey-style analysis arguing deployment outcomes come from interactions among compression, compiler and serving layers, with a reporting protocol (hardware, versions, batch semantics, thermal state) to make benchmarks comparable.", "relevance": 4, "relevance_reason": "Speaks directly to both composition (Q1) and comparable reporting (Q3).", "foundational_technique": false, "verified": true, "found_via": "gap check: Q3: arXiv abs:roofline AND abs:Jetson"}
{"id": "2602.15882", "title": "PACT-WAM: Predicting Actions and Visual Foresight with Compact Temporal Encoding for Robot Manipulation", "authors": ["Yushan Liu", "Jingjing Fan", "Shoujie Li", "Yifan Xie", "Xiao-Ping Zhang", "Wenbo Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.15882", "pdf_url": "https://arxiv.org/pdf/2602.15882", "code_url": "", "clusters": ["C8"], "keywords": ["compact temporal encoding", "hierarchical history", "proposal review"], "model_class": "world-action model", "techniques": ["coarse-to-fine history tokens (75% fewer)", "joint action/visual flow sampling", "VLM proposal rejection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Keeps 16 frames of history with 75% fewer tokens via hierarchical encoding and jointly samples actions with visual forecasts that a VLM can use to reject bad proposals.", "relevance": 4, "relevance_reason": "Token-efficient WAM with forecast-based rejection.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2602.00475", "title": "Parallel Stochastic Gradient-Based Planning for World Models", "authors": ["Michael Psenka", "Michael Rabbat", "Aditi Krishnapriyan", "Yann LeCun", "Amir Bar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.00475", "pdf_url": "https://arxiv.org/pdf/2602.00475", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["gradient-based planning", "lifted trajectory optimization", "parallel planning", "world model"], "model_class": "world model", "techniques": ["lifted parallel optimization", "action-only gradients", "stochastic state exploration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "Push-T H=40: 59.0% success, median 8.5 s planning", "metric_measures": "other", "benchmarks": ["Push-T"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "GRASP optimizes states and actions jointly in parallel over time and uses only action gradients, making long-horizon planning with learned world models faster and more robust than CEM.", "relevance": 4, "relevance_reason": "Efficient planning with world models; featured on BAIR blog.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: BAIR blog"}
{"id": "2609.18732", "title": "PASSAGE: Scaling Scene-Aligned Motion Learning for Perceptive Humanoid Traversal in Cluttered Environments", "authors": ["Yuxuan Ma", "Zi-Cheng Zeng", "Chun-Li Peng", "Zhou-Jian Li", "Ze-Tong Zhao", "Zhikai Zhang", "Yun-Rui Lian", "Han Xue", "Sikai Liang", "Wei-Yi Zhu", "Mu-Lin Chen", "Cheng-Hua Lin", "Jia-Yu Zeng", "Yan-Wei An", "Song Zhang", "Jia-Yuan Gu", "Ji-Long Wang", "Jing-Bo Wang", "He Wang", "Li Yi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18732", "pdf_url": "https://arxiv.org/pdf/2609.18732", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["humanoid", "onboard"], "model_class": "diffusion/flow policy", "techniques": ["flow-matching planner", "real-time chunking"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin", "reported_metric": "6.25 Hz planning, 50 Hz control", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "PASSAGE runs a flow-matching motion planner and whole-body tracker fully onboard a humanoid for traversal in clutter.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2505.22094, 2511.07820"}
{"id": "2607.18236", "title": "Patch Policy: Efficient Embodied Control via Dense Visual Representations", "authors": ["Gaoyue Zhou", "Zichen Jeff Cui", "Ada Langford", "Bowen Tan", "Yann LeCun", "Lerrel Pinto"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.18236", "pdf_url": "https://arxiv.org/pdf/2607.18236", "code_url": "", "clusters": ["C1"], "keywords": ["efficient policy", "lightweight policy", "high-frequency control"], "model_class": "other", "techniques": ["dense patch features", "block-causal attention", "dense patch tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~0.7% of OpenVLA-OFT parameters; +18% over it | ~0.7% of OpenVLA-OFT parameters while outperforming it by 18%", "metric_measures": "relative memory", "benchmarks": ["4 sim suites", "3 real suites"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Block-causal transformer policy that consumes dense pretrained ViT patch tokens directly without a VLM backbone, beating OpenVLA-OFT with about 0.7% of its parameters.", "relevance": 4, "relevance_reason": "Lightweight alternative to VLAs with large parameter savings.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]; snowball round 2: cites 2205.14135, 2303.04137"}
{"id": "2606.16690", "title": "PATCH: Action-Chunk-Conditioned Latent Patch Innovation Monitoring for Robot Manipulation", "authors": ["Yanan Zhou", "Ranpeng Qiu", "Yincong Chen", "Jiajie Cui", "Weiming Zhi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.16690", "pdf_url": "https://arxiv.org/pdf/2606.16690", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["PATCH", "latent patch innovation", "execution corridor", "intervention"], "model_class": "other", "techniques": ["action-conditioned latent prediction", "monitor-driven routing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Monitors residual latent patch changes inside the corridor swept by the active action chunk that cannot be explained by robot motion, triggering pause/recovery and resumption.", "relevance": 4, "relevance_reason": "Runtime monitor tied to open-loop chunk execution.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: runtime monitoring robot policy"}
{"id": "2608.15285", "title": "PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies", "authors": ["Yufei Guo", "Yinan Wu", "Haoran Duan", "Guiguang Ding", "Jungong Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15285", "pdf_url": "https://arxiv.org/pdf/2608.15285", "code_url": "", "clusters": ["C3"], "keywords": ["phase-conditioned LoRA", "action expert"], "model_class": "VLA", "techniques": ["LoRA", "dynamic adapters"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "LoRA variant whose update direction varies with manipulation phase, applied to the action expert of flow VLAs.", "relevance": 4, "relevance_reason": "New PEFT design for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.15594", "title": "Pixels to Proofs: Probabilistically-Safe Latent World Model Control via Parallel Conformal Robust MPC", "authors": ["Devesh Nath", "Anutam Srinivasan", "Haoran Yin", "Ruitong Jiang", "Jeffrey Fang", "Glen Chou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.15594", "pdf_url": "https://arxiv.org/pdf/2606.15594", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["robust MPC", "conformal prediction", "JEPA latent dynamics", "GPU-parallel SLS"], "model_class": "world model", "techniques": ["system level synthesis robust MPC in latent space", "conformalized latent error bounds", "GPU acceleration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Safe feedback planning from pixels via GPU-accelerated robust MPC in a compact JEPA latent space with conformally calibrated latent error bounds and a conformalized constraint checker.", "relevance": 4, "relevance_reason": "Probabilistic safety with efficient latent planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.07464", "title": "Planning-Aligned Token Compression for Long-Context Autonomous Driving", "authors": ["Zhi-Xuan Liang", "Yuxiao Chen", "Yurong You", "Peter Karkus", "Wenhao Ding", "Boyi Li", "Alexander Popov", "Yan Wang", "M. Igl", "Yi-Ming Li", "Danfei Xu", "Nikolai Smolyanskiy", "B. Ivanovic", "Ping Luo", "Marco Pavone"], "year": 2026, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2606.07464", "pdf_url": "https://arxiv.org/pdf/2606.07464", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["long-context", "autonomous driving", "vision-action model"], "model_class": "VLA", "techniques": ["token compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "3.3x speedup and 2.7x memory reduction vs uncompressed baseline", "metric_measures": "relative latency", "benchmarks": ["closed-loop driving"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Planning-aligned token compression for long-context driving vision-action models, keeping decision-critical history under tight budgets.", "relevance": 4, "relevance_reason": "Token compression for real-time budgets", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2503.14734"}
{"id": "2601.22018", "title": "PocketDP3: Efficient Pocket-Scale 3D Visuomotor Policy", "authors": ["Jinhao Zhang", "Zhexuan Zhou", "Huizhe Li", "Yichen Lai", "Wenlong Xia", "Haoming Song", "Youmin Gong", "Jie Mei"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.22018", "pdf_url": "https://arxiv.org/pdf/2601.22018", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["pocket-scale", "MLP-Mixer decoder", "3D diffusion policy", "compact decoder", "MLP-Mixer", "3D policy"], "model_class": "diffusion/flow policy", "techniques": ["compact architecture", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "<1% of parameters of prior methods; 2-step inference", "metric_measures": "relative memory", "benchmarks": ["RoboTwin 2.0", "Adroit", "Meta-World", "real robot", "MetaWorld"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Replaces DP3's heavy U-Net decoder with an MLP-Mixer diffusion mixer, cutting parameters by two orders of magnitude and enabling two-step inference.", "relevance": 4, "relevance_reason": "Shows decoder over-parameterization in diffusion policies.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy distillation; snowball from 2410.21257"}
{"id": "2604.20834", "title": "PokeVLA: Empowering Pocket-Sized Vision-Language-Action Model with Comprehensive World Knowledge Guidance", "authors": ["Yupeng Zheng", "Xiang Li", "Songen Gu", "Yuhang Zheng", "Shuai Tian", "Weize Li", "Linbo Wang", "Senyu Fei", "Pengfei Li", "Yinfeng Gao", "Zebin Xing", "Yilun Chen", "Qichao Zhang", "Haoran Li", "Wenchao Ding", "Yu-Peng Zheng", "Song-En Gu", "Yu-Hang Zheng", "Wei-Ze Li", "Lin-Bo Wang", "Yin-Feng Gao", "Yi-Lun Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.20834", "pdf_url": "https://arxiv.org/pdf/2604.20834", "code_url": "", "clusters": ["C1"], "keywords": ["lightweight VLA", "small VLA", "lightweight"], "model_class": "VLA", "techniques": ["compact VLM", "two-stage training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "PokeVLA is a pocket-sized VLA that pretrains a compact VLM on embodied reasoning data and then injects manipulation-relevant semantics and geometry into an action expert, targeting efficiency with strong performance.", "relevance": 4, "relevance_reason": "direct: lightweight VLA design", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2106.09685, 2303.04137; cited by 2608.09381"}
{"id": "2606.22540", "title": "PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models", "authors": ["Xianghui Wang", "Feng Chen", "Wenbo Zhang", "Hua Yan", "Zixuan Wang", "Changsheng Li", "Yinjie Lei"], "year": 2026, "venue": "ECCV 2026", "type": "preprint", "url": "https://arxiv.org/abs/2606.22540", "pdf_url": "https://arxiv.org/pdf/2606.22540", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["policy efficiency", "chunk length", "RL post-training", "PolicyTrim", "fewer inference calls"], "model_class": "VLA", "techniques": ["RL post-training", "action chunk extension", "post-training for chunk length/step reduction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 5.83x end-to-end deployment speedup; 51.4% fewer execution steps", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "RL post-training that rewards reliably executing longer chunks and completing tasks in fewer steps, reducing the number of forward passes per task.", "relevance": 4, "relevance_reason": "Reframes efficiency as fewer inference calls.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; arXiv API: abs:\"post-training\" AND abs:\"vision-language-action\" AND abs:\"compute\""}
{"id": "2606.30686", "title": "Position: Vision-Language-Action Models Cannot Be Verified to Perform Physical Reasoning", "authors": ["Taozhao Chen", "Ian Manchester", "Huaming Chen", "Ian R. Manchester", "Hua-Min Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.30686", "pdf_url": "https://arxiv.org/pdf/2606.30686", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["verification", "VLA evaluation", "position paper", "evaluation"], "model_class": "VLA", "techniques": ["position", "evaluation design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Position paper arguing task success rate cannot separate semantic matching from physical reasoning in VLAs, so physical generalization is unverified; proposes controlled-variation evaluations.", "relevance": 4, "relevance_reason": "Limits of current evidence for certifying VLA behavior.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=100]; snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2605.22446", "title": "Pre-VLA: Preemptive Runtime Verification for Reliable Vision-Language-Action and World-Model Rollouts", "authors": ["Zhen Sun", "Yongjian Guo", "Haoran Sun", "Luqiao Wang", "Wei Lu", "Jiachi Ji", "Shengzhe Ji", "Junwu Xiong", "Zhijun Meng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.22446", "pdf_url": "https://arxiv.org/pdf/2605.22446", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["Pre-VLA", "preemptive verification", "resampling under budget", "action validity", "world-model rollout gating"], "model_class": "other", "techniques": ["lightweight verifier head", "budgeted resampling scheduler", "lightweight dual-branch safety/advantage head", "preemptive resampling scheduler"], "base_model": "", "params": "", "hardware": "", "reported_metric": "183.9 ms avg verification per action chunk; success 30.79% -> 37.62% over RynnVLA-002", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Verifies candidate action chunks before execution or world-model imagination with a lightweight safety/advantage head and resamples low-quality chunks within a compute budget.", "relevance": 4, "relevance_reason": "Runtime verification with reported per-chunk latency cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"action chunk\" execution horizon adaptive; search query"}
{"id": "2608.26673", "title": "PredVLA: Predictive Sensorimotor Modeling for Sub-Million-Parameter Robot Manipulation", "authors": ["Hiroki Sawada", "Shunichi Kasahara"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.26673", "pdf_url": "https://arxiv.org/pdf/2608.26673", "code_url": "", "clusters": ["C1"], "keywords": ["sub-million-parameter policy", "compact VLA"], "model_class": "other", "techniques": ["predictive coding", "recurrent dynamics", "tiny policy", "hierarchical recurrent dynamics"], "base_model": "", "params": "0.68M", "hardware": "", "reported_metric": "86.9% mean success (3 short suites), 75.4% (all 4 LIBERO suites)", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "0.68M-parameter predictive-coding policy with hierarchical recurrent dynamics where observations update latent state only via prediction errors; strong LIBERO results without robot pretraining.", "relevance": 4, "relevance_reason": "Extreme parameter efficiency for language-conditioned manipulation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2605.12160", "title": "Premover: Fast Vision-Language-Action Control by Acting Before Instructions Are Complete", "authors": ["Joonha Park", "Jiseung Jeong", "Taesik Gong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.12160", "pdf_url": "https://arxiv.org/pdf/2605.12160", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["latency hiding", "streaming instructions"], "model_class": "VLA", "techniques": ["precomputation", "streaming instructions", "precomputation during instruction input"], "base_model": "", "params": "", "hardware": "", "reported_metric": "mean wall-clock 34.0 s -> 29.4 s (13.6%) at matched success", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lets a frozen VLA start acting while the user is still typing/speaking, using small heads that build a focus map from partial instructions and a readiness threshold.", "relevance": 4, "relevance_reason": "Hides instruction latency to cut wall-clock time.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "nvidia-blog-rise-of-world-action-models", "title": "Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models", "authors": ["Moritz Reuss"], "year": 2026, "venue": "NVIDIA Technical Blog", "type": "blog", "url": "https://developer.nvidia.com/blog/pretrained-to-imagine-fine-tuned-to-act-the-rise-of-world-action-models/", "pdf_url": "", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["world action model taxonomy", "inverse dynamics", "joint prediction", "representation-only WAM"], "model_class": "world-action model", "techniques": ["survey/taxonomy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "pi0.5 ~190 ms per chunk vs 590-800 ms for WAMs with full video generation (3-4x slowdown)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Overview that splits WAMs into inverse-dynamics, joint video-action, and representation-only designs, and points out that full video generation makes them 3-4x slower than VLAs, which is why skip-the-video variants are appearing.", "relevance": 4, "relevance_reason": "Practitioner overview of the latency cost of WAMs relative to VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.17423", "title": "Prism-GRPO: Faster VLA Policy Optimization via Splitting Same-outcome Groups", "authors": ["Zeyun Deng", "Yuzhe Lu", "Yawei Wang", "Linbo Liu", "Qing Ping", "Han Ding", "Guande Wu", "Panpan Xu", "Jun Huan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.17423", "pdf_url": "https://arxiv.org/pdf/2608.17423", "code_url": "", "clusters": ["C3"], "keywords": ["Prism-GRPO", "same-outcome groups", "rollout budget"], "model_class": "VLA", "techniques": ["GRPO", "quality-weighted rewards"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Splits all-success/all-fail GRPO groups by execution quality to recover signal and save rollouts.", "relevance": 4, "relevance_reason": "Cuts rollout cost of GRPO VLA RL.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:GRPO"}
{"id": "2604.25859", "title": "Privileged Foresight Distillation: Zero-Cost Future Correction for World Action Models", "authors": ["Pengcheng Fang", "Hongli Chen", "Xiaohao Cai"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.25859", "pdf_url": "https://arxiv.org/pdf/2604.25859", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["privileged foresight", "residual correction", "zero-cost"], "model_class": "world-action model", "techniques": ["distill privileged-future action-denoising residual into current-only policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Models what the future branch contributes as a residual correction to action denoising and distills it into a current-only WAM, recovering the benefit at zero inference cost.", "relevance": 4, "relevance_reason": "Zero-cost future correction for WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2609.09148", "title": "Proxy Policy Steering", "authors": ["Chuanruo Ning", "Tianrui Wang", "Wei-Chiu Ma", "Kuan Fang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.09148", "pdf_url": "https://arxiv.org/pdf/2609.09148", "code_url": "", "clusters": ["C3"], "keywords": ["inference-time steering", "proxy policies", "frozen base", "pi0.5"], "model_class": "VLA", "techniques": ["residual steering", "lightweight proxy policies"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "+53% absolute success on average", "metric_measures": "other", "benchmarks": ["8 real-world tasks", "4 sim tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Adapts a frozen flow VLA by steering its velocity field with the difference of two small trained proxy policies; outperforms LoRA fine-tuning.", "relevance": 4, "relevance_reason": "Lightweight adaptation without modifying base weights.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025"}
{"id": "2606.08015", "title": "Q-VGM: Q-Guided Value-Gradient Matching for Offline-to-Online RL of Flow-Matching VLA Policies", "authors": ["Ziqian Wang", "Rui Zhang", "Yitian Liu", "Xingjian Mao", "Minqian Wang", "Yao Mu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08015", "pdf_url": "https://arxiv.org/pdf/2606.08015", "code_url": "", "clusters": ["C3"], "keywords": ["Q-VGM", "offline-to-online RL", "velocity matching"], "model_class": "VLA", "techniques": ["offline-to-online RL", "critic-guided velocity matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Critic-guided velocity matching at late denoising steps avoids backprop through the denoising chain in RL fine-tuning of flow VLAs.", "relevance": 4, "relevance_reason": "Removes costly BPTT in RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.26425", "title": "QuantWM: Temporally Consistent 2-Bit KV Cache Quantization for Video World Models", "authors": ["Jiaqi Zhao", "Xiaobin Hu", "Bo Yin", "Junpeng Jiang", "Miao Zhang", "Shuicheng Yan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.26425", "pdf_url": "https://arxiv.org/pdf/2609.26425", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["KV cache quantization", "2-bit", "video world model"], "model_class": "world model", "techniques": ["KV-cache quantization", "training-free INT2 KV quantization", "sensitivity-aware key clustering", "attention-logit preservation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["video world model generation quality"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows standard 2-bit KV quantization causes flicker in video world models because key errors shift attention, and fixes it with sensitivity-aware key centroids so KV memory can be cut to 2 bits.", "relevance": 4, "relevance_reason": "Compression of world models (not robot-evaluated).", "foundational_technique": false, "verified": true, "found_via": "search: compression world model; search query"}
{"id": "2603.14498", "title": "R3DP: Real-Time 3D-Aware Policy for Embodied Manipulation", "authors": ["Yuhao Zhang", "Wan-Xi Dong", "Yue Shi", "Yingshi Liang", "Jing-Nan Gao", "Qiaochu Yang", "Yaxing Lyu", "Zhi-Xuan Liang", "Yibin Liu", "Cong-Sheng Xu", "Xian-Da Guo", "Wei Sui", "Yaohui Jin", "Xiao-Kang Yang", "Yanyan Xu", "Yao Mu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.14498", "pdf_url": "https://arxiv.org/pdf/2603.14498", "code_url": "", "clusters": ["C1"], "keywords": ["real-time", "fast-slow", "3D priors"], "model_class": "other", "techniques": ["asynchronous fast-slow", "keyframe querying", "feature prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "R3DP brings large 3D foundation-model priors (VGGT) into manipulation policies in real time by querying the slow model only on sparse keyframes and predicting intermediate features with a lightweight network.", "relevance": 4, "relevance_reason": "direct: real-time integration of heavy 3D priors", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2607.26657"}
{"id": "2608.15924", "title": "RAPAC-DP: Response-Aligned Pending-Action Compensation for Diffusion Policies under Delayed Execution", "authors": ["Tao Wang", "Wei Wang", "Jianhui Wang", "Qi Wang", "Weidi Huang", "Bing Xu", "Wei-Di Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15924", "pdf_url": "https://arxiv.org/pdf/2608.15924", "code_url": "", "clusters": ["C2", "C5", "C6", "C7"], "keywords": ["cloud-side inference", "delay compensation", "pending actions", "RAPAC-DP", "pending-action compensation", "cloud inference delay"], "model_class": "diffusion/flow policy", "techniques": ["pending-action conditioning", "parameter-efficient compensation pathway", "delay-augmented training", "parameter-efficient delay compensation pathway", "delay-conditioned training samples"], "base_model": "", "params": "", "hardware": "", "reported_metric": "retains 81.4% of delay-free performance at largest Kinetix delay", "metric_measures": "other", "benchmarks": ["Kinetix", "RoboMimic"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Conditions a parameter-efficient compensation pathway on actions already scheduled before the cloud response arrives, recovering the frozen policy exactly when delay is negligible.", "relevance": 4, "relevance_reason": "Explicitly targets cloud-deployed policies under delay.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778; snowball: citations of 2506.07339 (S2)"}
{"id": "2609.34170", "title": "RAVEL: Asynchronous Rolling Inference for Flow-Based Vision-Language-Action Models", "authors": ["Yuhan Chen", "Ke Yu", "Pengfei Liu", "Shuxun Wang", "Yi Yang", "Linchao Zhu", "Yu-Han Chen", "Peng-Fei Liu", "Shu-Xun Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34170", "pdf_url": "https://arxiv.org/pdf/2609.34170", "code_url": "", "clusters": ["C1", "C2", "C3", "C6", "C7"], "keywords": ["rolling denoising", "asynchronous", "fast observation pathway", "rolling inference", "flow VLA", "RAVEL", "partially denoised buffer"], "model_class": "VLA", "techniques": ["asynchronous inference", "rolling diffusion", "rolling denoising buffer", "decoupled VLM encoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Asynchronous rolling inference for flow VLAs that executes near-term actions after one denoising step while carrying partially denoised future actions forward, also addressing VLM encoding latency.", "relevance": 4, "relevance_reason": "Hierarchical slow/fast async design applicable to cloud/edge split.", "foundational_technique": false, "verified": true, "found_via": "search query: fast slow VLA; search query; snowball: citations of 2506.07339 (S2)"}
{"id": "2603.11106", "title": "RC-NF: Robot-Conditioned Normalizing Flow for Real-Time Anomaly Detection in Robotic Manipulation", "authors": ["Shijie Zhou", "Bin Zhu", "Jia-Rui Yang", "Xiangyu Zhao", "Jing-Jing Chen", "Yu-Gang Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.11106", "pdf_url": "https://arxiv.org/pdf/2603.11106", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["RC-NF", "normalizing flow", "real-time anomaly detection", "LIBERO-Anomaly-10"], "model_class": "VLA", "techniques": ["robot-conditioned normalizing flow"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "response latency under 100 ms", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Unsupervised normalizing-flow monitor over robot and object states giving a real-time OOD score for VLAs to trigger rollback or replanning, with a new anomaly benchmark.", "relevance": 4, "relevance_reason": "Real-time monitor with reported latency.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2510.09459 (S2)"}
{"id": "2609.22587", "title": "React When You Need To: Event-Triggered Asynchronous Inference for VLA Policies", "authors": ["Yansong Wu", "Huaqing Li", "Tianding Hou", "Lingyun Chen", "Alois Knoll"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22587", "pdf_url": "https://arxiv.org/pdf/2609.22587", "code_url": "", "clusters": ["C2", "C4", "C6", "C7"], "keywords": ["event-triggered inference", "asynchronous", "reactivity", "asynchronous VLA", "adaptive inference gap", "asynchronous inference", "inference gap"], "model_class": "VLA", "techniques": ["adaptive inference gap", "asynchronous inference", "event-triggered inference scheduling", "event-triggered scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "95% average success, +55 pp over strongest baseline", "metric_measures": "other", "benchmarks": ["real robot", "real-world static/dynamic tasks", "real world static and dynamic"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Triggers the next asynchronous VLA inference based on observed scene change rather than a fixed gap, balancing smoothness and reactivity.", "relevance": 4, "relevance_reason": "Adaptive inference-call scheduling reduces compute/network calls.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA inference performance; snowball from 2506.07339; arXiv search: \"action chunk\" asynchronous inference"}
{"id": "2609.36540", "title": "Reactive Real-Time Flow Policies via Asynchronous Distribution Alignment", "authors": ["Moritz Zoellner", "Reece O'Mahoney", "Ioannis Havoutis", "Rohan Paleja"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36540", "pdf_url": "https://arxiv.org/pdf/2609.36540", "code_url": "", "clusters": ["C1", "C2", "C6", "C7"], "keywords": ["asynchronous distribution alignment", "flow-field distillation", "propose-resolve", "reactivity", "recursive flow-field distillation"], "model_class": "VLA", "techniques": ["asynchronous inference", "distillation", "recursive flow-field distillation", "candidate selection by likelihood"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~80% of original success on RoboMimic (~30 pts over async baselines)", "metric_measures": "other", "benchmarks": ["LIBERO", "RoboMimic"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows asynchronous execution can distort a VLA's action distribution and restores it via recursive flow-field distillation plus selecting among pre-generated candidates with the latest observation.", "relevance": 4, "relevance_reason": "Correctness of async execution vs original policy.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk inference delay; snowball from 2506.07339; arXiv search: real-time chunking"}
{"id": "2606.14255", "title": "ReactVLA: Fast and Lightweight Reactive Robot Manipulation via Improved Mean Flow Action Generation", "authors": ["Yanzhao Guo", "Wenkai Chen", "Jianwei Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.14255", "pdf_url": "https://arxiv.org/pdf/2606.14255", "code_url": "", "clusters": ["C1"], "keywords": ["improved mean flow", "one-to-few-step", "attention residuals"], "model_class": "VLA", "techniques": ["mean flow", "dynamic depth routing"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">4x inference speed vs leading VLAs", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboIMI", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight VLA with an improved MeanFlow action generator for one-to-few-step sampling and attention-residual depth routing.", "relevance": 4, "relevance_reason": "Few-step lightweight VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: mean flow policy"}
{"id": "2606.13355", "title": "Real-Time Execution with Autoregressive Policies", "authors": ["Sangkyun Lee", "S. Park", "Tackgeun You", "Avi Caciularu", "Idan Szpektor", "Hwasup Lim", "Youngjae Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.13355", "pdf_url": "https://arxiv.org/pdf/2606.13355", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["autoregressive policy", "real-time execution", "constrained decoding", "latency bounds"], "model_class": "VLA", "techniques": ["tokenization horizon tuning", "constrained decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Shows autoregressive VLAs can run with asynchronous real-time execution by adjusting tokenization horizon and constrained decoding that guarantees strict latency bounds.", "relevance": 4, "relevance_reason": "Latency guarantees for autoregressive VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "modal-pi-remote-inference-2026", "title": "Real-time inference for robots at Physical Intelligence (Modal case study)", "authors": ["Modal"], "year": 2026, "venue": "Modal blog (case study)", "type": "blog", "url": "https://modal.com/blog/physical-intelligence-runs-real-time-remote-inference-for-robotic-control-on-modal", "pdf_url": "", "code_url": "", "clusters": ["C6", "C9"], "keywords": ["remote inference", "QUIC", "NAT traversal", "cloud GPU", "QUIC over UDP", "regional pinning"], "model_class": "VLA", "techniques": ["persistent QUIC/UDP streaming channel", "region-pinned GPU containers", "cloud offloading", "custom low-jitter transport"], "base_model": "", "params": "", "hardware": "datacenter-class cloud GPUs (Modal)", "reported_metric": "~10-15 ms added network overhead", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Case study of Physical Intelligence serving VLA control from cloud GPUs through a persistent QUIC-over-UDP channel with NAT hole punching, avoiding head-of-line blocking in the control loop.", "relevance": 4, "relevance_reason": "Concrete edge-cloud deployment engineering for robot foundation models. Year inferred from search (publication date not verified).", "foundational_technique": false, "verified": true, "found_via": "search query; blog sweep: Physical Intelligence"}
{"id": "nvidia-forum-flashrt-thor-pi05-groot", "title": "Real-Time Inference on Thor & RTX: Pi0.5/GR00T N1.6/1.7 (Thor 23 Hz, RTX 5090 50-80 Hz) — FlashRT", "authors": ["7thuniversels", "LiangSu8899 (7thuniversels)", "LiangSu8899 (forum user 7thuniversels)"], "year": 2026, "venue": "NVIDIA Developer Forums / GitHub", "type": "blog", "url": "https://forums.developer.nvidia.com/t/real-time-inference-on-thor-rtx-pi0-5-gr00t-n1-6-1-7-thor-23-hz-rtx-5090-50-80hz/368788", "pdf_url": "", "code_url": "https://github.com/LiangSu8899/FlashRT", "clusters": ["C2", "C4", "C5", "C9"], "keywords": ["FlashRT", "custom CUDA kernels", "Jetson Thor", "RTX 5090", "batch-1 latency", "small-batch inference", "MLIR-TRT", "NVFP4"], "model_class": "VLA", "techniques": ["hand-written CUDA kernels", "MLIR-TRT", "NVFP4", "hand-written kernels", "graph compilation", "FP4 quantization"], "base_model": "pi0.5, pi0, GR00T N1.6, pi0-FAST", "params": "", "hardware": "Jetson AGX Thor, RTX 5090", "reported_metric": "pi0.5: 44 ms (23 Hz) on Thor, 17.58 ms (57 Hz) on RTX 5090; GR00T N1.6: ~41-45 ms on Thor, ~12.5-13 ms on RTX 5090 | pi0.5: 44 ms (23 Hz) on Thor, 17.58 ms (57 Hz) on RTX 5090; GR00T N1.6: 45 ms (22 Hz) on Thor; pi0-FAST 8.1 ms/token on Thor", "metric_measures": "open-loop chunk rate", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Community report of a small-batch inference engine with custom kernels giving per-inference latency for pi0/pi0.5/GR00T on Thor and RTX 5090.", "relevance": 4, "relevance_reason": "Systems-level VLA acceleration numbers on onboard hardware; community source, not peer reviewed.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: NVIDIA Jetson Thor GR00T inference latency; web: Jetson Thor GR00T N1 deployment TensorRT; blog sweep: NVIDIA GEAR / Isaac"}
{"id": "2601.20130", "title": "Real-Time Robot Execution with Masked Action Chunking", "authors": ["Haoxuan Wang", "Gengyu Zhang", "Yan Yan", "Yuzhang Shang", "Ramana Rao Kompella", "Gaowen Liu", "Wang, Haoxuan", "Zhang, Gengyu", "Yan, Yan", "Shang, Yuzhang", "Kompella, Ramana Rao", "Liu, Gaowen", "Ramana Kompella"], "year": 2026, "venue": "ICLR 2026", "type": "preprint", "url": "https://arxiv.org/abs/2601.20130", "pdf_url": "https://arxiv.org/pdf/2601.20130", "code_url": "", "clusters": ["C1", "C2", "C4", "C6", "C7", "C9"], "keywords": ["asynchronous inference", "masked action chunking", "intra-chunk inconsistency", "REMAC"], "model_class": "diffusion/flow policy", "techniques": ["asynchronous inference", "masked chunk training", "masked action chunking", "masked action chunking fine-tuning", "prefix-preserved sampling", "asynchronous execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Identifies intra-chunk misalignment between executed actions and current perception as a failure cause under asynchronous inference and fine-tunes policies with masked action chunking plus prefix-preserving sampling for robustness across delays.", "relevance": 4, "relevance_reason": "Real-time chunk execution method.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; WebSearch: action chunk size inference latency trade-off; arXiv API all:\"action chunking\" AND latency; web search: real-time chunking PI blog"}
{"id": "2603.05449", "title": "RealWonder: Real-Time Physical Action-Conditioned Video Generation", "authors": ["Wei Liu", "Ziyu Chen", "Zizhang Li", "Yue Wang", "Hong-Xing Yu", "Jiajun Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.05449", "pdf_url": "https://arxiv.org/pdf/2603.05449", "code_url": "https://liuwei283.github.io/RealWonder/", "clusters": ["C3", "C7", "C8", "C9"], "keywords": ["real-time action-conditioned video", "physics simulation bridge", "4-step distillation", "action-conditioned video generation", "real-time world model", "few-step distillation"], "model_class": "world model", "techniques": ["single-image 3D reconstruction", "physics simulation to optical flow", "distilled 4-step video generator", "diffusion step distillation (4 steps)", "physics-simulation intermediate representation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "13.2 FPS at 480x832", "metric_measures": "throughput", "benchmarks": ["interactive rigid/deformable/fluid/granular scenes"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Real-time action-conditioned video world model from a single image: actions go through a physics simulator to optical flow/RGB, which a 4-step distilled video generator renders interactively.", "relevance": 4, "relevance_reason": "Real-time world model via step distillation, a CFP topic; from speaker Jiajun Wu.", "foundational_technique": false, "verified": true, "found_via": "search query; author sweep: Jiajun Wu"}
{"id": "10.1109/fccm68464.2026.00075", "title": "Reconfigurable Computing Challenge: Deploying SmolVLA Model on an AMD XDNA NPU", "authors": ["Xiao-Yu Liang", "Elaine Cao", "Chun-Ning Kao", "D. Lin", "Jia-Jie Li", "Zhiru Zhang"], "year": 2026, "venue": "IEEE Symposium on Field-Programmable Custom Computing Machines", "type": "paper", "url": "https://doi.org/10.1109/FCCM68464.2026.00075", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["SmolVLA", "AMD XDNA NPU", "edge deployment"], "model_class": "VLA", "techniques": ["NPU kernel tiling", "dataflow DSL"], "base_model": "SmolVLA", "params": "", "hardware": "AMD XDNA1 NPU", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Initial prototype deploying SmolVLA on an AMD XDNA NPU with profiled GEMM tiling and optimized kernels.", "relevance": 4, "relevance_reason": "VLA deployment on edge NPU hardware", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2504.16054"}
{"id": "2602.07845", "title": "Recurrent-Depth VLA: Implicit Test-Time Compute Scaling of Vision-Language-Action Models via Latent Iterative Reasoning", "authors": ["Yalcin Tur", "Jalal Naghiyev", "Haoquan Fang", "Wei-Chuan Tsai", "Jiafei Duan", "Dieter Fox", "Ranjay Krishna"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.07845", "pdf_url": "https://arxiv.org/pdf/2602.07845", "code_url": "", "clusters": ["C1"], "keywords": ["test-time compute", "RD-VLA"], "model_class": "VLA", "techniques": ["recurrent depth", "adaptive stopping", "latent reasoning", "adaptive compute", "early stopping"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 80x inference speedup vs prior reasoning-based VLAs; constant memory", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Weight-tied recurrent action head refines latents iteratively with adaptive stopping on convergence, giving variable test-time compute at constant memory instead of CoT tokens.", "relevance": 4, "relevance_reason": "Adaptive-depth inference with large speedups over reasoning VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]; snowball round 2: cites 2303.04137; cited by 2605.09948"}
{"id": "2605.24044", "title": "RED: Adaptive Real-Time DAG Scheduling for Robotic Inference under Environmental Dynamics", "authors": ["Zexin Li", "Tao Ren", "Johnathan Liu", "Xiaoxi He", "Cong Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24044", "pdf_url": "https://arxiv.org/pdf/2605.24044", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["RED", "real-time DAG scheduling", "deadline-aware", "multi-task DNN"], "model_class": "other", "techniques": ["deadline-aware scheduling", "sub-deadline assignment"], "base_model": "", "params": "", "hardware": "NVIDIA Jetson family; Apple M-series", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Deadline-aware scheduler for multi-task DNN inference on robots whose computation graph changes with the environment, preserving end-to-end timing guarantees under modeling assumptions.", "relevance": 4, "relevance_reason": "Timing guarantees for robot inference workloads.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"deadline\" robot inference neural network real-time scheduling"}
{"id": "2607.27782", "title": "RedFlow: Redirect Failure into Action-level Corrections for Flow-matching VLA Policy", "authors": ["Zhengyang Yan", "Junhao Li", "Fangqi Zhu", "Zijun Wang", "Quanxin Shou", "Yikun Miao", "Xiaoyi Pang", "Zicong Hong", "Song Guo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.27782", "pdf_url": "https://arxiv.org/pdf/2607.27782", "code_url": "", "clusters": ["C3"], "keywords": ["offline post-training", "flow matching", "failure correction"], "model_class": "VLA", "techniques": ["offline RL", "action redirection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "online methods need 8.7-16x more rollouts to reach same LIBERO-Spatial threshold", "metric_measures": "other", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Offline post-training for flow VLAs that turns failed chunks into corrective targets from fixed rollouts, very rollout-efficient.", "relevance": 4, "relevance_reason": "Sample-efficient post-training without online data.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025"}
{"id": "2607.12287", "title": "Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference", "authors": ["Yuzhou Wu", "Yuxin Zheng", "Muchun Niu", "Yishan Yang", "Tianhao Liu", "hanwen kang", "Jiajian Jing", "Linfeng Zhang", "Chuan Wen", "Yu-Zhou Wu", "Mu-Chun Niu", "Yi-Shan Yang", "Jia-Jian Jing", "Lin-Feng Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.12287", "pdf_url": "https://arxiv.org/pdf/2607.12287", "code_url": "", "clusters": ["C1", "C2", "C3", "C9"], "keywords": ["temporal redundancy", "incremental visual encoding", "2-step diffusion", "feature reuse", "step reduction"], "model_class": "VLA", "techniques": ["incremental token update", "few-step sampling", "visual encoding reuse", "diffusion step reduction"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">2x speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Re-encodes only dynamic image regions between frames and trains the diffusion head to sample in two steps, attacking both perception and action-generation redundancy.", "relevance": 4, "relevance_reason": "Combined perception+action acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2609.18207", "title": "Reinforcement Learning for Real-Time Vision-Language-Action Policies", "authors": ["Perry Dong", "Kuo-Han Hung", "Dorsa Sadigh", "Chelsea Finn", "D. Sadigh"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18207", "pdf_url": "https://arxiv.org/pdf/2609.18207", "code_url": "", "clusters": ["C4", "C6", "C7", "C9"], "keywords": ["Real-Time EXPO-FT", "RL fine-tuning", "action edits", "stale observation", "latency-aware", "stale observations", "asynchronous"], "model_class": "VLA", "techniques": ["slow VLA proposals + fast edit policy", "RL fine-tuning", "latency-aware RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "avg success 42% -> 97% with 10 min online data", "metric_measures": "other", "benchmarks": ["Kinetix", "4 dynamic real tasks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Pairs a slow VLA chunk proposer with a fast RL edit policy conditioned on the latest observation, making RL fine-tuning work under inference delay on dynamic tasks.", "relevance": 4, "relevance_reason": "RL that accounts for inference delay in VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: real-time chunking; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2608.09492", "title": "Rethink Before You Execute: Adaptive Execution for World Action Models", "authors": ["Feng Ye", "Yiming Zhao", "Yong Yu", "Hongxu Zhou", "Yong Pan", "Yuan Xue", "Peng Jia", "Chuanmin Jia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09492", "pdf_url": "https://arxiv.org/pdf/2608.09492", "code_url": "", "clusters": ["C6", "C7", "C8", "C9"], "keywords": ["adaptive execution horizon", "progress monitor", "replanning", "WAM"], "model_class": "world-action model", "techniques": ["recurrent progress monitor", "adaptive replanning protocol", "online calibration", "adaptive prefix execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "26.9% fewer WAM inferences on easy real tasks at equal success; +13.3 pp on hard tasks", "metric_measures": "other", "benchmarks": ["LIBERO", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Plug-in monitor (TempoWAM) that estimates task progress online and replans only when a chunk stops advancing the task, cutting inference calls and improving success on difficult tasks.", "relevance": 4, "relevance_reason": "Monitoring-driven compute reduction for WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2608.30144", "title": "Rethinking Language's Role in Efficient VLA for Autonomous Vehicles: Toward Smarter, Trustworthy Driving", "authors": ["Tongfei Guo", "Lili Su"], "year": 2026, "venue": "EMNLP 2026", "type": "preprint", "url": "https://arxiv.org/abs/2608.30144", "pdf_url": "https://arxiv.org/pdf/2608.30144", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["efficient VLA", "autonomous driving", "survey"], "model_class": "VLA", "techniques": ["survey", "taxonomy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["nuScenes", "NAVSIM", "Bench2Drive"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey/position on efficient driving VLAs organized by how much language is used at inference (train-only to per-frame generation), tagging methods on latency, parameters, memory, FLOPs and tokens.", "relevance": 4, "relevance_reason": "direct: efficiency of VLAs under onboard budgets (driving)", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]; snowball round 2: cites 2210.17323, 2306.03310"}
{"id": "2607.08029", "title": "Rethinking Small VLM Quantization: From Component-Wise Analysis to Hardware-Aware Edge Deployment", "authors": ["Hyeju Shin", "Chorwon Kim", "Ryangsoo Kim", "Hark Yoo", "Jaein Kim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.08029", "pdf_url": "https://arxiv.org/pdf/2607.08029", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["small VLM quantization", "Jetson Orin NX", "Jetson AGX Orin", "SigLIP INT8", "dequantization overhead", "intelligence per joule"], "model_class": "other", "techniques": ["component-wise quantization analysis"], "base_model": "", "params": "", "hardware": "Jetson Orin NX, Jetson AGX Orin", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Component-wise quantization study of sub-3B VLMs on Jetson finding INT4 LLMs can decode slower due to dequantization, SigLIP INT8 is disproportionately slow on Ampere Jetsons, and memory bandwidth drives energy efficiency.", "relevance": 4, "relevance_reason": "Hardware-kernel interactions directly relevant to VLA backbones on Jetson.", "foundational_technique": true, "verified": true, "found_via": "web: edge VLA Jetson AGX Orin latency"}
{"id": "2608.15938", "title": "Revisiting Open-Loop Execution in Robotics: Toward Reactive, Higher-Performing Policies", "authors": ["Michael Zeng", "Abhinav Agarwal", "Ajay Bati", "Brian Lee", "Siddharth Ancha", "Russ Tedrake"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15938", "pdf_url": "https://arxiv.org/pdf/2608.15938", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["open-loop execution", "reactivity", "non-Markovian demonstrations", "execution horizon"], "model_class": "diffusion/flow policy", "techniques": ["analysis", "long-context policies"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["4 sim tasks", "2 real tasks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Controlled study arguing long open-loop chunk execution mainly helps short-context policies imitate non-Markovian demos; with long enough context, fully closed-loop reactive policies perform best.", "relevance": 4, "relevance_reason": "Questions open-loop execution, central to reactivity/safety tradeoff.", "foundational_technique": false, "verified": true, "found_via": "arXiv API all:\"action chunking\" AND latency"}
{"id": "2606.31382", "title": "Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation", "authors": ["Fengnian Zhang", "Tao Huang", "Siyu Xu", "Zhong Jin", "Chang Xu", "Feng-Nian Zhang"], "year": 2026, "venue": "ECCV 2026", "type": "preprint", "url": "https://arxiv.org/abs/2606.31382", "pdf_url": "https://arxiv.org/pdf/2606.31382", "code_url": "", "clusters": ["C1", "C2", "C3"], "keywords": ["parameter redundancy", "pruning", "VLM-to-VLA divergence", "VLM-to-VLA adaptation", "joint pruning", "VLM-to-VLA"], "model_class": "VLA", "techniques": ["structured pruning", "pruning analysis"], "base_model": "OpenVLA; pi0.5", "params": "", "hardware": "", "reported_metric": "12-30% parameter reduction at ~90% performance without fine-tuning", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses how much each module's weights drifted during VLM-to-VLA adaptation to decide what can be pruned, yielding a multi-module pruning scheme needing no recovery training.", "relevance": 4, "relevance_reason": "Analysis of where redundancy lives in VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA-Adapter; snowball from 2506.07530; arXiv API: abs:\"vision-language-action\" AND abs:adapter AND abs:\"parameter\""}
{"id": "2604.16683", "title": "Rewind-IL: Online Failure Detection and State Respawning for Imitation Learning", "authors": ["Gehan Zheng", "Sanjay Seenivasan", "Matthew Johnson-Roberson", "Weiming Zhi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.16683", "pdf_url": "https://arxiv.org/pdf/2604.16683", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["Rewind-IL", "inter-chunk discrepancy", "state respawning", "conformal prediction"], "model_class": "diffusion/flow policy", "techniques": ["zero-shot chunk-consistency detector", "VLM checkpoints"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Training-free safeguard that detects failure from inter-chunk discrepancy with conformal calibration and rewinds the robot to the latest verified safe checkpoint before re-running the policy.", "relevance": 4, "relevance_reason": "Monitor + fallback for chunked policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure prediction generative robot policy"}
{"id": "2608.23565", "title": "ReWorld: An Interactive World Model with Long-Horizon Memory", "authors": ["Zhifei Chen", "Luozhou Wang", "Gui-Bao Shen", "Dongyu Yan", "Shuai Yang", "Tian-Shuo Xu", "Yi-Hua Du", "Wei Wang", "T. Gui", "Liang-Hua Huang", "Ying-Cong Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.23565", "pdf_url": "https://arxiv.org/pdf/2608.23565", "code_url": "", "clusters": ["C8"], "keywords": ["interactive world model", "long-horizon memory"], "model_class": "world model", "techniques": ["bounded KV cache", "distribution-matching distillation", "LoRA", "4-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "4-step sampling", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Interactive world model with bounded-memory attention and LoRA-based distillation to four sampling steps for real-time streaming.", "relevance": 4, "relevance_reason": "Efficient real-time world model with bounded memory.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2402.15391"}
{"id": "2604.23073", "title": "RL Token: Bootstrapping Online RL with Vision-Language-Action Models", "authors": ["Charles Xu", "Jost Tobias Springenberg", "Michael Equi", "Ali Amin", "Adnan Esmail", "Sergey Levine", "Liyiming Ke", "J. Springenberg", "M. Equi", "A. Esmail"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.23073", "pdf_url": "https://arxiv.org/pdf/2604.23073", "code_url": "", "clusters": ["C3"], "keywords": ["real-world RL", "precision tasks", "RL fine-tuning", "sample efficiency"], "model_class": "VLA", "techniques": ["RL token", "small actor-critic head", "online RL", "lightweight actor-critic"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 3x faster on hardest task segment", "metric_measures": "other", "benchmarks": ["real robot (4 tasks)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RL Token exposes a compact token from a pretrained VLA and trains a small actor-critic on top of it, enabling sample-efficient real-world RL fine-tuning in a few hours for more precise and faster manipulation.", "relevance": 4, "relevance_reason": "Very sample-efficient real-world VLA RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2303.04137, 2304.13705; cited by 2510.24795, 2605.10821"}
{"id": "2602.05765", "title": "RL-VLA$^3$: A Flexible and Asynchronous Reinforcement Learning Framework for VLA Training", "authors": ["Haoran Sun", "Yongjian Guo", "Zhong Guan", "Shuai Di", "Xiaodong Bai", "Jing Long", "Tianyun Zhao", "Mingxi Luo", "Hongke Zhao", "Likang Wu", "Xiaotie Deng", "Xu Chu", "Xi Xiao", "Sheng Wen", "Yicheng Gong", "Junwu Xiong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.05765", "pdf_url": "https://arxiv.org/pdf/2602.05765", "code_url": "", "clusters": ["C3"], "keywords": ["asynchronous RL", "distributed", "VLA training framework"], "model_class": "VLA", "techniques": ["asynchronous distributed RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Fully asynchronous RL framework for VLAs handling variable simulator latencies for higher training throughput.", "relevance": 4, "relevance_reason": "RL post-training system efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2602.07837", "title": "RLinf-USER: A Unified and Extensible System for Real-World Online Policy Learning in Embodied AI", "authors": ["Hongzhi Zang", "Shu'ang Yu", "Hao Lin", "Tianxing Zhou", "Zefang Huang", "Zhen Guo", "Xin Xu", "Jiakai Zhou", "Yuze Sheng", "Shizhe Zhang", "Feng Gao", "Wenhao Tang", "Yufeng Yue", "Quanlu Zhang", "Xinlei Chen", "Chao Yu", "Yu Wang"], "year": 2026, "venue": "RSS 2026", "type": "preprint", "url": "https://arxiv.org/abs/2602.07837", "pdf_url": "https://arxiv.org/pdf/2602.07837", "code_url": "https://github.com/RLinf/RLinf", "clusters": ["C3", "C5", "C6"], "keywords": ["real-world online learning", "cloud-edge training", "asynchronous RL", "system", "real-world online RL", "asynchronous training", "multi-robot"], "model_class": "VLA", "techniques": ["asynchronous training", "cloud-edge split", "replay buffer", "adaptive communication plane", "cloud-edge training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot", "simulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "USER system for real-world online policy learning with a hardware abstraction layer and adaptive communication plane that splits large-model training/inference between cloud and edge across multiple robots.", "relevance": 4, "relevance_reason": "Infrastructure for deployment-time adaptation with edge/cloud split.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RLinf VLA reinforcement learning infrastructure embodied arXiv; arXiv API: asynchronous cloud robot"}
{"id": "2609.19104", "title": "rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inference", "authors": ["Kaijun Zhou", "Zhiyang Li", "Le Chen", "Jinyu Gu", "Kai-Jun Zhou", "Zhi-Yang Li", "Jin-Yu Gu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19104", "pdf_url": "https://arxiv.org/pdf/2609.19104", "code_url": "", "clusters": ["C1", "C2", "C3", "C5", "C9"], "keywords": ["cross-execution caching", "activation sparsity", "muscle memory", "caching", "factory workloads", "Jetson Thor", "inference framework", "repetitive tasks"], "model_class": "VLA", "techniques": ["context caching", "activation caching", "visual-token cache reuse", "activation-pattern cache to reduce weight access", "workload-aware inference system"], "base_model": "", "params": "", "hardware": "RTX 4090, Jetson Thor", "reported_metric": "1.29-1.42x speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Exploits repetition across factory executions by caching visual-token outputs and neuron activation masks to cut compute and weight memory traffic during VLA inference.", "relevance": 4, "relevance_reason": "Memory-access-aware inference system evaluated on Jetson Thor.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2607.04434", "title": "RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies", "authors": ["Tianxing Chen", "Yue Chen", "Zixuan Li", "Junyuan Tang", "Kailun Su", "Haoran Lu", "Weijie Wan", "Baijun Chen", "Songling Liu", "Haowen Yan", "Honghao Su", "Zhiyang Dou", "Kaixuan Wang", "Dandan Zhang", "Yunze Liu", "Yan Qin", "Qiwei Liang", "Qiwei Wu", "Zijian Lin", "Wenwei Lin", "Yuran Wang", "Minghua He", "Tianshu Wu", "Ruihai Wu", "Jingquan Zhou"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2607.04434", "pdf_url": "https://arxiv.org/pdf/2607.04434", "code_url": "https://github.com/RoboDojo-Benchmark", "clusters": ["C4"], "keywords": ["RoboDojo", "sim-and-real", "remote real-world evaluation", "leaderboard", "XPolicyLab"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboDojo", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Unified sim (42 tasks, Isaac Sim) and real (18 tasks) benchmark with a remote, standardized real-world evaluation service and a policy integration layer; 30 policies on a public leaderboard.", "relevance": 4, "relevance_reason": "Reproducible real-robot evaluation protocol.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RoboArena distributed real-world evaluation"}
{"id": "2605.17522", "title": "RoboFlow4D: A Lightweight Flow World Model Toward Real-Time Flow-Guided Robotic Manipulation", "authors": ["Sixu Lin", "Junliang Chen", "Huaiyuan Xu", "Zhuohao Li", "Guangming Wang", "Yixiong Jing", "Sheng Xu", "Runyi Zhao", "Brian Sheil", "Lap-Pui Chau", "Guiliang Liu", "Si-Xu Lin", "Guang-Ming Wang", "Yi-Xiong Jing", "Gui-Liang Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.17522", "pdf_url": "https://arxiv.org/pdf/2605.17522", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["3D flow world model", "slow-fast", "lightweight", "lightweight flow world model", "4D flow", "real-time"], "model_class": "world model", "techniques": ["end-to-end multi-frame 3D flow prediction", "slow-fast flow/action loop", "unified perception+planning flow model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Lightweight world model predicting multi-frame 3D scene flow from images and text to guide an action policy, run in a slow-fast loop for real-time manipulation.", "relevance": 4, "relevance_reason": "Compact flow-based world model for real-time planning.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2607.01060", "title": "RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation", "authors": ["Byeongguk Jeon", "Seonghyeon Ye", "JaeHyeok Doo", "Sungdong Kim", "Minjoon Seo", "Hyungmok Son", "Kimin Lee"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.01060", "pdf_url": "https://arxiv.org/pdf/2607.01060", "code_url": "", "clusters": ["C3", "C4", "C8"], "keywords": ["RoboWorld", "neural simulator", "Step Forcing", "sim-real correlation", "policy evaluation", "VLM scoring"], "model_class": "world model", "techniques": ["fast autoregressive video world model", "Step Forcing (anchored + self-forwarded context)", "VLM progress scoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "Pearson r = 0.989, Spearman 0.970 vs real evaluation", "metric_measures": "other", "benchmarks": ["real robot", "real-world policy evaluation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Automated policy-evaluation pipeline pairing a fast autoregressive video world model with a progress-aware VLM judge; Step Forcing reduces train-test mismatch for long rollouts.", "relevance": 4, "relevance_reason": "Fast world model for scalable evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query: policy evaluation real robot scalable; search query"}
{"id": "10.3390/s26154757", "title": "ROEP: A Robotics-Oriented Evaluation Protocol for Deployment-Facing Vision–Language–Action Manipulation Policies", "authors": ["Sangwoo Han", "Hyunguk Choi"], "year": 2026, "venue": "Italian National Conference on Sensors", "type": "paper", "url": "https://doi.org/10.3390/s26154757", "pdf_url": "https://doi.org/10.3390/s26154757", "code_url": "", "clusters": ["C4"], "keywords": ["evaluation protocol", "runtime fidelity", "deployment"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "ROEP is a deployment-oriented VLA evaluation protocol that reports runtime-interface fidelity, repeatability and failure semantics beyond success rate.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2608.20208", "title": "RoMAN-Flow: Taming Autoregressive Normalizing Flows for Offline Reinforcement Learning in Robotic Manipulation", "authors": ["Shao-Xuan Wang", "Guangting Zheng", "Rui Huang", "Zhi-Peng Tang", "Sha Zhang", "Jia-Jun Deng", "Yanyong Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.20208", "pdf_url": "https://arxiv.org/pdf/2608.20208", "code_url": "https://github.com/konnyaku28/RoMAN-Flow", "clusters": ["C1", "C2"], "keywords": ["normalizing flow", "one-step distillation"], "model_class": "diffusion/flow policy", "techniques": ["autoregressive normalizing flow", "one-step distillation", "offline RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RoMAN-Flow makes autoregressive normalizing-flow policies practical with a sampling-free offline RL objective and distillation into a one-step action generator for low-latency deployment.", "relevance": 4, "relevance_reason": "direct: distillation for low-latency action prediction", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2603.16195", "title": "S-VAM: Shortcut Video-Action Model by Self-Distilling Geometric and Semantic Foresight", "authors": ["Hao-Dong Yan", "Zhi-De Zhong", "Jia-Guang Zhu", "Junjie He", "Wei-Lin Yuan", "Wen-Xuan Song", "Xinlong Gong", "Yingjie Cai", "Guan-Yi Zhao", "Xu Yan", "Bingbing Liu", "Ying-Cong Chen", "Hao-Ang Li", "Haodong Yan", "Zhide Zhong", "Jiaguan Zhu", "Weilin Yuan", "Wenxuan Song", "Xin Gong", "Guanyi Zhao", "Haoang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.16195", "pdf_url": "https://arxiv.org/pdf/2603.16195", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["shortcut", "self-distillation", "one forward pass foresight", "shortcut video-action model", "one-pass foresight"], "model_class": "world-action model", "techniques": ["self-distillation", "self-distill multi-step VFM foresight into one step", "lightweight decouplers"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Self-distills geometric and semantic foresight from multi-step generated videos into a single forward pass so the video-action model is both real-time and high-fidelity.", "relevance": 4, "relevance_reason": "Distillation to speed video-action models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246; snowball from 2602.15922 (citations)"}
{"id": "2607.14488", "title": "Safe Execution of RL Policies Via Acceleration-Based CBF-QP Constraint Enforcement for Real-World Robotic Deployments", "authors": ["Bastien Muraccioli", "Alice Cariou", "Pierre-Alexandre Leziart", "Mathieu Celerier", "Arnaud Demont", "Gentiane Venture", "Mehdi Benallegue"], "year": 2026, "venue": "IROS 2026", "type": "paper", "url": "https://arxiv.org/abs/2607.14488", "pdf_url": "https://arxiv.org/pdf/2607.14488", "code_url": "", "clusters": ["C7"], "keywords": ["Acc-CBF-QP", "safety filter", "RL policy", "humanoid", "manipulator"], "model_class": "other", "techniques": ["acceleration-based CBF-QP"], "base_model": "", "params": "", "hardware": "Kinova Gen3; Unitree H1", "reported_metric": "violations 10.04/s -> 0.80/s (-92%) on H1", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Acceleration-level CBF-QP safety filter enforcing joint, torque and collision limits on arbitrary RL policies at runtime with tunable deviation costs; open-source.", "relevance": 4, "relevance_reason": "Runtime safety filter for learned policies on real hardware.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"control barrier function\" learned policy robot manipulation"}
{"id": "2609.21220", "title": "Safe Real-Time Policy Steering via Noise-Space Trajectory Optimization for One-Step Generative Policies", "authors": ["Qingyi Chen", "Joseph Ruan", "Zachary Kingston"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21220", "pdf_url": "https://arxiv.org/pdf/2609.21220", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["INSPO", "noise-space trajectory optimization", "one-step generative policy", "constraint steering"], "model_class": "diffusion/flow policy", "techniques": ["noise-space particle optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Push-T", "Can", "LIBERO-Spatial"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Steers one-step generative policies at runtime to satisfy constraints like collision avoidance by optimizing their input noise against induced state trajectories, cheaper than gradient-guided multi-step sampling.", "relevance": 4, "relevance_reason": "Real-time constraint enforcement for generative/VLA policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"collision avoidance\" \"vision-language-action\" real-time"}
{"id": "2607.29569", "title": "Safe Vision Language Action Models via Barrier Enhanced Flow Matching", "authors": ["Kasra Sinaei", "Hung-Chieh Wu", "Donald Ebeigbe"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.29569", "pdf_url": "https://arxiv.org/pdf/2607.29569", "code_url": "", "clusters": ["C7"], "keywords": ["barrier-enhanced flow matching", "CBF in denoising", "log-sum-exp barrier", "action chunk safety"], "model_class": "VLA", "techniques": ["CBF-guided flow sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Integrates a smooth aggregate CBF into the flow-matching denoising process so VLAs generate chunk-wide safe trajectories with small overhead and bounded distribution shift.", "relevance": 4, "relevance_reason": "In-sampler safety filter for VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety\" \"action chunking\""}
{"id": "2605.29662", "title": "SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation", "authors": ["Shilin Ma", "Chubin Zhang", "Changyuan Wang", "Yuji Wang", "Yue Wu", "Zixuan Wang", "Jingqi Tian", "Zheng Zhu", "Yansong Tang"], "year": 2026, "venue": "ECCV 2026", "type": "paper", "url": "https://arxiv.org/abs/2605.29662", "pdf_url": "https://arxiv.org/pdf/2605.29662", "code_url": "", "clusters": ["C1"], "keywords": ["future-aware token pruning", "attention consistency"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.89x speedup, <1.5% success drop", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Forecasts deep-layer token saliency using attention that persists on the same semantic entities over time, avoiding premature removal by shallow-layer pruning.", "relevance": 4, "relevance_reason": "VLA token pruning.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning"}
{"id": "2606.20698", "title": "SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model", "authors": ["Kai Tang", "Peidong Jia", "Zhong Chu", "Jixian Wu", "Rui Ma", "Jiajun Cao", "Fangyuan Zhao", "Sixiang Chen", "Yichen Guo", "Xiaowei Chi", "Chun-Kai Fan", "Kevin Zhang", "Jinchang Xu", "Fubing Yang", "Weishi Mi", "Xiaozhu Ju", "Jian Tang", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20698", "pdf_url": "https://arxiv.org/pdf/2606.20698", "code_url": "", "clusters": ["C3", "C4", "C6", "C7", "C8"], "keywords": ["safe RL", "world model", "imagination", "SafeDojo", "model-based safe RL", "video world model", "VLA", "interactive world model", "safety cost head"], "model_class": "world model", "techniques": ["model-based safe RL", "world-model imagination", "safety cost head", "online safe RL inside video world model", "per-step safety cost prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Model-based safe RL for VLAs that learns safe behavior in an interactive video world model with learned progress and per-step safety-cost predictors.", "relevance": 4, "relevance_reason": "Avoids costly real exploration in post-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations); snowball: citations of 2506.09937 (S2); search query"}
{"id": "2605.27947", "title": "SANTS: A State-Adaptive Scheduler for World Action Models", "authors": ["Yirui Sun", "Guangyu Zhuge", "Keliang Liu", "Jie Gu", "Shiqin Dai", "Xinyu Bing", "Zhongxue Gan", "Chunxu Tian"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.27947", "pdf_url": "https://arxiv.org/pdf/2605.27947", "code_url": "", "clusters": ["C8"], "keywords": ["SANTS"], "model_class": "world-action model", "techniques": ["adaptive denoising schedule", "early stopping"], "base_model": "", "params": "", "hardware": "", "reported_metric": "latency -81.7% (RoboTwin 2.0), -79.0% (real) vs full video denoising", "metric_measures": "relative latency", "benchmarks": ["RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight scheduler that decides when to stop video denoising in pixel-space WAMs based on state, trained with downstream action reward.", "relevance": 4, "relevance_reason": "Large latency cuts by adaptive denoising depth.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=100]"}
{"id": "2605.11817", "title": "See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model", "authors": ["Yixu Feng", "Zinan Zhao", "Yanxiang Ma", "Chenghao Xia", "Chengbin Du", "Yunke Wang", "Chang Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.11817", "pdf_url": "https://arxiv.org/pdf/2605.11817", "code_url": "https://github.com/Fediory/Grid-Sampler", "clusters": ["C1"], "keywords": ["differentiable grid sampling", "token resampling"], "model_class": "VLA", "techniques": ["token resampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "76% FLOPs reduction with <10% visual tokens", "metric_measures": "other", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Predicts a few salient coordinates and resamples features via differentiable interpolation, reaching very low token counts without success loss.", "relevance": 4, "relevance_reason": "Visual token reduction.", "foundational_technique": false, "verified": true, "found_via": "search query: visual token reduction robot manipulation"}
{"id": "2607.09520", "title": "Seeing is Free, Speaking is Not: Uncovering the True Energy Bottleneck in Edge VLM Inference", "authors": ["Junfei Zhan", "Haoxun Shen", "Mingang Guo", "Zixuan Huang", "Tengjiao He"], "year": 2026, "venue": "ACM MM 2026", "type": "preprint", "url": "https://arxiv.org/abs/2607.09520", "pdf_url": "https://arxiv.org/pdf/2607.09520", "code_url": "https://github.com/Junfei-Z/seeing-is-free", "clusters": ["C4", "C5", "C6"], "keywords": ["energy profiling", "edge VLM", "decode vs prefill", "Jetson Orin NX", "output tokens dominate", "decode dominance"], "model_class": "other", "techniques": ["energy characterization", "energy profiling"], "base_model": "", "params": "", "hardware": "RTX 3070, Jetson Orin NX", "reported_metric": "output token costs 11-39x more wall-clock time than input token; power varies <5% across inputs | each output token costs 11-39x more time than each input token; average power varies <5% across inputs", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Energy profiling of edge VLMs on Jetson Orin NX finds power is roughly constant per model, so energy tracks latency, and decoded output tokens, not visual tokens, dominate cost.", "relevance": 4, "relevance_reason": "Energy measurement methodology relevant to VLA backbones; challenges visual-token-pruning focus.", "foundational_technique": false, "verified": true, "found_via": "search query: embodied inference energy; arXiv: memory-bound AND (robot OR VLA OR embodied); search query"}
{"id": "2602.08167", "title": "Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning", "authors": ["Milan Ganai", "Katie Luo", "Jonas Frey", "Clark W. Barrett", "Marco Pavone"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2602.08167", "pdf_url": "https://arxiv.org/pdf/2602.08167", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["embodied reasoning", "concise CoT", "R&B-EnCoRe", "self-improvement"], "model_class": "VLA", "techniques": ["reasoning primitive filtering", "importance-weighted variational training"], "base_model": "", "params": "", "hardware": "WidowX setup", "reported_metric": "inference ~5 s → ~3 s per step via ~half-length reasoning", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns which reasoning primitives actually help action prediction and retrains on concise traces, halving reasoning tokens and reducing per-step latency while improving success.", "relevance": 4, "relevance_reason": "Reasoning-length efficiency for reasoning VLAs; featured on SAIL blog.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Stanford AI Lab blog"}
{"id": "2609.23305", "title": "Shared Execution-Clock Drifting Policy for Dynamic Precision Manipulation", "authors": ["Zhen Dong", "Qing-Ran Wu", "Jin-Na Fu", "Jia-Min Wu", "Fu-Lin Chen", "Hong-Yu Yu", "Yi-De Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.23305", "pdf_url": "https://arxiv.org/pdf/2609.23305", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["time-constrained manipulation", "dynamic"], "model_class": "diffusion/flow policy", "techniques": ["one-step drifting policy", "execution clock"], "base_model": "", "params": "", "hardware": "NVIDIA Thor", "reported_metric": "91% success intercepting cups on a 16 m/min conveyor", "metric_measures": "closed-loop control", "benchmarks": ["real robot", "RoboMimic"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "One-step drifting policy with a shared execution clock for deadline- and rhythm-sensitive manipulation, run on NVIDIA Thor.", "relevance": 4, "relevance_reason": "Onboard one-step inference for dynamic tasks", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2505.11494", "title": "SHIELD: Safety on Humanoids via CBFs In Expectation on Learned Dynamics", "authors": ["Lizhi Yang", "Blake Werner", "Ryan K. Cosner", "David Fridovich-Keil", "Preston Culbertson", "Aaron D. Ames"], "year": 2026, "venue": "IROS 2025", "type": "paper", "url": "https://arxiv.org/abs/2505.11494", "pdf_url": "https://arxiv.org/pdf/2505.11494", "code_url": "", "clusters": ["C7"], "keywords": ["SHIELD", "stochastic CBF", "learned dynamics residual", "humanoid"], "model_class": "other", "techniques": ["generative dynamics residual", "stochastic discrete-time CBF"], "base_model": "", "params": "", "hardware": "Unitree G1 (onboard perception)", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Layers a probabilistic CBF safety filter over a black-box RL humanoid controller using a learned stochastic dynamics residual, giving risk-bounded obstacle avoidance.", "relevance": 4, "relevance_reason": "Safety filter for learned humanoid policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"shielding\" learned robot policy safety"}
{"id": "2603.11653", "title": "Simple Recipe Works: Vision-Language-Action Models are Natural Continual Learners with Reinforcement Learning", "authors": ["Jiaheng Hu", "Jay Shim", "Chen Tang", "Yoonchang Sung", "Bo Liu", "Peter Stone", "Roberto Martin-Martin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.11653", "pdf_url": "https://arxiv.org/pdf/2603.11653", "code_url": "https://github.com/UT-Austin-RobIn/continual-vla-rl", "clusters": ["C3"], "keywords": ["continual RL", "LoRA", "sequential fine-tuning"], "model_class": "VLA", "techniques": ["LoRA", "on-policy RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["lifelong RL benchmarks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Finds that plain sequential fine-tuning with LoRA and on-policy RL gives little forgetting for large pretrained VLAs, beating complex continual RL methods.", "relevance": 4, "relevance_reason": "Shows PEFT enables cheap continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: continual learning vision-language-action model catastrophic forgetting arXiv"}
{"id": "2606.00664", "title": "SKIP: Sparse Keyframe Interpolation Paradigm for Efficient Embodied World Models", "authors": ["Ziheng He", "Yixiang Chen", "Ning Yang", "Zhanqian Wu", "Qisen Ma", "Yuan Xu", "Jiabing Yang", "Peiyan Li", "Xiangnan Wu", "Xiaofeng Wang", "Zheng Zhu", "Jing Liu", "Nianfeng Liu", "Yan Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.00664", "pdf_url": "https://arxiv.org/pdf/2606.00664", "code_url": "", "clusters": ["C8"], "keywords": ["sparse keyframes", "interpolation", "embodied world model"], "model_class": "world model", "techniques": ["robot-aware keyframe selection", "sparse-to-dense interpolation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Generates only task-relevant keyframes (approach, contact, grasp, release) and interpolates the rest, avoiding frame-by-frame rollout cost.", "relevance": 4, "relevance_reason": "Temporal sparsity for robot world models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2609.20648", "title": "SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation", "authors": ["Kaivalya Agrawal", "Md Ashiqur Rahman", "Raymond A. Yeh", "Zachary Kingston"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20648", "pdf_url": "https://arxiv.org/pdf/2609.20648", "code_url": "", "clusters": ["C1"], "keywords": ["classical planning", "VLA step skipping", "energy"], "model_class": "VLA", "techniques": ["hybrid planning", "reduced VLA queries"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 2.5x faster task completion", "metric_measures": "other", "benchmarks": ["LIBERO", "real YAM arm"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses a motion planner for free-space moves and queries the VLA only for contact-rich skills, reusing its backbone to predict planner targets.", "relevance": 4, "relevance_reason": "Reduces VLA invocations.", "foundational_technique": false, "verified": true, "found_via": "search query: fast slow VLA"}
{"id": "2606.23444", "title": "SkyJEPA: Learning Long-Horizon World Models for Zero-Shot Sim-to-Real Control of Quadrotors", "authors": ["P. Rao", "Wan-Cong Zhang", "Randall Balestriero", "Yann LeCun", "Giuseppe Loianno"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.23444", "pdf_url": "https://arxiv.org/pdf/2606.23444", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["quadrotor", "JEPA", "embedded"], "model_class": "world model", "techniques": ["JEPA", "sampling-based optimal control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "SkyJEPA learns a JEPA-style latent dynamics model with a physics-inspired prober and pairs it with sampling-based optimal control for real-time quadrotor control on embedded hardware, trained with simulated data for zero-shot sim-to-real.", "relevance": 4, "relevance_reason": "direct: latent world model running real-time on embedded hardware", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2506.09985, 2510.14783; cited by 2607.26789"}
{"id": "2609.26007", "title": "Skytopia: Monocular Drone Navigation with Action-Conditioned Latent World Models", "authors": ["Yu-Hang Zhang", "Rangya Zhang", "Yu-Jing Shang", "Zhuo-Yuan Yu", "Wei-Ying Wang", "Steven Yang", "Q. Yan", "Chao Yan", "Mir Feroskhan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.26007", "pdf_url": "https://arxiv.org/pdf/2609.26007", "code_url": "", "clusters": ["C8"], "keywords": ["drone navigation", "latent world model"], "model_class": "world model", "techniques": ["latent world model", "predictor discarded at inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "59.4% inference cost removed by discarding predictor", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Skytopia trains a drone navigation policy on an action-conditioned latent world model but drops the predictor at deployment, keeping the learned representation while cutting inference cost.", "relevance": 4, "relevance_reason": "direct: world-model training benefit without inference cost", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2504.16054"}
{"id": "10.1109/icassp55912.2026.11464271", "title": "Sliding-Cache VLA: Training-Free Acceleration of Vision Language Action Models via Foreground-Background Decoupling", "authors": ["Yunqing Jiang", "Xiangyuan Chen", "Huijia Zhao", "Tian Liang", "Yan-Yan Liang", "Chio-in Ieong"], "year": 2026, "venue": "ICASSP 2026", "type": "paper", "url": "https://doi.org/10.1109/icassp55912.2026.11464271", "pdf_url": "", "code_url": "", "clusters": ["C1"], "keywords": ["Sliding-Cache VLA"], "model_class": "VLA", "techniques": ["KV/feature caching", "training-free acceleration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free VLA acceleration that separates foreground from background visual input and reuses cached computation for the redundant background, cutting real-time latency without retraining.", "relevance": 4, "relevance_reason": "verified 2026-09-30 by orchestrator: abstract via OpenAlex (DOI). Directly on training-free VLA inference acceleration.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.09771", "title": "SLIM-0.5B: Learning Action-Grounded Predictive Latents for Robot Manipulation", "authors": ["Jingkai Wang", "Zihan Tang", "Gu Zhang", "Mingyu Cao", "Jiapeng Chen", "Jingjiao Zhao", "Xiansheng Chen", "Pengwei Wang", "Lemao Liu", "Dejing Dou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09771", "pdf_url": "https://arxiv.org/pdf/2608.09771", "code_url": "", "clusters": ["C8"], "keywords": ["compact latent interaction policy", "0.5B", "action-grounded latents"], "model_class": "world-action model", "techniques": ["self-supervised masked trajectory prediction", "future-latent + action reconstruction"], "base_model": "", "params": "0.5B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "0.5B latent interaction model learning action-grounded predictive latents instead of using a large VLM backbone or pixel world model.", "relevance": 4, "relevance_reason": "Compact world-model-style policy.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.33575", "title": "SLIP-VLA: Single-Step Latent Imagination for Policy Learning in Vision-Language-Action Models", "authors": ["Tianfu Li", "Haoxuan Xu", "Wenbo Chen", "Haitian Li", "Changchuan Yang", "Xinhu Zheng", "Jun Ma", "Yuan Liu", "Lujia Wang", "Haoang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33575", "pdf_url": "https://arxiv.org/pdf/2609.33575", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["single-step latent imagination", "future-aware VLA"], "model_class": "VLA", "techniques": ["one denoising update for dense future latents", "perceptual sufficiency regularization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Gives a VLA future awareness with a single denoising update producing dense future latents, avoiding multi-step imagination cost while closing the gap to multi-step methods.", "relevance": 4, "relevance_reason": "Cheap imagination inside VLA policies.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.20458", "title": "Slow Brain, Fast Planner: Latency-Resilient VLM-Augmented Urban Navigation", "authors": ["Z. Peng", "Hong-Lin He", "Quanyi Li", "Yu-Kai Ma", "Bo-Lei Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20458", "pdf_url": "https://arxiv.org/pdf/2606.20458", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["VLM", "navigation", "network latency"], "model_class": "other", "techniques": ["latency-resilient fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">80% success with delays up to 5 s (simulation)", "metric_measures": "closed-loop control", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Fuses stale VLM trajectory selections into a real-time planner score with decay, keeping navigation robust under large VLM delays.", "relevance": 4, "relevance_reason": "Explicitly handles slow VLM latency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2502.01828"}
{"id": "2609.22854", "title": "SmoLSTM: A Compact Vision-Language-Action Model with Recurrent Memory that Persists", "authors": ["Jan-Gerrit Habekost", "Parsa Mastouri Kashani", "Connor Gäde", "Matthias Kerzel", "Philipp Allgeuer", "Cornelius Weber", "Stefan Wermter", "Jae Hee Lee", "C. Weber"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22854", "pdf_url": "https://arxiv.org/pdf/2609.22854", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["SmolVLM", "memory", "compact VLA"], "model_class": "VLA", "techniques": ["recurrent memory", "small backbone", "frozen small backbone"], "base_model": "SmolVLM-256M", "params": "0.04B trainable", "hardware": "", "reported_metric": "0.04B trainable parameters; 77.5% full-task success on LIBERO-Mem", "metric_measures": "other", "benchmarks": ["LIBERO-Mem", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Frozen 256M SmolVLM backbone with a matrix-memory LSTM layer that never resets, giving O(1) memory in episode length and a flow-matching chunk head.", "relevance": 4, "relevance_reason": "Compact memory VLA with 0.04B trainable parameters.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=0]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2605.24433", "title": "Smoother Action Chunking Flow Policy via Prior-Corrected Orthogonal Trust-Region Guidance", "authors": ["Kai Fang", "Hailong Pei", "Xuemin Chi", "Hai-Long Pei", "Xue-Min Chi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24433", "pdf_url": "https://arxiv.org/pdf/2605.24433", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["RTC", "chunk boundary", "smoothness", "action chunking", "chunk-boundary smoothness", "real-time chunking"], "model_class": "VLA", "techniques": ["real-time chunking", "flow matching guidance", "trust region", "RTC guidance"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Improves real-time chunking guidance for flow policies by rescaling the guidance weight with a data prior and trust-region constraining the orthogonal correction, reducing chunk-boundary jerk.", "relevance": 4, "relevance_reason": "Chunk-boundary continuity under RTC-style execution.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2601.03044", "title": "SOP: A Scalable Online Post-Training System for Vision-Language-Action Models", "authors": ["Mingjie Pan", "Siyuan Feng", "Qinglin Zhang", "Xinchen Li", "Jianheng Song", "Chendi Qu", "Yi Wang", "Chuankang Li", "Ziyu Xiong", "Zhi Chen", "Yi Liu", "Jianlan Luo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.03044", "pdf_url": "https://arxiv.org/pdf/2601.03044", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["fleet online post-training", "cloud learner", "HG-DAgger", "RECAP"], "model_class": "VLA", "techniques": ["fleet-to-cloud experience streaming", "asynchronous policy updates", "online RL/IL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "performance scales near-linearly with number of robots", "metric_measures": "other", "benchmarks": ["real-world cloth folding, box assembly, grocery restocking"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "A fleet of robots streams on-policy experience and interventions to a central cloud learner and asynchronously receives updated VLA weights, enabling online multi-task post-training within hours.", "relevance": 4, "relevance_reason": "Fleet-cloud learning loop architecture.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud AND abs:vision-language-action"}
{"id": "2604.19710", "title": "SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery Samples for Vision-Language-Action Model", "authors": ["Zewei Zhou", "Ruining Yang", "Xuewei", "Qi", "Yiluan Guo", "Sherry X. Chen", "Tao Feng", "Kateryna Pistunova", "Yishan Shen", "Lili Su", "Jiaqi Ma", "Xue-Wei Qi", "Yi-Luan Guo", "Sherry X Chen", "Jia-Qi Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.19710", "pdf_url": "https://arxiv.org/pdf/2604.19710", "code_url": "", "clusters": ["C1"], "keywords": ["driving VLA", "autonomous driving", "latency"], "model_class": "VLA", "techniques": ["flow-matching action expert", "GRPO", "trajectory initialization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SpanVLA pairs autoregressive reasoning with a flow-matching action expert initialized from historical trajectories to cut action-generation latency in driving VLAs, plus GRPO post-training on negative/recovery samples.", "relevance": 4, "relevance_reason": "direct: reduces driving-VLA inference latency", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2106.09685, 2502.19645"}
{"id": "2608.16172", "title": "SparkVLA: Stop-Aware Hierarchical VLA with Adaptive Action Chunking for Long-Horizon Manipulation", "authors": ["Xunyao Lei", "Renjun Wu", "Tianlin Huo", "Xuesong Li", "Xun-Yao Lei", "Xue-Song Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.16172", "pdf_url": "https://arxiv.org/pdf/2608.16172", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["stop-aware", "action prefix selection", "hierarchical VLA"], "model_class": "VLA", "techniques": ["adaptive chunk execution", "stop prediction", "visual token pruning", "caching", "adaptive action chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "47.12% success on RoboCerebra", "metric_measures": "other", "benchmarks": ["RoboCerebra", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Hierarchical VLA that ranks \"stop\" against every executable action-prefix length in one candidate set, jointly deciding subtask termination and how much of a chunk to execute, with anchor-guided token pruning.", "relevance": 4, "relevance_reason": "Directly learns how far into a chunk to commit before re-observing.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]; snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2601.12894", "title": "Sparse ActionGen: Accelerating Diffusion Policy with Real-time Pruning", "authors": ["Kangye Ji", "Yuan Meng", "Jian-Bo Zhou", "Ye Li", "Han-Yun Cui", "Zhi Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.12894", "pdf_url": "https://arxiv.org/pdf/2601.12894", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["runtime pruning", "activation reuse", "observation-conditioned pruner"], "model_class": "diffusion/flow policy", "techniques": ["dynamic computation pruning", "feature caching"], "base_model": "Diffusion Policy", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns an observation-conditioned pruner that decides which diffusion computations to skip and reuses cached activations, adapting sparsity to the rollout.", "relevance": 4, "relevance_reason": "Structured runtime pruning of diffusion policies.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2410.21257"}
{"id": "2603.20658", "title": "Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied Manipulation", "authors": ["Zhi-Chao Wu", "Junyin Ye", "Zhilong Zhang", "Yihao Sun", "Hao-Chu Lin", "Jia Luo", "Hao-Xiang Ren", "Lei Yuan", "Yang Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.20658", "pdf_url": "https://arxiv.org/pdf/2603.20658", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["execution speedup", "plug-and-play"], "model_class": "other", "techniques": ["action-chunk downsampling", "offline RL scheduler", "world-model surrogate"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.8x execution speedup", "metric_measures": "other", "benchmarks": ["LIBERO", "BiGym"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Policy-agnostic scheduler learned offline that adaptively downsamples action chunks to speed up execution, using a world model to constrain state deviation.", "relevance": 4, "relevance_reason": "Directly accelerates embodied policy execution.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.22888", "title": "Stable and Efficient Real-World Online VLA Post-Training via Asynchronous Replay-Anchored Policy Improvement", "authors": ["Jiarui Yang", "Jiajin Zhang", "Bin Zhu", "Jingjing Chen", "Yu-Gang Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22888", "pdf_url": "https://arxiv.org/pdf/2609.22888", "code_url": "", "clusters": ["C3"], "keywords": ["RAPolicy", "asynchronous replay-anchored", "one-step flow actor"], "model_class": "VLA", "techniques": ["asynchronous online RL", "advantage-weighted likelihood"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Concurrent rollout and learning for real-world online VLA post-training with replay-anchored critic/actor updates that cut compute.", "relevance": 4, "relevance_reason": "Efficient real-world online post-training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"post-training\" AND abs:\"vision-language-action\" AND abs:\"compute\""}
{"id": "2609.36471", "title": "Staircase Policy: Streaming Inference for World-Action Models with Large Action Chunks", "authors": ["Guoheng Sun", "Chen Chen", "Jin Wang", "Ang Li", "Teresa Lv"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36471", "pdf_url": "https://arxiv.org/pdf/2609.36471", "code_url": "", "clusters": ["C2", "C3", "C4", "C7"], "keywords": ["world-action model", "streaming", "time-to-first-action", "executed actions per second", "Staircase Policy", "staggered denoising", "streaming WAM", "large chunks"], "model_class": "world-action model", "techniques": ["staggered denoising", "sub-chunk streaming", "sub-chunks at staggered noise levels", "future latent re-prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "292.7 executed actions/s (3.62x); time-to-first-action 123.6 to 73.3 ms", "metric_measures": "throughput", "benchmarks": ["LIBERO", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Splits large action chunks into sub-chunks at staggered denoising stages, executing near-term actions immediately and re-predicting future latents at each boundary to refresh stale actions.", "relevance": 4, "relevance_reason": "Illustrates distinct throughput vs latency metrics.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA inference performance; arXiv search: real-time chunking"}
{"id": "2606.19774", "title": "Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection", "authors": ["Trong-Bao Ho", "Quang-Tan Nguyen", "Thien-Loc Ha", "Gia-Binh Nguyen", "Viet-Thanh Nguyen", "Long Dinh", "Minh N. Vu", "Duy M. H. Nguyen", "An Thai Le", "Ngo Anh Vien"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.19774", "pdf_url": "https://arxiv.org/pdf/2606.19774", "code_url": "", "clusters": ["C1", "C2", "C6", "C7"], "keywords": ["initial noise selection", "asynchronous execution", "training-free", "chunk consistency", "PAINT", "flow policy"], "model_class": "diffusion/flow policy", "techniques": ["noise inversion", "asynchronous inference", "training-free"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["12 simulated benchmarks", "6 real-world tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Achieves chunk-boundary consistency for async flow policies by choosing the initial noise (via backward Euler inversion) so the unmodified ODE reproduces the executed prefix, avoiding gradient-based guidance.", "relevance": 4, "relevance_reason": "Training-free alternative to RTC inpainting.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; snowball from 2506.07339; arXiv API all:\"action chunking\" AND latency"}
{"id": "10.1109/isoirs70157.2026.11545245", "title": "State-Space Modeling for Action Generation in Embodied Agents", "authors": ["Wei-Xuan Liu", "Jin-Tao Li", "Ao Li", "Shuxue Quan", "Shui Yu", "Yun Li"], "year": 2026, "venue": "2026 6th International Symposium on Intelligent Robotics and Systems (ISoIRS)", "type": "paper", "url": "https://doi.org/10.1109/ISoIRS70157.2026.11545245", "pdf_url": "", "code_url": "", "clusters": ["C1"], "keywords": ["SSM", "linear time", "on-device"], "model_class": "VLA", "techniques": ["state space model", "linear-time action expert", "cache-free"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Meta-World", "LIBERO", "RoboTwin"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cross-SSM replaces the cross-attention action expert of VLAs with a cache-free linear-time state-space model, giving predictable memory and lower per-step latency for on-device use.", "relevance": 4, "relevance_reason": "direct: efficient action-expert architecture for on-device deployment", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.28927", "title": "Streaming-WAM: Action-Conditioned World-Action Model for Asynchronous Robot Manipulation", "authors": ["Xuyao Huang", "Yixuan Wang", "Zengyao Ye", "Boyuan Zhao", "Chenyang Yu", "Haoran Wen", "Zhijie Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28927", "pdf_url": "https://arxiv.org/pdf/2609.28927", "code_url": "", "clusters": ["C2", "C4", "C7", "C8"], "keywords": ["Streaming-WAM", "committed-action conditioning", "asynchronous world-action model", "streaming WAM", "committed actions", "asynchronous control"], "model_class": "world-action model", "techniques": ["action-conditioned future prediction on committed prefix", "joint streaming update"], "base_model": "", "params": "", "hardware": "", "reported_metric": "98.35% avg LIBERO success | episode time 2.93x shorter than Fast-WAM on LIBERO; real task 90 s -> 38 s", "metric_measures": "other", "benchmarks": ["LIBERO", "real robot (Stamp Paper)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Makes asynchronous WAM execution consistent by conditioning future visual prediction on the actions already committed during inference, then generating the rest of the chunk from those predictions.", "relevance": 4, "relevance_reason": "Correct asynchronous inference for high-latency WAMs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: real-time chunking; search query"}
{"id": "2603.28565", "title": "StreamingVLA: Streaming Vision-Language-Action Model with Action Flow Matching and Adaptive Early Observation", "authors": ["Yiran Shi", "Dongqi Guo", "Tianchen Zhao", "Feng Gao", "Liangzhi Shi", "Chao Yu", "ZhiJian Mo", "Qihua Xiao", "XiaoShuai Peng", "Qingmin Liao", "Yu Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.28565", "pdf_url": "https://arxiv.org/pdf/2603.28565", "code_url": "", "clusters": ["C1", "C2", "C3", "C7"], "keywords": ["streaming VLA", "action flow matching", "adaptive early observation", "StreamingVLA", "early observation", "stage overlap"], "model_class": "VLA", "techniques": ["streaming inference", "asynchronous stages", "action flow matching", "adaptive early observation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Overlaps observation, generation and execution by learning action flows instead of chunk denoising and triggering observations early based on action saliency.", "relevance": 4, "relevance_reason": "Latency hiding for edge VLA deployment.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk latency VLA; arXiv API all:\"action chunking\" AND latency"}
{"id": "2602.01100", "title": "StreamVLA: Breaking the Reason-Act Cycle via Completion-State Gating", "authors": ["Tongqing Chen", "Hang Wu", "Jiasen Wang", "Xiaotao Li", "Lu Fang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.01100", "pdf_url": "https://arxiv.org/pdf/2602.01100", "code_url": "", "clusters": ["C1"], "keywords": ["completion-state gating", "System 2 triggering", "flow action head"], "model_class": "VLA", "techniques": ["gated reasoning", "dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "skips AR decoding for 72% of timesteps; 48% lower latency", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Only triggers slow reasoning and goal imagination at sub-task transitions, locking intent to condition a flow head the rest of the time.", "relevance": 4, "relevance_reason": "Adaptive reasoning frequency.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "2606.00267", "title": "StressDream: Steering Video World Models for Robust Policy Evaluation and Improvement", "authors": ["Junwon Seo", "Sushant Veer", "Ran Tian", "Wenhao Ding", "Apoorva Sharma", "Karen Leung", "Edward Schmerling", "Marco Pavone", "Andrea Bajcsy"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.00267", "pdf_url": "https://arxiv.org/pdf/2606.00267", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["stress testing", "noise optimization", "robust policy evaluation"], "model_class": "world model", "techniques": ["initial-noise optimization of diffusion world model toward target events", "OOD-noise regularization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Steers diffusion world-model imaginations toward plausible high-impact outcomes by optimizing initial noise, surfacing rare failures without drawing huge numbers of samples.", "relevance": 4, "relevance_reason": "Efficient rare-failure search with world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2605.19319", "title": "SWEET: Sparse World Modeling with Image Editing for Embodied Task Execution", "authors": ["Yiren Song", "Yihan Wang", "Xiyao Deng", "Zhuoran Yan", "Mike Zheng Shou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.19319", "pdf_url": "https://arxiv.org/pdf/2605.19319", "code_url": "", "clusters": ["C8"], "keywords": ["sparse world model", "image editing", "task keyframes"], "model_class": "world model", "techniques": ["image-editing model predicts task-level future keyframes"], "base_model": "FLUX-Kontext (vs Wan2.2)", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows an image-editing model predicting a few task-level keyframes is a cheaper and more reliable visual world model for manipulation than dense video generation.", "relevance": 4, "relevance_reason": "Sparse visual foresight for efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2603.16666 (citations)"}
{"id": "2609.36720", "title": "T$^2$Mem: Learning Test-Time Memory for Robotics", "authors": ["Yize Liu", "Huang Huang", "Yining Hong", "Zijian Du", "Zhi Cao", "Li Fei-Fei", "Jiajun Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36720", "pdf_url": "https://arxiv.org/pdf/2609.36720", "code_url": "", "clusters": ["C1", "C4", "C9"], "keywords": ["test-time training", "fast weights", "memory", "VLA"], "model_class": "VLA", "techniques": ["test-time training memory", "compact fast-weight history encoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "T2Mem gives a memory-free VLA memory by compressing observation history into fast weights via online self-supervised test-time updates, avoiding reprocessing the full history.", "relevance": 4, "relevance_reason": "Compute-efficient memory for VLAs; speaker Jiajun Wu.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "2609.05266", "title": "TacPAC: Tactile Prediction and Real-Time Action Correction in World-Action Models for Contact-Rich Manipulation", "authors": ["Zipei Ma", "Xiaofei Wei", "Junzhe Jiang", "Shunlin Lu", "Li Zhang", "Zi-Pei Ma", "Xiao-Fei Wei", "Jun-Zhe Jiang", "Shun-Lin Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.05266", "pdf_url": "https://arxiv.org/pdf/2609.05266", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["tactile correction", "cached plan", "contact-rich WAM", "tactile prediction", "real-time correction", "world-action model"], "model_class": "world-action model", "techniques": ["cached predicted-contact representation", "single-pass tactile correction of remaining actions", "real-time tactile action correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "one correction pass 20.7x cheaper than regenerating the chunk", "metric_measures": "relative latency", "benchmarks": ["5 real contact-rich tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Corrects not-yet-executed actions in real time by comparing incoming tactile images with the contact the WAM plan anticipated, at 1/20 the cost of re-planning.", "relevance": 4, "relevance_reason": "Cheap reactive correction for slow WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations); venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2606.07895", "title": "TBD-VLA: Temporal Block Diffusion Vision Language Action Model", "authors": ["Sung-Wook Lee", "Xuhui Kang", "Yen-Ling Kuo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.07895", "pdf_url": "https://arxiv.org/pdf/2606.07895", "code_url": "", "clusters": ["C1", "C3", "C9"], "keywords": ["temporal block diffusion", "parallel decoding", "RTC inpainting", "block diffusion", "discrete VLA"], "model_class": "VLA", "techniques": ["block diffusion decoding", "asynchronous execution", "temporal block diffusion decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "TBD-VLA decodes discrete action tokens in temporal blocks via diffusion, combining parallel speed with modeled token dependencies.", "relevance": 4, "relevance_reason": "Faster discrete action decoding.", "foundational_technique": false, "verified": true, "found_via": "search query: parallel decoding action; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2607.04171", "title": "Teaching Tiny VLA Models Where to Look and How to Move", "authors": ["Iok Tong Lei", "Ying Jie Yap", "Wei Huang", "Qingchen Xie", "Qianzhi Li", "Yujie Zhang", "Xiaolong Liu", "Zhidong Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.04171", "pdf_url": "https://arxiv.org/pdf/2607.04171", "code_url": "", "clusters": ["C1", "C2", "C9"], "keywords": ["tiny VLA", "spatial distillation", "latent flow matching", "SmolVLM2-0.25B", "spatial grounding", "XS-VLA"], "model_class": "VLA", "techniques": ["distillation", "small backbone", "knowledge distillation from VLM", "auxiliary grounding supervision", "action coherence"], "base_model": "SmolVLM2-0.25B (teacher Qwen3-VL-4B)", "params": "~0.25B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills coarse spatial grounding labels from a 4B VLM into a 0.25B VLA and adds latent flow matching, lifting tiny-VLA LIBERO-Long success substantially.", "relevance": 4, "relevance_reason": "Teacher-student recipe to recover capability in tiny VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: tiny VLA; web search: distilling VLA into small student; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2607.29482", "title": "Temporal Policy: History-Initialized Action Generation for Robotic Learning from Demonstration", "authors": ["Dylan Miller", "Martin Jägersand"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.29482", "pdf_url": "https://arxiv.org/pdf/2607.29482", "code_url": "", "clusters": ["C1"], "keywords": ["stochastic interpolants", "low-latency"], "model_class": "diffusion/flow policy", "techniques": ["history-initialized transport"], "base_model": "", "params": "", "hardware": "NVIDIA RTX 4080", "reported_metric": "19.1 ms inference latency", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Initializes generative action flow from recent history to straighten transport and cut inference cost.", "relevance": 4, "relevance_reason": "Low-latency generative policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "jetson-ai-lab-tensorrt-edge-llm", "title": "TensorRT Edge-LLM on Jetson", "authors": ["Aditya Sahu"], "year": 2026, "venue": "NVIDIA Jetson AI Lab tutorial", "type": "blog", "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm/", "pdf_url": "", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["TensorRT Edge-LLM", "C++ runtime", "Jetson Orin", "Jetson Thor", "NVFP4", "FP8", "INT4 AWQ", "Alpamayo R1"], "model_class": "other", "techniques": ["TensorRT engine compilation", "INT4 AWQ", "FP8", "NVFP4"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin, Orin NX, Orin Nano, Jetson Thor", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "NVIDIA's Python-free C++ TensorRT runtime for LLMs/VLMs on Jetson, supporting INT4 AWQ on Orin and FP8/NVFP4 on Thor, and listing the Alpamayo R1 VLA among supported models.", "relevance": 4, "relevance_reason": "Official edge runtime used for VLA backbones on Jetson; year inferred from site context (no date on page).", "foundational_technique": false, "verified": true, "found_via": "web: Jetson Thor NVFP4 FP8 humanoid onboard"}
{"id": "2605.13316", "title": "Test-time Sparsity for Extreme Fast Action Diffusion", "authors": ["Kangye Ji", "Yuan Meng", "Jianbo Zhou", "Ye Li", "Chen Tang", "Zhi Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.13316", "pdf_url": "https://arxiv.org/pdf/2605.13316", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["test-time sparsity", "feature reuse", "parallel pruner", "residual pruning", "DiT policy"], "model_class": "diffusion/flow policy", "techniques": ["dynamic sparsity", "feature caching", "dynamic pruning", "caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Predicts prunable residual computations per forward pass of an action diffusion transformer, with a parallel pipeline and omnidirectional feature reuse for aggressive sparsity.", "relevance": 4, "relevance_reason": "Dynamic pruning of action diffusion.", "foundational_technique": false, "verified": true, "found_via": "search query: feature caching diffusion policy; snowball from 2410.21257"}
{"id": "2609.34319", "title": "Text-Vision Synergistic Token Caching: A Training-Free Framework for Efficient Vision-Language-Action Inference", "authors": ["Qianer Li", "Chengjie Zhang", "Jingwen Chen", "Zanjia Tong", "Jiyuan Zhang", "Hong Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34319", "pdf_url": "https://arxiv.org/pdf/2609.34319", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["token caching", "attention head filtering", "training-free", "TVCache", "text-vision synergy"], "model_class": "VLA", "techniques": ["token caching", "KV cache reuse", "attention head filtering"], "base_model": "OpenVLA-OFT and 3 others", "params": "", "hardware": "", "reported_metric": "2.45x FLOPs reduction at 12.5% retention; +14.5pp over VLA-Cache", "metric_measures": "other", "benchmarks": ["LIBERO", "SIMPLER", "real robot", "two sim benchmarks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Improves VLA token caching by selecting reliable attention heads via text-vision focus and choosing which layers to reuse via entropy differences.", "relevance": 4, "relevance_reason": "Follow-up to VLA-Cache.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; search query: token caching VLA"}
{"id": "2607.27933", "title": "The Geometry of Flow-Matching Uncertainty: A Cost-free Uncertainty Proxy and Its Application in Flow-based VLA Failure Detection", "authors": ["Zi-Yang Rao", "Yi-Ren Zhao", "Wei-Yu Guo", "Ben Fei", "Yan-Dong Guo", "Hui Xiong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.27933", "pdf_url": "https://arxiv.org/pdf/2607.27933", "code_url": "https://github.com/rrrrrrzy/fm", "clusters": ["C7", "C8"], "keywords": ["flow-matching uncertainty", "denoising acceleration", "cost-free proxy"], "model_class": "VLA", "techniques": ["single-pass trajectory-bending uncertainty"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Interprets flow-matching uncertainty geometrically and proposes denoising acceleration, a free single-pass proxy that detects failing VLA rollouts early without resampling or training.", "relevance": 4, "relevance_reason": "Zero-overhead uncertainty for real-time monitoring.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2503.08558 (S2)"}
{"id": "2026.00044", "title": "The On-Robot Computing Needs of Embodied AI", "authors": ["Shao-Shan Liu"], "year": 2026, "venue": "IEEE MOST 2026", "type": "paper", "url": "https://doi.org/10.1109/MOST69733.2026.00044", "pdf_url": "https://arxiv.org/pdf/DOI:10.1109/MOST69733.2026.00044", "code_url": "", "clusters": ["C5"], "keywords": ["on-robot computing", "heterogeneous workloads", "thermal budget", "cross-layer co-design"], "model_class": "other", "techniques": ["position paper"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Position paper decomposing on-robot compute into servo loops, sensing, whole-body control and foundation models, arguing onboard compute architecture and distribution is the central bottleneck.", "relevance": 4, "relevance_reason": "Framing of onboard compute requirements across layers.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2407.04292 (citations)"}
{"id": "2602.19260", "title": "The Price Is Not Right: Neuro-Symbolic Methods Outperform VLAs on Structured Long-Horizon Manipulation Tasks with Significantly Lower Energy Consumption", "authors": ["T. Duggan", "Pierrick Lorang", "Hong Lu", "Matthias Scheutz"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.19260", "pdf_url": "https://arxiv.org/pdf/2602.19260", "code_url": "", "clusters": ["C4"], "keywords": ["energy consumption", "neuro-symbolic"], "model_class": "VLA", "techniques": [], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "~two orders of magnitude more training energy for VLA", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Compares fine-tuned pi0 with a neuro-symbolic planner on Towers of Hanoi, measuring success and energy; VLA fine-tuning uses ~100x more energy.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2607.15621", "title": "Think at 5 Hz, Act at 20 Hz: Asynchronous Fast-Slow Vision-Language-Action Inference for Closed-Loop Driving", "authors": ["Yun Li", "Jiachen Gong", "Simon Thompson", "Ehsan Javanmardi", "Qunli Zhang", "Zifan Zeng", "Shiming Liu", "Peng Wang", "Zixuan Guo", "Manabu Tsukada"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.15621", "pdf_url": "https://arxiv.org/pdf/2607.15621", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["autonomous driving", "LLM latency", "per-tick freshness"], "model_class": "VLA", "techniques": ["asynchronous fast-slow inference"], "base_model": "", "params": "", "hardware": "single consumer GPU", "reported_metric": "32 ms per-tick model cost; 20 Hz action / 5 Hz reasoning", "metric_measures": "closed-loop control", "benchmarks": ["closed-loop driving simulation"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Runs an LLM driving reasoner at 5 Hz and a lightweight action expert at 20 Hz so every control tick uses fresh observations.", "relevance": 4, "relevance_reason": "Async fast-slow VLA for real-time control", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.15099", "title": "Think Less, Act Early: Reinforced Latent Reasoning with Early Exit in Vision-Language-Action Models", "authors": ["Dianqiao Lei", "Lianlei Shan"], "year": 2026, "venue": "ICML 2026", "type": "preprint", "url": "https://arxiv.org/abs/2606.15099", "pdf_url": "https://arxiv.org/pdf/2606.15099", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["latent reasoning", "early exit", "RL"], "model_class": "VLA", "techniques": ["latent reasoning", "early exit", "latent CoT", "adaptive early exit"], "base_model": "", "params": "", "hardware": "", "reported_metric": "6x inference speedup vs explicit CoT", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces text CoT with RL-denoised latent reasoning variables and stops reasoning early when state confidence is high.", "relevance": 4, "relevance_reason": "Reasoning compute reduction.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA early exit; venue sweep: arXiv API VLA AND (MoE OR early exit)"}
{"id": "2602.06575", "title": "Think Proprioceptively: State-Grounded Visual Token Selection for VLA Policies", "authors": ["Fangyuan Wang", "Peng Zhou", "Jiaming Qi", "Shipeng Lyu", "Chengyang He", "David Navarro-Alarcon", "Guodong Guo", "Fang-Yuan Wang", "Jia-Ming Qi", "Cheng-Yang He", "D. Navarro-Alarcón"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.06575", "pdf_url": "https://arxiv.org/pdf/2602.06575", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["token pruning", "ThinkProprio", "proprioception tokens"], "model_class": "VLA", "techniques": ["visual token selection", "proprioceptive gating"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~12% visual tokens retained; lower end-to-end latency", "metric_measures": "other", "benchmarks": ["CALVIN", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Turns proprioception into VLM vocabulary tokens that, with the instruction, gate which visual patches enter the VLM; keeping ~12% of tokens still improves performance.", "relevance": 4, "relevance_reason": "State-grounded early token selection.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]; snowball round 2: cites 2306.03310, 2411.19650"}
{"id": "2602.02459", "title": "TIC-VLA: A Think-in-Control Vision-Language-Action Model for Robot Navigation in Dynamic Environments", "authors": ["Zhiyu Huang", "Yun Zhang", "Johnson Liu", "Rui Song", "Chen Tang", "Jia-Qi Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.02459", "pdf_url": "https://arxiv.org/pdf/2602.02459", "code_url": "", "clusters": ["C1", "C6", "C9"], "keywords": ["delayed reasoning", "latency-aware", "navigation"], "model_class": "VLA", "techniques": ["delayed semantic conditioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "TIC-VLA explicitly models delayed semantic reasoning when generating actions so navigation stays reactive in dynamic scenes.", "relevance": 4, "relevance_reason": "Latency-aware VLA design.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2601.14945", "title": "TIDAL: Temporally Interleaved Diffusion and Action Loop for High-Frequency VLA Control", "authors": ["Yuteng Sun", "Haoran Wang", "Ruofei Bai", "Zhengguo Li", "Jun Li", "Meng Yee Michael Chuah", "Wei Yun Yau"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.14945", "pdf_url": "https://arxiv.org/pdf/2601.14945", "code_url": "", "clusters": ["C1", "C2", "C4", "C7"], "keywords": ["dual-frequency", "interleaved diffusion", "staleness-aware", "open-loop blind spot", "feedback frequency", "TIDAL", "dual-frequency loop", "staleness-aware compensation", "dynamic interception"], "model_class": "VLA", "techniques": ["dual-rate control", "single-step flow integration", "interleaved flow integration", "semantic caching", "staleness-aware training", "hierarchical dual-frequency scheduling", "temporally misaligned training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "4x higher feedback frequency | 4x feedback frequency increase; 2.5x gain on dynamic interception", "metric_measures": "closed-loop control", "benchmarks": ["dynamic interception tasks"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Caches semantic embeddings in a slow loop while a fast loop interleaves single flow-integration steps with execution, and trains with deliberately stale intents so the policy compensates for latency shift.", "relevance": 4, "relevance_reason": "Makes the open-loop vs closed-loop rate distinction explicit.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; WebSearch: closed-loop vs open-loop control frequency VLA; arXiv API all:\"action chunking\" AND latency"}
{"id": "2607.16806", "title": "Token-Wise Latent Streaming from Slow Reasoners to Fast Planners for Dynamic Vision Language Navigation", "authors": ["Tianshuai Hu", "Yangyi Zhong", "Zeying Gong", "Ling-Dong Kong", "Xiaodong Mei", "Guo-Yang Zhao", "Xiaolu Liu", "Song Wang", "Rong Li", "Junwei Liang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.16806", "pdf_url": "https://arxiv.org/pdf/2607.16806", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["SPARK-VLN", "latent streaming", "slow reasoner fast planner", "observation staleness"], "model_class": "VLA", "techniques": ["streaming intermediate VLM hidden states to fast planner"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Streams intermediate hidden states of a slow VLM reasoner to a fast flow-matching planner during token generation so social navigation acts on fresh guidance, reducing safety-critical staleness.", "relevance": 4, "relevance_reason": "Explicitly frames staleness as safety-critical.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2608.00315", "title": "Towards General Language-Conditioned Latent Safety Filters", "authors": ["Ihab Tabbara", "Yuxuan Yang", "Hussein Sibai"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00315", "pdf_url": "https://arxiv.org/pdf/2608.00315", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["language-conditioned safety filter", "latent HJ reachability", "HJ actor-critic", "latent", "latent safety filter", "language-conditioned constraints"], "model_class": "world model", "techniques": ["language-conditioned HJ actor-critic", "language-conditioned HJ safety actor/critic in latent space", "latent reachability safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["pick-and-place", "table wiping"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Conditions latent HJ safety filters on natural-language constraints so one filter can enforce different user-specified safety requirements.", "relevance": 4, "relevance_reason": "Generalizable latent safety filters.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" \"vision-language-action\"; snowball from 2502.00935 (citations); venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2607.14852", "title": "Towards Human-like Physical Intelligence: Lifelong Vision-Language-Action Learning for Robotic Manipulation", "authors": ["Yao He", "Gan Sun", "Wenqi Liang", "Fazeng Li", "Yang Cong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.14852", "pdf_url": "https://arxiv.org/pdf/2607.14852", "code_url": "", "clusters": ["C3"], "keywords": ["LifelongVLA", "dual-timescale LoRA", "cache-efficient replay"], "model_class": "VLA", "techniques": ["LoRA gating", "replay"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Lifelong VLA learning with short- and long-term LoRA adapters gated per task and a cache-efficient replay for low-cost deployment.", "relevance": 4, "relevance_reason": "PEFT continual learning with deployment cost focus.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:LoRA"}
{"id": "2604.26694", "title": "Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising", "authors": ["Jun Guo", "Qiwei Li", "Peiyan Li", "Zilong Chen", "Nan Sun", "Yifei Su", "Heyun Wang", "Yuan Zhang", "Xinghang Li", "Huaping Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.26694", "pdf_url": "https://arxiv.org/pdf/2604.26694", "code_url": "", "clusters": ["C8"], "keywords": ["X-WAM", "4D world model", "asynchronous noise sampling"], "model_class": "world-action model", "techniques": ["asynchronous noise sampling (few steps for actions, full for video)", "lightweight depth branch"], "base_model": "", "params": "", "hardware": "", "reported_metric": "action latency 4665 ms -> 1033 ms (4.5x) with asynchronous inference (per search summary)", "metric_measures": "relative latency", "benchmarks": ["RoboCasa", "RoboTwin 2.0"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Unified RGB-D video + action model that decodes actions in a few steps for real-time execution while the video stream keeps its full denoising schedule for 4D synthesis.", "relevance": 4, "relevance_reason": "Asynchronous denoising to decouple action latency from world-model fidelity.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.22794", "title": "UniFS: Unified Fast-to-Slow Hierarchical Architecture for Vision-Language-Action Models", "authors": ["Lin Sun", "Zhiwei Guan", "Conglin Wang", "Zihong Chen", "Jianhai Yu", "Zongsheng Li", "Boyong He", "Tao Sun", "Jiale Cao", "Lige Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.22794", "pdf_url": "https://arxiv.org/pdf/2606.22794", "code_url": "https://github.com/linsun449/UniFS", "clusters": ["C1"], "keywords": ["multi-rate layers", "fast-slow", "hierarchical"], "model_class": "VLA", "techniques": ["multi-frequency layer updates", "feature caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "latency 36.5 ms -> 17.8 ms (2.1x)", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real Franka"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Groups VLM layers into tiers updated at decreasing frequencies so shallow layers track fast dynamics while deep layers cache stable context for the action expert.", "relevance": 4, "relevance_reason": "Fast-slow architecture within one backbone.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "2608.03244", "title": "UniNav: A Unified World-Action Diffusion Model for Visual Navigation", "authors": ["Changqing Zhou", "Yueru Luo", "Zeyu Jiang", "Changhao Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.03244", "pdf_url": "https://arxiv.org/pdf/2608.03244", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["UniNav", "navigation", "world-action diffusion"], "model_class": "world-action model", "techniques": ["joint diffusion", "one-step inference", "token dropping at inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~0.1 s latency with one-step inference (UniNav-Fast)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Unified world-action diffusion model for image-goal navigation; the Fast variant drops future-image tokens at inference and runs one-step.", "relevance": 4, "relevance_reason": "Shows WAM inference can drop video tokens.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2508.05635, 2601.16163"}
{"id": "2605.10821", "title": "UniSteer: Unified Noise Steering for Efficient Human-Guided VLA Adaptation", "authors": ["Junjie Lu", "Xinyao Qin", "Yuhua Jiang", "Kaixin Wang", "Chuheng Zhang", "Bin Liang", "Jun Yang", "Min Xu", "Li Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.10821", "pdf_url": "https://arxiv.org/pdf/2605.10821", "code_url": "", "clusters": ["C3"], "keywords": ["noise-space RL", "human corrections", "diffusion VLA"], "model_class": "VLA", "techniques": ["noise-space RL", "human-in-the-loop"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Maps human action-space corrections into noise space so a lightweight noise actor adapts a frozen diffusion VLA efficiently on-robot.", "relevance": 4, "relevance_reason": "Budget-limited real-world adaptation with frozen VLA.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.37772", "title": "Urgent Actions Go First: Urgency-Aware Denoising for Real-Time VLA Control", "authors": ["Zibo Wang", "Haochen Han", "Pengzhen Ren", "Mingtong Dai", "Fangming Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.37772", "pdf_url": "https://arxiv.org/pdf/2609.37772", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C7"], "keywords": ["urgency-aware denoising", "early release", "receding horizon", "action availability latency"], "model_class": "VLA", "techniques": ["non-uniform denoising schedule", "asynchronous execution", "heterogeneous denoising steps", "overlapped refinement", "heterogeneous denoising budgets", "trajectory reconciliation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.89x speedup in average action availability latency", "metric_measures": "relative latency", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Allocates fewer denoising steps to near-term actions and keeps refining tail actions during execution, with reconciliation and ghost-action corrections to handle early-release errors.", "relevance": 4, "relevance_reason": "Latency-aware denoising for real-time VLA control.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; search query: tail latency robot policy; arXiv API all:\"action chunking\" AND latency"}
{"id": "2609.37250", "title": "V-JEPA Policy: Building Effective World-Action Models on Predictive Visual Latents", "authors": ["Yang Zhang", "Jiangyuan Zhao", "Chenyou Fan", "Jiayu Hu", "Xiu Yuan", "Chenjia Bai", "Xiu Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.37250", "pdf_url": "https://arxiv.org/pdf/2609.37250", "code_url": "", "clusters": ["C8"], "keywords": ["V-JEPA Policy", "frozen predictive latents", "small WAM"], "model_class": "world-action model", "techniques": ["future-latent predictor on frozen V-JEPA 2.1", "KV-conditioned flow-matching action expert"], "base_model": "", "params": "0.9B total (0.6B trainable)", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Builds a 0.9B WAM on a frozen V-JEPA 2.1 latent space without inheriting a video generator, matching larger generator-based WAMs.", "relevance": 4, "relevance_reason": "Latent-JEPA route to small WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2608.08558", "title": "Vid2WAM: Distilling Video Diffusion Priors into World Action Models", "authors": ["Chenhao Qiu", "Ruixiang Wang", "Runyi Zhao", "Sixu Lin", "Songen Gu", "Shufeng Nan", "Guiliang Liu", "Kui Jia", "Yanwei Fu", "Simo Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.08558", "pdf_url": "https://arxiv.org/pdf/2608.08558", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["video prior distillation", "compact WAM student", "inverse dynamics pseudo-actions"], "model_class": "world-action model", "techniques": ["offline distillation from video foundation model", "IDM pseudo-labeling", "source-aware residual action adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Distills future rollouts from a large video model plus IDM pseudo-actions into a compact WAM student; teacher and IDM are dropped at deployment, keeping inference low-latency.", "relevance": 4, "relevance_reason": "Teacher-student distillation from video foundation models into small WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.28489", "title": "Video Generation Models as World Models: Efficient Paradigms, Architectures and Algorithms", "authors": ["Muyang He", "Hanzhong Guo", "Junxiong Lin", "Yizhou Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.28489", "pdf_url": "https://arxiv.org/pdf/2603.28489", "code_url": "https://github.com/Isaachhh/Efficient-VWM-Survey", "clusters": ["C8"], "keywords": ["survey", "efficient video world models"], "model_class": "world model", "techniques": ["survey: efficient paradigms, architectures, inference algorithms"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey of efficiency techniques for video-generation world models organized into modeling paradigms, network architectures and inference algorithms, with embodied AI and driving as target applications.", "relevance": 4, "relevance_reason": "Directly maps the efficient world-model design space.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2604.23775", "title": "Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms", "authors": ["Qi Li", "Bo Yin", "Weiqi Huang", "Ruhao Liu", "Bo Zou", "Runpeng Yu", "Jingwen Ye", "Wei-Hao Yu", "Xinchao Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.23775", "pdf_url": "https://arxiv.org/pdf/2604.23775", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["VLA safety survey", "attacks", "defenses", "real-time defense constraints", "real-time defense"], "model_class": "VLA", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of VLA safety organized by attack and defense timing (training vs inference), including runtime defenses and the real-time latency constraints on them.", "relevance": 4, "relevance_reason": "Survey explicitly noting latency constraints on runtime defenses.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2); venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2607.01804", "title": "VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon", "authors": ["Yi Pan", "Miao Pan", "Qi Lu", "Jiaming Huang", "Man Zhang", "Siteng Huang", "Xin Li", "Jie Zhang", "Yongliang Shen", "Xuhong Zhang", "Wenqi Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.01804", "pdf_url": "https://arxiv.org/pdf/2607.01804", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["VLA-Corrector", "latent vision monitor", "detect-and-correct", "adaptive horizon"], "model_class": "VLA", "techniques": ["latent predicted-vs-actual monitor", "triggered replanning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Monitors predicted versus actual visual feature evolution during open-loop chunk execution and triggers correction/replanning when persistent deviation appears, without modifying the VLA.", "relevance": 4, "relevance_reason": "Runtime monitor for open-loop blind spot.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2603.13966", "title": "vla-eval: A Unified Evaluation Harness for Vision-Language-Action Models", "authors": ["Suhwan Choi", "Yunsung Lee", "Yubeen Park", "Chris Dongjoo Kim", "Ranjay Krishna", "Dieter Fox", "Youngjae Yu"], "year": 2026, "venue": "arXiv", "type": "repo", "url": "https://arxiv.org/abs/2603.13966", "pdf_url": "https://arxiv.org/pdf/2603.13966", "code_url": "https://github.com/allenai/vla-evaluation-harness", "clusters": ["C4"], "keywords": ["vla-eval", "evaluation harness", "Docker isolation", "reproduction", "parallel evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 47x wall-clock speedup; 2,000 LIBERO episodes in ~18 min", "metric_measures": "throughput", "benchmarks": ["14 simulation benchmarks incl. LIBERO"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Harness decoupling model servers from benchmark environments via a WebSocket protocol with Docker isolation, covering 14 sim benchmarks and reproducing published scores.", "relevance": 4, "relevance_reason": "Reproducibility infrastructure for cross-benchmark VLA evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA leaderboard"}
{"id": "2603.26666", "title": "VLA-OPD: Bridging Offline SFT and Online RL for Vision-Language-Action Models via On-Policy Distillation", "authors": ["Zhide Zhong", "Haodong Yan", "Junfeng Li", "Junjie He", "Tianran Zhang", "Haoang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.26666", "pdf_url": "https://arxiv.org/pdf/2603.26666", "code_url": "", "clusters": ["C3"], "keywords": ["on-policy distillation", "dense supervision"], "model_class": "VLA", "techniques": ["on-policy distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Teacher provides token-level labels on the student VLA's own rollouts, combining SFT efficiency with RL-like robustness.", "relevance": 4, "relevance_reason": "Efficient post-training alternative to sparse-reward RL.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.04355", "title": "VLA-Precision: Asymmetric Co-Bootstrapping for Efficient Real-World Online RL of Vision-Language-Action Models", "authors": ["Chenyu Su", "Zhaolong Shen", "Yuan Qian", "Chen Qian", "Rui Zhang", "Feng Yan", "Weixing Chen", "Fei Zhang", "Jiamin Wang", "Shuang Cong", "Weiwei Shang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.04355", "pdf_url": "https://arxiv.org/pdf/2609.04355", "code_url": "", "clusters": ["C3"], "keywords": ["real-world online RL", "asymmetric co-bootstrapping", "throughput"], "model_class": "VLA", "techniques": ["online RL", "streaming architecture", "interventions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Real-world online RL framework for VLAs addressing value unreliability and large-model throughput via ACoB-Stream.", "relevance": 4, "relevance_reason": "Explicitly tackles large-VLA overhead in real-world RL.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: ti:VLA AND ti:RL"}
{"id": "2605.20774", "title": "VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models", "authors": ["Alex S. Huang", "Jiahui Zhang", "Shiqing Tang", "Yu Xiang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2605.20774", "pdf_url": "https://arxiv.org/pdf/2605.20774", "code_url": "", "clusters": ["C4"], "keywords": ["VLA-REPLICA", "low-cost real-world benchmark", "cross-lab reproducibility"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["VLA-REPLICA", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Off-the-shelf, low-cost real-world VLA benchmark designed to be replicated across labs, with ID/OOD protocols; results are consistent across independently built setups.", "relevance": 4, "relevance_reason": "Real-robot reproducibility methodology.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RoboCasa VLABench RoboTwin"}
{"id": "2601.11250", "title": "VLAgents: A Policy Server for Efficient VLA Inference", "authors": ["Tobias Jülg", "Khaled Gamal", "Nisarga Nilavadi", "Pierre Krack", "Seongjin Bien", "Michael Krawez", "Florian Walter", "Wolfram Burgard", "Jülg, Tobias", "Gamal, Khaled", "Nilavadi, Nisarga", "Krack, Pierre", "Bien, Seongjin", "Krawez, Michael", "Walter, Florian", "Burgard, Wolfram"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.11250", "pdf_url": "https://arxiv.org/pdf/2601.11250", "code_url": "https://github.com/RobotControlStack/vlagents", "clusters": ["C1", "C3", "C4", "C5", "C6", "C9"], "keywords": ["policy server", "shared memory", "compressed streaming", "VLAgents", "communication latency", "zero-copy shared memory", "remote inference", "Gymnasium protocol", "distributed inference", "interfaces"], "model_class": "VLA", "techniques": ["serving infrastructure", "policy serving", "zero-copy IPC", "compressed streaming", "unified policy server", "zero-copy shared memory", "compressed remote streaming", "modular policy server"], "base_model": "OpenVLA, pi0 and 5 others", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["local and remote communication benchmark"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Modular VLA policy server with a Gymnasium-style protocol that switches between zero-copy shared memory locally and compressed streaming for remote hardware, outperforming OpenVLA, openpi and LeRobot default servers.", "relevance": 4, "relevance_reason": "Directly optimizes the remote-inference communication layer for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: efficient VLA; WebSearch: RLBench VLA latency efficient 3D policy; web: VLA accelerator FPGA ASIC design; search query; venue sweep: arXiv API VLA AND benchmark AND latency AND hardware"}
{"id": "10.1109/iccc70295.2026.11680263", "title": "VTP: A Task-Aware Transmission Protocol for Edge Vision-Language-Action Robotic Systems", "authors": ["Richeng Huo", "Yan-Sen Wang", "Ye Wu", "Yan Dong", "Peng Yang"], "year": 2026, "venue": "2026 IEEE/CIC International Conference on Communications in China (ICCC)", "type": "paper", "url": "https://doi.org/10.1109/ICCC70295.2026.11680263", "pdf_url": "", "code_url": "", "clusters": ["C6"], "keywords": ["VTP", "edge offloading", "packet loss"], "model_class": "VLA", "techniques": ["transmission protocol"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Task-aware transmission protocol for edge-offloaded VLAs that prioritizes critical frames to cut visual delivery delay under packet loss.", "relevance": 4, "relevance_reason": "Edge-cloud VLA communication.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2502.19645"}
{"id": "2609.18193", "title": "WAVE-Go: World-Model Navigation with Adaptive Execution for Wheel-Legged Robots", "authors": ["Ming-Yi Li", "Ji Li", "Zhi-Hao Ouyang", "Ya-Ge He", "Börje F. Karlsson"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18193", "pdf_url": "https://arxiv.org/pdf/2609.18193", "code_url": "", "clusters": ["C7"], "keywords": ["wheel-legged navigation"], "model_class": "world model", "techniques": ["interruptible execution", "risk-bounded prefix selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "51.2% fewer replanning calls", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "WAVE-Go executes world-model action plans with adaptive prefix lengths and cancellation when observations invalidate them.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2506.07339"}
{"id": "2609.13984", "title": "What Makes an Efficient VLA? Navigating Action-Head Design, Scaling, and Latency", "authors": ["Luoyang Sun", "Guoyang Xia", "Fengfa Li", "Lei Ren", "Xinyu Cui", "Haifeng Zhang", "Fangxiang Feng", "Kaike Zhang", "Kun Zhan", "Yan Xie", "Jun Wang", "Cheng Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.13984", "pdf_url": "https://arxiv.org/pdf/2609.13984", "code_url": "", "clusters": ["C1", "C4", "C5", "C9"], "keywords": ["action head design", "latency-paired study", "initialization", "latency-paired ablation", "action head", "scaling", "action-head design", "on-device latency"], "model_class": "VLA", "techniques": ["analysis", "latency-paired architecture study"], "base_model": "SigLIP2 + Qwen2.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Controlled latency-paired sweep of action-head designs finding that initializing the head from the backbone's last layers matters most, with flow matching or extra passes helping only when misaligned.", "relevance": 4, "relevance_reason": "Latency-aware architecture guidance measured on device.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA; search query: VLA latency; arXiv: on-device AND (robot policy OR VLA OR diffusion policy)"}
{"id": "2609.34981", "title": "What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling", "authors": ["Renping Zhou", "Zanlin Ni", "Zihao Fan", "Guohao Fu", "Zeyu Liu", "Hao Shi", "Jie Zhang", "Chi Bene Chen", "Yang Yue", "Xueyang Fu", "Gao Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34981", "pdf_url": "https://arxiv.org/pdf/2609.34981", "code_url": "", "clusters": ["C8"], "keywords": ["test-time future modeling", "latent vs explicit WAM", "first denoising step"], "model_class": "world-action model", "techniques": ["controlled comparison of explicit vs latent WAMs"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Finds latent WAMs that drop future generation lose generalization, and that nearly all of the benefit comes from the first denoising step of the future, suggesting cheap partial imagination.", "relevance": 4, "relevance_reason": "Clarifies how little test-time imagination compute is needed.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2608.01428", "title": "When Replanning Becomes the Bottleneck: Budgeted Replanning for Embodied Agents", "authors": ["Shuai-Jun Liu", "Fei-Yang You", "Xing-Wei Chen", "Ning-Xin Su"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.01428", "pdf_url": "https://arxiv.org/pdf/2608.01428", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["LLM replanning", "SLO", "tail latency"], "model_class": "other", "techniques": ["token pruning", "budgeted replanning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "62-92% fewer replanning tokens; SLO violations from 85.5-100% to 4.7-50%", "metric_measures": "other", "benchmarks": ["Habitat", "RoboFactory", "AirSim"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Budgets LLM replanning calls for embodied agents and prunes replanning context to cut tail latency and deadline misses.", "relevance": 4, "relevance_reason": "Real-time deadlines for foundation-model planners", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2502.02175"}
{"id": "2602.11882", "title": "Where Bits Matter in World Model Planning: A Paired Mixed-Bit Study for Efficient Spatial Reasoning", "authors": ["Suraj Ranganath", "Anish Patnaik", "Vaishak Menon"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.11882", "pdf_url": "https://arxiv.org/pdf/2602.11882", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["mixed-bit", "encoder precision", "planning under low bits"], "model_class": "world model", "techniques": ["mixed/asymmetric/layerwise quantization"], "base_model": "DINO-WM", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Wall planning task"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Finds 8/6-bit DINO-WM planning is near FP16, 3-bit collapses, and at 4-bit keeping the encoder at higher precision matters most.", "relevance": 4, "relevance_reason": "Where to spend bits in world-model planners.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.20781", "title": "World Action Models: A Survey", "authors": ["Qiuhong Shen", "Shihua Zhang", "Yue Liao", "Qi Li", "Zhenxiong Tan", "Shizun Wang", "Shuicheng Yan", "Xinchao Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20781", "pdf_url": "https://arxiv.org/pdf/2606.20781", "code_url": "", "clusters": ["C8"], "keywords": ["world action models"], "model_class": "world-action model", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey defining world-action models and organizing them by what is generated and by substrate, backbone, action coupling and deployment regime; notes the field is moving toward generating less of the future.", "relevance": 4, "relevance_reason": "Central reference for C8 with explicit compute/latency trade-off framing.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2603.06331", "title": "WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching", "authors": ["Wei-Lun Feng", "Guo-Xin Fan", "Haotong Qin", "Chuan Yang", "Mingqiang Wu", "Yuqi Li", "Xiang-Qi Li", "Zhu-Lin An", "Li-Bo Huang", "Dingrui Wang", "Long-Long Liao", "Michele Magno", "Yong-Jun Xu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.06331", "pdf_url": "https://arxiv.org/pdf/2603.06331", "code_url": "", "clusters": ["C8"], "keywords": ["WorldCache", "diffusion world model"], "model_class": "world model", "techniques": ["feature caching", "token caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 3.7x end-to-end speedup at 98% rollout quality", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free heterogeneous token caching tailored to diffusion world models to skip redundant computation.", "relevance": 4, "relevance_reason": "Direct world-model acceleration.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104; cited by 2603.28489"}
{"id": "2603.22286", "title": "WorldCache: Content-Aware Caching for Accelerated Video World Models", "authors": ["Umair Nawaz", "Ahmed Heakl", "Ufaq Khan", "Abdelrahman Shaker", "Salman Khan", "Fahad Shahbaz Khan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.22286", "pdf_url": "https://arxiv.org/pdf/2603.22286", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["feature caching", "video world model", "motion-adaptive"], "model_class": "world model", "techniques": ["motion-adaptive cache thresholds", "saliency-weighted drift", "warp/blend approximation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Training-free caching for DiT video world models that decides when and how to reuse features based on motion and saliency, avoiding ghosting of naive caching.", "relevance": 4, "relevance_reason": "Inference acceleration for video world models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.06949 (citations)"}
{"id": "2604.20289", "title": "X-Cache: Cross-Chunk Block Caching for Few-Step Autoregressive World Models Inference", "authors": ["Yixiao Zeng", "Jianlei Zheng", "Chaoda Zheng", "Shijia Chen", "Mingdian Liu", "Tongping Liu", "Tengwei Luo", "Yu Zhang", "Boyang Wang", "Linkun Xu", "Siyuan Lu", "Bo Tian", "Xianming Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.20289", "pdf_url": "https://arxiv.org/pdf/2604.20289", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["cross-chunk caching", "few-step AR world model", "driving simulation"], "model_class": "world model", "techniques": ["per-block residual caches across chunks", "structure/action-aware gating"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["multi-camera driving world model"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Caches block residuals across consecutive generation chunks rather than denoising steps, accelerating already few-step autoregressive driving world models without training.", "relevance": 4, "relevance_reason": "Caching tailored to distilled interactive world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.27449", "title": "X2Real: an eXtensive simulation benchmark for real-world generalist policies", "authors": ["Ruan, Lian", "Yang, Jade", "Gao, Sherphylan", "Gao, Felix", "Liang, Kyson", "Liu, Galen", "Wu, Ligo", "Jin, Lane", "Gu, Guu", "Xie, Bevan", "Yan, Cloud", "Yuan, Zongzi", "Luo, Kino", "Chen, Emma", "Chen, Shuwen", "Ping, Yang", "Guo, Miles", "Sun, Rain", "Zhang, Kayden", "Du, Alex", "Wu, Ruihai", "Hao, Liang", "Li, Zhaoshuo", "Gan, Roy", "Wang, Hao", "Wang, Qian"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.27449", "pdf_url": "https://arxiv.org/pdf/2609.27449", "code_url": "", "clusters": ["C4"], "keywords": ["X2Real", "Isaac Lab-Arena", "sim-real correlation", "fairness", "disjoint train/eval"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "0.84 linear sim-real correlation", "metric_measures": "other", "benchmarks": ["X2Real"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Evolvable Isaac Lab-Arena benchmark with 44 long-horizon tasks over 10 capability dimensions, calibrated visuals/physics and strictly disjoint train/eval pipelines to prevent benchmark exploitation.", "relevance": 4, "relevance_reason": "Addresses benchmark fairness and sim-real validity.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: SimplerEnv evaluating real-world robot manipulation policies"}
{"id": "2602.12684", "title": "Xiaomi-Robotics-0: An Open-Sourced Vision-Language-Action Model with Real-Time Execution", "authors": ["Ruisi Cai", "Jun Guo", "Xin He", "Piao-Piao Jin", "Jie Li", "Bin Lin", "Fu-Teng Liu", "Wei Liu", "Fei Ma", "Kun Ma", "Feng Qiu", "Heng-Chao Qu", "Yifei Su", "Qiao Sun", "Dong Wang", "Dong-Hao Wang", "Yun-Hao Wang", "Ru-Jie Wu", "Di-Yun Xiang", "Yu Yang", "Hang Ye", "Yuan Zhang", "Quan Zhou"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2602.12684", "pdf_url": "https://arxiv.org/pdf/2602.12684", "code_url": "", "clusters": ["C1", "C2", "C5", "C6", "C9"], "keywords": ["consumer GPU", "real-time execution", "asynchronous training", "open VLA", "deployment strategy"], "model_class": "VLA", "techniques": ["asynchronous-execution-aware post-training", "chunk timestep alignment", "asynchronous execution", "training recipe for smooth real-time control"], "base_model": "", "params": "", "hardware": "consumer-grade GPU", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Xiaomi's open VLA trained for asynchronous execution so it rolls out smoothly on real bimanual robots with a consumer GPU.", "relevance": 4, "relevance_reason": "Industrial deployment report; limited hardware analysis.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2510.26742 (citations); venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2603.15757", "title": "You've Got a Golden Ticket: Improving Generative Robot Policies With A Single Noise Vector", "authors": ["Omkar Patil", "Ondrej Biza", "Thomas Weng", "Karl Schmeckpeper", "Wil Thomason", "Xiaohan Zhang", "Kausik Sivakumar", "Robin Walters", "Nakul Gopalan", "Sebastian Castro", "Stephen Hart", "Eric Rosen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.15757", "pdf_url": "https://arxiv.org/pdf/2603.15757", "code_url": "", "clusters": ["C3"], "keywords": ["golden ticket", "initial noise optimization", "frozen policy"], "model_class": "diffusion/flow policy", "techniques": ["noise-space policy improvement", "no-training adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to +28% real-world success within 60 search episodes", "metric_measures": "other", "benchmarks": ["simulation manipulation suites", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Improves frozen diffusion/flow policies (incl. VLAs) by searching for a single constant initial noise vector via Monte Carlo rollouts, with zero weight updates.", "relevance": 4, "relevance_reason": "Extremely cheap adaptation: no gradient updates or new networks.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: LoRA fine-tuning robot policy VLA adapter arXiv 2025"}
{"id": "2606.31846", "title": "Z-1: Efficient Reinforcement Learning for Vision-Language-Action Models", "authors": ["Lang Cao", "Renhong Chen", "Luyi Li", "Peng Wang", "Mofan Peng", "Yitong Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.31846", "pdf_url": "https://arxiv.org/pdf/2606.31846", "code_url": "", "clusters": ["C3"], "keywords": ["Z-1", "GRPO", "shared-prefix rollouts", "RoboCasa"], "model_class": "VLA", "techniques": ["GRPO", "tree-structured rollouts"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Efficient GRPO post-training on pi0.5 with shared-prefix rollouts and tree branching on RoboCasa.", "relevance": 4, "relevance_reason": "RL post-training efficiency tricks.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:GRPO"}
{"id": "2510.25889", "title": "$π_\\texttt{RL}$: Online RL Fine-tuning for Flow-based Vision-Language-Action Models", "authors": ["Kang Chen", "Zhihao Liu", "Tonghe Zhang", "Zhen Guo", "Si Xu", "Hao Lin", "Hongzhi Zang", "Xiang Li", "Quanlu Zhang", "Zhaofei Yu", "Guoliang Fan", "Tiejun Huang", "Yu Wang", "Chao Yu", "Zhi-Hao Liu"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://doi.org/10.48550/arXiv.2510.25889", "pdf_url": "https://arxiv.org/pdf/2510.25889", "code_url": "https://github.com/RLinf/RLinf", "clusters": ["C3"], "keywords": ["piRL", "flow-noise", "flow-SDE", "pi0.5", "RL fine-tuning", "flow VLA"], "model_class": "VLA", "techniques": ["online RL", "PPO", "online RL fine-tuning"], "base_model": "pi0; pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "ManiSkill"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "piRL: framework for online reinforcement-learning fine-tuning of flow-based VLAs such as pi0/pi0.5 (title only; no abstract available).", "relevance": 4, "relevance_reason": "context: RL fine-tuning of flow VLAs, heavily linked to core items", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025; snowball round 2: cites 2303.04137, 2306.03310; cited by 2506.17639, 2510.06710"}
{"id": "1x-world-model", "title": "1X World Model: Evaluating Bits, not Atoms", "authors": ["1X World Model Team"], "year": 2025, "venue": "1X Technologies tech report / blog", "type": "tech-report", "url": "https://www.1x.tech/discover/1x-world-model", "pdf_url": "https://www.1x.tech/1x-world-model.pdf", "code_url": "", "clusters": ["C4", "C7", "C8", "C9"], "keywords": ["humanoid world model", "policy evaluation", "action-controllable video", "world model evaluator", "action-conditioned video"], "model_class": "world model", "techniques": ["action-conditioned video generation", "world-model-based policy ranking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NEO humanoid tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Action-conditioned video world model for a full-body humanoid whose predicted success rates correlate with real-world evaluations, so candidate policies can be ranked in the model rather than on hardware.", "relevance": 4, "relevance_reason": "Industrial world model used as a cheap evaluation and monitoring substitute for real rollouts.", "foundational_technique": false, "verified": true, "found_via": "search query; blog sweep: 1X"}
{"id": "2507.05331", "title": "A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation", "authors": ["TRI LBM Team", "Jose Barreiros", "Andrew Beaulieu", "Aditya Bhat", "Rick Cory", "Eric Cousineau", "Hongkai Dai", "Ching-Hsin Fang", "Kunimatsu Hashimoto", "Muhammad Zubair Irshad", "Masha Itkina", "Naveen Kuppuswamy", "Kuan-Hui Lee", "Katherine Liu", "Dale McConachie", "Ian McMahon", "Haruki Nishimura", "Calder Phillips-Grafflin", "Charles Richter", "Paarth Shah", "Krishnan Srinivasan", "Blake Wulfe", "Chen Xu", "Mengchao Zhang", "Alex Alspach"], "year": 2025, "venue": "arXiv / tech report (TRI)", "type": "tech-report", "url": "https://toyotaresearchinstitute.github.io/lbm1/", "pdf_url": "https://arxiv.org/pdf/2507.05331", "code_url": "", "clusters": ["C3", "C4", "C9"], "keywords": ["large behavior models", "blind randomized trials", "statistical confidence", "TRI", "diffusion policy", "statistical evaluation"], "model_class": "diffusion/flow policy", "techniques": ["multitask diffusion transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot", "~1,800 real rollouts", "47k+ sim rollouts"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "TRI scales diffusion policies into multitask LBMs on ~1,700 h of data and evaluates them with blind, statistically rigorous trials, finding pretraining improves robustness and data efficiency.", "relevance": 4, "relevance_reason": "Benchmarking methodology and diffusion-based foundation policy from TRI.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation statistical; blog sweep: TRI"}
{"id": "2510.24795", "title": "A Survey on Efficient Vision-Language-Action Models", "authors": ["Zhaoshu Yu", "Bo Wang", "Pengpeng Zeng", "Haonan Zhang", "Ji Zhang", "Zheng Wang", "Lianli Gao", "Jingkuan Song", "Nicu Sebe", "Heng Tao Shen", "Zhaoshuo Yu", "N. Sebe", "Heng-Tao Shen"], "year": 2025, "venue": "IEEE TPAMI", "type": "preprint", "url": "https://arxiv.org/abs/2510.24795", "pdf_url": "https://arxiv.org/pdf/2510.24795", "code_url": "https://evla-survey.github.io/", "clusters": ["C1", "C2", "C3", "C5"], "keywords": ["survey", "efficient VLA", "taxonomy", "efficient training", "efficient data collection"], "model_class": "VLA", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey organizing efficient-VLA work across model design (architectures, compression), training, and data collection, with a living project page.", "relevance": 4, "relevance_reason": "Broad map of efficient VLA techniques; useful for coverage checking.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; snowball from 2506.07530; WebSearch: parameter-efficient fine-tuning vision-language-action model arXiv"}
{"id": "2509.22093", "title": "Action-aware Dynamic Pruning for Efficient Vision-Language-Action Manipulation", "authors": ["Xiaohuan Pei", "Yuxing Chen", "Siyu Xu", "Yunke Wang", "Yuheng Shi", "Chang Xu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.22093", "pdf_url": "https://arxiv.org/pdf/2509.22093", "code_url": "", "clusters": ["C1"], "keywords": ["dynamic token pruning", "action-aware gating"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "OpenVLA-OFT, OpenVLA", "params": "", "hardware": "", "reported_metric": "1.35x speedup on OpenVLA-OFT", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adjusts visual-token retention by manipulation phase, pruning more during coarse motion and less during fine operations, gated by recent action trajectories.", "relevance": 4, "relevance_reason": "VLA token pruning.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient"}
{"id": "2509.22195", "title": "Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting", "authors": ["Asher J. Hancock", "Xindi Wu", "Lihan Zha", "Olga Russakovsky", "Anirudha Majumdar"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.22195", "pdf_url": "https://arxiv.org/pdf/2509.22195", "code_url": "", "clusters": ["C3"], "keywords": ["VLM2VLA", "actions as language", "LoRA only", "forgetting"], "model_class": "VLA", "techniques": ["LoRA", "action-as-text"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot", "VQA"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Represents actions as natural language so a VLM can be turned into a VLA with LoRA alone, avoiding forgetting and costly co-training.", "relevance": 4, "relevance_reason": "PEFT-only VLA training that removes co-training cost.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: memory-efficient training VLA model single GPU fine-tuning arXiv"}
{"id": "2509.02055", "title": "Align-Then-stEer: Adapting the Vision-Language Action Models through Unified Latent Guidance", "authors": ["Yang Zhang", "Chenwei Wang", "Ouyang Lu", "Yuan Zhao", "Yunfei Ge", "Zhenglong Sun", "Xiu Li", "Chi Zhang", "Chenjia Bai", "Xuelong Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.02055", "pdf_url": "https://arxiv.org/pdf/2509.02055", "code_url": "", "clusters": ["C3"], "keywords": ["ATE", "latent guidance", "action space alignment"], "model_class": "VLA", "techniques": ["latent alignment", "guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Aligns disparate action spaces in a unified latent then steers the VLA, enabling data-efficient cross-embodiment fine-tuning.", "relevance": 4, "relevance_reason": "Reduces data/compute for adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2509.19555", "title": "AnySafe: Adapting Latent Safety Filters at Runtime via Safety Constraint Parameterization in the Latent Space", "authors": ["Sankalp Agrawal", "Junwon Seo", "Kensuke Nakamura", "Ran Tian", "Andrea Bajcsy"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.19555", "pdf_url": "https://arxiv.org/pdf/2509.19555", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["runtime-adaptable safety filter", "constraint parameterization", "conformal calibration"], "model_class": "world model", "techniques": ["image-conditioned constraint encoding", "training inside world-model imagination"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["vision-based manipulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Lets latent safety filters accept new user-specified constraints at runtime by conditioning on an image of the constraint, calibrated with conformal prediction and trained entirely in imagination.", "relevance": 4, "relevance_reason": "Adaptable world-model safety monitor.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2512.20188", "title": "Asynchronous Fast-Slow Vision-Language-Action Policies for Whole-Body Robotic Manipulation", "authors": ["Teqiang Zou", "Hongliang Zeng", "Yuxuan Nong", "Yifan Li", "Kehui Liu", "Haotian Yang", "Xinyang Ling", "Xin Li", "Lianyang Ma", "Hao-Tian Yang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.20188", "pdf_url": "https://arxiv.org/pdf/2512.20188", "code_url": "", "clusters": ["C1", "C2", "C5", "C7"], "keywords": ["asynchronous fast-slow", "whole-body", "latent buffer", "fast-slow", "asynchronous", "Astribot", "DuoCore-FS", "fast-slow VLA"], "model_class": "VLA", "techniques": ["asynchronous execution", "action tokenization", "asynchronous dual-system", "latent buffer", "whole-body action tokenizer", "latent representation buffer"], "base_model": "", "params": "3B VLM", "hardware": "", "reported_metric": "30 Hz whole-body action-chunk generation (~3x comparable VLAs)", "metric_measures": "open-loop chunk rate", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Truly asynchronous fast-slow VLA where a slow VLM writes guidance into a latent buffer and a fast action pathway generates whole-body chunks at high rate, jointly trained end-to-end.", "relevance": 4, "relevance_reason": "Primarily C2 (asynchrony); relevant to compute budgeting.", "foundational_technique": false, "verified": true, "found_via": "search query: fast slow VLA; web: VLA accelerator FPGA ASIC design; arXiv search: \"action chunk\" asynchronous inference"}
{"id": "lerobot-async-inference-docs", "title": "Asynchronous Inference (LeRobot documentation)", "authors": ["Hugging Face"], "year": 2025, "venue": "Hugging Face docs / blog", "type": "blog", "url": "https://huggingface.co/docs/lerobot/en/async", "pdf_url": "", "code_url": "https://github.com/huggingface/lerobot", "clusters": ["C6"], "keywords": ["policy server", "robot client", "gRPC", "asynchronous inference"], "model_class": "VLA", "techniques": ["remote policy server", "asynchronous chunk execution"], "base_model": "SmolVLA, pi0/pi0.5, ACT", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Practical guide to LeRobot's PolicyServer (GPU machine) and RobotClient streaming observations and action chunks over gRPC so the robot keeps moving while the next chunk is computed.", "relevance": 4, "relevance_reason": "Widely used open remote/async inference tooling.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "hf-async-robot-inference-2025", "title": "Asynchronous Robot Inference: Decoupling Action Prediction and Execution", "authors": ["Francesco Capuano", "Steven Palma", "Michel Aractingi", "Mustafa Shukor", "Dana Aubakirova", "Adil Zouitine", "Simon Alibert", "Remi Cadene"], "year": 2025, "venue": "blog (Hugging Face)", "type": "blog", "url": "https://huggingface.co/blog/async-robot-inference", "pdf_url": "", "code_url": "https://github.com/huggingface/lerobot", "clusters": ["C1", "C2", "C4", "C5", "C6", "C9"], "keywords": ["asynchronous inference", "policy server", "action queue", "gRPC", "LeRobot", "async inference", "robot client", "SmolVLA", "chunk threshold"], "model_class": "VLA", "techniques": ["asynchronous inference", "chunk aggregation", "remote GPU policy server", "asynchronous chunk execution", "remote policy server"], "base_model": "SmolVLA (also ACT, OpenVLA, pi0)", "params": "", "hardware": "policy server on RTX 4090; robot client onboard", "reported_metric": "~2x faster task completion at comparable success | ~2x speed-up in task completion time with comparable success", "metric_measures": "other", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "LeRobot's split into an onboard RobotClient with an action queue and a GPU PolicyServer that precomputes the next chunk once the queue drops below a threshold (g≈0.5–0.7).", "relevance": 4, "relevance_reason": "Practical reference for measuring end-to-end latency including network round-trip.", "foundational_technique": false, "verified": true, "found_via": "search query: SmolVLA asynchronous inference blog; WebSearch: LeRobot asynchronous inference blog; web: Hugging Face blog async inference LeRobot; blog sweep: Hugging Face LeRobot"}
{"id": "2506.06072", "title": "BEAST: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning", "authors": ["Hongyi Zhou", "Weiran Liao", "Xi Huang", "Yucheng Tang", "Fabian Otto", "Xiaogang Jia", "Xinkai Jiang", "Simon Hilber", "Ge Li", "Qian Wang", "Ömer Erdinç Yağmurlu", "Nils Blank", "Moritz Reuss", "Rudolf Lioutikov"], "year": 2025, "venue": "NeurIPS 2025", "type": "paper", "url": "https://arxiv.org/abs/2506.06072", "pdf_url": "https://arxiv.org/pdf/2506.06072", "code_url": "", "clusters": ["C1"], "keywords": ["B-spline tokenizer", "parallel decoding", "smooth actions"], "model_class": "VLA", "techniques": ["action tokenization", "parallel action decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Encodes action chunks as B-spline control points giving fixed-length tokens that decode in parallel and stay smooth across chunks, with no tokenizer training.", "relevance": 4, "relevance_reason": "Efficient action tokenizer.", "foundational_technique": false, "verified": true, "found_via": "search query: parallel decoding action"}
{"id": "2504.12285", "title": "BitNet b1.58 2B4T Technical Report", "authors": ["Shu-Ming Ma", "Hongyu Wang", "Shao-Han Huang", "Xingxing Zhang", "Ying Hu", "Ting Song", "Yan Xia", "Furu Wei"], "year": 2025, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2504.12285", "pdf_url": "https://arxiv.org/pdf/2504.12285", "code_url": "", "clusters": ["C2"], "keywords": ["open 1-bit LLM", "2B params", "4T tokens"], "model_class": "other", "techniques": ["ternary quantization-aware training"], "base_model": "", "params": "2B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Releases an open-weight native 1.58-bit 2B LLM trained on 4T tokens with efficient inference kernels.", "relevance": 4, "relevance_reason": "Backbone of BitVLA.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2506.07530 (references)"}
{"id": "10.24963/ijcai.2025/766", "title": "Boost Embodied AI Models with Robust Compression Boundary", "authors": ["Chong Yu", "Tao Chen", "Zhongxue Gan"], "year": 2025, "venue": "IJCAI 2025", "type": "paper", "url": "https://doi.org/10.24963/ijcai.2025/766", "pdf_url": "", "code_url": "", "clusters": ["C2"], "keywords": ["compression boundary", "embodied AI", "robustness"], "model_class": "other", "techniques": ["model compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Studies how far standard compression can be pushed on embodied AI models (driving, robotics) before robustness breaks, proposing a robust compression boundary.", "relevance": 4, "relevance_reason": "Embodied-specific compression robustness study.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246"}
{"id": "2508.02219", "title": "CO-RFT: Efficient Fine-Tuning of Vision-Language-Action Models through Chunked Offline Reinforcement Learning", "authors": ["Dongchi Huang", "Zhirui Fang", "Tianle Zhang", "Yihang Li", "Lin Zhao", "Chunhe Xia"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.02219", "pdf_url": "https://arxiv.org/pdf/2508.02219", "code_url": "", "clusters": ["C3"], "keywords": ["CO-RFT", "chunked offline RL", "TD learning"], "model_class": "VLA", "techniques": ["offline RL", "action chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Chunked TD-learning offline RL for fine-tuning VLAs from limited demonstrations with better sample efficiency than SFT.", "relevance": 4, "relevance_reason": "Efficient offline RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"offline reinforcement learning\""}
{"id": "2508.08144", "title": "COMponent-Aware Pruning for Accelerated Control Tasks in Latent Space Models", "authors": ["Ganesh Sundaram", "Jonas Ulmen", "Amjad Haider", "Daniel Görges"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.08144", "pdf_url": "https://arxiv.org/pdf/2508.08144", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["structured pruning", "TD-MPC", "Lyapunov stability", "edge controllers"], "model_class": "world model", "techniques": ["component-aware structured pruning", "Lyapunov-criterion stability checks"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Compresses TD-MPC latent world-model controllers with component-aware structured pruning while checking Lyapunov stability, targeting resource-constrained robots.", "relevance": 4, "relevance_reason": "Compression of a latent world-model planner for edge deployment.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828"}
{"id": "2511.18950", "title": "Compressor-VLA: Instruction-Guided Visual Token Compression for Efficient Robotic Manipulation", "authors": ["Juntao Gao", "Feiyang Ye", "Jing Zhang", "Wenjing Qian"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.18950", "pdf_url": "https://arxiv.org/pdf/2511.18950", "code_url": "", "clusters": ["C1"], "keywords": ["token compression", "instruction-conditioned"], "model_class": "VLA", "techniques": ["token compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "59% fewer FLOPs, >3x fewer visual tokens", "metric_measures": "other", "benchmarks": ["LIBERO", "real dual-arm robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Instruction-conditioned compressors distill task context and fine spatial detail into a few visual tokens.", "relevance": 4, "relevance_reason": "Visual token reduction for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning"}
{"id": "2502.05450", "title": "ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy", "authors": ["Yuhui Chen", "Shuai Tian", "Shugao Liu", "Yingting Zhou", "Haoran Li", "Dongbin Zhao"], "year": 2025, "venue": "RSS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2502.05450", "pdf_url": "https://arxiv.org/pdf/2502.05450", "code_url": "", "clusters": ["C3"], "keywords": ["reinforced fine-tuning", "consistency policy", "human intervention", "real-world RL"], "model_class": "VLA", "techniques": ["offline RL", "online RL", "consistency policy"], "base_model": "Octo", "params": "", "hardware": "", "reported_metric": "96.3% avg success after 45-90 min online fine-tuning", "metric_measures": "other", "benchmarks": ["8 real-world manipulation tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Offline BC+Q-learning then human-in-the-loop online RL with a consistency-policy head to fine-tune a VLA in the real world in under 90 minutes.", "relevance": 4, "relevance_reason": "Sample/time-efficient real-world RL adaptation of VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2510.04246", "title": "ContextVLA: Vision-Language-Action Model with Amortized Multi-Frame Context", "authors": ["Huiwon Jang", "Sihyun Yu", "Heeseung Kwon", "Hojin Jeon", "Younggyo Seo", "Jinwoo Shin"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.04246", "pdf_url": "https://arxiv.org/pdf/2510.04246", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["multi-frame VLA", "amortized context"], "model_class": "VLA", "techniques": ["context compression", "temporal context compression", "context token"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Compresses past observations into a single context token so a VLA gets multi-frame benefits with reduced training and inference time.", "relevance": 4, "relevance_reason": "Token compression for efficient temporal context in VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=100]; snowball round 2: cited by 2602.03983, 2602.15882"}
{"id": "2508.06804", "title": "D3P: Dynamic Denoising Diffusion Policy via Reinforcement Learning", "authors": ["Shu-Ang Yu", "Feng Gao", "Yi Wu", "Chao Yu", "Yu Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.06804", "pdf_url": "https://arxiv.org/pdf/2508.06804", "code_url": "", "clusters": ["C1"], "keywords": ["dynamic denoising steps", "RL adaptor"], "model_class": "diffusion/flow policy", "techniques": ["adaptive sampling steps", "RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.2x (sim), 1.9x (real) inference speedup", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "A state-aware adaptor trained with RL allocates more denoising steps to crucial actions and fewer to routine ones.", "relevance": 4, "relevance_reason": "Adaptive-compute diffusion policy.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy real-time inference speed"}
{"id": "2511.20906", "title": "DASIP: Dynamic Test-Time Compute Scaling for Robot Control with Stochastic Interpolant Policies", "authors": ["Inkook Chun", "Seungjae Lee", "M. Albergo", "Saining Xie", "Eric Vanden-Eijnden"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2511.20906", "pdf_url": "https://arxiv.org/pdf/2511.20906", "code_url": "", "clusters": ["C1"], "keywords": ["test-time compute scaling", "difficulty-aware"], "model_class": "diffusion/flow policy", "techniques": ["adaptive compute", "stochastic interpolants"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.6-4.4x reduction in total computation time", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Difficulty-aware stochastic interpolant policy that adapts the inference budget per control step.", "relevance": 4, "relevance_reason": "Adaptive inference compute for generative policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2510.13714", "title": "DeDelayed: Deleting Remote Inference Delay via On-Device Correction", "authors": ["Dan Jacobellis", "Mateen Ulhaq", "Fabien Racapé", "Hyomin Choi", "Neeraja J. Yadwadkar"], "year": 2025, "venue": "CVPR 2026", "type": "paper", "url": "https://arxiv.org/abs/2510.13714", "pdf_url": "https://arxiv.org/pdf/2510.13714", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["remote inference delay", "on-device correction", "video"], "model_class": "other", "techniques": ["remote model on delayed frames + local model on current frame"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Remote model trained to predict ahead from delayed video frames is fused with a small local model seeing the current frame, cancelling round-trip delay for real-time perception.", "relevance": 4, "relevance_reason": "General remote-delay compensation pattern applicable to robot perception/policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: remote inference robot"}
{"id": "2511.20720", "title": "DeeAD: Dynamic Early Exit of Vision-Language Action for Efficient Autonomous Driving", "authors": ["Haibo HU", "Lianming Huang", "Nan Guan", "Chun Jason Xue"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.20720", "pdf_url": "https://arxiv.org/pdf/2511.20720", "code_url": "", "clusters": ["C1", "C2", "C8", "C9"], "keywords": ["early exit", "autonomous driving", "training-free", "dynamic early exit", "driving VLA"], "model_class": "VLA", "techniques": ["early exit", "layer skipping", "action-guided early exit"], "base_model": "ORION", "params": "", "hardware": "", "reported_metric": "up to 29% latency reduction, 28% layer sparsity", "metric_measures": "relative latency", "benchmarks": ["Bench2Drive"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DeeAD exits the VLA transformer early when intermediate trajectory predictions are good enough, accelerating driving VLA planning without training.", "relevance": 4, "relevance_reason": "Early exit for VLA, driving domain.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA early exit; search: layer skipping VLA; venue sweep: arXiv API VLA AND (MoE OR early exit)"}
{"id": "2512.07697", "title": "Delay-Aware Diffusion Policy: Bridging the Observation-Execution Gap in Dynamic Tasks", "authors": ["Aileen Liao", "Dong-Ki Kim", "Max Olan Smith", "Ali-akbar Agha-mohammadi", "Shayegan Omidshafiei"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.07697", "pdf_url": "https://arxiv.org/pdf/2512.07697", "code_url": "", "clusters": ["C2", "C6", "C7"], "keywords": ["delay-aware policy", "observation-execution gap", "delay-aware diffusion policy", "delay conditioning"], "model_class": "diffusion/flow policy", "techniques": ["delay conditioning", "delay-compensated training targets", "delay-compensated targets"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains diffusion policies on delay-compensated trajectories and conditions on measured delay, and argues performance should be reported as a function of measured latency.", "relevance": 4, "relevance_reason": "Delay-robust policy learning relevant to offloaded inference.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.07339; arXiv search: \"inference delay\" robot policy"}
{"id": "2505.05787", "title": "Demystifying Diffusion Policies: Action Memorization and Simple Lookup Table Alternatives", "authors": ["Cheng-Yang He", "Xu Liu", "Gadiel Sznaier Camps", "G. Sartoretti", "Mac Schwager"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.05787", "pdf_url": "https://arxiv.org/pdf/2505.05787", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["memorization", "OOD monitor"], "model_class": "diffusion/flow policy", "techniques": ["lookup table policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "0.0034x inference time and 0.0085x memory of diffusion policy", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Argues diffusion policies memorize an action lookup table and replaces them with a tiny lookup policy plus an OOD runtime flag.", "relevance": 4, "relevance_reason": "Massive inference/memory reduction for small-data regimes", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2508.20072", "title": "Discrete Diffusion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies", "authors": ["Zhixuan Liang", "Yizhuo Li", "Tianshuo Yang", "Chengyue Wu", "Sitong Mao", "Liuao Pei", "Tian Nian", "Shunbo Zhou", "Xiaokang Yang", "Jiangmiao Pang", "Yao Mu", "Ping Luo"], "year": 2025, "venue": "ICML 2026", "type": "paper", "url": "https://arxiv.org/abs/2508.20072", "pdf_url": "https://arxiv.org/pdf/2508.20072", "code_url": "", "clusters": ["C1"], "keywords": ["discrete diffusion", "parallel decoding", "adaptive order"], "model_class": "VLA", "techniques": ["discrete diffusion decoding", "parallel action decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "SIMPLER", "AgileX real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Models discretized action chunks with discrete diffusion inside the VLM backbone, decoding easy elements first in parallel with re-masking.", "relevance": 4, "relevance_reason": "Parallel decoding VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: parallel decoding action"}
{"id": "2510.07865", "title": "DM1: MeanFlow with Dispersive Regularization for 1-Step Robotic Manipulation", "authors": ["Guo-Wei Zou", "Hai-Tao Wang", "He-Jun Wu", "Yu-Kun Qian", "Yuhang Wang", "Weibin Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.07865", "pdf_url": "https://arxiv.org/pdf/2510.07865", "code_url": "", "clusters": ["C1"], "keywords": ["one-step policy"], "model_class": "diffusion/flow policy", "techniques": ["one-step MeanFlow", "dispersive regularization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "0.07 s vs 2-3.5 s inference (20-40x faster)", "metric_measures": "relative latency", "benchmarks": ["RoboMimic"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "MeanFlow policy with dispersive regularization for strong one-step manipulation.", "relevance": 4, "relevance_reason": "One-step generation with reported speedup.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2506.05985", "title": "Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Lifelong Robot Learning", "authors": ["Yuheng Lei", "Sitong Mao", "Shunbo Zhou", "Hongyuan Zhang", "Xuelong Li", "Ping Luo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.05985", "pdf_url": "https://arxiv.org/pdf/2506.05985", "code_url": "", "clusters": ["C3"], "keywords": ["DMPEL", "expert library", "lifelong learning", "router"], "model_class": "VLA", "techniques": ["low-rank experts", "routing", "coefficient replay"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Progressively grows a low-rank expert library with a lightweight router for lifelong robot learning without task IDs.", "relevance": 4, "relevance_reason": "PEFT-based continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: TAIL task-specific adapters imitation learning LoRA robot continual"}
{"id": "2509.21930", "title": "DynaNav: Dynamic Feature and Layer Selection for Efficient Visual Navigation", "authors": ["Jia-Hui Wang", "Chang-Hao Chen"], "year": 2025, "venue": "NeurIPS 2025", "type": "paper", "url": "https://arxiv.org/abs/2509.21930", "pdf_url": "https://arxiv.org/pdf/2509.21930", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["visual navigation", "early exit", "feature selection", "ViNT"], "model_class": "other", "techniques": ["early exit", "dynamic feature selection"], "base_model": "ViNT", "params": "", "hardware": "", "reported_metric": "2.26x fewer FLOPs, 42.3% lower inference time, 32.8% lower memory vs ViNT", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a hard feature selector and Bayesian-optimized early exits to a navigation foundation model so compute scales with scene complexity.", "relevance": 4, "relevance_reason": "Dynamic depth compression of a navigation foundation model.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2411.02359"}
{"id": "2505.21567", "title": "EaqVLA: Encoding-aligned Quantization for Vision-Language-Action Models", "authors": ["Feng Jiang", "Zi-Hao Zheng", "Xiu-Ping Cui", "Mao-Liang Li", "Jia-Yu Chen", "Xiang Chen", "Zihao Zheng", "Xiuping Cui", "Maoliang Li", "JIayu Chen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.21567", "pdf_url": "https://arxiv.org/pdf/2505.21567", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["encoding alignment", "mixed precision", "encoding-aligned quantization"], "model_class": "VLA", "techniques": ["post-training quantization", "mixed-precision quantization", "encoding-aligned PTQ"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Identifies token/encoding misalignment in VLAs as a barrier for off-the-shelf quantizers and proposes an alignment-aware mixed-precision scheme.", "relevance": 4, "relevance_reason": "VLA-specific PTQ; abstract lacks concrete speed numbers.", "foundational_technique": false, "verified": true, "found_via": "search: vision-language-action quantization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2507.00523", "title": "Edge Computing and its Application in Robotics: A Survey", "authors": ["Nazish Tahir", "Ramviyas Parasuraman"], "year": 2025, "venue": "arXiv", "type": "paper", "url": "https://arxiv.org/abs/2507.00523", "pdf_url": "https://arxiv.org/pdf/2507.00523", "code_url": "", "clusters": ["C6"], "keywords": ["edge robotics", "survey"], "model_class": "other", "techniques": ["edge computing", "offloading"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey of edge computing for robotics covering motivations, applications, offloading approaches and open challenges for latency-critical robot workloads.", "relevance": 4, "relevance_reason": "Recent survey anchoring edge/fog robotics literature.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2507.14049", "title": "EdgeVLA: Efficient Vision-Language-Action Models", "authors": ["Paweł Budzianowski", "Wesley Maa", "Matthew Freed", "Jingxiang Mo", "Winston Hsiao", "Aaron Xie", "Tomasz Młoduchowski", "Viraj Tipnis", "Benjamin Bolte"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.14049", "pdf_url": "https://arxiv.org/pdf/2507.14049", "code_url": "https://github.com/kscalelabs/evla", "clusters": ["C1", "C5", "C6"], "keywords": ["edge VLA", "small language model", "non-autoregressive"], "model_class": "VLA", "techniques": ["small backbone", "parallel action decoding", "small language model backbone", "parallel (non-autoregressive) action decoding", "non-autoregressive action head", "small language model"], "base_model": "OpenVLA-style architecture with SLM", "params": "", "hardware": "", "reported_metric": "7x speedup from removing autoregressive EE prediction | 7x inference speedup from removing autoregressive EE prediction", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "K-Scale's EdgeVLA swaps the large LLM for a small language model and predicts end-effector actions jointly rather than autoregressively to reach real-time rates on edge devices.", "relevance": 4, "relevance_reason": "Seed 'EdgeVLA' verified; onboard efficiency rather than edge-cloud split.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; arXiv: vision-language-action AND Jetson/edge/onboard; search query"}
{"id": "2505.06861", "title": "Efficient Robotic Policy Learning via Latent Space Backward Planning", "authors": ["Dongxiu Liu", "Haoyi Niu", "Zhihao Wang", "Jinliang Zheng", "Yinan Zheng", "Zhonghong Ou", "Jianming Hu", "Jianxiong Li", "Xianyuan Zhan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.06861", "pdf_url": "https://arxiv.org/pdf/2505.06861", "code_url": "", "clusters": ["C8"], "keywords": ["latent backward planning", "subgoals", "real-time long-horizon"], "model_class": "world model", "techniques": ["backward subgoal planning in latent space from final goal"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Plans coarse latent subgoals backward from the task goal instead of generating full future frames, keeping planning cheap and on-task for real-time long-horizon control.", "relevance": 4, "relevance_reason": "Efficient latent planning for policies.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2510.17111", "title": "Efficient Vision-Language-Action Models for Embodied Manipulation: A Systematic Survey", "authors": ["Weifan Guan", "Qinghao Hu", "Aosheng Li", "Jian Cheng", "Guan, Weifan", "Hu, Qinghao", "Li, Aosheng", "Cheng, Jian"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.17111", "pdf_url": "https://arxiv.org/pdf/2510.17111", "code_url": "", "clusters": ["C1", "C2", "C3", "C4", "C5", "C9"], "keywords": ["survey", "efficient VLA", "latency", "memory", "edge", "edge deployment"], "model_class": "VLA", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Systematic survey of efficient VLAs organized by architecture, perception features, action generation, and training/inference strategies, with edge-deployment framing.", "relevance": 4, "relevance_reason": "Covers efficient training/adaptation as one axis of VLA efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient; snowball from 2506.07530; WebSearch: parameter-efficient fine-tuning vision-language-action model arXiv; WebSearch: embodied-efficiency-bench; arXiv: vision-language-action AND Jetson/edge/onboard; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2512.07371", "title": "ESPADA: Execution Speedup via Semantics Aware Demonstration Data Downsampling for Imitation Learning", "authors": ["Byung-Ju Kim", "Jinu Pahk", "Chungwoo Lee", "Jaejoong Kim", "Jangha Lee", "Theo Taeyeong Kim", "Kyuhwan Shim", "Jun-Ki Lee", "Byoung-Tak Zhang"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2512.07371", "pdf_url": "https://arxiv.org/pdf/2512.07371", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["execution speedup", "imitation learning"], "model_class": "other", "techniques": ["demonstration downsampling"], "base_model": "ACT, Diffusion Policy", "params": "", "hardware": "", "reported_metric": "about 2x speed-up while maintaining success", "metric_measures": "throughput", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Semantics-aware downsampling of non-critical demo segments so ACT/DP policies execute faster without retraining changes.", "relevance": 4, "relevance_reason": "Execution speedup for deployed policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2512.10675", "title": "Evaluating Gemini Robotics Policies in a Veo World Simulator", "authors": ["Gemini Robotics Team", "Krzysztof Choromanski", "Coline Devin", "Yilun Du", "Debidatta Dwibedi", "Ruiqi Gao", "Abhishek Jindal", "Thomas Kipf", "Sean Kirmani", "Isabel Leal", "Fangchen Liu", "Anirudha Majumdar", "Andrew Marmon", "Carolina Parada", "Yulia Rubanova", "Dhruv Shah", "Vikas Sindhwani", "Jie Tan", "Fei Xia", "Ted Xiao", "Sherry Yang", "Wenhao Yu", "Allan Zhou"], "year": 2025, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2512.10675", "pdf_url": "https://arxiv.org/pdf/2512.10675", "code_url": "", "clusters": ["C4", "C7", "C8", "C9"], "keywords": ["Veo world simulator", "policy evaluation", "safety probing", "OOD evaluation", "Veo", "video world simulator"], "model_class": "world model", "techniques": ["action-conditioned frontier video model", "multi-view consistency", "generative scene editing", "action-conditioned multi-view video generation"], "base_model": "Veo", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Gemini Robotics policies (real vs simulated)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Google DeepMind adapts Veo into an action-conditioned multi-view simulator to evaluate Gemini Robotics policies on nominal, OOD and physical/semantic safety scenarios produced by editing real scenes.", "relevance": 4, "relevance_reason": "World models for safety/evaluation at Google DeepMind; Dhruv Shah co-author.", "foundational_technique": false, "verified": true, "found_via": "search query; author sweep: Dhruv Shah"}
{"id": "2511.04555", "title": "Evo-1: Lightweight Vision-Language-Action Model with Preserved Semantic Alignment", "authors": ["Tao Lin", "Yilei Zhong", "Yuxin Du", "Jingjing Zhang", "Jiting Liu", "Yinxinyu Chen", "Encheng Gu", "Ziyan Liu", "Hongyi Cai", "Yanwen Zou", "Lixing Zou", "Zhaoye Zhou", "Gen Li", "Bo Zhao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.04555", "pdf_url": "https://arxiv.org/pdf/2511.04555", "code_url": "https://github.com/MINT-SJTU/Evo-1", "clusters": ["C1"], "keywords": ["lightweight VLA", "no robot pretraining", "cross-modulated DiT"], "model_class": "VLA", "techniques": ["small backbone", "diffusion transformer head"], "base_model": "", "params": "0.77B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Meta-World", "RoboTwin", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "0.77B VLA with a cross-modulated diffusion transformer and two-stage alignment, competitive without robot-data pretraining.", "relevance": 4, "relevance_reason": "Lightweight VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: lightweight VLA"}
{"id": "2511.06202", "title": "ExpReS-VLA: Specializing Vision-Language-Action Models Through Experience Replay and Retrieval", "authors": ["Shahram Najam Syed", "Yatharth Ahuja", "Arthur Jakobsson", "Jeff Ichnowski"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.06202", "pdf_url": "https://arxiv.org/pdf/2511.06202", "code_url": "", "clusters": ["C3"], "keywords": ["ExpReS-VLA", "compressed replay", "on-device adaptation"], "model_class": "VLA", "techniques": ["embedding replay", "retrieval augmentation"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Rapid on-device specialization of OpenVLA using compressed embedding replay and retrieval to avoid forgetting.", "relevance": 4, "relevance_reason": "Memory-efficient on-device adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"catastrophic forgetting\""}
{"id": "2506.07639", "title": "Fast ECoT: Efficient Embodied Chain-of-Thought via Thoughts Reuse", "authors": ["Zhekai Duan", "Yuan Zhang", "Shikai Geng", "Gaowen Liu", "Joschka Boedecker", "Chris Xiaoxuan Lu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.07639", "pdf_url": "https://arxiv.org/pdf/2506.07639", "code_url": "", "clusters": ["C1"], "keywords": ["embodied chain-of-thought", "reasoning reuse", "asynchronous"], "model_class": "VLA", "techniques": ["reasoning caching", "parallel generation", "asynchronous scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 7.5% latency reduction", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Caches high-level ECoT reasoning across timesteps, parallelizes reasoning modules, and decouples reasoning from action decoding.", "relevance": 4, "relevance_reason": "Efficient reasoning VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference"}
{"id": "2505.01179", "title": "Fast Flow-based Visuomotor Policies via Conditional Optimal Transport Couplings", "authors": ["Andreas Sochopoulos", "Nikolay Malkin", "Nikolaos Tsagkas", "João Moura", "Michael Gienger", "Sethu Vijayakumar"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.01179", "pdf_url": "https://arxiv.org/pdf/2505.01179", "code_url": "", "clusters": ["C1"], "keywords": ["optimal transport coupling", "straight flows", "few-step"], "model_class": "diffusion/flow policy", "techniques": ["OT coupling", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "10x speedup vs Diffusion Policy", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Conditional optimal-transport couplings between noise and actions straighten flow ODEs so policies work in 1-2 steps without distillation.", "relevance": 4, "relevance_reason": "Few-step flow policy.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy distillation"}
{"id": "2510.12483", "title": "Fast Visuomotor Policy for Robotic Manipulation", "authors": ["Jing-Kai Jia", "Tong Yang", "Xue-Yao Chen", "Chen-Huan Liu", "Wen-Qiang Zhang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2510.12483", "pdf_url": "https://arxiv.org/pdf/2510.12483", "code_url": "", "clusters": ["C1"], "keywords": ["Energy Policy"], "model_class": "diffusion/flow policy", "techniques": ["single-pass generation", "energy score"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MimicGen", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Multimodal action policy trained with an energy score that predicts actions in one forward pass, targeting high-frequency, resource-constrained control.", "relevance": 4, "relevance_reason": "Fast single-pass policy for resource-constrained systems.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2602.13718"}
{"id": "2505.15659", "title": "FLARE: Robot Learning with Implicit World Modeling", "authors": ["Ruijie Zheng", "Jing Wang", "Scott Reed", "Johan Bjorck", "Yu Fang", "Fengyuan Hu", "Joel Jang", "Kaushil Kundalia", "Zongyu Lin", "Loic Magne", "Avnish Narayan", "You Liang Tan", "Guanzhi Wang", "Qi Wang", "Jiannan Xiang", "Yinzhen Xu", "Seonghyeon Ye", "Jan Kautz", "Furong Huang", "Yuke Zhu", "Linxi Fan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.15659", "pdf_url": "https://arxiv.org/pdf/2505.15659", "code_url": "", "clusters": ["C8"], "keywords": ["FLARE", "implicit world modeling", "future latent alignment"], "model_class": "VLA", "techniques": ["future latent alignment tokens in diffusion policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["multitask sim", "humanoid tabletop"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "NVIDIA method that adds a few tokens to a VLA trained to predict future observation embeddings, gaining world-model foresight with negligible inference cost.", "relevance": 4, "relevance_reason": "Implicit world modeling at near-zero runtime cost.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2511.12101", "title": "Freeze, Share, Shrink: Rethinking the Action Backbone in Diffusion Policies", "authors": ["Jian Zhou", "Si-Hao Lin", "Shuai Fu", "Ze-Rui Li", "Geng-Ze Zhou", "Qi Wu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.12101", "pdf_url": "https://arxiv.org/pdf/2511.12101", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["action backbone", "over-parameterization"], "model_class": "diffusion/flow policy", "techniques": ["frozen shared backbone", "model downsizing"], "base_model": "", "params": "5M vs 244M", "hardware": "", "reported_metric": "5M-parameter MLP backbone matches 244M U-Net", "metric_measures": "other", "benchmarks": ["MimicGen", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows the diffusion action backbone can be frozen, shared across tasks and shrunk drastically, since actions are low-dimensional targets.", "relevance": 4, "relevance_reason": "Evidence that action heads are over-parameterized", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2502.01828", "title": "From Foresight to Forethought: VLM-In-the-Loop Policy Steering via Latent Alignment", "authors": ["Yi-Lin Wu", "Ran Tian", "Gokul Swamy", "A. Bajcsy", "Yilin Wu", "Andrea Bajcsy"], "year": 2025, "venue": "RSS 2025", "type": "paper", "url": "https://arxiv.org/abs/2502.01828", "pdf_url": "https://arxiv.org/pdf/2502.01828", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["FOREWARN", "policy steering", "latent world model", "VLM verifier"], "model_class": "world model", "techniques": ["latent alignment of world model to VLM", "sample-and-verify", "latent world model predicts action outcomes", "VLM aligned to latent space verifies"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Uses a latent world model to predict outcomes of candidate action plans and a VLM aligned to that latent space to pick safe/correct ones at runtime.", "relevance": 4, "relevance_reason": "Runtime verification of candidate actions (latency cost).", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2410.04640 (S2); search query"}
{"id": "2508.05635", "title": "Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation", "authors": ["Yue Liao", "Pengfei Zhou", "Siyuan Huang", "Donglin Yang", "Shengcong Chen", "Yuxin Jiang", "Yue Hu", "Jingbin Cai", "Si Liu", "Jianlan Luo", "Liliang Chen", "Shuicheng Yan", "Maoqing Yao", "Guanghui Ren"], "year": 2025, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2508.05635", "pdf_url": "https://arxiv.org/pdf/2508.05635", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["Genie Envisioner", "GE-Act", "GE-Sim", "EWMBench"], "model_class": "world-action model", "techniques": ["instruction-conditioned video diffusion base", "lightweight flow-matching action decoder", "action-conditioned neural simulator"], "base_model": "", "params": "", "hardware": "", "reported_metric": "slow-fast asynchronous mode: video at 5 Hz, actions at 30 Hz; 54-step trajectory within 200 ms (per search summary)", "metric_measures": "open-loop chunk rate", "benchmarks": ["EWMBench", "real robot (AgiBot)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "AgiBot platform unifying a video world foundation model, a lightweight action decoder for policies, and an action-conditioned neural simulator for closed-loop evaluation, with a benchmark for embodied world models.", "relevance": 4, "relevance_reason": "Industrial WAM + neural simulator with slow-fast inference.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2510.19430", "title": "GigaBrain-0: A World Model-Powered Vision-Language-Action Model", "authors": ["GigaBrain Team", "Angen Ye", "Boyuan Wang", "Chaojun Ni", "Guan Huang", "Zheng Zhu", "Guosheng Zhao", "Haoyun Li", "Jie Li", "Jiagang Zhu", "Lv Feng", "Peng Li", "Qiuping Deng", "Runqi Ouyang", "Wenkang Qin", "Xinze Chen", "Xiaofeng Wang", "Yang Wang", "Yifan Li", "Yilong Li", "Yiran Ding", "Yuan Xu", "Yun Ye", "Yukun Zhou", "Zhehao Dong", "Zhenan Wang", "Zhichao Liu", "An-Gen Ye", "Run-Qi Ouyang", "Wen-Kang Qin", "Xiao-Feng Wang", "Yi-Fan Li", "Yi-Long Li", "Yi-Ran Ding", "Yu-Kun Zhou", "Zhi-Chao Liu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.19430", "pdf_url": "https://arxiv.org/pdf/2510.19430", "code_url": "", "clusters": ["C1", "C2", "C5", "C8"], "keywords": ["edge VLA", "Jetson AGX Orin", "small VLM backbone", "torch.compile", "GigaBrain-0", "world-model data engine", "VLA foundation model"], "model_class": "VLA", "techniques": ["architecture downsizing (SmolVLM2 backbone, ~100M action expert)", "removing CPU-GPU transfers/dtype casts", "cached RoPE tables", "torch.compile", "world-model data generation", "small variant", "world-model generated data", "embodied CoT"], "base_model": "GigaBrain-0 / GigaBrain-0-Small vs pi0", "params": "402M (Small) vs 3.2B (pi0)", "hardware": "NVIDIA Jetson AGX Orin (small variant)", "reported_metric": "pi0: 17.5 GB VRAM, 1.28 s/inference, 80% success; GigaBrain-0-Small: 1.9 GB, 0.13 s, 80% success (G1 table bussing, 4400 vs 840 GFLOPs)", "metric_measures": "other", "benchmarks": ["real robot (Unitree G1 table bussing)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "World-model-data-trained VLA whose lightweight Small variant pairs a compact VLM with a ~100M action expert plus system tweaks, running ~10x faster than pi0 on AGX Orin at equal success on one humanoid task.", "relevance": 4, "relevance_reason": "Rare side-by-side on-Orin memory/latency for a pi0-class model vs. a compact variant; stacks architectural and systems optimizations.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: arXiv abs:\"vision-language-action\" AND abs:Orin; recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight); snowball round 2: cited by 2512.00903, 2606.27755"}
{"id": "nvidia-gr00t-n1.6-2025", "title": "GR00T N1.6", "authors": ["NVIDIA GEAR Lab"], "year": 2025, "venue": "tech report / project page (NVIDIA Research)", "type": "tech-report", "url": "https://research.nvidia.com/labs/gear/gr00t-n1_6/", "pdf_url": "", "code_url": "https://github.com/NVIDIA/Isaac-GR00T", "clusters": ["C1"], "keywords": ["GR00T", "dual-system", "DiT action head", "Cosmos VLM"], "model_class": "VLA", "techniques": ["dual-system architecture", "diffusion transformer action head"], "base_model": "Cosmos-2B VLM variant + 32-layer DiT", "params": "3B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "GR00T update doubling the DiT action head to 32 layers, dropping the post-VLM adapter in favor of unfreezing the top 4 VLM layers, and predicting state-relative action chunks.", "relevance": 4, "relevance_reason": "Widely used dual-system VLA baseline for acceleration work.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2503.14734"}
{"id": "2508.17588", "title": "HERO: Hierarchical Extrapolation and Refresh for Efficient World Models", "authors": ["Quanjian Song", "Xinyu Wang", "Dong-Hao Zhou", "Jingyu Lin", "Cunjian Chen", "Yue Ma", "Xiu Li"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2508.17588", "pdf_url": "https://arxiv.org/pdf/2508.17588", "code_url": "", "clusters": ["C8"], "keywords": ["training-free acceleration"], "model_class": "world model", "techniques": ["feature extrapolation", "token refresh"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.73x speedup", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "HERO accelerates diffusion world models by refreshing only volatile shallow tokens and extrapolating deep features.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2301.04104"}
{"id": "2502.19417", "title": "Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models", "authors": ["Lucy Xiaoyang Shi", "Brian Ichter", "Michael Equi", "Liyiming Ke", "Karl Pertsch", "Quan Vuong", "James Tanner", "Anna Walling", "Haohuan Wang", "Niccolo Fusai", "Adrian Li-Bell", "Danny Driess", "Lachy Groom", "Sergey Levine", "Chelsea Finn"], "year": 2025, "venue": "ICML 2025", "type": "paper", "url": "https://arxiv.org/abs/2502.19417", "pdf_url": "https://arxiv.org/pdf/2502.19417", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["hierarchical VLA", "high-level VLM", "System 2", "System 2/System 1"], "model_class": "VLA", "techniques": ["hierarchical policy", "hierarchical inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "A high-level VLM interprets complex prompts and feedback into low-level commands for a pi0 policy, a slow/fast hierarchy for open-ended instructions.", "relevance": 4, "relevance_reason": "Hierarchical split of reasoning vs control rate.", "foundational_technique": false, "verified": true, "found_via": "seed; blog sweep: Physical Intelligence"}
{"id": "2511.18606", "title": "How to Train Your Latent Control Barrier Function: Smooth Safety Filtering Under Hard-to-Model Constraints", "authors": ["Kensuke Nakamura", "Arun L. Bishop", "Steven Man", "Aaron M. Johnson", "Zachary Manchester", "Andrea Bajcsy"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.18606", "pdf_url": "https://arxiv.org/pdf/2511.18606", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["LatentCBF", "smooth safety filtering", "latent reachability", "latent control barrier function", "world model"], "model_class": "world model", "techniques": ["gradient-penalized margins", "mixed-distribution value training", "latent CBF training from reachability values", "QP-based smooth filtering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~2x task completion vs switching filters", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Makes latent reachability values usable as smooth CBFs via gradient penalties on margin functions and mixed nominal/safety-policy data, avoiding performance-harming switching.", "relevance": 4, "relevance_reason": "Minimally invasive filters for visuomotor policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"hamilton-jacobi\" reachability learned policy safety filter; snowball from 2502.00935 (citations)"}
{"id": "2510.06492", "title": "How Well Do Latent World Models Understand Partially Observable Safety Constraints?", "authors": ["Matthew Kim", "Kensuke Nakamura", "Andrea Bajcsy"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.06492", "pdf_url": "https://arxiv.org/pdf/2510.06492", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["partial observability", "safety observability", "latent world model failure modes"], "model_class": "world model", "techniques": ["mutual-information safety observability diagnostic", "rollout predictability diagnostic", "privileged multimodal supervision"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Identifies estimation and prediction gaps where latent world models fail to encode safety-relevant information and proposes diagnostics and mitigations for latent safety filters.", "relevance": 4, "relevance_reason": "When imagination-based safety monitors can fail.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2510.00600", "title": "Hybrid Training for Vision-Language-Action Models", "authors": ["Pietro Mazzaglia", "Cansu Sancaktar", "Markus Peschl", "D. Dijkman"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2510.00600", "pdf_url": "https://arxiv.org/pdf/2510.00600", "code_url": "", "clusters": ["C1"], "keywords": ["Hybrid Training", "chain-of-thought"], "model_class": "VLA", "techniques": ["CoT dropout at inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains VLAs with chain-of-thought but allows skipping thought generation at inference, keeping CoT gains without the action delay.", "relevance": 4, "relevance_reason": "Removes reasoning latency at inference.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2501.09747"}
{"id": "2502.12371", "title": "IMLE Policy: Fast and Sample Efficient Visuomotor Policy Learning via Implicit Maximum Likelihood Estimation", "authors": ["Krishan Rana", "Robert Lee", "David Pershouse", "Niko Suenderhauf"], "year": 2025, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2502.12371", "pdf_url": "https://arxiv.org/pdf/2502.12371", "code_url": "", "clusters": ["C1"], "keywords": ["sample efficiency", "fast inference"], "model_class": "diffusion/flow policy", "techniques": ["IMLE", "single-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "97.3% faster inference than Diffusion Policy", "metric_measures": "relative latency", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Implicit maximum likelihood policy that generates actions in one step and learns from few demonstrations.", "relevance": 4, "relevance_reason": "Single-step policy with large speedup", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2501.16664", "title": "Improving Vision-Language-Action Model with Online Reinforcement Learning", "authors": ["Yanjiang Guo", "Jianke Zhang", "Xiaoyu Chen", "Xiang Ji", "Yen-Jen Wang", "Yucheng Hu", "Jianyu Chen"], "year": 2025, "venue": "ICRA 2025", "type": "preprint", "url": "https://arxiv.org/abs/2501.16664", "pdf_url": "https://arxiv.org/pdf/2501.16664", "code_url": "", "clusters": ["C3"], "keywords": ["iRe-VLA", "iterative RL and SFT"], "model_class": "VLA", "techniques": ["online RL", "supervised fine-tuning", "partial parameter updates"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MetaWorld", "FrankaKitchen", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Alternates RL phases and supervised phases to improve VLAs stably, addressing instability and compute burden of directly applying online RL to large models.", "relevance": 4, "relevance_reason": "Addresses compute cost of RL fine-tuning of VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2502.00622", "title": "Inference-Time Enhancement of Generative Robot Policies via Predictive World Modeling", "authors": ["Han Qi", "Haocheng Yin", "Aris Zhu", "Yilun Du", "Heng Yang"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "preprint", "url": "https://arxiv.org/abs/2502.00622", "pdf_url": "https://arxiv.org/pdf/2502.00622", "code_url": "", "clusters": ["C3", "C7", "C8", "C9"], "keywords": ["generative predictive control", "inference-time planning", "diffusion policy", "inference-time enhancement", "world model"], "model_class": "world model", "techniques": ["action-conditioned world model ranking policy proposals", "online lookahead", "world-model lookahead for frozen diffusion policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Improves a frozen diffusion policy at deployment by predicting outcomes of its action proposals with a world model and ranking/refining them; rollouts dominate runtime, motivating faster world models.", "relevance": 4, "relevance_reason": "Canonical test-time world-model planning baseline (rollout cost bottleneck cited by DriftWorld).", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2507.05674", "title": "Integrating Diffusion-based Multi-task Learning with Online Reinforcement Learning for Robust Quadruped Robot Control", "authors": ["Xinyao Qin", "Xiaoteng Ma", "Yang Qi", "Qihan Liu", "Chuanyi Xue", "Ning Gui", "Qinyu Dong", "Jun Yang", "Bin Liang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2507.05674", "pdf_url": "https://arxiv.org/pdf/2507.05674", "code_url": "", "clusters": ["C5"], "keywords": ["quadruped", "onboard"], "model_class": "diffusion/flow policy", "techniques": ["DDIM sampling", "TensorRT"], "base_model": "", "params": "", "hardware": "onboard (TensorRT)", "reported_metric": "50 Hz onboard", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "DMLoco pretrains a language-conditioned diffusion locomotion policy, fine-tunes with PPO, and deploys onboard with TensorRT.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2505.17016", "title": "Interactive Post-Training for Vision-Language-Action Models", "authors": ["Shuhan Tan", "Kairan Dou", "Yue Zhao", "Philipp Krähenbühl"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.17016", "pdf_url": "https://arxiv.org/pdf/2505.17016", "code_url": "", "clusters": ["C3"], "keywords": ["RIPT-VLA", "interactive post-training", "leave-one-out advantage"], "model_class": "VLA", "techniques": ["online RL", "sparse rewards"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Interactive RL post-training with sparse binary rewards and dynamic rollout sampling; strong gains from very few demos.", "relevance": 4, "relevance_reason": "Data-efficient post-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2503.10966", "title": "Is Your Imitation Learning Policy Better than Mine? Policy Comparison with Near-Optimal Stopping", "authors": ["David Snyder", "Asher James Hancock", "Apurva Badithela", "Emma Dixon", "Patrick Miller", "Rares Andrei Ambrus", "Anirudha Majumdar", "Masha Itkina", "Haruki Nishimura"], "year": 2025, "venue": "RSS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2503.10966", "pdf_url": "https://arxiv.org/pdf/2503.10966", "code_url": "", "clusters": ["C4"], "keywords": ["policy comparison", "sequential test", "near-optimal stopping", "p-hacking"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Sequential statistical test for comparing two imitation policies with small trial budgets that allows adaptive stopping without p-hacking.", "relevance": 4, "relevance_reason": "Statistical evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation statistical"}
{"id": "2509.21027", "title": "KeyWorld: Key Frame Reasoning Enables Effective and Efficient World Models", "authors": ["Sibo Li", "Qianyue Hao", "Yu Shang", "Yong Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.21027", "pdf_url": "https://arxiv.org/pdf/2509.21027", "code_url": "", "clusters": ["C8"], "keywords": ["key frames", "efficient robotic world model", "frame interpolation"], "model_class": "world model", "techniques": ["transformer on semantic key frames", "lightweight conv interpolation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Spends transformer compute only on a few semantically important key frames of a robot world-model rollout and fills the rest with a light convolutional model, speeding up generation.", "relevance": 4, "relevance_reason": "Temporal sparsity for efficient robot world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "bd-tri-lbm-atlas-blog-2025", "title": "Large Behavior Models and Atlas Find New Footing", "authors": ["Eric Cousineau", "Scott Kuindersma", "Lucas Manuelli", "Pat Marion", "Boston Dynamics and TRI teams"], "year": 2025, "venue": "blog (Boston Dynamics / TRI)", "type": "blog", "url": "https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/", "pdf_url": "", "code_url": "", "clusters": ["C3", "C6", "C9"], "keywords": ["LBM", "Atlas", "diffusion transformer", "flow matching", "test-time speedup"], "model_class": "diffusion/flow policy", "techniques": ["flow matching DiT", "inference-time action retiming"], "base_model": "", "params": "450M", "hardware": "", "reported_metric": "30 Hz image input; 48-step chunks (1.6 s), ~24 executed; 1.5–2x test-time speedup without retraining", "metric_measures": "other", "benchmarks": ["real robot (Atlas humanoid)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "450M flow-matching DiT LBM controls Atlas whole-body from language; policies can be executed 1.5–2x faster at test time by retiming predicted actions.", "relevance": 4, "relevance_reason": "Disclosed size/rates and test-time speedup for a humanoid foundation policy.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: TRI"}
{"id": "2504.19322", "title": "Learned Perceptive Forward Dynamics Model for Safe and Platform-aware Robotic Navigation", "authors": ["Pascal Roth", "Jonas Frey", "Cesar Cadena", "Marco Hutter"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2504.19322", "pdf_url": "https://arxiv.org/pdf/2504.19322", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["forward dynamics model", "traversability", "safe navigation"], "model_class": "world model", "techniques": ["learned perceptive forward dynamics", "sampling-based planning with predicted states"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["legged robot navigation (real)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Learned perceptive forward-dynamics world model that predicts a legged robot's future states from geometry and proprioception for real-time safe planning onboard.", "relevance": 4, "relevance_reason": "Onboard real-time world model for safe navigation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2508.10259", "title": "Leveraging OS-Level Primitives for Robotic Action Management", "authors": ["Wen-Xin Zheng", "Boyang Li", "Bin Xu", "Er-Hu Feng", "Jin-Yu Gu", "Hai-Bo Chen"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2508.10259", "pdf_url": "https://arxiv.org/pdf/2508.10259", "code_url": "", "clusters": ["C1", "C5", "C7"], "keywords": ["OS primitives", "execution efficiency"], "model_class": "VLA", "techniques": ["action exception", "computation reuse", "record-and-replay"], "base_model": "", "params": "", "hardware": "", "reported_metric": "29%-74% end-to-end execution time saved", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "AMS borrows OS primitives (exceptions, context switches, replay) to interrupt faulty actions and remove redundant VLA computation.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2509.01819", "title": "ManiFlow: A General Robot Manipulation Policy via Consistency Flow Training", "authors": ["Ge Yan", "Jiyue Zhu", "Yuquan Deng", "Shiqi Yang", "Ri-Zhao Qiu", "Xuxin Cheng", "Marius Memmel", "Ranjay Krishna", "Ankit Goyal", "Xiaolong Wang", "Dieter Fox"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.01819", "pdf_url": "https://arxiv.org/pdf/2509.01819", "code_url": "", "clusters": ["C1"], "keywords": ["consistency flow training", "DiT-X", "1-2 steps"], "model_class": "diffusion/flow policy", "techniques": ["consistency flow training", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Flow policy with consistency training generating dexterous actions in 1-2 steps, using a DiT with adaptive cross-attention conditioning.", "relevance": 4, "relevance_reason": "Few-step general manipulation policy.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching robot policy efficient"}
{"id": "2512.09101", "title": "Masked Generative Policy for Robotic Control", "authors": ["Lipeng Zhuang", "Shiyu Fan", "Florent P. Audonnet", "Yingdong Ru", "Edmond S. L. Ho", "Gerardo Aragon Camarasa", "Paul Henderson"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.09101", "pdf_url": "https://arxiv.org/pdf/2512.09101", "code_url": "", "clusters": ["C1"], "keywords": ["masked generation", "parallel decoding"], "model_class": "other", "techniques": ["masked generative transformer", "parallel token generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 35x lower per-sequence inference time", "metric_measures": "relative latency", "benchmarks": ["Meta-World", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Masked Generative Policy generates discrete action tokens in parallel and refines only low-confidence ones, giving much faster inference than diffusion or autoregressive policies.", "relevance": 4, "relevance_reason": "direct: parallel action decoding", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2511.22697", "title": "Mechanistic Finetuning of Vision-Language-Action Models via Few-Shot Demonstrations", "authors": ["Chancharik Mitra", "Yusen Luo", "Raj Saravanan", "Dantong Niu", "Anirudh Pai", "Jesse Thomason", "Trevor Darrell", "Abrar Anwar", "Deva Ramanan", "Roei Herzig"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.22697", "pdf_url": "https://arxiv.org/pdf/2511.22697", "code_url": "", "clusters": ["C3"], "keywords": ["Robotic Steering", "attention heads", "mechanistic interpretability", "sparse finetuning"], "model_class": "VLA", "techniques": ["sparse fine-tuning", "head selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Franka real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Identifies task-relevant attention heads from few demos and fine-tunes only them, beating LoRA with less compute.", "relevance": 4, "relevance_reason": "Sparse PEFT alternative to LoRA for VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: few-shot adaptation vision-language-action model new embodiment efficient"}
{"id": "2510.05580", "title": "MetaVLA: Unified Meta Co-training For Efficient Embodied Adaption", "authors": ["Chen Li", "Zhantao Yang", "Han Zhang", "Fangyi Chen", "Chenchen Zhu", "Anudeepsekhar Bolimera", "Marios Savvides"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.05580", "pdf_url": "https://arxiv.org/pdf/2510.05580", "code_url": "", "clusters": ["C3"], "keywords": ["meta co-training", "auxiliary tasks", "attentive neural process"], "model_class": "VLA", "techniques": ["meta-learning", "multi-task post-training"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "training steps 240K->75K, ~76% less GPU time", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Consolidates target tasks into one post-training stage with auxiliary tasks and a lightweight ANP-style meta-learner, cutting GPU time substantially.", "relevance": 4, "relevance_reason": "Reduces post-training compute.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: knowledge insulation co-training VLA training efficiency pi0.5 arXiv"}
{"id": "2505.09430", "title": "Mini Diffuser: Fast Multi-Task Diffusion Policy Training Using Two-Level Mini-Batches", "authors": ["Yu-Tong Hu", "Pin-Hao Song", "Kehan Wen", "R. Detry"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2505.09430", "pdf_url": "https://lirias.kuleuven.be/retrieve/8bb84ea3-8330-43d2-94d6-bfce513492fd", "code_url": "", "clusters": ["C3"], "keywords": ["Mini Diffuser", "training efficiency"], "model_class": "diffusion/flow policy", "techniques": ["mini-batching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "5% training time, 7% memory at 95% performance", "metric_measures": "relative memory", "benchmarks": ["RLBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-level mini-batching that cuts multi-task diffusion-policy training time and memory by about an order of magnitude.", "relevance": 4, "relevance_reason": "Training compute/memory efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2511.19433", "title": "Mixture of Horizons in Action Chunking", "authors": ["Dong Jing", "Gang Wang", "Jiaqi Liu", "Weiliang Tang", "Zelong Sun", "Yunchao Yao", "Zhenyu Wei", "Yunhui Liu", "Zhiwu Lu", "Mingyu Ding"], "year": 2025, "venue": "ICML 2026", "type": "paper", "url": "https://arxiv.org/abs/2511.19433", "pdf_url": "https://arxiv.org/pdf/2511.19433", "code_url": "", "clusters": ["C1"], "keywords": ["mixture of horizons", "chunk length", "dynamic inference"], "model_class": "VLA", "techniques": ["multi-horizon action chunking", "adaptive horizon"], "base_model": "pi0, pi0.5", "params": "", "hardware": "", "reported_metric": "2.5x higher throughput than baselines", "metric_measures": "throughput", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Processes several chunk horizons in parallel with a shared action transformer and gated fusion, enabling adaptive horizons via cross-horizon consensus.", "relevance": 4, "relevance_reason": "Chunk-horizon design.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking"}
{"id": "2510.16617", "title": "MoS-VLA: A Vision-Language-Action Model with One-Shot Skill Adaptation", "authors": ["Ruihan Zhao", "Tyler Ingebrand", "Sandeep Chinchali", "Ufuk Topcu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.16617", "pdf_url": "https://arxiv.org/pdf/2510.16617", "code_url": "", "clusters": ["C3"], "keywords": ["mixture of skills", "one-shot adaptation", "basis functions"], "model_class": "VLA", "techniques": ["skill basis", "one-shot adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Represents policies as linear combinations of learned skill bases so one demo suffices to adapt via a lightweight optimization.", "relevance": 4, "relevance_reason": "Gradient-free one-shot adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2508.16845", "title": "NinA: Normalizing Flows in Action. Training VLA Models with Normalizing Flows", "authors": ["Denis Tarasov", "Alexander Nikulin", "Ilya Zisman", "Albina Klepach", "Nikita Lyubaykin", "Andrei Polubarov", "Alexander Derevyagin", "Vladislav Kurenkov"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.16845", "pdf_url": "https://arxiv.org/pdf/2508.16845", "code_url": "https://github.com/dunnolab/NinA/", "clusters": ["C1"], "keywords": ["normalizing flow", "one-shot sampling", "action decoder"], "model_class": "VLA", "techniques": ["normalizing flow action head"], "base_model": "FLOWER", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces a VLA's diffusion decoder with a normalizing flow for one-shot invertible sampling at matched performance.", "relevance": 4, "relevance_reason": "Single-pass action decoder.", "foundational_technique": false, "verified": true, "found_via": "search query: denoising steps VLA"}
{"id": "2504.19854", "title": "NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied Tasks", "authors": ["Chia-Yu Hung", "Qi Sun", "Pengfei Hong", "Amir Zadeh", "Chuan Li", "U-Xuan Tan", "Navonil Majumder", "Soujanya Poria"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2504.19854", "pdf_url": "https://arxiv.org/pdf/2504.19854", "code_url": "", "clusters": ["C1"], "keywords": ["small VLA", "Qwen2.5-VL-3B", "real-time"], "model_class": "VLA", "techniques": ["small backbone", "FAST+ action tokenization"], "base_model": "Qwen-2.5-VL-3B", "params": "3B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A 3B-parameter generalist VLA on a Qwen2.5-VL-3B backbone with FAST+ action tokens, aimed at cutting the compute overhead of 7B-class VLAs while keeping task performance.", "relevance": 4, "relevance_reason": "Directly targets smaller, cheaper VLAs for real-time robot use.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2512.19347", "title": "OMP: One-step Meanflow Policy with Directional Alignment", "authors": ["Han Fang", "Yize Huang", "Yuheng Zhao", "Paul Weng", "Xiao Li", "Yutong Ban"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.19347", "pdf_url": "https://arxiv.org/pdf/2512.19347", "code_url": "", "clusters": ["C1"], "keywords": ["one-step generation", "MeanFlow"], "model_class": "diffusion/flow policy", "techniques": ["one-step MeanFlow"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Adroit", "Meta-World"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "One-step MeanFlow policy with directional alignment and a memory-light JVP computation, beating multi-step baselines on precise tasks.", "relevance": 4, "relevance_reason": "Single-step generative policy for low latency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2507.23675", "title": "One-Step Flow Policy Mirror Descent", "authors": ["Tianyi Chen", "Hai-Tong Ma", "Na Li", "Kai Wang", "Bo Dai"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.23675", "pdf_url": "https://arxiv.org/pdf/2507.23675", "code_url": "", "clusters": ["C1"], "keywords": ["online RL", "flow policy"], "model_class": "diffusion/flow policy", "techniques": ["one-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MuJoCo", "DeepMind Control"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Online RL algorithm enabling one-step flow policy sampling without distillation.", "relevance": 4, "relevance_reason": "One-step policy inference.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2505.03912", "title": "OpenHelix: A Short Survey, Empirical Analysis, and Open-Source Dual-System VLA Model for Robotic Manipulation", "authors": ["Can Cui", "Pengxiang Ding", "Wenxuan Song", "Shuanghao Bai", "Xinyang Tong", "Zirui Ge", "Runze Suo", "Wanqi Zhou", "Yang Liu", "Bofang Jia", "Han Zhao", "Siteng Huang", "Donglin Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.03912", "pdf_url": "https://arxiv.org/pdf/2505.03912", "code_url": "", "clusters": ["C1"], "keywords": ["dual-system VLA", "open source", "survey"], "model_class": "VLA", "techniques": ["dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Surveys and empirically compares dual-system (slow VLM + fast policy) VLA designs and releases a low-cost open-source model.", "relevance": 4, "relevance_reason": "Dual-system VLA analysis.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "openpi-repo", "title": "openpi: open-source models and fine-tuning code for pi0, pi0-FAST and pi0.5", "authors": ["Physical Intelligence"], "year": 2025, "venue": "GitHub", "type": "repo", "url": "https://github.com/Physical-Intelligence/openpi", "pdf_url": "", "code_url": "https://github.com/Physical-Intelligence/openpi", "clusters": ["C3"], "keywords": ["LoRA fine-tuning", "FSDP", "memory requirements"], "model_class": "VLA", "techniques": ["LoRA", "FSDP", "mixed precision"], "base_model": "pi0 / pi0-FAST / pi0.5", "params": "", "hardware": "RTX 4090 (LoRA), A100/H100 80GB (full)", "reported_metric": "inference >8GB; LoRA FT >22.5GB; full FT >70GB GPU memory", "metric_measures": "relative memory", "benchmarks": ["LIBERO", "DROID", "ALOHA"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Official pi-model repo documenting LoRA vs full fine-tuning configs and their GPU memory needs.", "relevance": 4, "relevance_reason": "De-facto fine-tuning recipe and memory budget reference for flow VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: pi0 fine-tuning openpi LoRA recipe Physical Intelligence blog"}
{"id": "2509.20189", "title": "Pagoda: An Energy and Time Roofline Study for DNN Workloads on Edge Accelerators", "authors": ["Prashanthi S. K.", "Kunal Kumar Sahoo", "Amartya Ranjan Saikia", "Pranav Gupta", "Atharva Vinay Joshi", "Priyanshu Pansari", "Yogesh Simmhan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.20189", "pdf_url": "https://arxiv.org/pdf/2509.20189", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["energy roofline", "power modes", "Jetson Orin AGX", "DVFS"], "model_class": "other", "techniques": ["roofline modelling", "power-mode tuning"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin", "reported_metric": "power-mode tuning gives up to 15% lower energy with minimal latency loss; MAXN not most energy-efficient", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Builds time and energy roofline models for Jetson AGX Orin across power modes and uses them to explain and tune DNN inference energy.", "relevance": 4, "relevance_reason": "Methodology for hardware-normalized (roofline) and energy comparison on Jetson.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q3: arXiv abs:roofline AND abs:Jetson"}
{"id": "2512.16881", "title": "PolaRiS: Scalable Real-to-Sim Evaluations for Generalist Robot Policies", "authors": ["Jain, Arhan", "Zhang, Mingtong", "Arora, Kanav", "Chen, William", "Torne, Marcel", "Irshad, Muhammad Zubair", "Zakharov, Sergey", "Wang, Yue", "Levine, Sergey", "Finn, Chelsea", "Ma, Wei-Chiu", "Shah, Dhruv", "Gupta, Abhishek", "Pertsch, Karl"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2512.16881", "pdf_url": "https://arxiv.org/pdf/2512.16881", "code_url": "", "clusters": ["C4"], "keywords": ["PolaRiS", "neural reconstruction", "real-to-sim evaluation", "DROID"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["PolaRiS", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Turns short video scans of real scenes into interactive sim environments via neural reconstruction for high-fidelity generalist-policy evaluation that tracks real performance.", "relevance": 4, "relevance_reason": "Real-to-sim evaluation proxy.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RobotArena infinity"}
{"id": "2503.04051", "title": "RA-DP: Rapid Adaptive Diffusion Policy for Training-Free High-frequency Robotics Replanning", "authors": ["Xiaohan Ye", "Rui Yang", "Jun Jin", "Yin-Chuan Li", "Amir Rasouli"], "year": 2025, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "type": "paper", "url": "https://arxiv.org/abs/2503.04051", "pdf_url": "https://arxiv.org/pdf/2503.04051", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["high-frequency replanning"], "model_class": "diffusion/flow policy", "techniques": ["action queue per denoising step", "training-free guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RA-DP emits replanned actions at each denoising step with guidance, enabling high-frequency diffusion replanning without retraining.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2503.02881", "title": "Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation", "authors": ["Han Xue", "Jieji Ren", "Wendi Chen", "Gu Zhang", "Yuan Fang", "Guoying Gu", "Huazhe Xu", "Cewu Lu", "Guo-Ying Gu"], "year": 2025, "venue": "Robotics", "type": "preprint", "url": "https://arxiv.org/abs/2503.02881", "pdf_url": "https://arxiv.org/pdf/2503.02881", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["reactive diffusion policy", "action chunking", "contact-rich", "tactile", "reactive control"], "model_class": "diffusion/flow policy", "techniques": ["slow-fast hierarchy", "latent diffusion", "tactile feedback"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Slow-fast visual-tactile policy: a slow latent diffusion policy emits chunks while a fast asymmetric tokenizer-based controller reacts to tactile feedback within each chunk; comes with an AR tactile teleoperation system.", "relevance": 4, "relevance_reason": "Seminal slow-fast design for reacting inside open-loop chunks.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "pi-real-time-chunking-blog-2025", "title": "Real-Time Action Chunking with Large Models", "authors": ["Physical Intelligence", "Kevin Black", "Manuel Y. Galliker", "Sergey Levine"], "year": 2025, "venue": "blog (Physical Intelligence)", "type": "blog", "url": "https://www.pi.website/research/real_time_chunking", "pdf_url": "https://www.pi.website/download/real_time_chunking.pdf", "code_url": "", "clusters": ["C1", "C2", "C6", "C7", "C9"], "keywords": ["real-time chunking", "pi0.5", "asynchronous execution", "RTC", "latency", "demo videos"], "model_class": "VLA", "techniques": ["asynchronous inference", "action inpainting", "inpainting", "real-time chunking"], "base_model": "pi0 / pi0.5", "params": "", "hardware": "", "reported_metric": "precise tasks succeed with >300 ms inference delay", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Physical Intelligence research post (June 2025) accompanying RTC, explaining with videos how freezing/inpainting overlapping chunks lets pi0.5 act smoothly despite hundreds of ms of inference latency.", "relevance": 4, "relevance_reason": "Official lab write-up of the RTC seed.", "foundational_technique": false, "verified": true, "found_via": "search query: Physical Intelligence blog; WebSearch: Physical Intelligence blog real-time action chunking; blog sweep: Physical Intelligence"}
{"id": "2508.05396", "title": "Real-Time Iteration Scheme for Diffusion Policy", "authors": ["Yufei Duan", "Hang Yin", "Danica Kragic"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.05396", "pdf_url": "https://arxiv.org/pdf/2508.05396", "code_url": "", "clusters": ["C1"], "keywords": ["real-time iteration", "warm start", "training-free"], "model_class": "diffusion/flow policy", "techniques": ["warm-start denoising", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Borrows the real-time iteration scheme from optimal control, warm-starting denoising from the previous step's solution to cut steps without retraining.", "relevance": 4, "relevance_reason": "Training-free diffusion acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy distillation"}
{"id": "10.1145/3676641.3716016", "title": "ReCA: Integrated Acceleration for Real-Time and Efficient Cooperative Embodied Autonomous Agents", "authors": ["Zishen Wan", "Yuhang Du", "Jiayi Qian"], "year": 2025, "venue": "ASPLOS 2025", "type": "paper", "url": "https://dl.acm.org/doi/abs/10.1145/3676641.3716016", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["cooperative embodied agents", "heterogeneous GPU + accelerator", "planning acceleration"], "model_class": "other", "techniques": ["local model processing", "dual memory", "hierarchical cooperative planning", "heterogeneous hardware (GPU + low-level planning accelerator)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "10.2x speedup with 4.3% more successful missions vs SOTA cooperative agent systems", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cross-stack co-design for LLM-based cooperative embodied agents pairing a GPU for high-level planning with a dedicated accelerator for low-level planning.", "relevance": 4, "relevance_reason": "Heterogeneous accelerator co-design for embodied agents. Author list truncated to the first three named in the search result.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2608.04428 (references)"}
{"id": "1x-redwood-ai-blog-2025", "title": "Redwood AI", "authors": ["1X Technologies"], "year": 2025, "venue": "blog (1X)", "type": "blog", "url": "https://www.1x.tech/discover/redwood-ai", "pdf_url": "", "code_url": "", "clusters": ["C1", "C3", "C9"], "keywords": ["onboard VLA", "whole-body control", "diffusion policy head", "humanoid NEO"], "model_class": "VLA", "techniques": ["small cross-embodiment transformer", "diffusion action decoding", "learning from failures"], "base_model": "", "params": "160M", "hardware": "NEO onboard embedded GPU", "reported_metric": "~5 Hz onboard", "metric_measures": "closed-loop control", "benchmarks": ["real robot (NEO Gamma)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Compact 160M vision-language-proprioception transformer with a diffusion action decoder that runs entirely on NEO's embedded GPU for whole-body mobile manipulation.", "relevance": 4, "relevance_reason": "Small fully-onboard humanoid policy with disclosed size and rate.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: 1X"}
{"id": "2503.05833", "title": "Refined Policy Distillation: From VLA Generalists to RL Experts", "authors": ["Tobias Jülg", "Wolfram Burgard", "Florian Walter"], "year": 2025, "venue": "IROS 2026 (accepted)", "type": "preprint", "url": "https://arxiv.org/abs/2503.05833", "pdf_url": "https://arxiv.org/pdf/2503.05833", "code_url": "", "clusters": ["C2", "C3"], "keywords": ["RL distillation", "VLA teacher", "compact expert", "RPD", "policy distillation", "VLA to RL expert"], "model_class": "VLA", "techniques": ["policy distillation", "RL", "RL distillation", "BC-guided RL"], "base_model": "Octo; OpenVLA teachers", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill3"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses VLA actions to guide on-policy RL of a compact student, producing task experts that outperform their VLA teachers.", "relevance": 4, "relevance_reason": "Compresses generalist into efficient specialist via RL.", "foundational_technique": false, "verified": true, "found_via": "web search: distilling VLA into small student; arXiv API: ti:VLA AND ti:RL"}
{"id": "2505.22094", "title": "ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning", "authors": ["Tonghe Zhang", "Chao Yu", "Sichang Su", "Yu Wang"], "year": 2025, "venue": "NeurIPS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2505.22094", "pdf_url": "https://arxiv.org/pdf/2505.22094", "code_url": "https://github.com/ReinFlow/ReinFlow", "clusters": ["C3"], "keywords": ["flow policy RL", "learnable noise", "few-step"], "model_class": "diffusion/flow policy", "techniques": ["online RL", "few-step flow"], "base_model": "", "params": "", "hardware": "", "reported_metric": "82.63% wall-time saving vs DPPO (locomotion)", "metric_measures": "other", "benchmarks": ["legged locomotion", "manipulation"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Injects learnable noise into flow policies for exact likelihoods, enabling RL fine-tuning even at 1-4 denoising steps with less wall time than diffusion RL.", "relevance": 4, "relevance_reason": "Compute-efficient RL fine-tuning of flow policies; extended to VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025"}
{"id": "2510.04354", "title": "Reliable and Scalable Robot Policy Evaluation with Imperfect Simulators", "authors": ["Apurva Badithela", "David Snyder", "Lihan Zha", "Joseph Mikhail", "Matthew O'Kelly", "Anushri Dixit", "Anirudha Majumdar"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.04354", "pdf_url": "https://arxiv.org/pdf/2510.04354", "code_url": "", "clusters": ["C4"], "keywords": ["SureSim", "prediction-powered inference", "confidence intervals", "sim-real"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "Diffusion Policy, pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Frames combining many sim evaluations with few real ones as prediction-powered inference to get bias-corrected confidence intervals on real performance.", "relevance": 4, "relevance_reason": "Statistical evaluation protocol.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation statistical"}
{"id": "2509.19301", "title": "Residual Off-Policy RL for Finetuning Behavior Cloning Policies", "authors": ["Lars Ankile", "Zhenyu Jiang", "Rocky Duan", "Guanya Shi", "Pieter Abbeel", "Anusha Nagabandi"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.19301", "pdf_url": "https://arxiv.org/pdf/2509.19301", "code_url": "", "clusters": ["C3"], "keywords": ["ResFiT", "residual off-policy RL", "high-DoF"], "model_class": "VLA", "techniques": ["residual RL", "off-policy RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Recipe for sample-efficient real-world residual off-policy RL on top of frozen BC policies, including a high-DoF humanoid.", "relevance": 4, "relevance_reason": "Sample-efficient real-world fine-tuning with frozen base.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2502.12724", "title": "Responsive Noise-Relaying Diffusion Policy: Responsive and Efficient Visuomotor Control", "authors": ["Zhuoqun Chen", "Xiu Yuan", "Tong-Zhou Mu", "Hao Su"], "year": 2025, "venue": "Trans. Mach. Learn. Res.", "type": "paper", "url": "https://arxiv.org/abs/2502.12724", "pdf_url": "https://arxiv.org/pdf/2502.12724", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["responsive control", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["noise relaying", "denoising step reuse"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+18% success vs Diffusion Policy on response-sensitive tasks; +6.9% vs DDIM", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Keeps a noise-relaying buffer so each step reuses partially denoised actions, making diffusion policies more responsive and cheaper.", "relevance": 4, "relevance_reason": "Responsiveness and compute efficiency for diffusion policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2510.14830", "title": "RL-100: Performant Robotic Manipulation with Real-World Reinforcement Learning", "authors": ["Kun Lei", "Huan-Yu Li", "Dongjie Yu", "Zhenyu Wei", "Lingxiao Guo", "Zhennan Jiang", "Ziyu Wang", "Shiyu Liang", "Huazhe Xu"], "year": 2025, "venue": "Sci. Robotics", "type": "paper", "url": "https://arxiv.org/abs/2510.14830", "pdf_url": "https://arxiv.org/pdf/2510.14830", "code_url": "", "clusters": ["C2", "C3"], "keywords": ["real-world RL", "one-step controller", "consistency distillation"], "model_class": "diffusion/flow policy", "techniques": ["real-world RL", "PPO", "consistency distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RL-100 fine-tunes diffusion visuomotor policies with a unified imitation+RL PPO-style objective and uses consistency distillation to compress multi-step diffusion into a one-step controller for high-frequency real-robot control.", "relevance": 4, "relevance_reason": "direct: consistency distillation for deployment latency plus efficient real-world RL", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2601.03044, 2605.25477"}
{"id": "2510.06710", "title": "RLinf-VLA: A Unified and Efficient Framework for Reinforcement Learning of Vision-Language-Action Models", "authors": ["Hongzhi Zang", "Mingjie Wei", "Si Xu", "Yongji Wu", "Zhen Guo", "Yuanqing Wang", "Hao Lin", "Peihong Wang", "Liangzhi Shi", "Yuqing Xie", "Zhexuan Xu", "Zhihao Liu", "Kang Chen", "Wenhao Tang", "Quanlu Zhang", "Weinan Zhang", "Chao Yu", "Yu Wang", "Ming-Jie Wei", "Yong-Ji Wu", "Yuan-Qing Wang", "Liang-Zhi Shi", "Yu-Qing Xie", "Zhe-Xuan Xu", "Zhi-Hao Liu", "Wei-Nan Zhang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2510.06710", "pdf_url": "https://arxiv.org/pdf/2510.06710", "code_url": "https://github.com/RLinf/RLinf", "clusters": ["C3", "C6"], "keywords": ["RLinf-VLA", "RL framework", "pipeline allocation", "VLA+RL", "training framework"], "model_class": "VLA", "techniques": ["online RL", "resource allocation", "pipelining", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.61x-1.88x training speedup on ManiSkill", "metric_measures": "throughput", "benchmarks": ["LIBERO", "ManiSkill", "RoboTwin"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "RLinf-VLA: a unified, efficiency-oriented framework for reinforcement-learning training of VLA models (title-only; no abstract available).", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RLinf VLA reinforcement learning infrastructure embodied arXiv; snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2507.00435", "title": "RoboEval: Where Robotic Manipulation Meets Structured and Scalable Evaluation", "authors": ["Yi Ru Wang", "Carter Ung", "Christopher Tan", "Grant Tannert", "Jiafei Duan", "Josephine Li", "Anh Le", "Rishabh Oswal", "Markus Grotz", "Wilbert Pumacay", "Yuquan Deng", "Ranjay Krishna", "Dieter Fox", "Siddhartha Srinivasa"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2507.00435", "pdf_url": "https://arxiv.org/pdf/2507.00435", "code_url": "", "clusters": ["C4"], "keywords": ["RoboEval", "behavioral metrics", "efficiency", "coordination", "stagewise progress"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboEval"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Bimanual benchmark instrumenting every task with standardized efficiency, coordination and safety/stability metrics plus stagewise progress, validated for discriminative power among policies with similar success.", "relevance": 4, "relevance_reason": "Provides execution-quality metrics beyond binary success, needed to assess efficiency techniques.", "foundational_technique": false, "verified": true, "found_via": "search query: RoboEval"}
{"id": "2505.03673", "title": "RoboOS: A Hierarchical Embodied Framework for Cross-Embodiment and Multi-Agent Collaboration", "authors": ["Huajie Tan", "Xiaoshuai Hao", "Cheng Chi", "Minglan Lin", "Yaoxu Lyu", "Mingyu Cao", "Dong Liang", "Zhuo Chen", "Mengsi Lyu", "Cheng Peng", "Chenrui He", "Yulong Ao", "Yonghua Lin", "Pengwei Wang", "Zhongyuan Wang", "Shanghang Zhang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.03673", "pdf_url": "https://arxiv.org/pdf/2505.03673", "code_url": "", "clusters": ["C6"], "keywords": ["brain-cerebellum hierarchy", "multi-agent", "edge-cloud communication"], "model_class": "other", "techniques": ["hierarchical cloud brain/onboard skills", "cloud distributed inference", "shared memory"], "base_model": "RoboBrain", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Hierarchical embodied OS where a cloud-hosted multimodal brain plans for multiple robots while onboard skill libraries execute, with edge-cloud communication and distributed inference for scalability.", "relevance": 4, "relevance_reason": "Hierarchical onboard/cloud architecture for multi-robot fleets.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2510.23571", "title": "RobotArena $\\infty$: Scalable Robot Benchmarking via Real-to-Sim Translation", "authors": ["Jangir, Yash", "Zhang, Yidi", "Lo, Pang-Chi", "Yamazaki, Kashu", "Zhang, Chenyu", "Tu, Kuan-Hsun", "Ke, Tsung-Wei", "Ke, Lei", "Bisk, Yonatan", "Fragkiadaki, Katerina"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2510.23571", "pdf_url": "https://arxiv.org/pdf/2510.23571", "code_url": "", "clusters": ["C4"], "keywords": ["RobotArena Infinity", "real-to-sim translation", "human preference", "VLM scoring"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RobotArena Infinity"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Automatically converts robot-dataset videos into simulated digital twins and scores VLAs with VLM-guided judgments plus crowd-sourced human preferences, with controlled perturbations for robustness.", "relevance": 4, "relevance_reason": "Scalable, reproducible benchmarking framework.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RobotArena infinity"}
{"id": "10.1088/1674-4926/25020034", "title": "Robotic computing system and embodied AI evolution: an algorithm-hardware co-design perspective", "authors": ["Longke Yan", "Xin Zhao", "Bohan Yang", "Yongkun Wu", "Guangnan Dai", "Jiancong Li", "Chi-Ying Tsui", "Kwang-Ting Cheng", "Yi-Han Zhang", "Feng-Bin Tu"], "year": 2025, "venue": "Journal of Semiconductors", "type": "paper", "url": "https://doi.org/10.1088/1674-4926/25020034", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["robotic computing", "embodied AI hardware"], "model_class": "other", "techniques": ["algorithm-hardware co-design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Review of algorithm-hardware co-design for robotic computing systems as robot algorithms move to end-to-end models.", "relevance": 4, "relevance_reason": "Hardware-aware design survey for embodied AI", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2505.09040", "title": "RT-Cache: Training-Free Retrieval for Real-Time Manipulation", "authors": ["O.-Kil Kwon", "Abraham George", "Alison Bartsch", "A. Farimani"], "year": 2025, "venue": "IEEE-RAS International Conference on Humanoid Robots", "type": "paper", "url": "https://arxiv.org/abs/2505.09040", "pdf_url": "https://arxiv.org/pdf/2505.09040", "code_url": "", "clusters": ["C1"], "keywords": ["retrieval", "caching", "real-time"], "model_class": "other", "techniques": ["retrieval-as-control", "trajectory caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~30% faster task completion vs retrieval baselines", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RT-Cache replaces per-step model calls with retrieval and replay of cached multi-step image-action snippets from a vector memory, shifting cost from compute to storage for real-time control.", "relevance": 4, "relevance_reason": "direct: training-free real-time control via caching", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2410.24164; cited by 2603.17573, 2609.18663"}
{"id": "2509.01728", "title": "SafeDec: Constrained Decoding for Safe Autoregressive Generalist Robot Navigation Policies", "authors": ["Parv Kapoor", "Akila Ganlath", "Michael Clifford", "Changliu Liu", "Sebastian Scherer", "Eunsuk Kang"], "year": 2025, "venue": "ICML 2026", "type": "paper", "url": "https://arxiv.org/abs/2509.01728", "pdf_url": "https://arxiv.org/pdf/2509.01728", "code_url": "https://constrained-robot-fms.github.io/", "clusters": ["C8", "C9"], "keywords": ["constrained decoding", "signal temporal logic", "runtime safety", "autoregressive policy"], "model_class": "VLA", "techniques": ["constrained decoding", "STL specification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CHORES"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Enforces STL safety specs at decoding time on transformer navigation policies, both masking unsafe action tokens and conditioning generation on spec satisfaction, without retraining.", "relevance": 4, "relevance_reason": "Runtime safety layer for robot foundation models by organizer Changliu Liu.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2509.24243", "title": "SafeFlowMatcher: Safe and Fast Planning using Flow Matching with Control Barrier Functions", "authors": ["Jeongyong Yang", "Seunghwan Jang", "SooJean Han"], "year": 2025, "venue": "ICLR 2026", "type": "paper", "url": "https://arxiv.org/abs/2509.24243", "pdf_url": "https://arxiv.org/pdf/2509.24243", "code_url": "", "clusters": ["C7"], "keywords": ["safe flow matching", "real-time planning"], "model_class": "diffusion/flow policy", "techniques": ["flow matching", "CBF-QP", "barrier certificate"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-phase flow-matching planner: a fast prediction step then a CBF-QP correction on the executed path only, with a proved barrier certificate and finite-time convergence to the safe set.", "relevance": 4, "relevance_reason": "Certified-safe fast generative planning; relevant template for certifying a chunk before execution.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\")"}
{"id": "2506.11948", "title": "SAIL: Faster-than-Demonstration Execution of Imitation Learning Policies", "authors": ["N. R. Arachchige", "Zhenyang Chen", "Wonsuhk Jung", "Woo-Chul Shin", "Rohan Bansal", "Pierre Barroso", "Yu-Hang He", "Y. Lin", "Benjamin Joffe", "Shreyas Kousik", "Danfei Xu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.11948", "pdf_url": "https://arxiv.org/pdf/2506.11948", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["faster-than-demonstration", "throughput"], "model_class": "other", "techniques": ["adaptive speed modulation", "latency-aware action scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 4x (sim) and 3.2x (real) speedup over demonstration speed", "metric_measures": "throughput", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Executes imitation policies faster than demonstration speed via consistent action inference, speed modulation and scheduling that handles system latency.", "relevance": 4, "relevance_reason": "Task throughput and latency handling for deployed policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2506.08009", "title": "Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion", "authors": ["Xun Huang", "Zhengqi Li", "Guande He", "Mingyuan Zhou", "Eli Shechtman"], "year": 2025, "venue": "NeurIPS 2025", "type": "paper", "url": "https://arxiv.org/abs/2506.08009", "pdf_url": "https://arxiv.org/pdf/2506.08009", "code_url": "https://self-forcing.github.io/", "clusters": ["C2", "C8"], "keywords": ["Self Forcing", "exposure bias", "few-step AR video diffusion", "rolling KV cache"], "model_class": "world model", "techniques": ["autoregressive self-rollout during training", "few-step diffusion", "rolling KV cache"], "base_model": "", "params": "", "hardware": "single GPU", "reported_metric": "real-time streaming with sub-second latency on a single GPU", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Trains autoregressive video diffusion on its own rollouts with few-step sampling and a rolling KV cache, enabling real-time streaming generation; now the standard recipe for distilling robot world models (e.g., Cosmos-H-Dreams).", "relevance": 4, "relevance_reason": "Foundational technique used by real-time robot world models.", "foundational_technique": true, "verified": true, "found_via": "search query"}
{"id": "2511.00091", "title": "Self-Improving Vision-Language-Action Models with Data Generation via Residual RL", "authors": ["Wenli Xiao", "Haotian Lin", "Andy Peng", "Haoru Xue", "Tairan He", "Yuqi Xie", "Fengyuan Hu", "Jimmy Wu", "Zhengyi Luo", "Linxi \"Jim\" Fan", "Guanya Shi", "Yuke Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.00091", "pdf_url": "https://arxiv.org/pdf/2511.00091", "code_url": "", "clusters": ["C3"], "keywords": ["PLD", "residual RL", "probe-learn-distill"], "model_class": "VLA", "techniques": ["residual RL", "distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Trains lightweight residual actors to probe VLA failures, collects deployment-aligned recovery data and distills it back into the generalist.", "relevance": 4, "relevance_reason": "Cheap self-improvement without human demos.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2410.15185", "title": "Semantically Safe Robot Manipulation: From Semantic Scene Understanding to Motion Safeguards", "authors": ["Lukas Brunke", "Yanni Zhang", "Ralf Römer", "Jack Naimer", "Nikola Staykov", "Siqi Zhou", "Angela P. Schoellig"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.15185", "pdf_url": "https://arxiv.org/pdf/2410.15185", "code_url": "", "clusters": ["C7"], "keywords": ["semantic safety filter", "LLM", "control barrier certification"], "model_class": "diffusion/flow policy", "techniques": ["CBF certification", "semantic map + LLM reasoning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Certifies robot inputs against LLM-inferred semantic constraints (e.g., no water above laptop) plus geometric constraints via control barrier certificates, demonstrated with learned diffusion policies.", "relevance": 4, "relevance_reason": "Safety filter for learned policies with semantic constraints.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" \"diffusion policy\""}
{"id": "2511.10518", "title": "SemanticVLA: Semantic-Aligned Sparsification and Enhancement for Efficient Robotic Manipulation", "authors": ["Wei Li", "Renshan Zhang", "Rui Shao", "Zhijian Fang", "Kaiwen Zhou", "Zhuotao Tian", "Liqiang Nie"], "year": 2025, "venue": "AAAI 2026", "type": "paper", "url": "https://arxiv.org/abs/2511.10518", "pdf_url": "https://arxiv.org/pdf/2511.10518", "code_url": "https://github.com/JiuTian-VL/SemanticVLA", "clusters": ["C1"], "keywords": ["semantic pruning", "dual vision encoder", "sparsification"], "model_class": "VLA", "techniques": ["token pruning", "feature fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.7x lower inference latency, 3.0x lower training cost vs OpenVLA", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Instruction-guided pruning of SigLIP tokens and geometric aggregation of DINOv2 tokens, fused hierarchically for an efficient VLA.", "relevance": 4, "relevance_reason": "Sparsified VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: hierarchical VLA latency"}
{"id": "2509.09674", "title": "SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning", "authors": ["Haozhan Li", "Yuxin Zuo", "Jiale Yu", "Yuhao Zhang", "Zhaohui Yang", "Kaiyan Zhang", "Xuekai Zhu", "Yuchen Zhang", "Tianxing Chen", "Ganqu Cui", "Dehui Wang", "Dingxiang Luo", "Yuchen Fan", "Youbang Sun", "Jia Zeng", "Jiangmiao Pang", "Shanghang Zhang", "Yu Wang", "Yao Mu", "Bowen Zhou", "Ning Ding"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.09674", "pdf_url": "https://arxiv.org/pdf/2509.09674", "code_url": "https://github.com/PRIME-RL/SimpleVLA-RL", "clusters": ["C3"], "keywords": ["RL fine-tuning", "GRPO", "veRL", "pushcut"], "model_class": "VLA", "techniques": ["online RL", "parallel rollouts"], "base_model": "OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboTwin 1.0", "RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Scalable RL framework on veRL with VLA-specific sampling and multi-env rendering; RL on OpenVLA-OFT reduces demo dependence and beats SFT.", "relevance": 4, "relevance_reason": "Key data-efficient RL post-training framework for VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2510.14783", "title": "SkyDreamer: Interpretable End-to-End Vision-Based Drone Racing with Model-Based Reinforcement Learning", "authors": ["Aderik Verraest", "Stavrow Bahnam", "Robin Ferede", "Guido de Croon", "Christophe De Wagter"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.14783", "pdf_url": "https://arxiv.org/pdf/2510.14783", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["drone racing", "informed Dreamer", "onboard execution"], "model_class": "world model", "techniques": ["Dreamer world model decoding privileged state", "end-to-end pixel-to-motor policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "First end-to-end vision drone-racing policy trained with an informed Dreamer world model that transfers sim-to-real and runs fully onboard.", "relevance": 4, "relevance_reason": "Onboard deployment of a world-model-trained policy.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "hf-smolvla-blog-2025", "title": "SmolVLA: Efficient Vision-Language-Action Model trained on LeRobot Community Data", "authors": ["Dana Aubakirova", "Andres Marafioti", "Merve Noyan", "Aritra Roy Gosthipaty", "Francesco Capuano", "Loubna Ben Allal", "Pedro Cuenca", "Mustafa Shukor", "Remi Cadene"], "year": 2025, "venue": "blog (Hugging Face)", "type": "blog", "url": "https://huggingface.co/blog/smolvla", "pdf_url": "", "code_url": "https://github.com/huggingface/lerobot", "clusters": ["C1", "C2", "C9"], "keywords": ["SmolVLA", "layer skipping", "64 visual tokens", "PixelShuffle", "async inference"], "model_class": "VLA", "techniques": ["layer skipping (half the VLM layers)", "visual token reduction (64 tokens/frame)", "interleaved cross/self attention", "asynchronous inference"], "base_model": "", "params": "450M (~100M action expert)", "hardware": "CPU / single consumer GPU / MacBook", "reported_metric": "async: 9.7 s vs 13.75 s per task; 19 vs 9 completions in fixed window", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Blog detailing SmolVLA's efficiency design: half-depth VLM features, 64 visual tokens per frame, a 100M flow-matching expert and async inference, trained on 487 community datasets.", "relevance": 4, "relevance_reason": "Concrete efficiency design choices for a small VLA.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Hugging Face LeRobot"}
{"id": "2511.07820", "title": "SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control", "authors": ["Zhengyi Luo", "Ye Yuan", "Tingwu Wang", "Chen-Ran Li", "Fernando Castañeda", "Sirui Chen", "Zi-ang Cao", "Jiefeng Li", "David Minor", "Qingwei Ben", "Jinhyung D. Park", "David Sami", "Zi Wang", "Xingye Da", "Runyu Ding", "Cyrus Hogg", "Lina Song", "Edy Lim", "Eugene Jeong", "Tairan He", "Haoru Xue", "Wenli Xiao", "S. Yuen", "Jan Kautz", "Yan Chang", "Umar Iqbal", "L. Fan", "Yu-Ke Zhu"], "year": 2025, "venue": "Science Robotics", "type": "paper", "url": "https://arxiv.org/abs/2511.07820", "pdf_url": "https://arxiv.org/pdf/2511.07820", "code_url": "", "clusters": ["C5"], "keywords": ["humanoid whole-body control", "motion tracking foundation model", "Unitree G1", "Jetson Orin", "TensorRT", "CUDA Graph", "multi-rate loops"], "model_class": "other", "techniques": ["TensorRT", "CUDA graphs", "multi-rate architecture"], "base_model": "", "params": "1.2M-42M", "hardware": "Jetson Orin (onboard)", "reported_metric": "1-2 ms per policy forward pass; ~12 ms kinematic planner on Jetson Orin; policy at 50 Hz, command streaming 500 Hz", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "NVIDIA's scaled motion-tracking humanoid controller runs fully onboard a Jetson Orin via TensorRT with CUDA graphs in a four-loop multi-rate architecture that can be driven by a VLA.", "relevance": 4, "relevance_reason": "Onboard multi-rate compute design for a humanoid foundation controller.", "foundational_technique": false, "verified": true, "found_via": "web: Jetson Orin quadruped/humanoid onboard inference"}
{"id": "2506.12723", "title": "SP-VLA: A Joint Model Scheduling and Token Pruning Approach for VLA Model Acceleration", "authors": ["Ye Li", "Yuan Meng", "Zewen Sun", "Kangye Ji", "Chen Tang", "Jiajun Fan", "Xinzhu Ma", "Shutao Xia", "Zhi Wang", "Wenwu Zhu", "Li, Ye", "Meng, Yuan", "Sun, Zewen", "Ji, Kangye", "Tang, Chen", "Fan, Jiajun", "Ma, Xinzhu", "Xia, Shutao", "Wang, Zhi", "Zhu, Wenwu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.12723", "pdf_url": "https://arxiv.org/pdf/2506.12723", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["model scheduling", "token pruning", "fast-slow", "SP-VLA", "joint acceleration"], "model_class": "VLA", "techniques": ["token pruning", "model switching", "model scheduling with lightweight generator"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.5x (LIBERO) / 2.4x (SimplerEnv) lossless acceleration | 1.5x lossless acceleration on LIBERO, 2.4x on SimplerEnv", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "SIMPLER", "SimplerEnv"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Routes 'intuitive' steps to a lightweight generator and 'deliberative' steps to the full VLA, combined with spatio-semantic token pruning.", "relevance": 4, "relevance_reason": "Composition of temporal and spatial redundancy reduction; benchmark-dependent speedups.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning; WebSearch: do VLA acceleration techniques compose"}
{"id": "2504.18792", "title": "STDArm: Transferring Visuomotor Policies From Static Data Training to Dynamic Robot Manipulation", "authors": ["Yifan Duan", "Heng Li", "Yilong Wu", "Wenhao Yu", "Xinran Zhang", "Yedong Shen", "Jianmin Ji", "Yanyong Zhang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2504.18792", "pdf_url": "https://arxiv.org/pdf/2504.18792", "code_url": "", "clusters": ["C7"], "keywords": ["STDArm", "mobile platforms"], "model_class": "other", "techniques": ["latency estimation", "delay compensation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Real-time action-correction layer that raises control frequency, compensates platform motion and estimates latency online so static-trained policies work on moving bases.", "relevance": 4, "relevance_reason": "Delay compensation under onboard compute limits.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2505.21851", "title": "Streaming Flow Policy: Simplifying diffusion/flow-matching policies by treating action trajectories as flow trajectories", "authors": ["Sunshine Jiang", "Xiaolin Fang", "Nicholas Roy", "Tomás Lozano-Pérez", "Leslie Pack Kaelbling", "Siddharth Ancha"], "year": 2025, "venue": "CoRL 2025", "type": "paper", "url": "https://arxiv.org/abs/2505.21851", "pdf_url": "https://arxiv.org/pdf/2505.21851", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["streaming flow policy", "receding horizon", "on-the-fly actions"], "model_class": "diffusion/flow policy", "techniques": ["flow over action trajectory", "streaming"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Treats the action trajectory itself as the flow trajectory starting near the last action, so actions stream to the robot during sampling for tighter sensorimotor loops.", "relevance": 4, "relevance_reason": "Removes wait-for-full-sample latency.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"streaming\" flow policy robot action"}
{"id": "2512.00903", "title": "SwiftVLA: Unlocking Spatiotemporal Dynamics for Lightweight VLA Models at Minimal Overhead", "authors": ["Chaojun Ni", "Cheng Chen", "Xiaofeng Wang", "Zheng Zhu", "Wenzhao Zheng", "Boyuan Wang", "Tianrun Chen", "Guosheng Zhao", "Haoyun Li", "Zhehao Dong", "Qiang Zhang", "Yun Ye", "Yang Wang", "Guan Huang", "Wenjun Mei"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.00903", "pdf_url": "https://arxiv.org/pdf/2512.00903", "code_url": "", "clusters": ["C1"], "keywords": ["lightweight VLA", "4D features", "mask-and-reconstruct"], "model_class": "VLA", "techniques": ["small backbone", "auxiliary branch dropped at inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "18x faster, 12x less memory vs larger VLAs on edge devices", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Gives a compact VLA 4D spatiotemporal understanding via a geometry transformer branch that is trained with masking so it can be dropped at inference.", "relevance": 4, "relevance_reason": "Lightweight VLA design.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference"}
{"id": "2510.23176", "title": "TARC: Time-Adaptive Robotic Control", "authors": ["Arnav Sukhija", "Lenart Treven", "Jin Cheng", "Florian Dorfler", "Stelian Coros", "Andreas Krause"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.23176", "pdf_url": "https://arxiv.org/pdf/2510.23176", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["control rate", "event-triggered"], "model_class": "other", "techniques": ["time-adaptive control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "matches high-frequency controllers at less than half their control frequency", "metric_measures": "other", "benchmarks": ["RC car", "Unitree Go1", "VLA simulation"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns when to switch control so policies run at adaptive rates, matching high-frequency controllers with fewer queries.", "relevance": 4, "relevance_reason": "Reduces inference queries while preserving performance", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2507.01823", "title": "TD-MPC-Opt: Distilling Model-Based Multi-Task Reinforcement Learning Agents", "authors": ["Dmytro Kuzmenko", "Nadiya Shvai"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.01823", "pdf_url": "https://arxiv.org/pdf/2507.01823", "code_url": "https://github.com/dmytro-kuzmenko", "clusters": ["C2", "C8"], "keywords": ["distillation", "TD-MPC2", "FP16 quantization"], "model_class": "world model", "techniques": ["teacher-student distillation 317M -> 1M", "FP16 PTQ"], "base_model": "", "params": "1M (student) from 317M teacher", "hardware": "", "reported_metric": "MT30 normalized score 28.45 vs 18.93 for native 1M; ~50% size reduction from FP16", "metric_measures": "relative memory", "benchmarks": ["MT30 (DMControl)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Distills a 317M multi-task TD-MPC2 world-model agent into a 1M-parameter model that beats a natively trained 1M model, then halves size with FP16 quantization.", "relevance": 4, "relevance_reason": "World-model agent distillation for constrained hardware.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828"}
{"id": "2509.12594", "title": "The Better You Learn, The Smarter You Prune: Towards Efficient Vision-language-action Models via Differentiable Token Pruning", "authors": ["Titong Jiang", "Xuefeng Jiang", "Yuan Ma", "Xin Wen", "Bailin Li", "Kun Zhan", "Peng Jia", "Yahui Liu", "Sheng Sun", "Xianpeng Lang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.12594", "pdf_url": "https://arxiv.org/pdf/2509.12594", "code_url": "", "clusters": ["C1"], "keywords": ["differentiable token pruning", "Gumbel softmax"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "59.1% fewer FLOPs, 38.2% lower latency", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LightVLA learns which visual tokens to keep with Gumbel-softmax selection during fine-tuning, improving both efficiency and success.", "relevance": 4, "relevance_reason": "Learned token pruning for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning"}
{"id": "2505.21200", "title": "Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models", "authors": ["Xudong Tan", "Yaoxin Yang", "Peng Ye", "Jialin Zheng", "Bizhe Bai", "Xinyi Wang", "Jia Hao", "Tao Chen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.21200", "pdf_url": "https://arxiv.org/pdf/2505.21200", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["action reuse", "token selection", "training-free", "FlashVLA", "token pruning", "FLOPs vs latency"], "model_class": "VLA", "techniques": ["action reuse", "token pruning", "visual token pruning"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "55.7% fewer FLOPs, 36.0% lower latency", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free combination of action reuse across stable steps and visual token pruning; FLOP reduction exceeds latency reduction.", "relevance": 4, "relevance_reason": "Composition of two techniques; FLOPs vs latency gap.", "foundational_technique": false, "verified": true, "found_via": "search query: visual token VLA compression; search query: vision-language-action edge deployment benchmark latency"}
{"id": "2512.09927", "title": "Token Expand-Merge: Training-Free Token Compression for Vision-Language-Action Models", "authors": ["Yi-Fan Ye", "Jia-Qi Ma", "Jun Cen", "Zhihe Lu"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2512.09927", "pdf_url": "https://arxiv.org/pdf/2512.09927", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["token compression", "training-free", "expand-merge"], "model_class": "VLA", "techniques": ["visual token merging"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "TEAM-VLA compresses visual tokens with a training-free expand-and-merge scheme to accelerate VLA inference.", "relevance": 4, "relevance_reason": "Training-free token compression for VLAs.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2512.11921", "title": "Towards Accessible Physical AI: LoRA-Based Fine-Tuning of VLA Models for Real-World Robot Control", "authors": ["Abdullah Yahya Abdullah Omaisan", "Ibrahim Sheikh Mohamed"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.11921", "pdf_url": "https://arxiv.org/pdf/2512.11921", "code_url": "", "clusters": ["C2", "C3", "C4", "C5", "C9"], "keywords": ["LoRA", "quantization", "SO101", "low-cost robot", "8GB VRAM", "affordable robots", "real-world fine-tuning"], "model_class": "VLA", "techniques": ["LoRA", "quantization", "frozen vision encoder"], "base_model": "", "params": "~3.1B", "hardware": "consumer GPU with 8GB VRAM", "reported_metric": "3.1B VLA fine-tuned/run within 8GB VRAM", "metric_measures": "relative memory", "benchmarks": ["SO101 real robot button pressing"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Practical study of LoRA plus quantization to adapt a ~3B VLA to a cheap SO101 arm on an 8GB GPU, analyzing frozen vs unfrozen vision encoders and data quantity.", "relevance": 4, "relevance_reason": "Direct memory-efficient VLA fine-tuning for affordable hardware.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: LoRA fine-tuning robot policy VLA adapter arXiv 2025; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2503.07404", "title": "Towards Safe Robot Foundation Models", "authors": ["Maximilian Tölle", "Theo Gruner", "Daniel Palenicek", "Jonas Günster", "Puze Liu", "Joe Watson", "Davide Tateo", "Jan Peters"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.07404", "pdf_url": "https://arxiv.org/pdf/2503.07404", "code_url": "", "clusters": ["C7"], "keywords": ["ATACOM", "safety layer", "generalist policy", "air hockey"], "model_class": "VLA", "techniques": ["constraint-manifold safety layer (ATACOM)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Wraps a generalist robot policy with an ATACOM safety layer that maps actions into a constraint-respecting action space, preventing collisions without safety fine-tuning (air hockey demo).", "relevance": 4, "relevance_reason": "Safety layer for robot foundation models.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: safety filter learned policy robot foundation model CBF VLA"}
{"id": "2505.10219", "title": "Towards Safe Robot Foundation Models Using Inductive Biases", "authors": ["Maximilian Tölle", "Theo Gruner", "Daniel Palenicek", "Tim Schneider", "Jonas Günster", "Joe Watson", "Davide Tateo", "Puze Liu", "Jan Peters"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.10219", "pdf_url": "https://arxiv.org/pdf/2505.10219", "code_url": "", "clusters": ["C7"], "keywords": ["ATACOM", "inductive biases", "formal safety", "robot foundation model"], "model_class": "VLA", "techniques": ["constraint-manifold safety layer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Extends the ATACOM safety layer after a foundation policy to give formal constraint satisfaction for manipulation and fast dynamic tasks without extra safe demonstrations.", "relevance": 4, "relevance_reason": "Formal safety for generalist policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"Towards Safe Robot Foundation Models\""}
{"id": "2509.24527", "title": "Training Agents Inside of Scalable World Models", "authors": ["Danijar Hafner", "Wilson Yan", "Timothy Lillicrap", "Timothy P. Lillicrap"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.24527", "pdf_url": "https://arxiv.org/pdf/2509.24527", "code_url": "", "clusters": ["C8"], "keywords": ["Dreamer 4", "world model", "imagination training", "real-time world model"], "model_class": "world model", "techniques": ["shortcut forcing", "imagination RL", "efficient transformer"], "base_model": "", "params": "", "hardware": "single GPU", "reported_metric": "real-time interactive inference on a single GPU", "metric_measures": "other", "benchmarks": ["Minecraft"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Dreamer 4: scalable world model with shortcut forcing and an efficient transformer achieving real-time interactive inference on one GPU; agents trained purely in imagination from offline data obtain Minecraft diamonds.", "relevance": 4, "relevance_reason": "Efficient, real-time world model for imagination-based policy learning.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight); snowball round 2: cited by 2512.10675, 2602.06949"}
{"id": "2512.15773", "title": "TS-DP: Reinforcement Speculative Decoding For Temporal Adaptive Diffusion Policy Acceleration", "authors": ["Ye Li", "Jiahe Feng", "Yuan Meng", "Kangye Ji", "Chen Tang", "Xinwan Wen", "Shutao Xia", "Zhi Wang", "Wenwu Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.15773", "pdf_url": "https://arxiv.org/pdf/2512.15773", "code_url": "", "clusters": ["C1"], "keywords": ["speculative decoding", "diffusion policy", "RL scheduler"], "model_class": "diffusion/flow policy", "techniques": ["speculative decoding", "distilled drafter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 4.17x faster; 25 Hz inference frequency", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Speculative decoding for diffusion policies with a distilled transformer drafter and an RL scheduler adjusting speculation to task difficulty.", "relevance": 4, "relevance_reason": "Speculative diffusion policy.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy distillation"}
{"id": "2503.00200", "title": "Unified Video Action Model", "authors": ["Shuang Li", "Yihuai Gao", "Dorsa Sadigh", "Shuran Song"], "year": 2025, "venue": "RSS 2025", "type": "paper", "url": "https://arxiv.org/abs/2503.00200", "pdf_url": "https://arxiv.org/pdf/2503.00200", "code_url": "", "clusters": ["C8"], "keywords": ["UVA", "unified video action", "decoupled decoding"], "model_class": "world-action model", "techniques": ["joint video-action latent", "decoupled lightweight action diffusion head", "skip video decoding at inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Learns a joint video-action latent but decodes actions with a separate light head, so policy inference can skip video generation and match direct policies in speed.", "relevance": 4, "relevance_reason": "Early decoupled-decoding WAM for fast inference.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2512.17661", "title": "Vidarc: Embodied Video Diffusion Model for Closed-loop Control", "authors": ["Yao Feng", "Chendong Xiang", "Xin-Yi Mao", "Heng-Kai Tan", "Zuyue Zhang", "Shu-He Huang", "Kaiwen Zheng", "Hai-Tian Liu", "Hang Su", "Jun Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.17661", "pdf_url": "https://arxiv.org/pdf/2512.17661", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["video diffusion policy", "closed-loop control"], "model_class": "world-action model", "techniques": ["cached autoregressive generation", "masked inverse dynamics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "91% latency reduction", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Autoregressive embodied video diffusion with a masked inverse-dynamics model and cached generation for fast closed-loop control.", "relevance": 4, "relevance_reason": "Latency reduction for video-diffusion policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2510.21817", "title": "VITA-E: Natural Embodied Interaction with Concurrent Seeing, Hearing, Speaking, and Acting", "authors": ["Xiaoyu Liu", "Chaoyou Fu", "Chi Yan", "Chu Wu", "Haihan Gao", "Yi-Fan Zhang", "Shaoqi Dong", "Cheng Qian", "Bin Luo", "Xiuyong Yang", "Guanwu Li", "Yusheng Cai", "Yunhang Shen", "Deqiang Jiang", "Haoyu Cao", "Xing Sun", "Caifeng Shan", "Ran He", "Shao-Qi Dong", "Chen Qian", "Yun-Hang Shen", "Hao-Yu Cao"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2510.21817", "pdf_url": "https://arxiv.org/pdf/2510.21817", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["interruption", "concurrency"], "model_class": "VLA", "techniques": ["dual-model architecture", "model-as-controller", "dual active/standby models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Runs an Active and a Standby VLA in parallel and has the VLM emit control tokens, enabling concurrent speaking/acting and near-real-time interruption of ongoing actions.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (interrupt* OR \"human intervention\" OR \"mid-execution\"); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2510.09607", "title": "VITA-VLA: Efficiently Teaching Vision-Language Models to Act via Action Expert Distillation", "authors": ["Shaoqi Dong", "Chaoyou Fu", "Haihan Gao", "Yi-Fan Zhang", "Chi Yan", "Chu Wu", "Xiaoyu Liu", "Yunhang Shen", "Jing Huo", "Deqiang Jiang", "Haoyu Cao", "Yang Gao", "Xing Sun", "Ran He", "Caifeng Shan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://ltbai.github.io/VITA-VLA/", "pdf_url": "https://arxiv.org/pdf/2510.09607", "code_url": "", "clusters": ["C2", "C3"], "keywords": ["action expert distillation", "small action model teacher", "VITA-VLA"], "model_class": "VLA", "techniques": ["knowledge distillation", "distillation", "lightweight alignment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Gives a VLM action ability by aligning its hidden states to a small pretrained action model's action space and reusing that model's decoder, avoiding costly VLA pretraining.", "relevance": 4, "relevance_reason": "Distillation for cheap VLA training rather than inference compression.", "foundational_technique": false, "verified": true, "found_via": "search: distill vision-language-action small model; arXiv API: abs:\"vision-language-action\" AND abs:\"training cost\""}
{"id": "2507.13231", "title": "VITA: Vision-to-Action Flow Matching Policy", "authors": ["De-Chen Gao", "Bo Zhao", "Andrew Lee", "Ian Chuang", "Han-Chu Zhou", "Hang Wang", "Zhe Zhao", "Jun-Shan Zhang", "Iman Soltani"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2507.13231", "pdf_url": "https://arxiv.org/pdf/2507.13231", "code_url": "", "clusters": ["C1"], "keywords": ["vision-to-action flow"], "model_class": "diffusion/flow policy", "techniques": ["noise-free flow matching", "latent action autoencoder"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.5x-2x faster inference", "metric_measures": "relative latency", "benchmarks": ["ALOHA", "Robomimic", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VITA flows directly from visual latents to action latents, removing noise sampling and repeated conditioning in flow policies.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2512.02902", "title": "VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling", "authors": ["Weiqi Li", "Quan Zhang", "Ruifeng Zhai", "Liang Lin", "Guang-Run Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.02902", "pdf_url": "https://arxiv.org/pdf/2512.02902", "code_url": "", "clusters": ["C3"], "keywords": ["one-shot adaptation", "viewpoint robustness"], "model_class": "VLA", "techniques": ["feature token modulation", "low-rank adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "FTM 4K params (48.5% to 87.1%); FLA 4.7M params (90.8%)", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "One-shot visual adaptation of VLAs with tiny affine or low-rank updates that restore viewpoint robustness at a fraction of LoRA cost.", "relevance": 4, "relevance_reason": "Extremely parameter-efficient VLA adaptation", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2306.03310"}
{"id": "2512.24673", "title": "VLA-RAIL: A Real-Time Asynchronous Inference Linker for VLA Models and Robots", "authors": ["Yongsheng Zhao", "Lei Zhao", "Baoping Cheng", "Gongxin Yao", "Xuanzhang Wen", "Han Gao", "Zhao, Yongsheng", "Zhao, Lei", "Cheng, Baoping", "Yao, Gongxin", "Wen, Xuanzhang", "Gao, Han"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.24673", "pdf_url": "https://arxiv.org/pdf/2512.24673", "code_url": "", "clusters": ["C1", "C2", "C4", "C6", "C7"], "keywords": ["asynchronous inference", "chunk fusion", "trajectory smoothing", "asynchronous inference linker", "VLA-RAIL", "chunk fuser"], "model_class": "VLA", "techniques": ["asynchronous inference", "chunk blending", "polynomial smoothing", "chunk fusion", "polynomial trajectory smoothing", "C2-continuous chunk fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["dynamic simulation tasks", "real robot", "real-world manipulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Systems-level linker that runs VLA inference and motion control asynchronously, smoothing each chunk with polynomial fitting and fusing successive chunks with position/velocity/acceleration continuity.", "relevance": 4, "relevance_reason": "Deployment middleware for async (possibly remote) VLA serving.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; WebSearch: closed-loop vs open-loop control frequency VLA; snowball from 2506.07339; arXiv search: \"action chunk\" asynchronous inference"}
{"id": "2510.00406", "title": "VLA-RFT: Vision-Language-Action Reinforcement Fine-tuning with Verified Rewards in World Simulators", "authors": ["Hengtao Li", "Pengxiang Ding", "Runze Suo", "Yihao Wang", "Zirui Ge", "Dongyuan Zang", "Kexian Yu", "Mingyang Sun", "Hongyin Zhang", "Donglin Wang", "Weihua Su"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.00406", "pdf_url": "https://arxiv.org/pdf/2510.00406", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["reinforcement fine-tuning", "world model simulator", "verified rewards"], "model_class": "VLA", "techniques": ["RL fine-tuning", "world-model rollouts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "surpasses SFT baselines with <400 fine-tuning steps", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Uses a learned video world model as a simulator to give dense rewards for RL fine-tuning of a VLA, cutting real-interaction and step requirements.", "relevance": 4, "relevance_reason": "Sample-efficient RL post-training without real-world rollouts.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: parameter-efficient fine-tuning vision-language-action model arXiv"}
{"id": "2505.18719", "title": "VLA-RL: Towards Masterful and General Robotic Manipulation with Scalable Reinforcement Learning", "authors": ["Guanxing Lu", "Wenkai Guo", "Chubin Zhang", "Yuheng Zhou", "Haonan Jiang", "Zifeng Gao", "Yansong Tang", "Ziwei Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.18719", "pdf_url": "https://arxiv.org/pdf/2505.18719", "code_url": "https://github.com/GuanxingLu/vlarl", "clusters": ["C3"], "keywords": ["online RL", "process reward model", "vectorized envs"], "model_class": "VLA", "techniques": ["online RL", "PPO", "batch decoding"], "base_model": "OpenVLA-7B", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Trajectory-level online RL for autoregressive VLAs with a VLM-based process reward model and systems tricks (GPU-balanced vector envs, batch decoding) for scalable training.", "relevance": 4, "relevance_reason": "Systems-level efficiency findings for RL post-training of VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2505.19789", "title": "What Can RL Bring to VLA Generalization? An Empirical Study", "authors": ["Jijia Liu", "Feng Gao", "Bingwen Wei", "Xinlei Chen", "Qingmin Liao", "Yi Wu", "Chao Yu", "Yu Wang"], "year": 2025, "venue": "NeurIPS 2025", "type": "preprint", "url": "https://arxiv.org/abs/2505.19789", "pdf_url": "https://arxiv.org/pdf/2505.19789", "code_url": "", "clusters": ["C3"], "keywords": ["RL4VLA", "PPO", "generalization", "empirical study"], "model_class": "VLA", "techniques": ["online RL", "PPO", "LoRA"], "base_model": "OpenVLA-7B", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill-based generalization benchmark"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Empirical study (RL4VLA) comparing PPO/GRPO/DPO fine-tuning of VLAs against SFT; finds PPO best and provides an efficient PPO recipe.", "relevance": 4, "relevance_reason": "Provides efficient RL fine-tuning recipe for VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2512.24497", "title": "What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?", "authors": ["Basile Terver", "Tsung-Yen Yang", "Jean Ponce", "Adrien Bardes", "Yann LeCun"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.24497", "pdf_url": "https://arxiv.org/pdf/2512.24497", "code_url": "", "clusters": ["C8"], "keywords": ["JEPA-WM", "planning in representation space", "design study"], "model_class": "world model", "techniques": ["study of architecture, objective, planner for JEPA world models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["navigation", "manipulation", "real robot data"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Systematic study of what makes planning in JEPA world model latent space work (architecture, loss, planner), yielding a model that beats DINO-WM and V-JEPA 2-AC.", "relevance": 4, "relevance_reason": "Design guidance for latent planners, the efficient alternative to pixel world models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.09985"}
{"id": "2503.02552", "title": "World Models for Anomaly Detection during Model-Based Reinforcement Learning Inference", "authors": ["Fabian Domberg", "Georg Schildbach"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.02552", "pdf_url": "https://arxiv.org/pdf/2503.02552", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["anomaly detection", "prediction discrepancy", "emergency stop", "world model", "deployment monitoring"], "model_class": "world model", "techniques": ["monitor world-model prediction error at inference", "threshold-triggered stop", "world-model prediction error monitoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulated robot control", "real-world"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Uses discrepancy between a model-based RL world model's predictions and observations at deployment as a task-agnostic anomaly detector that can trigger an emergency stop.", "relevance": 4, "relevance_reason": "World models for real-time safety monitoring (CFP topic).", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API world model AND safety AND robot"}
{"id": "2512.05927", "title": "World Models That Know When They Don't Know - Controllable Video Generation with Calibrated Uncertainty", "authors": ["Zhiting Mei", "Tenny Yin", "Micah Baker", "Ola Shorinwa", "Anirudha Majumdar"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.05927", "pdf_url": "https://arxiv.org/pdf/2512.05927", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["calibrated uncertainty", "hallucination detection", "controllable video model"], "model_class": "world model", "techniques": ["subpatch-level calibrated confidence estimation (C3)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Trains controllable video world models to output dense, calibrated confidence maps so hallucinated predictions can be detected during policy evaluation and planning.", "relevance": 4, "relevance_reason": "Uncertainty for trustworthy world-model monitoring.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2511.00062", "title": "World Simulation with Video Foundation Models for Physical AI", "authors": ["NVIDIA", ":", "Arslan Ali", "Junjie Bai", "Maciej Bala", "Yogesh Balaji", "Aaron Blakeman", "Tiffany Cai", "Jiaxin Cao", "Tianshi Cao", "Elizabeth Cha", "Yu-Wei Chao", "Prithvijit Chattopadhyay", "Mike Chen", "Yongxin Chen", "Yu Chen", "Shuai Cheng", "Yin Cui", "Jenna Diamond", "Yifan Ding", "Jiaojiao Fan", "Linxi Fan", "Liang Feng", "Francesco Ferroni", "Sanja Fidler", "Xiao Fu", "Ruiyuan Gao", "Yunhao Ge", "Jinwei Gu", "Aryaman Gupta", "Siddharth Gururani", "Imad El Hanafi", "Ali Hassani", "Zekun Hao", "Jacob Huffman", "Joel Jang", "Pooya Jannaty", "Jan Kautz", "Grace Lam", "Xuan Li", "Zhaoshuo Li", "Maosheng Liao", "Chen-Hsuan Lin", "Tsung-Yi Lin", "Yen-Chen Lin", "Huan Ling", "Ming-Yu Liu", "Xian Liu", "Yifan Lu", "Alice Luo", "Qianli Ma", "Hanzi Mao", "Kaichun Mo", "Seungjun Nah", "Yashraj Narang", "Abhijeet Panaskar", "Lindsey Pavao", "Trung Pham", "Morteza Ramezanali", "Fitsum Reda", "Scott Reed", "Xuanchi Ren", "Haonan Shao", "Yue Shen", "Stella Shi", "Shuran Song", "Bartosz Stefaniak", "Shangkun Sun", "Shitao Tang", "Sameena Tasmeen", "Lyne Tchapmi", "Wei-Cheng Tseng", "Jibin Varghese", "Andrew Z. Wang", "Hao Wang", "Haoxiang Wang", "Heng Wang", "Ting-Chun Wang", "Fangyin Wei", "Jiashu Xu", "Dinghao Yang", "Xiaodong Yang", "Haotian Ye", "Seonghyeon Ye", "Xiaohui Zeng", "Jing Zhang", "Qinsheng Zhang", "Kaiwen Zheng", "Andrew Zhu", "Yuke Zhu"], "year": 2025, "venue": "arXiv (NVIDIA tech report)", "type": "tech-report", "url": "https://arxiv.org/abs/2511.00062", "pdf_url": "https://arxiv.org/pdf/2511.00062", "code_url": "https://github.com/nvidia-cosmos/cosmos-predict2.5", "clusters": ["C8"], "keywords": ["Cosmos-Predict2.5", "flow matching", "world simulation", "timestep distillation"], "model_class": "world model", "techniques": ["flow-based video world model", "RL post-training", "smaller Transfer model"], "base_model": "", "params": "2B and 14B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Next-generation Cosmos world model unifying text/image/video-to-world generation with flow matching at 2B/14B scale; companion Cosmos-Transfer2.5 is 3.5x smaller than its predecessor; repo documents few-step distillation for faster rollout.", "relevance": 4, "relevance_reason": "Seed family (Cosmos); common backbone for efficient robot WAMs and distilled real-time simulators.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2510.18135", "title": "World-in-World: World Models in a Closed-Loop World", "authors": ["Jiahan Zhang", "Muqing Jiang", "Nanru Dai", "Taiming Lu", "Arda Uzunoglu", "Shunchi Zhang", "Yana Wei", "Jiahao Wang", "Vishal M. Patel", "Paul Pu Liang", "Daniel Khashabi", "Cheng Peng", "Rama Chellappa", "Tianmin Shu", "Alan Yuille", "Yilun Du", "Jieneng Chen"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2510.18135", "pdf_url": "https://arxiv.org/pdf/2510.18135", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["closed-loop benchmark", "embodied utility", "world model evaluation"], "model_class": "world model", "techniques": ["unified online planning API", "closed-loop evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["World-in-World (4 closed-loop environments)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Benchmark that tests world models by task success in closed-loop planning rather than visual quality, finding controllability matters more than fidelity and reporting data/compute scaling for inference-time planning.", "relevance": 4, "relevance_reason": "Evaluation of world models for planning utility (incl. inference compute).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2505.19017", "title": "WorldEval: World Model as Real-World Robot Policies Evaluator", "authors": ["Yaxuan Li", "Yichen Zhu", "Junjie Wen", "Chaomin Shen", "Yi Xu", "Yi-Chen Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.19017", "pdf_url": "https://arxiv.org/pdf/2505.19017", "code_url": "", "clusters": ["C4", "C7", "C8", "C9"], "keywords": ["WorldEval", "world model evaluator", "policy ranking", "Policy2Vec", "safety detector", "policy evaluation"], "model_class": "world model", "techniques": ["latent-action conditioning of video generator", "online policy evaluation", "action-conditioned video generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot", "real-world manipulation policy ranking"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Turns a video generator into a latent-action-following simulator to rank real-robot policies and checkpoints, and to act as a safety detector that flags dangerous actions from new models before deployment.", "relevance": 4, "relevance_reason": "World model as policy evaluator and pre-deployment safety screen.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: AutoEval autonomous evaluation; search query; venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2511.20095", "title": "WPT: World-to-Policy Transfer via Online World Model Distillation", "authors": ["Guangfeng Jiang", "Yueru Luo", "Jun Liu", "Yi Huang", "Yiyao Zhu", "Zhan Qu", "Dave Zhenyu Chen", "Bingbing Liu", "Xu Yan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.20095", "pdf_url": "https://arxiv.org/pdf/2511.20095", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["world-to-policy distillation", "lightweight student", "driving"], "model_class": "world model", "techniques": ["world-model-guided reward", "policy and reward distillation into student"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Distills world-model reasoning into a lightweight student driving policy via a world-model-aligned reward, removing world-model inference overhead at runtime.", "relevance": 4, "relevance_reason": "Removes runtime world-model cost by distillation.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2402.04292", "title": "AdaFlow: Imitation Learning with Variance-Adaptive Flow-Based Policies", "authors": ["Xixi Hu", "Bo Liu", "Xing-Chao Liu", "Qiang Liu"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2402.04292", "pdf_url": "https://arxiv.org/pdf/2402.04292", "code_url": "", "clusters": ["C1"], "keywords": ["adaptive step size", "fast inference"], "model_class": "diffusion/flow policy", "techniques": ["variance-adaptive ODE solver", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "AdaFlow uses flow-based policies with a variance-adaptive ODE solver that collapses to one step when the action distribution is unimodal.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2406.10093", "title": "BiKC: Keypose-Conditioned Consistency Policy for Bimanual Robotic Manipulation", "authors": ["Dongjie Yu", "Hang Xu", "Yizhou Chen", "Yi Ren", "Jia Pan"], "year": 2024, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2406.10093", "pdf_url": "https://arxiv.org/pdf/2406.10093", "code_url": "", "clusters": ["C1"], "keywords": ["keypose conditioning", "bimanual"], "model_class": "diffusion/flow policy", "techniques": ["consistency model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "BiKC pairs a keypose predictor with a consistency-model trajectory generator for fast bimanual manipulation.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2412.06782", "title": "CARP: Visuomotor Policy Learning via Coarse-to-Fine Autoregressive Prediction", "authors": ["Zhefei Gong", "Pengxiang Ding", "Shangke Lyu", "Siteng Huang", "Mingyang Sun", "Wei Zhao", "Zhaoxin Fan", "Dong-Lin Wang"], "year": 2024, "venue": "IEEE International Conference on Computer Vision", "type": "paper", "url": "https://arxiv.org/abs/2412.06782", "pdf_url": "https://arxiv.org/pdf/2412.06782", "code_url": "", "clusters": ["C1"], "keywords": ["next-scale prediction"], "model_class": "other", "techniques": ["coarse-to-fine autoregression", "multi-scale action tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "10x faster inference than SOTA policies", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "CARP generates action sequences coarse-to-fine with a GPT-style model, matching diffusion-policy accuracy at autoregressive speed.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2411.19650", "title": "CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation", "authors": ["Qixiu Li", "Yaobo Liang", "Zeyu Wang", "Lin Luo", "Xi Chen", "Mozheng Liao", "Fangyun Wei", "Yu Deng", "Sicheng Xu", "Yizhong Zhang", "Xiaofan Wang", "Bei Liu", "Jianlong Fu", "Jianmin Bao", "Dong Chen", "Yuanchun Shi", "Jiaolong Yang", "Baining Guo"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2411.19650", "pdf_url": "https://arxiv.org/pdf/2411.19650", "code_url": "", "clusters": ["C1"], "keywords": ["componentized VLA", "diffusion action transformer"], "model_class": "VLA", "techniques": ["diffusion action head"], "base_model": "", "params": "7B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SIMPLER", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VLA with a separate diffusion action transformer conditioned on VLM output, beating OpenVLA substantially; the base model for several acceleration papers.", "relevance": 4, "relevance_reason": "Seed; common acceleration target.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2409.00588", "title": "Diffusion Policy Policy Optimization", "authors": ["Allen Z. Ren", "Justin Lidard", "Lars L. Ankile", "Anthony Simeonov", "Pulkit Agrawal", "Anirudha Majumdar", "Benjamin Burchfiel", "Hongkai Dai", "Max Simchowitz"], "year": 2024, "venue": "ICLR 2025", "type": "preprint", "url": "https://arxiv.org/abs/2409.00588", "pdf_url": "https://arxiv.org/pdf/2409.00588", "code_url": "", "clusters": ["C3"], "keywords": ["DPPO", "policy gradient", "diffusion policy fine-tuning"], "model_class": "diffusion/flow policy", "techniques": ["online RL", "PPO over denoising MDP"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Gym", "Robomimic", "Furniture-Bench", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Treats diffusion denoising as an MDP and fine-tunes diffusion policies with PPO, finding it the most stable and efficient RL fine-tuning option.", "relevance": 4, "relevance_reason": "Foundational RL fine-tuning method for generative policies used by VLA RL works.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: ti:\"policy optimization\" AND abs:diffusion AND abs:\"fine-tuning\""}
{"id": "2411.15139", "title": "DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving", "authors": ["Bencheng Liao", "Shaoyu Chen", "Haoran Yin", "Bo Jiang", "Cheng Wang", "Sixu Yan", "Xinbang Zhang", "Xiangyu Li", "Ying Zhang", "Qian Zhang", "Xinggang Wang"], "year": 2024, "venue": "CVPR 2025", "type": "paper", "url": "https://arxiv.org/abs/2411.15139", "pdf_url": "https://arxiv.org/pdf/2411.15139", "code_url": "https://github.com/hustvl/DiffusionDrive", "clusters": ["C1"], "keywords": ["DiffusionDrive"], "model_class": "diffusion/flow policy", "techniques": ["truncated diffusion", "anchors"], "base_model": "", "params": "", "hardware": "NVIDIA RTX 4090", "reported_metric": "45 FPS; 10x fewer denoising steps", "metric_measures": "throughput", "benchmarks": ["NAVSIM"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Truncated diffusion driving policy starting from anchored Gaussians with a cascade decoder, needing 2 denoising steps.", "relevance": 4, "relevance_reason": "Widely cited real-time diffusion planner.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (\"diffusion policy\" OR \"flow matching\") AND robot* AND accelerat* AND (distill* OR \"denoising steps\")"}
{"id": "2405.20935", "title": "Effective Interplay between Sparsity and Quantization: From Theory to Practice", "authors": ["Simla Burcu Harma", "Ayan Chakraborty", "Elizaveta Kostenok", "Danila Mishin", "Dongho Ha", "Babak Falsafi", "Martin Jaggi", "Ming Liu", "Yunho Oh", "Suvinay Subramanian", "Amir Yazdanbakhsh"], "year": 2024, "venue": "arXiv", "type": "paper", "url": "https://arxiv.org/abs/2405.20935", "pdf_url": "https://arxiv.org/pdf/2405.20935", "code_url": "", "clusters": ["C2"], "keywords": ["sparsity", "quantization", "non-orthogonality", "order of compression"], "model_class": "other", "techniques": ["pruning", "quantization"], "base_model": "OPT, LLaMA (125M-8B), ViT, ResNet", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Proves and shows empirically that sparsity and quantization are not orthogonal: errors compound, and quantizing before pruning distorts importance scores, so order and joint design matter.", "relevance": 4, "relevance_reason": "Core foundational evidence against naive stacking of compression techniques.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q1: arXiv ti:\"interplay between sparsity and quantization\""}
{"id": "2412.04987", "title": "FlowPolicy: Enabling Fast and Robust 3D Flow-based Policy via Consistency Flow Matching for Robot Manipulation", "authors": ["Qinglun Zhang", "Zhen Liu", "Haoqiang Fan", "Guanghui Liu", "Bing Zeng", "Shuaicheng Liu"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.04987", "pdf_url": "https://arxiv.org/pdf/2412.04987", "code_url": "https://github.com/zql-kk/FlowPolicy", "clusters": ["C1"], "keywords": ["consistency flow matching", "3D policy", "one-step"], "model_class": "diffusion/flow policy", "techniques": ["consistency flow matching", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "7x faster inference", "metric_measures": "relative latency", "benchmarks": ["Adroit", "Meta-World"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Consistency flow matching on point-cloud policies to produce actions in one step.", "relevance": 4, "relevance_reason": "One-step flow policy.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching robot policy efficient"}
{"id": "10.1109/icra57147.2024.10610759", "title": "FogROS2-LS: A Location-Independent Fog Robotics Framework for Latency Sensitive ROS2 Applications", "authors": ["Kai-Peng Chen", "Michael Wang", "Marcus Gualtieri", "Nan Tian", "Christian Juette", "Liu Ren", "Jeffrey Ichnowski", "John Kubiatowicz", "Ken Goldberg"], "year": 2024, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://doi.org/10.1109/ICRA57147.2024.10610759", "pdf_url": "", "code_url": "", "clusters": ["C6"], "keywords": ["fog robotics", "ROS2", "latency-sensitive"], "model_class": "other", "techniques": ["anycast routing", "cloud/edge offloading"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 8.5x fewer collision failures", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "FogROS2-LS offloads latency-sensitive robot computation to the best available cloud/edge server transparently in ROS2.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 1902.05703, 1903.09589"}
{"id": "2410.05562", "title": "FogROS2-PLR: Probabilistic Latency-Reliability For Cloud Robotics", "authors": ["Kaiyuan Chen", "Nan Tian", "Christian Juette", "Tianshuang Qiu", "Liu Ren", "John Kubiatowicz", "Ken Goldberg"], "year": 2024, "venue": "ICRA 2025 (submitted)", "type": "preprint", "url": "https://arxiv.org/abs/2410.05562", "pdf_url": "https://arxiv.org/pdf/2410.05562", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["latency reliability", "5G", "multi-interface", "replication", "P99"], "model_class": "other", "techniques": ["server replication", "multi-network-interface routing"], "base_model": "", "params": "", "hardware": "two 5G carriers, cloud servers", "reported_metric": "up to 3.7x lower P99 latency vs single provider", "metric_measures": "relative latency", "benchmarks": ["cloud-connected driving experiment"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Formulates an impossibility triangle for latency reliability and sends each request over several network interfaces to replicated cloud servers, taking the first answer to cut deadline misses.", "relevance": 4, "relevance_reason": "Tail-latency reliability for cloud-offloaded robot compute.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2412.05408"}
{"id": "genie-2-deepmind-blog", "title": "Genie 2: A large-scale foundation world model", "authors": ["Google DeepMind"], "year": 2024, "venue": "Google DeepMind blog", "type": "blog", "url": "https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation-world-model/", "pdf_url": "", "code_url": "", "clusters": ["C8"], "keywords": ["foundation world model", "interactive 3D environments", "embodied agent training"], "model_class": "world model", "techniques": ["video autoencoder", "latent autoregressive transformer", "action-conditioned frame-by-frame sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Autoregressive latent video world model that turns a single image into a playable, action-controllable 3D environment with up to ~1 minute of consistency, used to train and evaluate embodied agents such as SIMA.", "relevance": 4, "relevance_reason": "Seed (Genie 2); foundation world model as an agent training environment.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2402.15391", "title": "Genie: Generative Interactive Environments", "authors": ["Jake Bruce", "Michael Dennis", "Ashley Edwards", "Jack Parker-Holder", "Yuge Shi", "Edward Hughes", "Matthew Lai", "Aditi Mavalankar", "Richie Steigerwald", "Chris Apps", "Yusuf Aytar", "Sarah Bechtle", "Feryal Behbahani", "Stephanie Chan", "Nicolas Heess", "Lucy Gonzalez", "Simon Osindero", "Sherjil Ozair", "Scott Reed", "Jingwei Zhang", "Konrad Zolna", "Jeff Clune", "Nando de Freitas", "Satinder Singh", "Tim Rocktäschel"], "year": 2024, "venue": "ICML 2024", "type": "paper", "url": "https://arxiv.org/abs/2402.15391", "pdf_url": "https://arxiv.org/pdf/2402.15391", "code_url": "", "clusters": ["C8"], "keywords": ["latent action model", "generative interactive environment", "foundation world model"], "model_class": "world model", "techniques": ["spatiotemporal video tokenizer", "autoregressive dynamics", "unsupervised latent actions"], "base_model": "", "params": "11B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Platformers", "RT-1 robotics videos"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Learns an action-controllable world model from unlabeled internet video by inferring discrete latent actions, allowing frame-by-frame interaction with generated environments and imitation from unseen videos.", "relevance": 4, "relevance_reason": "Seed (Genie 1); origin of latent-action world models reused in robot WAMs (DreamDojo, LAPA).", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2405.15223", "title": "iVideoGPT: Interactive VideoGPTs are Scalable World Models", "authors": ["Jialong Wu", "Shaofeng Yin", "Ningya Feng", "Xu He", "Dong Li", "Jianye Hao", "Mingsheng Long"], "year": 2024, "venue": "NeurIPS 2024", "type": "paper", "url": "https://arxiv.org/abs/2405.15223", "pdf_url": "https://arxiv.org/pdf/2405.15223", "code_url": "https://github.com/thuml/iVideoGPT", "clusters": ["C8"], "keywords": ["iVideoGPT", "compressive tokenization", "interactive world model"], "model_class": "world model", "techniques": ["compressive visual tokenization", "autoregressive transformer over obs/action/reward tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["action-conditioned video prediction", "visual planning", "MBRL"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Scalable autoregressive world model whose compressive tokenizer keeps token counts low enough for interactive use, pretrained on millions of human/robot trajectories.", "relevance": 4, "relevance_reason": "Token-compression approach to efficient interactive world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2409.07163", "title": "Mamba Policy: Towards Efficient 3D Diffusion Policy with Hybrid Selective State Models", "authors": ["Jia-Hang Cao", "Qiang Zhang", "Jing-Kai Sun", "Jia-Xu Wang", "Hao-Tai Cheng", "Yu-Ling Li", "Jun Ma", "Kun Wu", "Zhiyuan Xu", "Yecheng Shao", "Wen Zhao", "Gang Han", "Yi-Jie Guo", "Renjing Xu"], "year": 2024, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "type": "paper", "url": "https://arxiv.org/abs/2409.07163", "pdf_url": "https://arxiv.org/pdf/2409.07163", "code_url": "", "clusters": ["C1"], "keywords": ["3D diffusion policy", "lightweight"], "model_class": "diffusion/flow policy", "techniques": ["Mamba backbone"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">80% fewer parameters", "metric_measures": "relative memory", "benchmarks": ["Adroit", "DexArt", "MetaWorld"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Mamba-attention hybrid policy network cutting parameters substantially.", "relevance": 4, "relevance_reason": "Lightweight policy for resource-limited devices.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2303.04137"}
{"id": "2406.01586", "title": "ManiCM: Real-time 3D Diffusion Policy via Consistency Model for Robotic Manipulation", "authors": ["Zifeng Gao", "Guanxing Lu", "Tianxing Chen", "Wenxun Dai", "Ziwei Wang", "Chao Shang", "Wenbo Ding", "Yansong Tang", "Zi-Feng Gao", "Tian-Xing Chen", "Wen-Dao Dai", "Wen-Bo Ding"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2406.01586", "pdf_url": "https://arxiv.org/pdf/2406.01586", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["consistency model", "3D diffusion policy", "point cloud", "point clouds"], "model_class": "diffusion/flow policy", "techniques": ["consistency distillation", "one-step generation"], "base_model": "DP3", "params": "", "hardware": "", "reported_metric": "10x faster average inference", "metric_measures": "relative latency", "benchmarks": ["Adroit", "Meta-World", "MetaWorld"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Applies consistency distillation to 3D point-cloud diffusion policies for one-step action generation.", "relevance": 4, "relevance_reason": "Consistency distillation for 3D policies.", "foundational_technique": false, "verified": true, "found_via": "search query: one-step diffusion policy; search: one-step policy distillation robot"}
{"id": "2410.00425", "title": "ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI", "authors": ["Tao, Stone", "Xiang, Fanbo", "Shukla, Arth", "Qin, Yuzhe", "Hinrichsen, Xander", "Yuan, Xiaodi", "Bao, Chen", "Lin, Xinsong", "Liu, Yulin", "Chan, Tse-kai", "Gao, Yuan", "Li, Xuanlin", "Mu, Tongzhou", "Xiao, Nan", "Gurha, Arnav", "Rajesh, Viswesh Nagaswamy", "Choi, Yong Woo", "Chen, Yen-Ru", "Huang, Zhiao", "Calandra, Roberto", "Chen, Rui", "Luo, Shan", "Su, Hao"], "year": 2024, "venue": "RSS 2025", "type": "benchmark", "url": "https://arxiv.org/abs/2410.00425", "pdf_url": "https://arxiv.org/pdf/2410.00425", "code_url": "https://github.com/haosulab/ManiSkill", "clusters": ["C4"], "keywords": ["ManiSkill3", "GPU-parallel simulation", "rendering throughput"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "GPU", "reported_metric": "up to 30,000+ FPS with rendering; 10-1000x faster, 2-3x less GPU memory than other platforms", "metric_measures": "throughput", "benchmarks": ["ManiSkill3"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "GPU-parallelized simulation and rendering across 12 domains, enabling far faster evaluation/training and fast parallel policy evaluation.", "relevance": 4, "relevance_reason": "Seed simulator enabling cheap large-sample evaluation.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2410.12032", "title": "MLPerf Power: Benchmarking the Energy Efficiency of Machine Learning Systems from Microwatts to Megawatts for Sustainable AI", "authors": ["Arya Tschand", "Arun Tejusve Raghunath Rajan", "Sachin Idgunji", "Vijay Janapa Reddi"], "year": 2024, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2410.12032", "pdf_url": "https://arxiv.org/pdf/2410.12032", "code_url": "", "clusters": ["C4"], "keywords": ["energy efficiency", "power measurement rules", "edge to datacenter"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "1,841 reproducible measurements from 60 systems", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Defines MLPerf's power-measurement methodology spanning microwatt to megawatt systems and analyzes 1,841 measurements for energy-efficiency trade-offs.", "relevance": 4, "relevance_reason": "Standard energy-per-inference measurement methodology applicable to robot compute.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q3: arXiv id lookup MLPerf Power"}
{"id": "2410.22689", "title": "Multi-Task Interactive Robot Fleet Learning with Visual World Models", "authors": ["Huihan Liu", "Yu Zhang", "Vaarij Betala", "Evan Zhang", "James Liu", "Crystal Ding", "Yuke Zhu"], "year": 2024, "venue": "CoRL 2024", "type": "paper", "url": "https://arxiv.org/abs/2410.22689", "pdf_url": "https://arxiv.org/pdf/2410.22689", "code_url": "", "clusters": ["C6", "C7", "C8"], "keywords": ["Sirius-Fleet", "fleet learning", "visual world model anomaly prediction"], "model_class": "world model", "techniques": ["visual world model predicting future outcomes", "adaptive anomaly predictors", "human intervention requests"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot fleet"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Fleet-learning system that uses a visual world model to forecast whether upcoming actions will lead to anomalies and asks humans for help only when needed, reducing supervision over time.", "relevance": 4, "relevance_reason": "World-model safety monitoring at fleet scale.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2405.12213", "title": "Octo: An Open-Source Generalist Robot Policy", "authors": ["Octo Model Team", "Dibya Ghosh", "Homer Walke", "Karl Pertsch", "Kevin Black", "Oier Mees", "Sudeep Dasari", "Joey Hejna", "Tobias Kreiman", "Charles Xu", "Jianlan Luo", "You Liang Tan", "Lawrence Yunliang Chen", "Pannag Sanketi", "Quan Vuong", "Ted Xiao", "Dorsa Sadigh", "Chelsea Finn", "Sergey Levine"], "year": 2024, "venue": "RSS 2024", "type": "preprint", "url": "https://arxiv.org/abs/2405.12213", "pdf_url": "https://arxiv.org/pdf/2405.12213", "code_url": "https://github.com/octo-models/octo", "clusters": ["C1", "C3"], "keywords": ["generalist policy", "Open X-Embodiment", "diffusion head", "finetuning to new action spaces", "OXE"], "model_class": "diffusion/flow policy", "techniques": ["diffusion action head", "action chunking", "full fine-tuning", "modular tokenizers"], "base_model": "Octo", "params": "27M/93M", "hardware": "consumer GPU", "reported_metric": "finetuning within a few hours on consumer GPUs", "metric_measures": "other", "benchmarks": ["9 real robot platforms"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Transformer diffusion-head generalist policy trained on 800k OXE trajectories, designed to be finetuned cheaply to new sensors and action spaces.", "relevance": 4, "relevance_reason": "Reference small generalist policy for low-cost adaptation.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2410.10394", "title": "PIVOT-R: Primitive-Driven Waypoint-Aware World Model for Robotic Manipulation", "authors": ["Kai-Dong Zhang", "Peng-Zhen Ren", "Bingqian Lin", "Junfan Lin", "Shikui Ma", "Hang Xu", "Xiao-Dan Liang"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2410.10394", "pdf_url": "https://arxiv.org/pdf/2410.10394", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["waypoint world model"], "model_class": "world model", "techniques": ["asynchronous hierarchical execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SeaWave"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PIVOT-R predicts task waypoints with a world model and runs modules at different rates via an asynchronous executor to cut redundant computation.", "relevance": 4, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2405.05941"}
{"id": "2412.06685", "title": "Policy Agnostic RL: Offline RL and Online RL Fine-Tuning of Any Class and Backbone", "authors": ["Max Sobol Mark", "Tian Gao", "Georgia Gabriela Sampaio", "Mohan Kumar Srirama", "Archit Sharma", "Chelsea Finn", "Aviral Kumar"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.06685", "pdf_url": "https://arxiv.org/pdf/2412.06685", "code_url": "", "clusters": ["C3"], "keywords": ["policy-agnostic RL", "Cal-QL", "action re-ranking"], "model_class": "VLA", "techniques": ["offline RL", "online RL", "action optimization"], "base_model": "OpenVLA-7B; diffusion policies", "params": "", "hardware": "", "reported_metric": "OpenVLA 40%->70% real-world in 40 minutes", "metric_measures": "other", "benchmarks": ["real robot", "simulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Replaces the policy-improvement step with supervised learning on critic-optimized actions, enabling RL fine-tuning of any policy class including a 7B VLA.", "relevance": 4, "relevance_reason": "Sample-efficient RL fine-tuning applicable to large VLAs.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2412.13630", "title": "Policy Decorator: Model-Agnostic Online Refinement for Large Policy Model", "authors": ["Xiu Yuan", "Tongzhou Mu", "Stone Tao", "Yunhao Fang", "Mengke Zhang", "Hao Su"], "year": 2024, "venue": "ICLR 2025", "type": "preprint", "url": "https://arxiv.org/abs/2412.13630", "pdf_url": "https://arxiv.org/pdf/2412.13630", "code_url": "", "clusters": ["C3"], "keywords": ["Policy Decorator", "residual policy", "online refinement"], "model_class": "diffusion/flow policy", "techniques": ["residual RL", "controlled exploration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill", "Adroit"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Model-agnostic residual policy refines large imitation policies online with controlled exploration, leaving the base frozen.", "relevance": 4, "relevance_reason": "Lightweight adaptation of large policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2404.00456", "title": "QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs", "authors": ["Saleh Ashkboos", "Amirkeivan Mohtashami", "Maximilian L. Croci", "Bo Li", "Pashmina Cameron", "Martin Jaggi", "Dan Alistarh", "Torsten Hoefler", "James Hensman"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2404.00456", "pdf_url": "https://arxiv.org/pdf/2404.00456", "code_url": "", "clusters": ["C2"], "keywords": ["Hadamard rotation", "W4A4", "outlier-free"], "model_class": "other", "techniques": ["rotation-based quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Rotates hidden states with Hadamard transforms to remove outliers so weights, activations and KV cache can all be quantized to 4 bits.", "relevance": 4, "relevance_reason": "Rotation basis for W4A4 VLA methods (HoloQ-VLA, FoldQuantVLA).", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2410.07864", "title": "RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation", "authors": ["Songming Liu", "Lingxuan Wu", "Bangguo Li", "Hengkai Tan", "Huayu Chen", "Zhengyi Wang", "Ke Xu", "Hang Su", "Jun Zhu"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.07864", "pdf_url": "https://arxiv.org/pdf/2410.07864", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion foundation model", "bimanual", "unified action space"], "model_class": "diffusion/flow policy", "techniques": ["diffusion transformer"], "base_model": "", "params": "1.2B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "1.2B diffusion transformer foundation model for bimanual manipulation with a physically interpretable unified action space.", "relevance": 4, "relevance_reason": "Seed diffusion VLA.", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2407.08735", "title": "Real-Time Anomaly Detection and Reactive Planning with Large Language Models", "authors": ["Rohan Sinha", "Amine Elhafsi", "Christopher Agia", "Matthew Foutter", "E. Schmerling", "Marco Pavone"], "year": 2024, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2407.08735", "pdf_url": "https://arxiv.org/pdf/2407.08735", "code_url": "", "clusters": ["C7"], "keywords": ["runtime monitoring", "LLM", "anomaly detection"], "model_class": "other", "techniques": ["fast anomaly classifier", "fallback planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["quadrotor", "autonomous vehicle"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Two-stage monitor: a fast embedding-based anomaly classifier triggers a slower LLM reasoner, while fallback plans keep the system safe during reasoning latency.", "relevance": 4, "relevance_reason": "Safety under foundation-model latency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2410.04640, 2502.01828"}
{"id": "2412.09265", "title": "Score and Distribution Matching Policy: Advanced Accelerated Visuomotor Policies via Matched Distillation", "authors": ["Bofang Jia", "Pengxiang Ding", "Can Cui", "Mingyang Sun", "Pengfang Qian", "Siteng Huang", "Zhaoxin Fan", "Donglin Wang"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.09265", "pdf_url": "https://arxiv.org/pdf/2412.09265", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["score matching", "distribution matching", "dual teacher"], "model_class": "diffusion/flow policy", "techniques": ["diffusion distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "6x inference speedup", "metric_measures": "relative latency", "benchmarks": ["57-task simulation benchmark"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-stage score- and distribution-matching distillation with frozen and adversarial teachers turns diffusion policies into single-step generators.", "relevance": 4, "relevance_reason": "One-step distillation for visuomotor policies.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2410.21257"}
{"id": "2410.13816", "title": "Steering Your Generalists: Improving Robotic Foundation Models via Value Guidance", "authors": ["Mitsuhiko Nakamoto", "Oier Mees", "Aviral Kumar", "Sergey Levine"], "year": 2024, "venue": "CoRL 2024", "type": "preprint", "url": "https://arxiv.org/abs/2410.13816", "pdf_url": "https://arxiv.org/pdf/2410.13816", "code_url": "", "clusters": ["C3"], "keywords": ["V-GPS", "value-guided steering", "re-ranking"], "model_class": "VLA", "techniques": ["value-function re-ranking", "inference-time steering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Re-ranks action samples from a frozen generalist policy with an offline-RL value function, improving it without fine-tuning weights.", "relevance": 4, "relevance_reason": "Weight-free adaptation of generalists.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2406.04806", "title": "Streaming Diffusion Policy: Fast Policy Synthesis with Variable Noise Diffusion Models", "authors": ["Sigmund H. Høeg", "Yilun Du", "Olav Egeland"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2406.04806", "pdf_url": "https://arxiv.org/pdf/2406.04806", "code_url": "", "clusters": ["C1"], "keywords": ["streaming diffusion policy", "variable noise", "rolling"], "model_class": "diffusion/flow policy", "techniques": ["rolling diffusion", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Outputs trajectories with increasing noise over the horizon so each new observation needs only a few denoising steps on the shifted previous plan.", "relevance": 4, "relevance_reason": "Streaming diffusion policy.", "foundational_technique": false, "verified": true, "found_via": "search query: streaming diffusion policy"}
{"id": "2402.17764", "title": "The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits", "authors": ["Shu-Ming Ma", "Hongyu Wang", "Ling-Xiao Ma", "Lei Wang", "Wen-Hui Wang", "Shao-Han Huang", "Lifeng Dong", "Rui-Ping Wang", "Jilong Xue", "Furu Wei"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2402.17764", "pdf_url": "https://arxiv.org/pdf/2402.17764", "code_url": "", "clusters": ["C2"], "keywords": ["ternary", "1.58-bit", "BitNet"], "model_class": "other", "techniques": ["ternary quantization-aware training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains LLMs from scratch with ternary {-1,0,1} weights that match full-precision perplexity at scale with large efficiency gains.", "relevance": 4, "relevance_reason": "Direct foundation of BitVLA.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2506.07530 (references)"}
{"id": "2409.11403", "title": "UniLCD: Unified Local-Cloud Decision-Making via Reinforcement Learning", "authors": ["Kathakoli Sengupta", "Zhongkai Shangguan", "Sandesh Bharadwaj", "Sanjay Arora", "Eshed Ohn-Bar", "Renato Mancuso"], "year": 2024, "venue": "ECCV 2024", "type": "paper", "url": "https://arxiv.org/abs/2409.11403", "pdf_url": "https://arxiv.org/pdf/2409.11403", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["local-cloud routing", "RL", "navigation", "energy/latency/safety"], "model_class": "other", "techniques": ["learned local/cloud routing policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["crowded navigation (simulation)"], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Learns via RL when a mobile robot should use its local model versus a cloud model, balancing energy, latency and safety in crowded navigation.", "relevance": 4, "relevance_reason": "Learned onboard/cloud switching for end-to-end embodied policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: split computing robot"}
{"id": "2406.12538", "title": "Variational Distillation of Diffusion Policies into Mixture of Experts", "authors": ["Hongyi Zhou", "Denis Blessing", "Ge Li", "Onur Celik", "Xiaogang Jia", "Gerhard Neumann", "Rudolf Lioutikov", "Xiao-Gang Jia"], "year": 2024, "venue": "NeurIPS 2024", "type": "paper", "url": "https://arxiv.org/abs/2406.12538", "pdf_url": "https://arxiv.org/pdf/2406.12538", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["diffusion distillation", "mixture of experts", "variational", "variational distillation", "likelihood"], "model_class": "diffusion/flow policy", "techniques": ["distillation", "mixture of experts", "diffusion distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills diffusion policies into mixture-of-experts Gaussian policies via a variational bound, giving fast single-pass inference that keeps multimodality.", "relevance": 4, "relevance_reason": "Distilling diffusion policies into fast students.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy distillation; snowball from 2405.07503"}
{"id": "2412.14803", "title": "Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations", "authors": ["Yucheng Hu", "Yanjiang Guo", "Pengchao Wang", "Xiaoyu Chen", "Yen-Jen Wang", "Jianke Zhang", "Koushil Sreenath", "Chaochao Lu", "Jianyu Chen"], "year": 2024, "venue": "ICML 2025", "type": "paper", "url": "https://arxiv.org/abs/2412.14803", "pdf_url": "https://arxiv.org/pdf/2412.14803", "code_url": "", "clusters": ["C8"], "keywords": ["video prediction policy", "predictive visual representations", "single-step VDM features"], "model_class": "world-action model", "techniques": ["implicit inverse dynamics on video diffusion features", "single denoising step feature extraction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN ABC-D", "real dexterous manipulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Uses internal features from a single forward pass of a fine-tuned video diffusion model as predictive representations for an action head, avoiding full video generation while gaining future awareness.", "relevance": 4, "relevance_reason": "Early representation-only WAM that sidesteps costly video denoising.", "foundational_technique": false, "verified": true, "found_via": "snowball from NVIDIA WAM blog"}
{"id": "2306.00978", "title": "AWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration", "authors": ["Ji Lin", "Jiaming Tang", "Haotian Tang", "Shang Yang", "Wei-Ming Chen", "Wei-Chen Wang", "Guangxuan Xiao", "Xingyu Dang", "Chuang Gan", "Song Han", "Guang-Xuan Xiao"], "year": 2023, "venue": "MLSys 2024", "type": "paper", "url": "https://doi.org/10.48550/arXiv.2306.00978", "pdf_url": "https://arxiv.org/pdf/2306.00978", "code_url": "https://github.com/mit-han-lab/llm-awq", "clusters": ["C1", "C2", "C5"], "keywords": ["weight-only quantization", "activation-aware scaling", "4-bit", "AWQ", "TinyChat", "Jetson Orin", "on-device LLM", "weight quantization", "LLM"], "model_class": "other", "techniques": ["post-training quantization", "activation-aware weight quantization", "kernel fusion", "SIMD-aware packing", "quantization"], "base_model": "", "params": "", "hardware": "Jetson Orin, RTX 4090, desktop/mobile GPUs", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Protects the small fraction of weight channels that matter most (judged by activation magnitude) via per-channel scaling, enabling accurate 4-bit weight-only LLM quantization.", "relevance": 4, "relevance_reason": "Foundational on-device quantization + kernels used by edge VLA work.", "foundational_technique": true, "verified": true, "found_via": "seed; snowball round 2: cites 2210.17323; cited by 2404.00456, 2412.01034"}
{"id": "2307.08691", "title": "FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning", "authors": ["Tri Dao"], "year": 2023, "venue": "ICLR 2024", "type": "paper", "url": "https://arxiv.org/abs/2307.08691", "pdf_url": "https://arxiv.org/pdf/2307.08691", "code_url": "https://github.com/Dao-AILab/flash-attention", "clusters": ["C5"], "keywords": ["attention kernel", "work partitioning", "GPU parallelism"], "model_class": "other", "techniques": ["optimized attention kernel"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reworks FlashAttention's parallelization and warp-level work partitioning to reach much higher GPU utilization.", "relevance": 4, "relevance_reason": "Standard attention kernel in VLA training/inference stacks.", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2311.05600", "title": "FogROS2-Config: Optimizing Latency and Cost for Multi-Cloud Robot Applications", "authors": ["Kaiyuan Chen", "Kush Hari", "Rohil Khare", "Charlotte Le", "Trinity Chung", "Jaimyn Drake", "Jeffrey Ichnowski", "John Kubiatowicz", "Ken Goldberg"], "year": 2023, "venue": "ICRA 2024", "type": "paper", "url": "https://arxiv.org/abs/2311.05600", "pdf_url": "https://arxiv.org/pdf/2311.05600", "code_url": "https://github.com/BerkeleyAutomation/FogROS2", "clusters": ["C6"], "keywords": ["server selection", "cost-latency Pareto", "multi-cloud"], "model_class": "other", "techniques": ["cloud configuration search", "benchmark sampling"], "base_model": "", "params": "", "hardware": "cloud instances (multi-provider)", "reported_metric": "up to 20x cost reduction", "metric_measures": "other", "benchmarks": ["visual SLAM", "grasp planning", "motion planning"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Toolkit that benchmarks ROS 2 nodes on a small sample of cloud server types to present a latency-vs-cost menu, helping choose cost-efficient offload hardware.", "relevance": 4, "relevance_reason": "Directly on cost/latency trade-off (TCO) of cloud robot compute.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2310.06114", "title": "Learning Interactive Real-World Simulators", "authors": ["Sherry Yang", "Yilun Du", "Kamyar Ghasemipour", "Jonathan Tompson", "Leslie Kaelbling", "Dale Schuurmans", "Pieter Abbeel"], "year": 2023, "venue": "ICLR 2024", "type": "paper", "url": "https://arxiv.org/abs/2310.06114", "pdf_url": "https://arxiv.org/pdf/2310.06114", "code_url": "", "clusters": ["C8"], "keywords": ["UniSim", "universal simulator", "action-conditioned video generation"], "model_class": "world model", "techniques": ["action-conditioned video diffusion", "training policies in learned simulator", "sim-to-real zero-shot"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot", "Language Table"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains a single video-diffusion simulator on heterogeneous data (robot, navigation, human, image) that responds to both language and low-level actions, and trains VLA and RL policies purely inside it that transfer zero-shot to real robots.", "relevance": 4, "relevance_reason": "Seed; early demonstration of learned video simulators for robot policy training (compute cost is the key bottleneck).", "foundational_technique": false, "verified": true, "found_via": "seed"}
{"id": "2309.11453", "title": "Multi-Step Model Predictive Safety Filters: Reducing Chattering by Increasing the Prediction Horizon", "authors": ["Federico Pizarro Bejarano", "Lukas Brunke", "Angela P. Schoellig"], "year": 2023, "venue": "CDC 2023", "type": "paper", "url": "https://arxiv.org/abs/2309.11453", "pdf_url": "https://arxiv.org/pdf/2309.11453", "code_url": "https://github.com/Federico-PizarroBejarano/safe-control-gym/tree/smooth_mpsc_paper", "clusters": ["C7"], "keywords": ["chattering", "multi-step safety filter"], "model_class": "other", "techniques": ["model predictive safety filter", "robust MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Model predictive safety filter that corrects inputs over a longer horizon instead of the next step only, reducing chattering with recursive feasibility under bounded model error; quadrotor experiments.", "relevance": 4, "relevance_reason": "Horizon-level filtering matches correcting whole action chunks.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"predictive safety filter\""}
{"id": "2305.14314", "title": "QLoRA: Efficient Finetuning of Quantized LLMs", "authors": ["Tim Dettmers", "Artidoro Pagnoni", "Ari Holtzman", "Luke Zettlemoyer"], "year": 2023, "venue": "NeurIPS 2023", "type": "preprint", "url": "https://arxiv.org/abs/2305.14314", "pdf_url": "https://arxiv.org/pdf/2305.14314", "code_url": "https://github.com/artidoro/qlora", "clusters": ["C2", "C3"], "keywords": ["4-bit quantization", "LoRA", "NF4", "paged optimizers"], "model_class": "other", "techniques": ["QLoRA", "quantization", "LoRA"], "base_model": "LLaMA 65B", "params": "", "hardware": "single 48GB GPU", "reported_metric": "65B finetuning fits on a single 48GB GPU", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Backpropagates through a frozen 4-bit NF4-quantized model into LoRA adapters, making fine-tuning of very large models possible on a single GPU.", "relevance": 4, "relevance_reason": "Foundational memory-efficient fine-tuning technique adopted for VLA adaptation.", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2309.09212", "title": "RobotPerf: An Open-Source, Vendor-Agnostic, Benchmarking Suite for Evaluating Robotics Computing System Performance", "authors": ["Víctor Mayoral-Vilches", "Jason Jabbour", "Yu-Shun Hsiao", "Zishen Wan", "Martiño Crespo-Álvarez", "Matthew Stewart", "Juan Manuel Reina-Muñoz", "Prateek Nagras", "Gaurav Vikhe", "Mohammad Bakhshalipour", "Martin Pinzger", "Stefan Rass", "Smruti Panigrahi", "Giulio Corradi", "Niladri Roy", "Phillip B. Gibbons", "Sabrina M. Neuman", "Brian Plancher", "Vijay Janapa Reddi"], "year": 2023, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2309.09212", "pdf_url": "https://arxiv.org/pdf/2309.09212", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["ROS 2", "vendor-agnostic", "grey-box", "black-box benchmarking"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "multiple CPUs/GPUs/FPGAs", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Vendor-neutral ROS 2 benchmarking suite for robotics computing pipelines, with grey-box and black-box methods to compare hardware platforms fairly.", "relevance": 4, "relevance_reason": "Closest robotics analogue to MLPerf-style cross-hardware methodology.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q3: arXiv ti:RobotPerf"}
{"id": "2307.01928", "title": "Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners", "authors": ["Allen Z. Ren", "Anushri Dixit", "Alexandra Bodrova", "Sumeet Singh", "Stephen Tu", "Noah Brown", "Peng Xu", "Leila Takayama", "Fei Xia", "Jake Varley", "Zhenjia Xu", "Dorsa Sadigh", "Andy Zeng", "Anirudha Majumdar"], "year": 2023, "venue": "CoRL 2023", "type": "paper", "url": "https://arxiv.org/abs/2307.01928", "pdf_url": "https://arxiv.org/pdf/2307.01928", "code_url": "", "clusters": ["C7"], "keywords": ["ask for help", "LLM planner"], "model_class": "other", "techniques": ["conformal prediction", "uncertainty alignment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "KnowNo: conformal prediction over LLM planner options gives statistical task-completion guarantees and asks for help when the prediction set is ambiguous.", "relevance": 4, "relevance_reason": "Seminal conformal \"know when you do not know\" trigger for robot foundation models.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: seminal fetch by arXiv id (HJ reachability, KnowNo, DeepReach)"}
{"id": "2307.15818", "title": "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control", "authors": ["Anthony Brohan", "Noah Brown", "Justice Carbajal", "Yevgen Chebotar", "Xi Chen", "Krzysztof Choromanski", "Tianli Ding", "Danny Driess", "Avinava Dubey", "Chelsea Finn", "Pete Florence", "Chuyuan Fu", "Montse Gonzalez Arenas", "Keerthana Gopalakrishnan", "Kehang Han", "Karol Hausman", "Alexander Herzog", "Jasmine Hsu", "Brian Ichter", "Alex Irpan", "Nikhil Joshi", "Ryan Julian", "Dmitry Kalashnikov", "Yuheng Kuang", "Isabel Leal", "Lisa Lee", "Tsang-Wei Edward Lee", "Sergey Levine", "Yao Lu", "Henryk Michalewski", "Igor Mordatch", "Karl Pertsch", "Kanishka Rao", "Krista Reymann", "Michael Ryoo", "Grecia Salazar", "Pannag Sanketi", "Pierre Sermanet", "Jaspiar Singh", "Anikait Singh", "Radu Soricut", "Huong Tran", "Vincent Vanhoucke", "Quan Vuong", "Ayzaan Wahid", "Stefan Welker", "Paul Wohlhart", "Jialin Wu", "Fei Xia", "Ted Xiao", "Peng Xu", "Sichun Xu", "Tianhe Yu", "Brianna Zitkovich", "Kumar Avinava Dubey", "et al."], "year": 2023, "venue": "CoRL 2023", "type": "preprint", "url": "https://arxiv.org/abs/2307.15818", "pdf_url": "https://arxiv.org/pdf/2307.15818", "code_url": "", "clusters": ["C1", "C6", "C9"], "keywords": ["VLA", "actions as text tokens", "co-fine-tuning", "RT-2", "cloud TPU serving", "action as text tokens", "cloud TPU inference"], "model_class": "VLA", "techniques": ["action tokenization", "multi-TPU cloud serving", "co-fine-tuning"], "base_model": "PaLI-X 55B / PaLM-E 12B", "params": "55B (largest), 5B", "hardware": "multi-TPU cloud service", "reported_metric": "RT-2-PaLI-X-55B at 1-3 Hz; 5B model ~5 Hz", "metric_measures": "closed-loop control", "benchmarks": ["real robot (Everyday Robots)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Defines the VLA paradigm by co-fine-tuning large VLMs to emit discrete action tokens; very large models had to be served from cloud TPUs, motivating later efficiency work.", "relevance": 4, "relevance_reason": "Seminal VLA and the origin of the latency problem the workshop addresses.", "foundational_technique": false, "verified": true, "found_via": "seed; search query; venue sweep: CoRL 2023"}
{"id": "2302.09182", "title": "Safe Networked Robotics with Probabilistic Verification", "authors": ["Sai Shankar Narasimhan", "Sharachchandra Bhat", "Sandeep P. Chinchali"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2302.09182", "pdf_url": "https://arxiv.org/pdf/2302.09182", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["networked robotics safety", "stochastic latency", "shield", "formal verification"], "model_class": "other", "techniques": ["runtime safety shield", "probabilistic model checking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Builds a formally verified runtime shield that lists safe actions for delayed observations given expected and maximum network latency, keeping cloud-controlled robots safe.", "relevance": 4, "relevance_reason": "Safety under network delay for cloud-assisted control.", "foundational_technique": false, "verified": true, "found_via": "snowball from 1902.05703"}
{"id": "2306.00148", "title": "SafeDiffuser: Safe Planning with Diffusion Probabilistic Models", "authors": ["Wei Xiao", "Tsun-Hsuan Wang", "Chuang Gan", "Daniela Rus"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2306.00148", "pdf_url": "https://arxiv.org/pdf/2306.00148", "code_url": "", "clusters": ["C7"], "keywords": ["safe diffusion", "planning"], "model_class": "diffusion/flow policy", "techniques": ["control barrier functions", "diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["maze", "locomotion", "manipulation"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Embeds finite-time CBF-based invariance into each denoising step so diffusion planners output trajectories that satisfy specifications.", "relevance": 4, "relevance_reason": "Seminal method for certifying generated trajectories (chunks) against constraints.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\")"}
{"id": "10.1109/mm.2022.3219803", "title": "Data Centers on Wheels: Emissions From Computing Onboard Autonomous Vehicles", "authors": ["Soumya Sudhakar", "Vivienne Sze", "Sertaç Karaman"], "year": 2022, "venue": "IEEE Micro", "type": "paper", "url": "https://doi.org/10.1109/mm.2022.3219803", "pdf_url": "", "code_url": "", "clusters": ["C5", "C6"], "keywords": ["fleet-scale compute", "emissions", "onboard power budget", "hardware efficiency"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "onboard AV computers (modelled)", "reported_metric": "with ~95% AV adoption, onboard compute must stay under 1.2 kW per vehicle to keep emissions below 2018 data-center levels in 90% of scenarios; hardware efficiency must double every 1.1 years", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Probabilistic fleet-level model showing onboard AV computing could emit as much as all data centers, setting a per-vehicle compute power budget.", "relevance": 4, "relevance_reason": "Closest fleet-scale onboard-compute cost/energy analysis; directly transferable framing for robot fleets.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q2: OpenAlex title_and_abstract.search: data centers on wheels"}
{"id": "2205.14135", "title": "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness", "authors": ["Tri Dao", "Daniel Y. Fu", "S. Ermon", "A. Rudra", "Christopher Ré"], "year": 2022, "venue": "NeurIPS 2022", "type": "paper", "url": "https://arxiv.org/abs/2205.14135", "pdf_url": "https://arxiv.org/pdf/2205.14135", "code_url": "https://github.com/Dao-AILab/flash-attention", "clusters": ["C5"], "keywords": ["IO-aware attention", "tiling", "HBM/SRAM"], "model_class": "other", "techniques": ["IO-aware exact attention kernel"], "base_model": "", "params": "", "hardware": "NVIDIA GPUs (A100)", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Exact attention kernel that tiles computation to minimize reads/writes between GPU HBM and on-chip SRAM, cutting memory from quadratic to linear and speeding up transformers.", "relevance": 4, "relevance_reason": "Kernel used in VLA backbones (OpenVLA, pi0, GR00T training/inference).", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2210.17323", "title": "GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers", "authors": ["Elias Frantar", "Saleh Ashkboos", "Torsten Hoefler", "Dan Alistarh", "T. Hoefler"], "year": 2022, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://www.semanticscholar.org/paper/363668677c459ebc0ff494655f993a93a0251009", "pdf_url": "https://arxiv.org/pdf/2210.17323", "code_url": "", "clusters": ["C2"], "keywords": ["second-order PTQ", "3/4-bit", "one-shot", "GPTQ", "OPTQ"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "One-shot layer-wise weight quantization using approximate second-order information, scaling 3-4 bit quantization to 175B-parameter models.", "relevance": 4, "relevance_reason": "Foundational quantization technique. No abstract available.", "foundational_technique": true, "verified": true, "found_via": "seed; snowball round 2: cited by 2402.17764, 2412.01034"}
{"id": "2212.06817", "title": "RT-1: Robotics Transformer for Real-World Control at Scale", "authors": ["Anthony Brohan", "Noah Brown", "Justice Carbajal", "Yevgen Chebotar", "Joseph Dabis", "Chelsea Finn", "Keerthana Gopalakrishnan", "Karol Hausman", "Alex Herzog", "Jasmine Hsu", "Julian Ibarz", "Brian Ichter", "Alex Irpan", "Tomas Jackson", "Sally Jesmonth", "Nikhil J Joshi", "Ryan Julian", "Dmitry Kalashnikov", "Yuheng Kuang", "Isabel Leal", "Kuang-Huei Lee", "Sergey Levine", "Yao Lu", "Utsav Malla", "Deeksha Manjunath", "Igor Mordatch", "Ofir Nachum", "Carolina Parada", "Jodilyn Peralta", "Emily Perez", "Karl Pertsch", "Jornell Quiambao", "Kanishka Rao", "Michael Ryoo", "Grecia Salazar", "Pannag Sanketi", "Kevin Sayed", "Jaspiar Singh", "Sumedh Sontakke", "Austin Stone", "Clayton Tan", "Huong Tran", "Vincent Vanhoucke", "Steve Vega", "Quan Vuong", "Fei Xia", "Ted Xiao", "Peng Xu", "Sichun Xu", "Tianhe Yu", "Brianna Zitkovich", "et al."], "year": 2022, "venue": "RSS 2023", "type": "preprint", "url": "https://arxiv.org/abs/2212.06817", "pdf_url": "https://arxiv.org/pdf/2212.06817", "code_url": "", "clusters": ["C1", "C2", "C9"], "keywords": ["Robotics Transformer", "TokenLearner", "discretized actions", "real-world scale", "efficient robot transformer", "real-time control"], "model_class": "other", "techniques": ["token compression", "action discretization", "TokenLearner token compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Google's Robotics Transformer trained on large-scale real robot data, establishing the scalable transformer-policy baseline that later VLAs build on.", "relevance": 4, "relevance_reason": "Seminal foundation; efficient tokenization for real-time robot transformers.", "foundational_technique": false, "verified": true, "found_via": "seed; venue sweep: RSS 2023"}
{"id": "2211.10438", "title": "SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models", "authors": ["Guangxuan Xiao", "Ji Lin", "Mickael Seznec", "Hao Wu", "Julien Demouth", "Song Han"], "year": 2022, "venue": "ICML 2023", "type": "paper", "url": "https://arxiv.org/abs/2211.10438", "pdf_url": "https://arxiv.org/pdf/2211.10438", "code_url": "", "clusters": ["C2"], "keywords": ["W8A8", "outlier smoothing"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Migrates activation outlier difficulty into weights through an equivalent scaling, enabling accurate W8A8 LLM inference.", "relevance": 4, "relevance_reason": "Common baseline in VLA PTQ papers (e.g., QVLA).", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2203.04955", "title": "Temporal Difference Learning for Model Predictive Control", "authors": ["Nicklas Hansen", "Xiaolong Wang", "Hao Su"], "year": 2022, "venue": "ICML 2022", "type": "paper", "url": "https://arxiv.org/abs/2203.04955", "pdf_url": "https://arxiv.org/pdf/2203.04955", "code_url": "https://nicklashansen.github.io/td-mpc", "clusters": ["C8"], "keywords": ["TD-MPC", "latent MPC", "terminal value"], "model_class": "world model", "techniques": ["task-oriented latent dynamics", "short-horizon MPPI + learned terminal value"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DMControl", "Meta-World"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Original TD-MPC: short-horizon planning in a decoder-free latent model with a learned terminal value, cutting the cost of long-horizon planning.", "relevance": 4, "relevance_reason": "Foundation of efficient latent MPC.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2310.16828"}
{"id": "2209.05487", "title": "Understanding Time Variations of DNN Inference in Autonomous Driving", "authors": ["Liangkai Liu", "Yanzhi Wang", "Weisong Shi"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2209.05487", "pdf_url": "https://arxiv.org/pdf/2209.05487", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["inference time variation", "autonomous driving"], "model_class": "other", "techniques": ["timing analysis", "profiling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Measures and dissects execution-time variation of DNN inference in autonomous driving across data, I/O, model, runtime, hardware and full-pipeline levels, deriving six root-cause insights.", "relevance": 4, "relevance_reason": "Empirical basis for why WCET of neural inference on embedded GPUs is hard to bound.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2102.11907", "title": "A predictive safety filter for learning-based racing control", "authors": ["Ben Tearle", "Kim P. Wabersich", "Andrea Carron", "Melanie N. Zeilinger"], "year": 2021, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2102.11907", "pdf_url": "https://arxiv.org/pdf/2102.11907", "code_url": "", "clusters": ["C7"], "keywords": ["racing", "safety filter"], "model_class": "other", "techniques": ["predictive safety filter", "MPC", "invariant sets"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "MPC-based predictive safety filter that certifies or minimally replaces learning-based racing control inputs using a convex-approximated invariant set, extended in real time.", "relevance": 4, "relevance_reason": "Canonical predictive safety filter for learned controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"predictive safety filter\""}
{"id": "2112.03227", "title": "CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks", "authors": ["Mees, Oier", "Hermann, Lukas", "Rosete-Beas, Erick", "Burgard, Wolfram"], "year": 2021, "venue": "IEEE RA-L 2022", "type": "benchmark", "url": "https://arxiv.org/abs/2112.03227", "pdf_url": "https://arxiv.org/pdf/2112.03227", "code_url": "https://github.com/mees/calvin", "clusters": ["C4"], "keywords": ["CALVIN", "long-horizon", "language-conditioned", "ABC-D"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Simulated Franka benchmark of 34 tasks chained into long-horizon language-conditioned sequences, with zero-shot generalization splits to new environments; commonly reported as average chain length.", "relevance": 4, "relevance_reason": "Seed benchmark widely used for VLA and efficient-policy reporting.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2112.11490", "title": "Do Androids Dream of Electric Fences? Safety-Aware Reinforcement Learning with Latent Shielding", "authors": ["Chloe He", "Borja G. Leon", "Francesco Belardinelli"], "year": 2021, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2112.11490", "pdf_url": "https://arxiv.org/pdf/2112.11490", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["latent shielding", "safety-aware RL", "imagined trajectories"], "model_class": "world model", "techniques": ["shield using model-based agent's latent imagination"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Early work using a model-based agent's latent world model to imagine future trajectories and block those violating formal safety specifications.", "relevance": 4, "relevance_reason": "Origin of imagination-based latent safety shields.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2108.11355", "title": "FogROS: An Adaptive Framework for Automating Fog Robotics Deployment", "authors": ["Kaiyuan", "Chen", "Yafei Liang", "Nikhil Jha", "Jeffrey Ichnowski", "Michael Danielczuk", "Joseph Gonzalez", "John Kubiatowicz", "Ken Goldberg"], "year": 2021, "venue": "IEEE CASE 2021", "type": "paper", "url": "https://arxiv.org/abs/2108.11355", "pdf_url": "https://arxiv.org/pdf/2108.11355", "code_url": "https://github.com/BerkeleyAutomation/FogROS2", "clusters": ["C6"], "keywords": ["fog robotics", "ROS", "cloud deployment automation"], "model_class": "other", "techniques": ["cloud offloading", "automated provisioning"], "base_model": "", "params": "", "hardware": "cloud GPU / 96-core server", "reported_metric": "compute speedups 2.6x (ORB-SLAM2), 6.0x (Dex-Net), 34.2x (motion planning) with added network latency 1.2 s / 0.6 s / 0.5 s", "metric_measures": "relative latency", "benchmarks": ["ORB-SLAM2", "Dex-Net grasp planning", "motion planning"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Original FogROS: lets developers mark ROS components to run on specified cloud hardware, trading network latency for large compute speedups.", "relevance": 4, "relevance_reason": "Seed predecessor of FogROS2.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2106.09685", "title": "LoRA: Low-Rank Adaptation of Large Language Models", "authors": ["Edward J. Hu", "Yelong Shen", "Phillip Wallis", "Zeyuan Allen-Zhu", "Yuanzhi Li", "Shean Wang", "Lu Wang", "Weizhu Chen"], "year": 2021, "venue": "ICLR 2022", "type": "preprint", "url": "https://arxiv.org/abs/2106.09685", "pdf_url": "https://arxiv.org/pdf/2106.09685", "code_url": "https://github.com/microsoft/LoRA", "clusters": ["C3"], "keywords": ["low-rank adaptation", "PEFT"], "model_class": "other", "techniques": ["LoRA"], "base_model": "GPT-3 / RoBERTa", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Freezes pretrained weights and trains low-rank update matrices, cutting trainable parameters and optimizer memory by orders of magnitude with no added inference latency after merging.", "relevance": 4, "relevance_reason": "Foundational PEFT method used by nearly every VLA fine-tuning pipeline.", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2102.12981", "title": "The Black-Box Simplex Architecture for Runtime Assurance of Autonomous CPS", "authors": ["Usama Mehmood", "Sanaz Sheikhi", "Stanley Bak", "Scott A. Smolka", "Scott D. Stoller"], "year": 2021, "venue": "NASA Formal Methods (2022) 231-250", "type": "paper", "url": "https://arxiv.org/abs/2102.12981", "pdf_url": "https://arxiv.org/pdf/2102.12981", "code_url": "", "clusters": ["C7"], "keywords": ["black-box simplex"], "model_class": "other", "techniques": ["simplex architecture", "runtime assurance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows runtime checks can replace static verification of the Simplex baseline controller, with case studies using MPC and neural controllers for F-16 collision avoidance.", "relevance": 4, "relevance_reason": "Seminal black-box runtime assurance applicable to learned controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2004.06089", "title": "Thinking While Moving: Deep Reinforcement Learning with Concurrent Control", "authors": ["Ted Xiao", "Eric Jang", "Dmitry Kalashnikov", "S. Levine", "Julian Ibarz", "Karol Hausman", "Alexander Herzog"], "year": 2020, "venue": "ICLR 2020", "type": "paper", "url": "https://arxiv.org/abs/2004.06089", "pdf_url": "https://arxiv.org/pdf/2004.06089", "code_url": "", "clusters": ["C7"], "keywords": ["concurrent control", "thinking while moving", "delay-aware Bellman"], "model_class": "other", "techniques": ["continuous-time delay-aware Q-learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Formulates RL where the next action is chosen while the previous one is still executing, deriving delay-aware discretized Bellman updates and applying them to large-scale robotic grasping.", "relevance": 4, "relevance_reason": "Foundational concurrent (asynchronous) control; cited by RTC.", "foundational_technique": true, "verified": true, "found_via": "snowball: references of 2506.07339 (S2)"}
{"id": "1903.09589", "title": "A Fog Robotics Approach to Deep Robot Learning: Application to Object Recognition and Grasp Planning in Surface Decluttering", "authors": ["Ajay Kumar Tanwani", "Nitesh Mor", "John Kubiatowicz", "Joseph E. Gonzalez", "Ken Goldberg"], "year": 2019, "venue": "ICRA 2019", "type": "paper", "url": "https://arxiv.org/abs/1903.09589", "pdf_url": "https://arxiv.org/pdf/1903.09589", "code_url": "", "clusters": ["C6"], "keywords": ["fog robotics", "cloud training, edge adaptation", "grasp planning"], "model_class": "other", "techniques": ["cloud-edge federated training", "edge inference serving"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains deep models on synthetic data in the cloud, adapts them at a trusted edge on private real images, and serves low-latency inference to robots on the local network.", "relevance": 4, "relevance_reason": "Seminal fog-robotics deep-learning paper.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "1903.11199", "title": "Control Barrier Functions: Theory and Applications", "authors": ["Aaron D. Ames", "Samuel Coogan", "Magnus Egerstedt", "Gennaro Notomista", "Koushil Sreenath", "Paulo Tabuada"], "year": 2019, "venue": "European Control Conference (ECC) 2019", "type": "paper", "url": "https://arxiv.org/abs/1903.11199", "pdf_url": "https://arxiv.org/pdf/1903.11199", "code_url": "", "clusters": ["C7"], "keywords": ["control barrier functions", "safety-critical control", "CBF-QP"], "model_class": "other", "techniques": ["control barrier functions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Tutorial overview of control barrier functions for verifying and enforcing safety in optimization-based controllers, with robotics applications.", "relevance": 4, "relevance_reason": "Seminal CBF foundation used by VLA safety filters.", "foundational_technique": true, "verified": true, "found_via": "seed; WebSearch"}
{"id": "1902.05703", "title": "Network Offloading Policies for Cloud Robotics: a Learning-based Approach", "authors": ["Sandeep Chinchali", "Apoorva Sharma", "James Harrison", "Amine Elhafsi", "Daniel Kang", "Evgenya Pergament", "Eyal Cidon", "Sachin Katti", "Marco Pavone"], "year": 2019, "venue": "RSS 2019", "type": "paper", "url": "https://arxiv.org/abs/1902.05703", "pdf_url": "https://arxiv.org/pdf/1902.05703", "code_url": "", "clusters": ["C6"], "keywords": ["offloading policy", "deep RL", "cloud robotics", "bandwidth vs accuracy"], "model_class": "other", "techniques": ["learned offloading policy (RL)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Formulates when a robot should offload perception queries to a cloud model as an MDP and learns an RL policy trading accuracy against network and compute cost.", "relevance": 4, "relevance_reason": "Seminal learned offloading policy for cloud robotics.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "1811.01828", "title": "Verisig: verifying safety properties of hybrid systems with neural network controllers", "authors": ["Radoslav Ivanov", "James Weimer", "Rajeev Alur", "George J. Pappas", "Insup Lee"], "year": 2018, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1811.01828", "pdf_url": "https://arxiv.org/pdf/1811.01828", "code_url": "", "clusters": ["C7"], "keywords": ["Verisig"], "model_class": "other", "techniques": ["hybrid system reachability", "NN verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Verifies closed-loop safety of sigmoid neural network controllers by converting the network into an equivalent hybrid system and using hybrid-system reachability tools.", "relevance": 4, "relevance_reason": "Seminal closed-loop NN controller verification.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "kehoe-2015-cloud-robotics-survey", "title": "A Survey of Research on Cloud Robotics and Automation", "authors": ["Ben Kehoe", "Sachin Patil", "Pieter Abbeel", "Ken Goldberg"], "year": 2015, "venue": "IEEE Transactions on Automation Science and Engineering 12(2)", "type": "paper", "url": "https://rll.berkeley.edu/~sachin/papers/Kehoe-TASE2015a.pdf", "pdf_url": "https://rll.berkeley.edu/~sachin/papers/Kehoe-TASE2015a.pdf", "code_url": "", "clusters": ["C6"], "keywords": ["cloud robotics", "survey", "big data", "collective learning"], "model_class": "other", "techniques": ["cloud computing", "collective robot learning", "crowdsourcing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Foundational survey framing cloud robotics around big data, cloud compute, collective robot learning and human computation, with open challenges including latency and connectivity.", "relevance": 4, "relevance_reason": "Heavily cited seminal reference for the cloud robotics paradigm.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.16885", "title": "$\\tau_0$-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation", "authors": ["Xiaowei Cai", "Yunuo Cai", "Bing Chen", "Jingxi Chen", "Zhi Chen", "Si-Yuan Feng", "Teng Hou", "Jingshun Huang", "Han Jiang", "Runkun Ju", "Dong Li", "Mingxi Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.16885", "pdf_url": "https://arxiv.org/pdf/2608.16885", "code_url": "", "clusters": ["C1", "C7", "C9"], "keywords": ["hierarchical VLA", "world-model-guided test-time compute"], "model_class": "VLA", "techniques": ["compute-scalable subtask search with world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "tau0-VLA allocates extra test-time compute to hard high-level subtask decisions using world-model-guided search.", "relevance": 3, "relevance_reason": "Adaptive test-time compute in hierarchical VLA.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2607.23783", "title": "$N_0$-TWAM: Scaling Tactile-Native World-Action Model for Contact-Rich Manipulation", "authors": ["NeoteAI Team", "Fudan TEAI Team"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.23783", "pdf_url": "https://arxiv.org/pdf/2607.23783", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["tactile WAM", "tactile world-action model", "real-time efficiency"], "model_class": "world-action model", "techniques": ["asymmetric mixture of transformers", "asymmetric mixture-of-transformers"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Large-scale tactile-native world-action model predicting future vision and contact; uses an asymmetric Mixture-of-Transformers with slim action/tactile experts for real-time efficiency.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2402.15391, 2405.12213"}
{"id": "2603.02083", "title": "$π$-StepNFT: Wider Space Needs Finer Steps in Online RL for Flow-based VLAs", "authors": ["Siting Wang", "Xiaofeng Wang", "Zheng Zhu", "Minnan Pei", "Xinyu Cui", "Cheng Deng", "Jian Zhao", "Guan Huang", "Haifeng Zhang", "Jun Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.02083", "pdf_url": "https://arxiv.org/pdf/2603.02083", "code_url": "", "clusters": ["C3"], "keywords": ["critic-free", "likelihood-free", "step-wise fine-tuning"], "model_class": "VLA", "techniques": ["online RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Critic- and likelihood-free online RL for flow VLAs requiring one forward pass per optimization step.", "relevance": 3, "relevance_reason": "Cheaper RL fine-tuning per step.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "10.1109/iccar69571.2026.11549730", "title": "A Hard Real-Time Bilateral Teleoperation System for High-Fidelity Robot Learning Data Collection", "authors": ["Saurav Khadka", "Tomi Pitkäaho"], "year": 2026, "venue": "2026 12th International Conference on Control, Automation and Robotics (ICCAR)", "type": "paper", "url": "https://doi.org/10.1109/ICCAR69571.2026.11549730", "pdf_url": "", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["hard real-time teleoperation", "data collection"], "model_class": "other", "techniques": ["PREEMPT_RT scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "inter-packet jitter 0.046 ms at 500 Hz", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Hard real-time bilateral teleoperation stack with 1 kHz haptics and 500 Hz control for high-fidelity demonstration collection.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.03103", "title": "A hierarchical approach to imitation learning for manipulation tasks requiring time varying forces", "authors": ["Rishabh Shukla", "Adithya Santhosh", "Shaili Gandhi", "Samrudh Moode", "Satyandra K. Gupta"], "year": 2026, "venue": "Robotics Comput. Integr. Manuf.", "type": "paper", "url": "https://arxiv.org/abs/2608.03103", "pdf_url": "https://arxiv.org/pdf/2608.03103", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["high-frequency force control", "diffusion latency"], "model_class": "diffusion/flow policy", "techniques": ["hierarchical control", "neural impedance controller"], "base_model": "", "params": "", "hardware": "", "reported_metric": "5 Hz high level / 60 Hz low level", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Decouples a 5 Hz diffusion planner from a 60 Hz lightweight force-conditioned controller to handle contact transients despite diffusion latency.", "relevance": 3, "relevance_reason": "Addresses diffusion-policy latency for high-rate contact control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2602.14255", "title": "A Latency-Aware Framework for Visuomotor Policy Learning on Industrial Robots", "authors": ["Daniel Ruan", "Salma Mozaffari", "Sigrid Adriaenssens", "Arash Adel"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.14255", "pdf_url": "https://arxiv.org/pdf/2602.14255", "code_url": "", "clusters": ["C1"], "keywords": ["latency-aware execution", "industrial robot", "timestamps"], "model_class": "other", "techniques": ["asynchronous execution", "latency calibration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Timestamps predicted action sequences and executes only temporally feasible actions to handle observation/inference/execution latency on industrial arms.", "relevance": 3, "relevance_reason": "Latency-aware deployment.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot"}
{"id": "2601.18692", "title": "A Pragmatic VLA Foundation Model", "authors": ["Wei Wu", "Fan Lu", "Yunnan Wang", "Shuai Yang", "Shi Liu", "Fangjing Wang", "Qian Zhu", "He Sun", "Yong Wang", "Shuailei Ma", "Yiyu Ren", "Kejia Zhang", "Hui Yu", "Jingmei Zhao", "Shuai Zhou", "Zhenqi Qiu", "Houlong Xiong", "Ziyu Wang", "Zechen Wang", "Ran Cheng", "Yong-Lu Li", "Yongtao Huang", "Xing Zhu", "Yujun Shen", "Kecheng Zheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.18692", "pdf_url": "https://arxiv.org/pdf/2601.18692", "code_url": "", "clusters": ["C3"], "keywords": ["LingBot-VLA", "post-training cost", "efficient codebase"], "model_class": "VLA", "techniques": ["efficient training codebase"], "base_model": "", "params": "", "hardware": "8-GPU training setup", "reported_metric": "261 samples/s training throughput; 1.5-2.8x over existing VLA codebases", "metric_measures": "throughput", "benchmarks": ["4 real robot platforms, 100 tasks each"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Dual-arm VLA pretrained on ~20k hours emphasizing cheap post-training (130 episodes/task) and a faster training codebase.", "relevance": 3, "relevance_reason": "Reports adaptation data cost and training throughput.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: knowledge insulation co-training VLA training efficiency pi0.5 arXiv"}
{"id": "2606.21216", "title": "A scalar per patch from pre-trained ViTs enables fast moving navigation in the real world", "authors": ["Steeven Janny", "Leonid Antsfeld", "Christian Wolf"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.21216", "pdf_url": "https://arxiv.org/pdf/2606.21216", "code_url": "", "clusters": ["C2"], "keywords": ["visual encoders", "fast navigation"], "model_class": "other", "techniques": ["multi-teacher distillation", "feature bottleneck"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Large real-world navigation study of visual encoders showing distilled, heavily bottlenecked (one scalar per patch) features suffice for fast navigation.", "relevance": 3, "relevance_reason": "Compact visual representations for fast embodied control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "10.1109/iementech202669403.2026.11434122", "title": "A Survey on On-Device Small Language Models: Architectures, Compression, Applications, and Open Challenges", "authors": ["Ayan Patra", "Sayan Kundu", "Abhirup Dey", "T. Saha", "Samit Karmakar", "S. Bishnu"], "year": 2026, "venue": "2026 9th International Conference on Electronics, Materials Engineering & Nano-Technology (IEMENTech)", "type": "paper", "url": "https://doi.org/10.1109/IEMENTech202669403.2026.11434122", "pdf_url": "", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["on-device SLM", "survey"], "model_class": "other", "techniques": ["quantization", "pruning", "distillation", "PEFT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of on-device small language models covering compact design, compression, parameter-efficient adaptation and inference optimizations.", "relevance": 3, "relevance_reason": "Generic on-device efficiency reference.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2205.14135"}
{"id": "2607.19191", "title": "ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU", "authors": ["Fan Jiang", "Zhaoxu Sun", "Mengchao Wang", "Ziyu Zhu", "Chiyu Wang", "Yunpeng Zhang", "Wenlin Liu", "Yun Wang", "Xue Zheng", "Rui Sun", "Junfeng Ni", "Hongyu Pan", "Zhongxu Sun", "Fei Yu", "Zengye Ge", "Mengmeng Du", "Nianfei Fan", "Mingchao Sun", "Yu Liu", "Yongchang", "Yanqing Zhu", "Jiahang Wang", "Ning Ying", "Yuze Xuan", "Di Yang", "Zhicheng Liu", "Zhe Gao", "Tingbing Xu", "Jiacheng Sui", "Wenjin Yang", "Junnan Lai", "Shufeng Liu", "Yuan Liu", "Zheng Zhou", "Yingliang Peng", "Dawei Cao", "Kaifeng Sheng", "Yuxiang Cai", "Fei Lu", "Mu Xu", "Ning Guo"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2607.19191", "pdf_url": "https://arxiv.org/pdf/2607.19191", "code_url": "", "clusters": ["C8"], "keywords": ["single desktop GPU", "interactive world rollout", "LongForcing"], "model_class": "world model", "techniques": ["teacher forcing + ODE distillation to causal student", "long-horizon self-rollout alignment"], "base_model": "", "params": "", "hardware": "single desktop GPU", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Action-conditioned interactive video world model distilled into a causal student that runs real-time, long-horizon closed-loop rollouts on a single desktop GPU.", "relevance": 3, "relevance_reason": "Cheap real-time interactive world model.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.23901", "title": "ABPolicy: Asynchronous B-Spline Flow Policy for Real-Time and Smooth Robotic Manipulation", "authors": ["Fan Yang", "Peiguang Jing", "Kaihua Qu", "Ningyuan Zhao", "Yuting Su", "Ning-Yuan Zhao", "Yu-Ting Su"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.23901", "pdf_url": "https://arxiv.org/pdf/2602.23901", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["B-spline actions", "asynchronous flow policy", "smoothness", "ABPolicy", "B-spline action space"], "model_class": "diffusion/flow policy", "techniques": ["asynchronous inference", "spline action parameterization", "B-spline control points", "bidirectional prediction with refitting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Asynchronous flow policy in B-spline control-point space for intra-chunk smoothness with bidirectional prediction and refitting for inter-chunk continuity.", "relevance": 3, "relevance_reason": "Smooth real-time async execution.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; snowball: citations of 2506.07339 (S2)"}
{"id": "2602.02864", "title": "Accelerating Structured Chain-of-Thought in Autonomous Vehicles", "authors": ["Yi Gu", "Yan Wang", "Yuxiao Chen", "Yurong You", "Wenjie Luo", "Yue Wang", "Wenhao Ding", "Boyi Li", "Heng Yang", "Boris Ivanovic", "Marco Pavone"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.02864", "pdf_url": "https://arxiv.org/pdf/2602.02864", "code_url": "", "clusters": ["C1"], "keywords": ["FastDriveCoT"], "model_class": "VLA", "techniques": ["parallel decoding", "structured CoT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "3-4x faster CoT generation", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Decomposes template-structured driving CoT into a dependency graph and generates independent steps in parallel in one forward pass.", "relevance": 3, "relevance_reason": "Parallel decoding of reasoning to cut latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]"}
{"id": "2608.29937", "title": "AcrossWAM1.0:A Modular Latent World-Action Stack for Compact Robot Policies", "authors": ["Yafei Zhang", "Nan Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29937", "pdf_url": "https://arxiv.org/pdf/2608.29937", "code_url": "", "clusters": ["C8"], "keywords": ["latent WAM stack", "deployment export", "compact policies"], "model_class": "world-action model", "techniques": ["modular latent subgoal WAM", "training-only teachers separated from inference graph"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Modularizes the LaWAM latent world-action stack with explicit boundaries and a deployment export that excludes training-only teachers, for compact robot policies.", "relevance": 3, "relevance_reason": "Engineering of compact latent WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2603.16666 (citations)"}
{"id": "2603.05147", "title": "Act, Think or Abstain: Complexity-Aware Adaptive Inference for Vision-Language-Action Models", "authors": ["Riccardo Andrea Izzo", "Gianluca Bardaro", "Matteo Matteucci"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.05147", "pdf_url": "https://arxiv.org/pdf/2603.05147", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["adaptive inference", "abstain", "complexity-aware routing", "OOD"], "model_class": "VLA", "techniques": ["latent-embedding complexity detection", "routing act/think/abstain"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Routes VLA execution to act directly, invoke reasoning, or abstain based on detected state complexity/OOD-ness from backbone embeddings.", "relevance": 3, "relevance_reason": "Abstention fallback plus compute-aware routing.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2609.36250", "title": "Action Chunking Proximal Policy Optimization with Feedback Correction", "authors": ["Sanghyun Hahn", "Jonghyun Choi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36250", "pdf_url": "https://arxiv.org/pdf/2609.36250", "code_url": "https://github.com/hshhahn/ACPPO", "clusters": ["C3", "C7"], "keywords": ["within-chunk feedback", "reinforcement learning"], "model_class": "other", "techniques": ["action chunking", "PPO", "feedback correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["IsaacGym", "Bi-DexHands"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "PPO variant with a chunked actor and a per-step state-value critic, plus a stepwise feedback corrector that edits planned actions inside each chunk to restore within-chunk reactivity.", "relevance": 3, "relevance_reason": "Adds within-chunk feedback correction to chunked RL policies.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*)"}
{"id": "2603.18091", "title": "Action Draft and Verify: A Self-Verifying Framework for Vision-Language-Action Model", "authors": ["Chen Zhao", "Zhuoran Wang", "Haoyang Li", "Shifeng Bao", "Guanlin Li", "Youhe Feng", "Yang Li", "Jie Tang", "Jing Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.18091", "pdf_url": "https://arxiv.org/pdf/2603.18091", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["chunk selection", "VLA"], "model_class": "VLA", "techniques": ["draft-and-verify", "reranking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+4.3 pts sim, +19.7 pts real success", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Diffusion expert drafts several candidate action chunks and the VLM reranks them in one forward pass with a perplexity-style score.", "relevance": 3, "relevance_reason": "Cheap chunk-level verification before execution.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunk*\" AND (uncertaint* OR \"out-of-distribution\")"}
{"id": "2607.10706", "title": "Action Map Policy: Learning 3D Closed-loop Manipulation via Pixel Classification", "authors": ["Hao-zhe Huang", "Zhang-Chen Ye", "Lin-Feng Zhao", "Bo-Ce Hu", "Ming Jia", "Yu Qi", "Ahmed Agha", "Dian Wang", "Robert Platt", "Robin Walters"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.10706", "pdf_url": "https://arxiv.org/pdf/2607.10706", "code_url": "", "clusters": ["C1"], "keywords": ["Action Map Policy"], "model_class": "other", "techniques": ["single-pass classification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Casts 3D action prediction as pixel classification, producing full action chunks in one forward pass and inferring much faster than diffusion policies.", "relevance": 3, "relevance_reason": "Fast-inference policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2602.15397", "title": "ActionCodec: What Makes for Good Action Tokenizers", "authors": ["Zibin Dong", "Yicheng Liu", "Shiduo Zhang", "Baijun Ye", "Yifu Yuan", "Fei Ni", "Jingjing Gong", "Xipeng Qiu", "Hang Zhao", "Yinchuan Li", "Jianye Hao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.15397", "pdf_url": "https://arxiv.org/pdf/2602.15397", "code_url": "", "clusters": ["C1"], "keywords": ["action tokenizer design", "training efficiency"], "model_class": "VLA", "techniques": ["action tokenization"], "base_model": "SmolVLM2-2.2B", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Derives information-theoretic design principles for action tokenizers and builds ActionCodec, improving VLA training efficiency without robot pretraining.", "relevance": 3, "relevance_reason": "Action tokenization.", "foundational_technique": false, "verified": true, "found_via": "search query: action tokenization"}
{"id": "2609.33256", "title": "ActionGround: Training-Free Runtime Refinement of Frozen VLA Policies", "authors": ["Namai Chandra", "Madhur Thareja", "Shriram Damodaran", "Addison Lin Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33256", "pdf_url": "https://arxiv.org/pdf/2609.33256", "code_url": "", "clusters": ["C7"], "keywords": ["frozen VLA"], "model_class": "VLA", "techniques": ["finite-state machine", "dynamics prior", "runtime refinement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "<1 ms overhead per control step", "metric_measures": "other", "benchmarks": ["LIBERO-Spatial", "Robosuite"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Neuro-symbolic wrapper around frozen VLAs: a phase state machine applies rule-based corrections and an Euler-Lagrange term injects dynamics, with under 1 ms overhead per step.", "relevance": 3, "relevance_reason": "Very cheap runtime correction layer.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "10.1109/ipac69021.2026.11542954", "title": "Active Jitter Suppression in VLA-Driven Robotics via Cascaded Tracking Differentiators", "authors": ["Hao-Wen Tang", "Weilin Chen", "Wenbo Zhu", "Zi-Han Chen", "Jin-Li Pang"], "year": 2026, "venue": "2026 International Conference on Intelligent Perception and Autonomous Control (IPAC)", "type": "paper", "url": "https://doi.org/10.1109/IPAC69021.2026.11542954", "pdf_url": "", "code_url": "", "clusters": ["C7"], "keywords": ["jitter suppression", "VLA output smoothing"], "model_class": "VLA", "techniques": ["tracking differentiators"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Cascaded tracking differentiators filter high-frequency jitter in VLA outputs to match electromechanical hardware in assembly.", "relevance": 3, "relevance_reason": "Execution-level stability of VLA commands", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.11697", "title": "ActSafeGuard: Differentiable and Training-Aligned Constraint Enforcement for Flow-Matching Policies", "authors": ["Jianming Ma", "Rongjun Jin", "Xiaxi Si", "Yang Zhang", "Yiheng Li", "Yue Gao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.11697", "pdf_url": "https://arxiv.org/pdf/2609.11697", "code_url": "", "clusters": ["C7"], "keywords": ["ActSafeGuard", "hard constraints", "differentiable safeguard", "flow matching"], "model_class": "VLA", "techniques": ["training-aligned differentiable constraint layer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Differentiable safeguard layer integrated into flow-matching VLA/WAM training and inference to enforce hard per-step action feasibility constraints.", "relevance": 3, "relevance_reason": "Hard-constraint enforcement for generative policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: safe robot foundation models"}
{"id": "2606.32026", "title": "AdaJEPA: An Adaptive Latent World Model", "authors": ["Ying Wang", "Oumayma Bounou", "Yann LeCun", "Meng-Ye Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.32026", "pdf_url": "https://arxiv.org/pdf/2606.32026", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["test-time adaptation", "latent world model"], "model_class": "world model", "techniques": ["test-time adaptation", "JEPA", "MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "AdaJEPA adapts a latent world model online inside the MPC loop using observed transitions, improving planning under distribution shift with as little as one gradient step per replan.", "relevance": 3, "relevance_reason": "relevant: lightweight test-time adaptation of world models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2310.16828"}
{"id": "2606.31958", "title": "Adapting Generalist Robot Policies with Semantic Reinforcement Learning", "authors": ["Jagdeep Singh Bhatia", "Andrew Wagenmaker", "William Chen", "Sergey Levine"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.31958", "pdf_url": "https://arxiv.org/pdf/2606.31958", "code_url": "", "clusters": ["C3"], "keywords": ["efficient online adaptation", "semantic actions"], "model_class": "VLA", "techniques": ["RL in prompt space"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "SARL adapts generalist VLAs by running RL over language prompts rather than raw actions, yielding sample-efficient online improvement on long-horizon tasks.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2609.34944", "title": "Adjoint Guidance Flow: Amortized Critic Guidance for VLA Policies", "authors": ["Jeongsol Kim", "Youngjun Jun", "Kyumin Choi", "Youngmin Kim", "Seonghyun Jin", "Sunwoo Park", "Jangho Park", "Kwanyoung Kim", "Jong Chul Ye", "Young Min Kim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34944", "pdf_url": "https://arxiv.org/pdf/2609.34944", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["flow VLA", "critic guidance"], "model_class": "VLA", "techniques": ["critic guidance", "amortization", "optimal control", "amortized guidance network", "adjoint method"], "base_model": "", "params": "", "hardware": "", "reported_metric": "3.6x faster per guidance step, 7.0x fewer parameters vs QGF", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "RoboCasa", "LIBERO-Pro"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Amortizes trajectory-aware critic guidance for flow VLAs into a small guidance network regressed onto an optimal-control costate; one extra forward pass per step.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]; snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2604.00061", "title": "Advancing Multi-Robot Networks via MLLM-Driven Sensing, Communication, and Computation: A Comprehensive Survey", "authors": ["Hyun Jong Yang", "Howon Lee", "Kyuhong Shim", "Jeongho Kwak", "Hyunsoo Kim", "Donghoon Kim", "Khoa Anh Ngo", "Sehyun Ryu", "Jaehyun Choi", "Youbin Kim", "Chanjun Moon", "Michael Ryoo", "Byonghyo Shim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.00061", "pdf_url": "https://arxiv.org/pdf/2604.00061", "code_url": "", "clusters": ["C6"], "keywords": ["multi-robot networks", "MLLM", "communication-computation", "survey"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey on how multimodal LLMs interplay with sensing, communication and computation in multi-robot networks, including cloud-based orchestrators and bandwidth limits.", "relevance": 3, "relevance_reason": "Survey covering fleet-level communication/compute trade-offs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2601.20540", "title": "Advancing Open-source World Models", "authors": ["Ze-Lin Gao", "Qiuyu Wang", "Yanhong Zeng", "Jia-Peng Zhu", "K. Cheng", "Yixuan Li", "Hanlin Wang", "Yinghao Xu", "Shuailei Ma", "Yihan Chen", "Jie Liu", "Yansong Cheng", "Yao Yao", "Jia-Yi Zhu", "Yi-Hao Meng", "Ke-Cheng Zheng", "Qingyan Bai", "Jing-Ye Chen", "Zehong Shen", "Yue Yu", "Xing Zhu", "Yujun Shen", "Hao Ouyang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.20540", "pdf_url": "https://arxiv.org/pdf/2601.20540", "code_url": "", "clusters": ["C8"], "keywords": ["open-source world model", "real-time interactive", "long-term memory"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "<1 s latency while generating 16 fps", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open-source video-based world simulator with minute-level horizons and real-time interactive generation.", "relevance": 3, "relevance_reason": "Real-time interactive world model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2402.15391"}
{"id": "2603.22899", "title": "Agile-VLA: Few-Shot Industrial Pose Rectification via Implicit Affordance Anchoring", "authors": ["Teng Yan", "Zhengyang Pei", "Chengyu Shi", "Yue Yu", "Yikun Chen", "Zilong Zhu", "Zelin Fang", "Kaile Guo", "Zihang Wang", "Peigen Tian", "Bingzhuo Zhong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.22899", "pdf_url": "https://arxiv.org/pdf/2603.22899", "code_url": "", "clusters": ["C1", "C6", "C9"], "keywords": ["hierarchical edge VLA", "industrial"], "model_class": "VLA", "techniques": ["hierarchical slow-fast"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Agile-VLA uses a hierarchy to reconcile slow semantic inference with high-frequency control on edge platforms for industrial pose rectification.", "relevance": 3, "relevance_reason": "Edge dual-rate VLA.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: arXiv API VLA AND Jetson"}
{"id": "2609.34467", "title": "Alignment-Guided Flow Transformer for Efficient Vision-Language-Action Policy Learning", "authors": ["Shengchao Hu", "Peng Wang", "Qiyang Zhou", "Guodong Zheng", "Yuqi Huang", "Li Shen", "Ya Zhang", "Dacheng Tao", "Sheng-Chao Hu", "Qi-Yang Zhou", "Guo-Dong Zheng", "Yu-Qi Huang", "Da-Cheng Tao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34467", "pdf_url": "https://arxiv.org/pdf/2609.34467", "code_url": "", "clusters": ["C1"], "keywords": ["efficient VLA", "tri-modal alignment", "fewer inference steps"], "model_class": "VLA", "techniques": ["alignment loss", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a tri-modal vision-language-action alignment loss and uses flow matching for fewer inference steps, reporting lower latency than baselines.", "relevance": 3, "relevance_reason": "Reports lower inference latency vs baselines", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2609.22973", "title": "An Empirical Study and Open Testbed for Federated Fine-Tuning of Vision-Language-Action Models", "authors": ["Zhe-Kai Duan", "Kevin Xie", "Xin-Yu Tan", "Shi-Kai Geng", "Cheng-Xu Zhou", "R. Kompella", "Gao-Wen Liu", "C. X. Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22973", "pdf_url": "https://arxiv.org/pdf/2609.22973", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["federated learning", "fine-tuning"], "model_class": "VLA", "techniques": ["federated fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Systematic study and open testbed for federated fine-tuning of pretrained VLAs across sites, analyzing parameter scopes and aggregation choices in sim and on real robots.", "relevance": 3, "relevance_reason": "relevant: distributed adaptation of VLAs", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2306.03310"}
{"id": "2604.01567", "title": "AnchorVLA: Anchored Diffusion for Efficient End-to-End Mobile Manipulation", "authors": ["Jia Syuen Lim", "Zhizhen Zhang", "Peter Bohm", "Brendan Tidd", "Zi Huang", "Yadan Luo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.01567", "pdf_url": "https://arxiv.org/pdf/2604.01567", "code_url": "", "clusters": ["C1"], "keywords": ["anchored diffusion", "truncated schedule", "mobile manipulation"], "model_class": "VLA", "techniques": ["anchor initialization", "truncated diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Denoises locally around anchor trajectories with a truncated schedule to keep multimodality while cutting inference cost for mobile manipulation.", "relevance": 3, "relevance_reason": "Few-step diffusion VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: efficient diffusion policy action chunk"}
{"id": "2605.23733", "title": "Any2Any: Efficient Cross-Embodiment Transfer for Humanoid Whole-Body Tracking", "authors": ["Ming Yang", "Tao Yu", "Feng Li", "Hua Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.23733", "pdf_url": "https://arxiv.org/pdf/2605.23733", "code_url": "", "clusters": ["C3"], "keywords": ["humanoid whole-body tracking", "cross-embodiment", "PEFT"], "model_class": "other", "techniques": ["kinematic alignment", "PEFT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Transfers a humanoid whole-body tracking policy to a new humanoid with kinematic alignment plus lightweight PEFT.", "relevance": 3, "relevance_reason": "Cheap cross-embodiment transfer.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"robot\" AND abs:\"parameter-efficient fine-tuning\""}
{"id": "2609.32550", "title": "Are Vision-Language-Action Models Robust to One-Step Observation Perturbations?", "authors": ["Shojiro Yamabe", "J. Sakuma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.32550", "pdf_url": "https://arxiv.org/pdf/2609.32550", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["one-step observation perturbation", "CARE", "execution length", "robustness"], "model_class": "VLA", "techniques": ["consistency-based execution length selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows momentary single-step observation corruption substantially degrades VLAs with effect depending on chunk execution length, and proposes CARE to pick execution length dynamically from consistency.", "relevance": 3, "relevance_reason": "Transient input faults and open-loop horizon.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2607.24481", "title": "ArmnetBench v0.1: Parallel Real-World Evaluation of Manipulation Policies on a Low-Cost Arm Farm", "authors": ["Selvaraj, Praveen", "Uttini, Lorenzo", "Kuosmanen, Ville"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2607.24481", "pdf_url": "https://arxiv.org/pdf/2607.24481", "code_url": "", "clusters": ["C4"], "keywords": ["ArmnetBench", "arm farm", "SO-101", "parallel real-world evaluation", "LeRobot"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "SO-101 arms", "reported_metric": "", "metric_measures": "", "benchmarks": ["ArmnetBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Fleet of low-cost SO-101 cells evaluating 7 policies on 12 tasks (2,518 rollouts) with three-way quality labels, released in LeRobot format.", "relevance": 3, "relevance_reason": "Scalable low-cost real-robot evaluation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RoboArena distributed real-world evaluation"}
{"id": "2609.20034", "title": "Astronex-World 1.0: Real-Time Interactive World Model Foundation", "authors": ["Xin Zhou", "Cong-Wen Miao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20034", "pdf_url": "https://arxiv.org/pdf/2609.20034", "code_url": "", "clusters": ["C8"], "keywords": ["real-time world model"], "model_class": "world model", "techniques": ["few-step distillation", "DMD", "KV caching"], "base_model": "", "params": "5B", "hardware": "NVIDIA L20 48GB", "reported_metric": "832x480 at 24 fps", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Astronex-World is an open action/camera-controllable video world model distilled into a causal few-step model that streams in real time on one GPU.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2501.03575"}
{"id": "2605.07308", "title": "AT-VLA: Adaptive Tactile Injection for Enhanced Feedback Reaction in Vision-Language-Action Models", "authors": ["Xiaoqi Li", "Mu Cai", "Jiadong Xu", "Juan Zhu", "Hongwei Fan", "Yan-Kun Shen", "Guangrui Ren", "Hao Dong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.07308", "pdf_url": "https://arxiv.org/pdf/2605.07308", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["tactile VLA", "real-time response"], "model_class": "VLA", "techniques": ["dual-stream fast/slow", "adaptive tactile injection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "closed-loop tactile response within 0.04 s", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adds a fast tactile control stream beside the slow VLM stream so a VLA reacts to touch in real time.", "relevance": 3, "relevance_reason": "Dual-rate design addressing VLA inference latency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2608.07065", "title": "AutoIntervene: Calibrated Intervention for Action-Chunking Imitation Learning Policies", "authors": ["Jin-He Tang", "Wei-Ming Zhi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.07065", "pdf_url": "https://arxiv.org/pdf/2608.07065", "code_url": "", "clusters": ["C7"], "keywords": ["AutoIntervene", "calibrated intervention", "shared control", "action chunking"], "model_class": "other", "techniques": ["support-memory scoring", "quantile-calibrated switching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Switches control between a chunking policy and a human operator using calibrated visual-action support scores, with separate thresholds for handing over and returning control.", "relevance": 3, "relevance_reason": "Monitor-triggered fallback to human.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2503.08558 (S2)"}
{"id": "2607.01051", "title": "AutoSpeed: Annotation-Free Stage-Adaptive Motion Speed Learning for Robot Manipulation", "authors": ["Qingda Hu", "Ziheng Qiu", "Jie-Ru Zhao", "Zhongxue Gan", "Wen-Chao Ding"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.01051", "pdf_url": "https://arxiv.org/pdf/2607.01051", "code_url": "", "clusters": ["C1"], "keywords": ["stage-adaptive speed", "task throughput"], "model_class": "other", "techniques": ["speed modulation (DCT)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "AutoSpeed learns stage-dependent execution speeds for visuomotor policies without annotations, running easy stages faster to cut task time.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.27753", "title": "AWM-VLA: AlignedWorld Modeling for Efficient and Explainable Vision-Language-Action Policies", "authors": ["Lan-Ji An", "Da-Wei Liu", "Jin Li", "Hao-Ran Xu", "Mei Chen", "Yu-Ying Tian"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.27753", "pdf_url": "https://arxiv.org/pdf/2609.27753", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["aligned world modeling", "explainability"], "model_class": "diffusion/flow policy", "techniques": ["future latent alignment", "learnable future tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Embeds world modeling in a diffusion-transformer policy via a few future-aligned tokens rather than pixel prediction, adding little cost.", "relevance": 3, "relevance_reason": "Cheap world-model integration without pixel generation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2607.09648", "title": "B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations", "authors": ["Xiao-Shen Han", "Haoyu Xiong", "Hao-Nan Chen", "Chaoqi Liu", "A. Torralba", "Yu-Ke Zhu", "Yi-Lun Du"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.09648", "pdf_url": "https://arxiv.org/pdf/2607.09648", "code_url": "", "clusters": ["C1"], "keywords": ["time-scalable actions", "task speed"], "model_class": "other", "techniques": ["B-spline action parameterization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Represents action chunks as B-spline curves so trajectories can be time-scaled and executed faster by low-level controllers.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.27450", "title": "BEE: Intervention-Adaptive Real-World Reinforcement Learning with Vision-Language-Action Models", "authors": ["Weihui Zhao", "Xiaohan Yan", "Zunian Wan", "Xuan Du", "Zhaozhan Chi", "Jianbo Mao", "Ruipu Wu", "Rushuai Yang", "Houlin Li", "Shukai Yang", "Jing Wu", "Yuxiang Yan", "Yongcheng Liu", "Chuankang Li", "Guanghui Ren", "Wei Shan", "Maoqing Yao", "Wei-Hui Zhao", "Xiao Yan", "Zu-Nian Wan", "Zhao-Zhan Chi", "Jian-Bo Mao", "Rui-Pu Wu", "Ru-Shuai Yang", "Hou-Lin Li", "Shu-Kai Yang", "Yu-Xiang Yan", "Yong-Cheng Liu", "Chuan-Kang Li", "Shan Wei", "Mao-Qing Yao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.27450", "pdf_url": "https://arxiv.org/pdf/2609.27450", "code_url": "", "clusters": ["C3"], "keywords": ["human intervention", "real-world RL", "BEE", "online RL"], "model_class": "VLA", "techniques": ["residual RL", "correction modeling", "real-world RL", "human intervention"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Pro", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Real-world RL on a frozen VLA that treats human corrections as per-dimension constraints, improving success with fewer interventions at a matched online-data budget.", "relevance": 3, "relevance_reason": "Efficient human-in-the-loop RL for precision phases.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (interrupt* OR \"human intervention\" OR \"mid-execution\"); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.29201", "title": "Behavior Uncloning: Distilling Mode Redirection into Policy Weights without Inference-Time Steering", "authors": ["Hao Wang", "Jiu-Zhou Lei", "Dayou Li", "Bangya Liu", "Ming-Hui Zheng", "Manling Li", "Ruo-Han Zhang", "Zhi-Wen Fan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.29201", "pdf_url": "https://arxiv.org/pdf/2606.29201", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["mode redirection", "unsafe modes"], "model_class": "other", "techniques": ["distillation of steering signal"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills inference-time mode redirection into policy weights via a short 'uncloning' step, avoiding steering overhead at deployment.", "relevance": 3, "relevance_reason": "Removes inference-time steering overhead for safe behavior.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.33940", "title": "Behavioral Monitoring of JEPA World Models with Jacobian Centroids", "authors": ["Thomas Walker", "Randall Balestriero", "Richard Baraniuk"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33940", "pdf_url": "https://arxiv.org/pdf/2609.33940", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["behavioral monitoring", "Jacobian centroids", "JEPA WM"], "model_class": "world model", "techniques": ["Jacobian-vector-product centroids as monitoring signal"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Monitors whether a JEPA world model used for planning is behaviorally aligned with the task using cheap Jacobian-based centroids of its internal representations.", "relevance": 3, "relevance_reason": "Monitoring planning world models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2604.24018", "title": "Betting for Sim-to-Real Performance Evaluation", "authors": ["Zaid Mahboob", "Yujia Chen", "Bowen Weng"], "year": 2026, "venue": "RSS 2026", "type": "preprint", "url": "https://arxiv.org/abs/2604.24018", "pdf_url": "https://arxiv.org/pdf/2604.24018", "code_url": "https://github.com/ISUSAIL/Bet4Sim2Real", "clusters": ["C4"], "keywords": ["sim-to-real evaluation", "betting", "statistical estimation", "variance reduction"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Frames estimating real-world robot performance with few physical trials as a betting problem using cheap simulation, with conditions under which it provably beats Monte Carlo.", "relevance": 3, "relevance_reason": "Statistically efficient evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2605.19919", "title": "Beyond Action Residuals: Real-World Robot Policy Steering via Bottleneck Latent Reinforcement Learning", "authors": ["Dongjie Yu", "Kun Lei", "Zhennan Jiang", "Jia Pan", "Huazhe Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.19919", "pdf_url": "https://arxiv.org/pdf/2605.19919", "code_url": "", "clusters": ["C3"], "keywords": ["latent steering", "information bottleneck", "real-world RL"], "model_class": "VLA", "techniques": ["latent-space RL", "VIB module"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Steers a pretrained policy through a compact bottleneck latent via RL instead of weights or action residuals.", "relevance": 3, "relevance_reason": "Lightweight real-world adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.28339", "title": "Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control", "authors": ["Zan-Yi Wang", "Yu-Heng Lei", "Deng-Yang Jiang", "Ping Luo", "Meng-Di Wang", "Zhi-Xuan Liang", "Shi-Long Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28339", "pdf_url": "https://arxiv.org/pdf/2609.28339", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["DiT prior", "generative adaptation", "world-action"], "model_class": "world-action model", "techniques": ["token reduction", "co-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.8x faster training step (2.85 s -> 1.63 s); visual tokens 784 -> 392", "metric_measures": "other", "benchmarks": ["LIBERO-Plus"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts pretrained image/video DiTs to control by denoising the current observation jointly with actions instead of predicting the future, halving visual tokens during training.", "relevance": 3, "relevance_reason": "Reports training-efficiency gains (fewer tokens, faster step) for generative-prior policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2602.12628", "title": "Beyond Imitation: Reinforcement Learning-Based Sim-Real Co-Training for VLA Models", "authors": ["Liangzhi Shi", "Shuaihang Chen", "Feng Gao", "Yinuo Chen", "Kang Chen", "Tonghe Zhang", "Hongzhi Zang", "Jiakai Zhou", "Weinan Zhang", "Chao Yu", "Yu Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.12628", "pdf_url": "https://arxiv.org/pdf/2602.12628", "code_url": "", "clusters": ["C3"], "keywords": ["sim-real co-training", "RL-Co"], "model_class": "VLA", "techniques": ["RL co-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Two-stage sim-real co-training that adds interactive RL in simulation while preserving real-world capability.", "relevance": 3, "relevance_reason": "Reduces real demo cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2608.21204", "title": "Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning", "authors": ["Varun Giridhar", "Anant Khandelwal", "Jeremy A. Collins", "Ignat Georgiev", "Animesh Garg"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.21204", "pdf_url": "https://arxiv.org/pdf/2608.21204", "code_url": "", "clusters": ["C3"], "keywords": ["self-improvement", "frozen BC policy"], "model_class": "VLA", "techniques": ["small off-policy Q-function", "value-guided action selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adds a small Q-function to a large frozen BC policy for value-weighted action selection and online improvement that only updates the Q-function.", "relevance": 3, "relevance_reason": "Efficient adaptation of large robot policies without updating policy weights.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2609.11270", "title": "Beyond Noise Steering: Dual-Latent Space Reinforcement Learning for Generative Robot Policy", "authors": ["Pengfei Zhang", "Teng Sun", "Xianchao Xiu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.11270", "pdf_url": "https://arxiv.org/pdf/2609.11270", "code_url": "", "clusters": ["C3"], "keywords": ["noise steering", "latent-space RL", "frozen generator"], "model_class": "VLA", "techniques": ["latent-space RL", "adapters"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "RL actor steers both initial noise and injected intermediate adapter features of a frozen generative policy.", "relevance": 3, "relevance_reason": "Lightweight RL adaptation of frozen VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.33595", "title": "Beyond One-Step Accuracy: State-Affine Latent Transition for Reliable Visual Planning", "authors": ["Boyuan Zhang", "Yingjun Du", "Xiantong Zhen", "Ling Shao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33595", "pdf_url": "https://arxiv.org/pdf/2609.33595", "code_url": "", "clusters": ["C8"], "keywords": ["state-affine transitions", "rollout error propagation", "reliable latent planning"], "model_class": "world model", "techniques": ["state-affine latent dynamics (SALT)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows state-affine latent transitions remove nonlinear error propagation in recursive rollouts, improving reliability of JEPA visual planning.", "relevance": 3, "relevance_reason": "Reliable latent planning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2609.25820", "title": "Beyond Reconstruction Error: Analytical and Data-Driven Action Tokenization for Autoregressive Vision-Language-Action Models", "authors": ["Yu-Xin Yang", "Gao-Han He", "Chang-Xue Guan", "Hang Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.25820", "pdf_url": "https://arxiv.org/pdf/2609.25820", "code_url": "", "clusters": ["C1"], "keywords": ["action tokenization", "DCT"], "model_class": "VLA", "techniques": ["action tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Compares analytical, linear and neural action tokenizers for autoregressive VLAs, showing reconstruction error alone does not predict closed-loop control quality.", "relevance": 3, "relevance_reason": "relevant: action tokenization design", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2603.09513", "title": "Beyond Short-Horizon: VQ-Memory for Robust Long-Horizon Manipulation in Non-Markovian Simulation Benchmarks", "authors": ["Honghui Wang", "Zhihui Jing", "Jicong Ao", "Shiji Song", "Xue-Long Li", "Gao Huang", "Chen-Jia Bai"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.09513", "pdf_url": "https://arxiv.org/pdf/2603.09513", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["VQ-Memory", "RuleSafe"], "model_class": "VLA", "techniques": ["vector quantization", "compact memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a non-Markovian benchmark and a VQ-VAE memory of past proprioception that improves long-horizon VLA performance at reduced compute.", "relevance": 3, "relevance_reason": "Compact memory with lower compute.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.15940", "title": "Beyond Single-Axis Testing: Paired Evaluation of Compound Robustness in Vision-Language-Action Policies", "authors": ["Hiroki Sawada", "Shunichi Kasahara"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.15940", "pdf_url": "https://arxiv.org/pdf/2609.15940", "code_url": "", "clusters": ["C4"], "keywords": ["compound robustness", "paired evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proposes paired evaluation of VLA policies under simultaneous distribution shifts rather than one perturbation at a time.", "relevance": 3, "relevance_reason": "Benchmarking methodology for VLAs", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2406.09246"}
{"id": "2606.01095", "title": "Beyond Task Success: Behavioral and Representational Diagnostics for WAM and VLA", "authors": ["H. Mai", "Bin Zhu", "Tuan Do"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.01095", "pdf_url": "https://arxiv.org/pdf/2606.01095", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["runtime cost", "sparse autoencoder"], "model_class": "world-action model", "techniques": ["diagnostics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboTwin 2.0"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diagnostic framework comparing WAMs and VLAs beyond success rate, measuring behavior, internal representations and runtime cost; finds WAM gains come with higher inference cost.", "relevance": 3, "relevance_reason": "Evaluation methodology that explicitly accounts for inference cost.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2606.21775", "title": "Beyond the Next Step: Variable-Length Latent World Models for Long-Horizon Planning", "authors": ["Tianqi Du", "Qi Zhang", "Yifei Wang", "Yisen Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.21775", "pdf_url": "https://arxiv.org/pdf/2606.21775", "code_url": "", "clusters": ["C8"], "keywords": ["variable-length latent prediction", "long-horizon planning", "JEPA"], "model_class": "world model", "techniques": ["direct prediction from variable-length action sequences"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Trains JEPA latent world models to jump directly to states after variable-length action sequences, avoiding recursive one-step rollouts in planning.", "relevance": 3, "relevance_reason": "Cheaper, more accurate latent planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.26757", "title": "Beyond Viewpoint Generalization: What Multi-View Demonstrations Offer and How to Synthesize Them for Robot Manipulation?", "authors": ["Boyang Cai", "Qiwei Liang", "Jiawei Li", "Shihan Weng", "Zhaoxin Zhang", "Tao Lin", "Xiang-Yu Chen", "Wen-Jie Zhang", "Jiaqing Mao", "Wei-Sheng Xu", "Bingyu Yang", "Jiaming Liang", "Jun-Hao Cai", "Ren-Jing Xu"], "year": 2026, "venue": "arXiv", "type": "paper", "url": "https://arxiv.org/abs/2603.26757", "pdf_url": "https://arxiv.org/pdf/2603.26757", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["survey", "diffusion policy", "real-time"], "model_class": "diffusion/flow policy", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of diffusion policies organized around 3D geometric generalization, real-time physical/tactile interaction and generalist policies.", "relevance": 3, "relevance_reason": "relevant: survey covering real-time diffusion policy methods", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.13924", "title": "BICPO-VLA: Behavior-Identified Continuation Preference Optimization for Smooth Asynchronous Vision-Language-Action Control", "authors": ["Ming Shang", "Yuchen Huang", "Jiaoyang Chen", "Haoyuan Hu", "Hangzheng Yu", "Liping Song", "Luyun Feng", "Shuo Bao", "Wei Dong", "Xinzhou Wang", "Fu-Chun Sun"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.13924", "pdf_url": "https://arxiv.org/pdf/2608.13924", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["BICPO-VLA", "request-to-handoff gap", "continuation preference optimization"], "model_class": "VLA", "techniques": ["Haar subspace generation", "preference optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Addresses the gap between chunk request and handoff in async VLAs via behavior identification, faster two-stage Haar-subspace generation, and continuation preference optimization.", "relevance": 3, "relevance_reason": "Smooth async control.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2601.12116", "title": "BiKC+: Bimanual Hierarchical Imitation With Keypose-Conditioned Coordination-Aware Consistency Policies", "authors": ["Hang Xu", "Yizhou Chen", "Dongjie Yu", "Yi Ren", "Jia Pan"], "year": 2026, "venue": "IEEE Transactions on Automation Science and Engineering", "type": "paper", "url": "https://arxiv.org/abs/2601.12116", "pdf_url": "https://arxiv.org/pdf/2601.12116", "code_url": "", "clusters": ["C1"], "keywords": ["bimanual", "keypose hierarchy"], "model_class": "diffusion/flow policy", "techniques": ["consistency model", "one-step inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Hierarchical bimanual IL with a keypose predictor and a one-step consistency trajectory generator for fast inference.", "relevance": 3, "relevance_reason": "One-step generative action decoding.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.08684", "title": "BLUE: Toward Better Language Use in Efficient Vision-Language-Action Models for Autonomous Driving", "authors": ["George Ling", "Lijin Yang", "Hao Yang", "Zhongzhan Huang"], "year": 2026, "venue": "EMNLP 2026", "type": "paper", "url": "https://arxiv.org/abs/2606.08684", "pdf_url": "https://arxiv.org/pdf/2606.08684", "code_url": "https://github.com/George-Ling3/BLUE", "clusters": ["C1"], "keywords": ["language use", "driving VLA"], "model_class": "VLA", "techniques": ["gating", "conditional computation"], "base_model": "", "params": "0.11M gate", "hardware": "", "reported_metric": "2.54x inference speedup; 76.2% success Bench2Drive", "metric_measures": "relative latency", "benchmarks": ["Bench2Drive", "Longest6 v2"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A 0.11M-parameter gate on frozen driving-VLA hidden states decides per frame whether to generate language or output actions directly.", "relevance": 3, "relevance_reason": "Conditional reasoning to avoid per-frame language cost.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]"}
{"id": "2604.01570", "title": "Boosting Vision-Language-Action Finetuning with Feasible Action Neighborhood Prior", "authors": ["Haochen Niu", "Kanyu Zhang", "Shuyu Yin", "Qinghai Guo", "Peilin Liu", "Fei Wen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.01570", "pdf_url": "https://arxiv.org/pdf/2604.01570", "code_url": "", "clusters": ["C3"], "keywords": ["feasible action neighborhood", "regularizer", "sample efficiency"], "model_class": "VLA", "techniques": ["distributional regularization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Gaussian feasible-action-neighborhood prior regularizes VLA fine-tuning for better sample efficiency in SFT and RFT.", "relevance": 3, "relevance_reason": "Improves fine-tuning sample efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2605.30226", "title": "BORA: Bridging Offline Reinforcement Learning and Online Residual Adaptation for Real-World Dexterous VLA Models", "authors": ["Zhongxi Chen", "Yifan Han", "Yanming Shao", "Huanming Liu", "Congsheng Xu", "Xiaoyu Chen", "Yao Mu", "Wenzhao Lian"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30226", "pdf_url": "https://arxiv.org/pdf/2605.30226", "code_url": "", "clusters": ["C3"], "keywords": ["dexterous VLA"], "model_class": "VLA", "techniques": ["offline-to-online RL", "residual adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Offline-to-online RL for dexterous VLAs: an action-chunk-aware critic built on VLM tokens, then online residual adaptation to reduce exploration risk.", "relevance": 3, "relevance_reason": "Sample-efficient, safer real-world VLA RL.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]"}
{"id": "2602.15010", "title": "BPP: Long-Context Robot Imitation Learning by Focusing on Key History Frames", "authors": ["Max Sobol Mark", "Jacky Liang", "Maria Attarian", "C. Fu", "Debidatta Dwibedi", "Dhruv Shah", "Aviral Kumar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.15010", "pdf_url": "https://arxiv.org/pdf/2602.15010", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["long-context policy", "key-frame selection"], "model_class": "VLA", "techniques": ["history compression", "key-frame selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Shows naive history conditioning yields spurious correlations and conditions policies only on selected key past frames for robust long-context behavior.", "relevance": 3, "relevance_reason": "Compact history reduces context cost for memory-dependent policies.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2607.02195", "title": "Bridge-WA: Predicting Where and How the World Changes for Robotic Action", "authors": ["Yongjie Bai", "Hanting Wang", "Ming-Tong Dai", "Qi-Jun Zhong", "Yang Liu", "Liang Lin"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.02195", "pdf_url": "https://arxiv.org/pdf/2607.02195", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["Bridge-WA", "future-change teacher"], "model_class": "world-action model", "techniques": ["distillation", "lightweight priors"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["VLABench", "RoboTwin 2.0", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills a frozen future-prediction teacher into compact priors (future tokens, change and flow maps) that condition the action transformer, removing dense future generation at deployment.", "relevance": 3, "relevance_reason": "Avoids expensive world-model rollouts at inference via distillation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2606.27962", "title": "Building a Scalable, Reproducible, Evaluatable, and Closed-Loop Simulation Environment Foundation for Embodied Intelligence", "authors": ["Junwu Xiong", "Yong-Jian Guo", "Mingxia Luo", "Ning Qiao", "Lei Kang", "Song Wang", "Yince Gao", "Chen-Feng Gu", "Zhen Sun", "Haoran Li", "Wei Lu", "Yu-Cheng Guo", "Shuai Di", "Xiaodong Bai", "Haoran Sun", "Jing Long", "Jia-Xuan Gao", "Hui Zhang", "Peng Hao", "Luelue Lu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.27962", "pdf_url": "https://arxiv.org/pdf/2606.27962", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["evaluation infrastructure"], "model_class": "other", "techniques": ["cloud-native simulation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cloud-native simulation infrastructure for large-scale training, standardized evaluation and data collection for embodied models.", "relevance": 3, "relevance_reason": "Evaluation/serving infrastructure.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2607.29302", "title": "BWM: A Low-Cost High-Fidelity World Simulator for Robot Learning", "authors": ["BWM Team"], "year": 2026, "venue": "arXiv", "type": "repo", "url": "https://arxiv.org/abs/2607.29302", "pdf_url": "https://arxiv.org/pdf/2607.29302", "code_url": "", "clusters": ["C8"], "keywords": ["low-cost world simulator", "risk anticipation", "policy ranking"], "model_class": "world model", "techniques": ["stateful autoregressive action-conditioned prediction", "action-aligned clip construction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["WorldArena", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Open-source low-cost action-conditioned robot world model used both as a data engine and as a closed-loop evaluator that anticipates risky outcomes before execution.", "relevance": 3, "relevance_reason": "Cheap world simulator with risk anticipation use.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2605.26820", "title": "Can Vision-Language-Action Models Learn from Real-World Data Continually without Forgetting?", "authors": ["Jiarun Zhu", "Yijun Hong", "Xiaoquan Sun", "Zetian Xu", "Qijun He", "Haijier Chen", "Zhiyong Wang", "Mingqi Yuan", "Wenjun Zeng", "Jiayu Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.26820", "pdf_url": "https://arxiv.org/pdf/2605.26820", "code_url": "", "clusters": ["C3"], "keywords": ["continual learning", "experience replay", "real-world benchmark"], "model_class": "VLA", "techniques": ["experience replay", "sequential fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["10-task real-world continual benchmark"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Real-robot continual learning benchmark for VLAs showing well-tuned experience replay beats joint training at equal compute.", "relevance": 3, "relevance_reason": "Compute-matched continual adaptation study.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: continual learning vision-language-action model catastrophic forgetting arXiv"}
{"id": "2609.24118", "title": "CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies", "authors": ["Junlan Xiao", "Junwei Jiang", "Zaibin Zhang", "Yifan Wang", "Zhongbo Zhang", "Huchuan Lu", "Lijun Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24118", "pdf_url": "https://arxiv.org/pdf/2609.24118", "code_url": "https://github.com/xiaojunlan/care", "clusters": ["C7"], "keywords": ["CARE", "FSR-Bench"], "model_class": "VLA", "techniques": ["failure-driven data synthesis", "3D monitoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns failure deviation distributions from real rollouts to synthesize corrective demos and uses stage-wise 3D monitoring to trigger atomic corrections at inference.", "relevance": 3, "relevance_reason": "Monitoring-triggered corrective execution.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=0]"}
{"id": "2609.18016", "title": "Causal-History Test-Time Scaling for Failure Recovery in Autoregressive World-Action Models", "authors": ["Lin Li", "Long Chen", "Kwunhang", "Wong", "Jiaming Lei", "Song Jin", "Shucheng Du", "Chuhan Zhang", "Songchen Ma", "Weihao Zhang", "Jun Xiao", "Kwang-Ting", "Cheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18016", "pdf_url": "https://arxiv.org/pdf/2609.18016", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["failure recovery", "causal history", "test-time scaling"], "model_class": "world-action model", "techniques": ["training-free revision of autoregressive causal history"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Training-free failure recovery for autoregressive WAMs that decides when and where to rewrite the causal history after execution diverges.", "relevance": 3, "relevance_reason": "Runtime recovery for WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2606.15341", "title": "CausalDrive: Real-time Causal World Models for Autonomous Driving", "authors": ["Tianyi Yan", "Huan Zheng", "Dubing Chen", "Meizhi Qu", "Yingying Shen", "Lijun Zhou", "Mingfei Tu", "Bing Wang", "Guang Chen", "Hangjun Ye", "Haiyang Sun", "Cheng-zhong Xu", "Jianbing Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.15341", "pdf_url": "https://arxiv.org/pdf/2606.15341", "code_url": "", "clusters": ["C8"], "keywords": ["real-time driving world model", "Context-Forced DMD", "counterfactual reactions"], "model_class": "world model", "techniques": ["flow matching + self-correcting DMD distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Real-time controllable driving world renderer that predicts reactive agent behavior from ego trajectory and text, using a distillation scheme to remove diffusion latency for closed-loop policy learning.", "relevance": 3, "relevance_reason": "Real-time distilled world model for closed-loop driving.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.27468", "title": "CereVLA: Cerebellum-Inspired Consequence-Aware Residual Governance for Efficient Vision-Language-Action Execution", "authors": ["Shuai Zeng", "Yuxuan Liang", "Hangmiao Hu", "Fobao Zhou", "Zixiang Wang", "Wenxi Hong", "Hang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.27468", "pdf_url": "https://arxiv.org/pdf/2609.27468", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["CereVLA", "consequence-aware residual", "governor"], "model_class": "VLA", "techniques": ["flow residual refinement", "consequence prediction", "governor suppression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds residual corrections to frozen VLA chunks but predicts their short-horizon consequences with a state-space model and suppresses unfavorable ones via a lightweight governor.", "relevance": 3, "relevance_reason": "Guarded fast correction layer.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"action chunk\" execution horizon adaptive"}
{"id": "2603.11642", "title": "Chunk-Boundary Artifact in Action-Chunked Generative Policies: A Noise-Sensitive Failure Mechanism", "authors": ["Rui Wang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.11642", "pdf_url": "https://arxiv.org/pdf/2603.11642", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["chunk-boundary artifact", "noise-space attribution", "diffusion policy failure mechanism", "chunk boundary", "failure mechanism"], "model_class": "diffusion/flow policy", "techniques": ["analysis", "noise control", "action chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Treats chunk-boundary discontinuity as a measurable failure mechanism, showing it separates successful from failed episodes and can be modulated via the initial latent noise depending on sampler.", "relevance": 3, "relevance_reason": "Chunk execution/timing failure analysis per title; unverified.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.29172", "title": "CLIFT: Turning Gemini Robotics On-Device into Humanoid Specialists via Non-Invasive Closed-Loop Iterative Fine-Tuning", "authors": ["Yuxin Chen", "Hari Srikanth", "Nathan Jew", "Menglin Wu", "Pengcheng Wang", "Junli Ren", "Masayoshi Tomizuka", "Peng Xu", "Jinyu Xie", "Thomas Tian"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.29172", "pdf_url": "https://arxiv.org/pdf/2607.29172", "code_url": "", "clusters": ["C1"], "keywords": ["Gemini Robotics On-Device", "managed fine-tuning API", "humanoid"], "model_class": "VLA", "techniques": ["closed-loop fine-tuning"], "base_model": "Gemini Robotics On-Device", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Studies adapting the closed-weight on-device Gemini Robotics model to humanoid tasks via a managed SFT API with closed-loop iterative data collection.", "relevance": 3, "relevance_reason": "Evidence on the on-device Gemini model.", "foundational_technique": false, "verified": true, "found_via": "search query: Gemini Robotics"}
{"id": "2607.20345", "title": "Closing the Lab-to-Store Gap: A Data-Efficient Post-Training and Experience-Driven Learning VLA Framework for Retail Humanoids", "authors": ["Roger Sala Sisó", "Tiago Silvério", "Jakob Sand", "Tran Nguyen Le"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.20345", "pdf_url": "https://arxiv.org/pdf/2607.20345", "code_url": "", "clusters": ["C3"], "keywords": ["DEED", "retail humanoid", "GR00T N1.6", "data-efficient post-training"], "model_class": "VLA", "techniques": ["data curation", "experience-driven learning"], "base_model": "GR00T N1.6", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Data-efficient post-training pipeline plus experience-driven learning for a GR00T N1.6 humanoid on supermarket restocking.", "relevance": 3, "relevance_reason": "Practical post-training for deployment.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"single GPU\""}
{"id": "2609.25415", "title": "Cloud, Edge, or Split? Profiling Onboard and Split Vision-Language Model Deployment for Drone AI", "authors": ["Zoha Azimi", "Reza Farahani", "Schahram Dustdar", "Christian Timmerer"], "year": 2026, "venue": "iEdge 2026", "type": "paper", "url": "https://arxiv.org/abs/2609.25415", "pdf_url": "https://arxiv.org/pdf/2609.25415", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["UAV", "onboard vs cloud vs split", "profiling"], "model_class": "other", "techniques": ["split computing", "deployment profiling"], "base_model": "SmolVLM-256M", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Benchmarks onboard, cloud and split deployment of a small VLM for drones across resolutions and network conditions, finding no universally best option.", "relevance": 3, "relevance_reason": "Direct onboard/cloud/split trade-off measurement.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.19923", "title": "Co-VLA: Consensus-based Federated Training for Vision-Language-Action Models", "authors": ["Haolong Li", "Guner Dilsad Er", "Michael Muehlebach", "Joerg Stueckler"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19923", "pdf_url": "https://arxiv.org/pdf/2609.19923", "code_url": "", "clusters": ["C3"], "keywords": ["federated learning", "ADMM", "rank-adaptive adapters"], "model_class": "VLA", "techniques": ["federated PEFT", "LoRA"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Consensus (ADMM) federated training of VLAs supporting full-model and fixed/rank-adaptive adapter fine-tuning across robot clients.", "relevance": 3, "relevance_reason": "Distributed PEFT for VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"robot\" AND abs:\"parameter-efficient fine-tuning\""}
{"id": "2603.10436", "title": "COHORT: Hybrid RL for Collaborative Large DNN Inference on Multi-Robot Systems Under Real-Time Constraints", "authors": ["Mohammad Saeid Anwar", "Anuradha Ravi", "Indrajeet Ghosh", "Gaurav Shinde", "Carl Busart", "Nirmalya Roy"], "year": 2026, "venue": "IEEE WoWMoM 2026", "type": "paper", "url": "https://arxiv.org/abs/2603.10436", "pdf_url": "https://arxiv.org/pdf/2603.10436", "code_url": "", "clusters": ["C5", "C6", "C7"], "keywords": ["multi-robot collaborative inference", "RL scheduling", "ROS", "COHORT", "collaborative inference", "multi-robot", "real-time constraints"], "model_class": "other", "techniques": ["distributed DNN inference", "offline+online RL scheduling", "hybrid offline-online RL scheduling"], "base_model": "CLIP, SAM", "params": "", "hardware": "", "reported_metric": "15.4% lower battery use, 51.67% higher GPU utilization", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Schedules modules of large vision-language models across a team of robots using offline RL then online MAPPO fine-tuning, meeting deadlines without server support.", "relevance": 3, "relevance_reason": "Fleet-level distributed inference for large models on robots (not VLA).", "foundational_technique": false, "verified": true, "found_via": "search query; arXiv search: \"deadline\" robot inference neural network real-time scheduling"}
{"id": "2601.23087", "title": "CoLA-Flow Policy: Temporally Coherent Imitation Learning via Continuous Latent Action Flow Matching for Robotic Manipulation", "authors": ["Song Wu", "Zhi-Duo Jiang", "Wandong Sun", "Guang-Hu Xie", "Yuteng Xie", "Ruikang Zhao", "Yang Liu", "Hong Liu"], "year": 2026, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2601.23087", "pdf_url": "https://arxiv.org/pdf/2601.23087", "code_url": "", "clusters": ["C1"], "keywords": ["near-single-step", "temporal coherence"], "model_class": "diffusion/flow policy", "techniques": ["latent action flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Flow matching in a continuous latent action space for near single-step, smooth long-horizon control.", "relevance": 3, "relevance_reason": "Fast inference vs diffusion policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2402.15391"}
{"id": "2605.27759", "title": "Colosseum V2: Benchmarking Generalization for Vision-Language-Action Models", "authors": ["Jeremy Morgan", "Hyeonho Oh", "Prajwal Vijay", "Jincen Song", "Ashvin Arora", "Hojung Lim", "Alina Du", "Jesse Thomason", "Gaurav Sukhatme", "Ishika Singh"], "year": 2026, "venue": "IEEE RA-L", "type": "benchmark", "url": "https://arxiv.org/abs/2605.27759", "pdf_url": "https://arxiv.org/pdf/2605.27759", "code_url": "", "clusters": ["C4"], "keywords": ["Colosseum V2", "ManiSkill", "GPU-parallel evaluation", "sim-real correlation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Colosseum V2", "ManiSkill"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "28-task, two-morphology generalization benchmark built on ManiSkill with GPU-parallel evaluation, evaluating ACT and pi0.5 and reporting sim-real metric correlation.", "relevance": 3, "relevance_reason": "Fast, reproducible generalization benchmark.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA acceleration evaluation"}
{"id": "10.1109/tcsvt.2025.3637139", "title": "Complexity-Aware Policy via Heterogeneous Experts for Robotic Manipulation", "authors": ["Ming Yang", "Yunzhe Hu", "Ge Yuan", "Di-Fan Zou", "Jianxin Pang", "Dong Xu"], "year": 2026, "venue": "IEEE transactions on circuits and systems for video technology (Print)", "type": "paper", "url": "https://doi.org/10.1109/TCSVT.2025.3637139", "pdf_url": "", "code_url": "", "clusters": ["C1"], "keywords": ["complexity-aware compute"], "model_class": "diffusion/flow policy", "techniques": ["mixture of heterogeneous experts", "rectified flow"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN", "LIBERO", "SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "CAP scales model capacity to task complexity via heterogeneous experts in a flow policy, allocating compute efficiently.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2606.21562", "title": "Compressing Observation History into Agent Memory: Distilling Transformers into Recurrent Transformers", "authors": ["Philippe Weinzaepfel", "Christian Wolf", "Mert Bulent Sariyildiz", "G. Bono", "G. Monaci"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.21562", "pdf_url": "https://arxiv.org/pdf/2606.21562", "code_url": "", "clusters": ["C2"], "keywords": ["memory compression", "streaming"], "model_class": "other", "techniques": ["distillation", "recurrent transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills a full-history transformer's compression into a recurrent transformer for linear-time robotic memory.", "relevance": 3, "relevance_reason": "Distillation for efficient streaming robotics models.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2504.16054"}
{"id": "2603.02196", "title": "Conformal Policy Control", "authors": ["Drew Prinster", "Clara Fannjiang", "Ji Won Park", "Kyunghyun Cho", "Anqi Liu", "Suchi Saria", "Samuel Stanton"], "year": 2026, "venue": "ICML 2026", "type": "paper", "url": "https://arxiv.org/abs/2603.02196", "pdf_url": "https://arxiv.org/pdf/2603.02196", "code_url": "", "clusters": ["C7"], "keywords": ["safe exploration"], "model_class": "other", "techniques": ["conformal risk control", "policy control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses a safe reference policy and conformal calibration to bound how aggressively a new, untested policy may act while provably meeting a user risk tolerance.", "relevance": 3, "relevance_reason": "Principled regulator for deploying untested policies.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:conformal AND (ti:robot* OR ti:planning OR ti:control)"}
{"id": "2602.03799", "title": "Conformal Reachability for Safe Control in Unknown Environments", "authors": ["Xinhang Ma", "Junlin Wu", "Yiannis Kantaros", "Yevgeniy Vorobeychik"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.03799", "pdf_url": "https://arxiv.org/pdf/2602.03799", "code_url": "", "clusters": ["C7"], "keywords": ["safe control", "unknown dynamics"], "model_class": "other", "techniques": ["conformal prediction", "reachability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Combines per-step conformal uncertainty intervals on unknown dynamics with reachability to verify safety horizons and train policies maximizing the certified horizon.", "relevance": 3, "relevance_reason": "Probabilistic horizon certification, analogous to certifying a chunk length.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:conformal AND (ti:robot* OR ti:planning OR ti:control)"}
{"id": "2605.05126", "title": "ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation", "authors": ["Wei Li", "Jizhihui Liu", "Li Yixing", "Junwen Tong", "Rui Shao", "Liqiang Nie", "Yixin Li"], "year": 2026, "venue": "CVPR 2026", "type": "preprint", "url": "https://arxiv.org/abs/2605.05126", "pdf_url": "https://arxiv.org/pdf/2605.05126", "code_url": "https://github.com/iLearn-Lab/CVPR26-ConsisVLA-4D", "clusters": ["C1"], "keywords": ["3D perception", "4D reasoning", "efficient VLA"], "model_class": "VLA", "techniques": ["compact latent fusion", "token filtering", "compact latent representations"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "2.3x (LIBERO) and 2.4x (real) inference speedup vs OpenVLA", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Efficient VLA framework for cross-view and cross-object consistency using compact latents instead of extra sensors, improving success and speed over OpenVLA.", "relevance": 3, "relevance_reason": "Efficient 3D/4D perception with reported speedups.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.03483", "title": "Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution", "authors": ["Weichen Xu", "Zhenhua Liu", "Lin Luo", "Yaobo Liang", "Chengtang Yao", "Qingyu Mei", "Jian Cao", "Xixin Cao", "Xing Zhang", "Jiaolong Yang", "Baining Guo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.03483", "pdf_url": "https://arxiv.org/pdf/2608.03483", "code_url": "https://fleetfootwork.github.io/BCP/", "clusters": ["C1", "C2", "C7"], "keywords": ["continue-or-replan", "adaptive horizon", "RL", "Bernoulli-continuation policy", "continue or replan"], "model_class": "VLA", "techniques": ["adaptive action chunking", "RL", "continuation head trained with RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a plug-in continuation head that makes per-step continue-or-replan decisions for a frozen VLA, trained with RL from trajectory outcomes.", "relevance": 3, "relevance_reason": "Adaptive horizons reduce VLA calls.", "foundational_technique": false, "verified": true, "found_via": "search query: efficient VLA; arXiv search: \"action chunk\" execution horizon adaptive"}
{"id": "2609.32779", "title": "Copper-Policy: Focus on the Representation for Robust Robot Manipulation", "authors": ["Ze Feng", "Yi-Xu Feng", "Ling-Yu Xiao", "Shang Su", "Kexin Zheng", "Chang Xu", "Meng-Kai Shi", "Shuo Feng", "Xin-Tao Yan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.32779", "pdf_url": "https://arxiv.org/pdf/2609.32779", "code_url": "", "clusters": ["C8"], "keywords": ["training efficiency", "JEPA-style"], "model_class": "world-action model", "techniques": ["compact latent prediction", "joint-embedding prediction"], "base_model": "", "params": "2B", "hardware": "8x RTX 5090 (training); A100 comparison", "reported_metric": "6x faster training than Fast-WAM; 9.67 h training", "metric_measures": "other", "benchmarks": ["RoboTwin", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Copper-Policy learns a compact predictive world representation with the policy instead of reconstructing pixels, cutting training tokens and cost versus pixel/latent WAMs.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.21935", "title": "CoRDE: Concept-Prior Routed Diffusion Experts for Structural Generalization in Robot Manipulation", "authors": ["Haidong Huang", "Xixin Zhao", "Yao Zhou", "Jia Song", "Jia-Yi Zhang", "Jun Ma", "Hai-Yue Zhu", "Xiaocong Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.21935", "pdf_url": "https://arxiv.org/pdf/2606.21935", "code_url": "", "clusters": ["C3"], "keywords": ["parameter-efficient experts", "routing"], "model_class": "diffusion/flow policy", "techniques": ["LoRA experts", "mixture of experts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Concept-routed diffusion experts built as LoRA adapters on a shared frozen backbone to avoid parameter inflation.", "relevance": 3, "relevance_reason": "PEFT-based MoE for policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2609.16724", "title": "CorrRisk-WM: Corridor-Conditioned Risk World Modeling for Safety-Critical Trajectory Planning", "authors": ["Tingyu Guo", "Reza Langari"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.16724", "pdf_url": "https://arxiv.org/pdf/2609.16724", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["risk world model", "trajectory corridors", "near-miss prediction"], "model_class": "world model", "techniques": ["latent environment model + recurrent per-candidate risk module"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Planning-oriented partial world model predicting per-candidate-trajectory intrusion and near-miss probabilities for safety-critical local planning.", "relevance": 3, "relevance_reason": "World-model-based risk estimation for planning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2605.24463", "title": "Cost-Aware Adaptive Conformal Inference for Runtime Assurance in Dynamic Environments", "authors": ["Taoran Wu", "Jingduo Pan", "Luke Ong", "Bai Xue"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24463", "pdf_url": "https://arxiv.org/pdf/2605.24463", "code_url": "", "clusters": ["C7"], "keywords": ["distribution shift", "risk-aware control"], "model_class": "other", "techniques": ["adaptive conformal inference", "runtime assurance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adaptive conformal inference whose margins adapt to violation severity as well as frequency, bounding both violation rate and cumulative cost for runtime assurance under drift.", "relevance": 3, "relevance_reason": "Online-calibrated uncertainty margins for safety monitors.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2609.05832", "title": "CR-VLA-Force: Learning Control-aware Compliance VLA Model for Robust Contact-rich Robotic Manipulation", "authors": ["Zhao-Hong Mai", "Chao Wang", "Chao Zeng", "Sitong Mao", "Heng Zhang", "Shun-Bo Zhou", "Chen-Guang Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.05832", "pdf_url": "https://arxiv.org/pdf/2609.05832", "code_url": "", "clusters": ["C7"], "keywords": ["CR-VLA-Force", "force/torque"], "model_class": "VLA", "techniques": ["compliance control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Control-aware compliance VLA that mitigates the perception-to-control latency of chunk execution in contact-rich tasks for safer force control.", "relevance": 3, "relevance_reason": "Safety under chunk latency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.05185", "title": "Critic in the Loop: A Tri-System VLA Framework for Robust Long-Horizon Manipulation", "authors": ["Pengfei Yi", "Yingjie Ma", "Wenjiang Xu", "Yanan Hao", "Shuai Gan", "Wanting Li", "Shanlin Zhong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.05185", "pdf_url": "https://arxiv.org/pdf/2603.05185", "code_url": "", "clusters": ["C1"], "keywords": ["tri-system", "critic scheduling", "VLM replanning"], "model_class": "VLA", "techniques": ["hierarchical scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A lightweight visual critic routes control between a fast VLA and a slow VLM planner, calling the VLM only on anomalies.", "relevance": 3, "relevance_reason": "Minimizes expensive VLM calls.", "foundational_technique": false, "verified": true, "found_via": "search query: hierarchical VLA latency"}
{"id": "2606.29898", "title": "Critical Interval MSE: Toward Reliable Offline Validation for Robot Manipulation Policies", "authors": ["Hao-Xu Huang", "Tongsam Zheng", "Yifan Chen", "Jiacheng You", "Yang Gao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.29898", "pdf_url": "https://arxiv.org/pdf/2606.29898", "code_url": "", "clusters": ["C4"], "keywords": ["offline validation", "CI-MSE"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Offline validation metric restricting action error to task-critical segments to better predict real-world policy performance.", "relevance": 3, "relevance_reason": "Benchmarking methodology for policy selection.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2602.03445", "title": "CRL-VLA: Continual Vision-Language-Action Learning", "authors": ["Qixin Zeng", "Shuo Zhang", "Hongyin Zhang", "Renjie Wang", "Han Zhao", "Libang Zhao", "Runze Li", "Donglin Wang", "Chao Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.03445", "pdf_url": "https://arxiv.org/pdf/2602.03445", "code_url": "", "clusters": ["C3"], "keywords": ["continual RL", "stability-plasticity", "dual critic"], "model_class": "VLA", "techniques": ["continual RL", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Continual RL post-training for VLAs with a dual-critic design that constrains advantages on old tasks to limit forgetting.", "relevance": 3, "relevance_reason": "Continual adaptation of VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: continual learning vision-language-action model catastrophic forgetting arXiv"}
{"id": "2606.09572", "title": "CT-VAM: A Cerebello-Thalamic-Inspired Vision-Action Model for Efficient Visuomotor Control", "authors": ["Jiacheng Li", "Yize Guo", "Jiabin Guo", "Qingchen Liu", "Jiahu Qin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09572", "pdf_url": "https://arxiv.org/pdf/2606.09572", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["language processed once", "vision-action model", "high-frequency"], "model_class": "VLA", "techniques": ["language-free low-level loop"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "CT-VAM processes language only for intent and runs a lighter vision-action loop at high frequency.", "relevance": 3, "relevance_reason": "Architecture reducing per-step compute.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2601.02295", "title": "CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding", "authors": ["Chen-Yang Ma", "Kai Lu", "Guangyu Yang", "Jiu-Ming Liu", "Shi-Tong Xu", "Bill Byrne", "Ioannis Havoutis", "Niki Trigoni", "Andrew Markham"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2601.02295", "pdf_url": "https://arxiv.org/pdf/2601.02295", "code_url": "", "clusters": ["C7"], "keywords": ["self-correction", "MBR decoding"], "model_class": "VLA", "techniques": ["failure prediction", "test-time scaling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adds proactive failure anticipation to VLAs: a progress-aware policy flags risky subtask transitions, a VLM predicts failures and triggers backtracking, and MBR decoding improves retries.", "relevance": 3, "relevance_reason": "Runtime failure detection/recovery for VLAs (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2604.16683"}
{"id": "2609.34792", "title": "D$^2$-VLA: Dual-Memory Dual-Frequency Vision-Language-Action Model For Long Dynamic Manipulation", "authors": ["Zi-Jian Ye", "Chen Wei", "Wei Huang", "An-Lin Zheng", "Chun-Yu Zou", "Liang-Yu Wu", "Zi-Kang Zhao", "Zhen-Jie Peng", "Yu-Shuo Yang", "Shuo Zhao", "Zhong-Rui Wang", "Xiao-Juan Qi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34792", "pdf_url": "https://arxiv.org/pdf/2609.34792", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["dual-frequency", "KV cache", "dual memory", "dynamic manipulation"], "model_class": "VLA", "techniques": ["block-wise causal KV caching", "gated fast visual adapter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds dual memory and dual-frequency control at the KV-cache interface of a pretrained VLA so fresh visual features update the action expert between slow VLM passes.", "relevance": 3, "relevance_reason": "Fast control under slow VLM refresh.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2609.24749", "title": "D-JEPA: A Decision-Aligned Latent World Model", "authors": ["Shuaijun Liu", "Chengyu Wu", "Qifu Wen", "Feiyang You", "Chenglong Zhang", "Shuyang Hao", "Xi Lin", "Ningxin Su"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24749", "pdf_url": "https://arxiv.org/pdf/2609.24749", "code_url": "", "clusters": ["C8"], "keywords": ["decision-aligned latent world model", "action ranking"], "model_class": "world model", "techniques": ["permutation-equivariant ordinal operator over candidate futures"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["manipulation", "physical robots", "driving"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Aligns JEPA latent geometry with which candidate actually succeeds when executed, fixing cases where the predicted-closest future is the wrong choice while keeping native latent-distance planning.", "relevance": 3, "relevance_reason": "Improves reliability of cheap latent planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2607.10842", "title": "D-SafeMPC: Diffusion-Driven Safe Model Predictive Control with Discrete-Time Control Barrier Functions", "authors": ["Erdi Sayar", "Ersin Daş", "Joel W. Burdick", "Alois Knoll", "Erdal Kayacan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.10842", "pdf_url": "https://arxiv.org/pdf/2607.10842", "code_url": "https://github.com/erdiphd/D-SafeMPC", "clusters": ["C7"], "keywords": ["safe diffusion planning", "CBF"], "model_class": "diffusion/flow policy", "techniques": ["CBF-guided diffusion", "MPC projection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Franka sim", "real robot"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Guides reverse diffusion with CBF/CLF terms and iteratively projects each denoising step through MPC to produce safe, feasible trajectories and good MPC warm starts.", "relevance": 3, "relevance_reason": "Constraint enforcement inside trajectory denoising.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\")"}
{"id": "2605.19887", "title": "DAG-Based QoS-Aware Dynamic Task Placement for Networked Multi-Stage Control Pipelines", "authors": ["Thien Tran", "Jonathan Kua", "Thuong N. Hoang", "Minh Tran", "Yue-Min Ding", "Jiong Jin", "Thuong Hoang", "Yuemin Ding"], "year": 2026, "venue": "IEEE INDIN 2026", "type": "preprint", "url": "https://arxiv.org/abs/2605.19887", "pdf_url": "https://arxiv.org/pdf/2605.19887", "code_url": "", "clusters": ["C5", "C6"], "keywords": ["task placement", "industrial network", "3C co-design", "offloading", "dynamic task placement", "DAG", "QoS"], "model_class": "other", "techniques": ["DAG-based dynamic task placement", "dynamic task placement", "hysteresis switching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Models perception-planning-control pipelines as a DAG and switches stages among local, offloaded and hybrid placements using a tail-latency/deadline/utilization cost with switching penalties.", "relevance": 3, "relevance_reason": "Placement framework for networked robot control pipelines (theoretical).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2604.24447 (citations); snowball from 2205.09778"}
{"id": "2606.12105", "title": "DAM-VLA: Decoupled Asynchronous Multimodal Vision Language Action model", "authors": ["Pankhuri Vanjani", "Zhuo Li", "Jakub Suliga", "Moritz Reuss", "G. Geraci", "Xin-Kai Jiang", "Rudolf Lioutikov"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.12105", "pdf_url": "https://arxiv.org/pdf/2606.12105", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["DAM-VLA", "per-modality rates", "asynchronous multimodal", "latent buffers"], "model_class": "VLA", "techniques": ["per-modality latent buffers at sensor rates"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Maintains per-modality latent buffers refreshed at native sensor rates and read continuously by the action head, decoupling fast (force/proprio) from slow (vision/language) modalities.", "relevance": 3, "relevance_reason": "Asynchronous multi-rate control.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2602.05513", "title": "DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter", "authors": ["Xukun Li", "Yu Sun", "Lei Zhang", "Bo-Sheng Huang", "Yibo Peng", "Yuan Meng", "Haojun Jiang", "Shao-Xuan Xie", "Guo-Cai Yao", "A. Knoll", "Zhen-Shan Bing", "Xinlong Wang", "Zhenguo Sun"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.05513", "pdf_url": "https://arxiv.org/pdf/2602.05513", "code_url": "", "clusters": ["C3"], "keywords": ["tactile", "bimanual dexterous"], "model_class": "diffusion/flow policy", "techniques": ["parameter-efficient adapter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "DECO decouples vision/proprioception/tactile conditioning in a diffusion transformer and adds tactile via a small adapter tuning under 10% of parameters.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.13474", "title": "Decoding Task Progress from VLA Representations", "authors": ["Atiksh Bhardwaj", "E. W. Duan", "Prithwish Dan", "Wei-Chiu Ma", "Preston Culbertson"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.13474", "pdf_url": "https://arxiv.org/pdf/2608.13474", "code_url": "", "clusters": ["C7"], "keywords": ["task progress"], "model_class": "VLA", "techniques": ["linear probing", "runtime monitoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows task progress is linearly decodable from VLA internals, offering a lightweight runtime monitoring signal.", "relevance": 3, "relevance_reason": "Runtime monitoring for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2504.16054, 2505.23705"}
{"id": "2608.20114", "title": "DECOWAM: Decoupled Whole-Body World-Action Model for Legged Mobile Manipulation", "authors": ["Siyuan Ma", "Boshi Zhang", "Yutian Zhang", "Qinglian Wu", "Jiaqi Zhai", "Dong Wei", "Qiaojun Yu", "Si-Yuan Ma", "Bo-Shi Zhang", "Yu-Tian Zhang", "Qing-Lian Wu", "Jia-Qi Zhai", "Dong-Jun Wei", "Qiao-Jun Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.20114", "pdf_url": "https://arxiv.org/pdf/2608.20114", "code_url": "", "clusters": ["C2", "C3", "C8"], "keywords": ["DECOWAM", "mobile manipulation", "parameter-efficient", "legged mobile manipulation"], "model_class": "world-action model", "techniques": ["residual adapters", "distillation", "frozen backbone", "distilled bottleneck"], "base_model": "FastWAM", "params": "25.95M trainable adaptation parameters", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Whole-body WAM for legged mobile manipulation freezing FastWAM and training 25.95M adapter parameters with distilled future bottleneck and decoupled base/arm latents.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2106.09685, 2301.04104"}
{"id": "2608.16556", "title": "DeepInsight II: One Trace from Benchmark to Robot", "authors": ["Siyi Li", "Yuchen Kang", "Wuliang Wang", "Zhengjie Zhang", "Jiangpin Liu", "Jianhao Yao", "Jie Chen"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2608.16556", "pdf_url": "https://arxiv.org/pdf/2608.16556", "code_url": "", "clusters": ["C4"], "keywords": ["DeepInsight", "evaluation harness", "trace", "sim-to-real"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["navigation benchmarks", "manipulation benchmarks", "MotionBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Evaluation infrastructure report unifying traces from benchmark to physical robot; reproduces released checkpoints on six benchmarks and quantifies sim-to-real gap for whole-body controllers under shared trace identity.", "relevance": 3, "relevance_reason": "Reproducibility infrastructure spanning sim and real.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2608.22067", "title": "DELE-w0.5: Inferring Action from Future Latent State for Robotic Manipulation", "authors": ["Fenghao Lei", "Zhixiong Huang", "Long Yang", "Jiabao Chen", "Peilin Huang", "Han Fu", "Zhuo Li", "Xiaoxue Ren", "Jia-Bao Chen", "Pei-Lin Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.22067", "pdf_url": "https://arxiv.org/pdf/2608.22067", "code_url": "", "clusters": ["C8"], "keywords": ["DELE-w0.5", "world-action model", "no video generation", "low latency"], "model_class": "world-action model", "techniques": ["future latent state prediction", "latent future-state prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot (640 trials)", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Infers actions from a compact predicted future latent state instead of generating dense future video, claiming cheaper training and low-latency inference.", "relevance": 3, "relevance_reason": "Efficiency-motivated WAM design (no numbers on speed)", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2405.12213, 2411.19650"}
{"id": "2605.31286", "title": "DeMaVLA: A Vision-Language-Action Foundation Model for Generalizable Deformable Manipulation", "authors": ["Taiyi Su", "Jian Zhu", "Tianjian Wang", "Youzhang He", "Zi-Tai Huang", "Jian-Jun Zhang", "Chong Ma", "Han-Yang Wang", "Tian-jiao Zhang", "Munan Yin", "Wei-Hao Ding", "Yi Xu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.31286", "pdf_url": "https://arxiv.org/pdf/2605.31286", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["deformable manipulation"], "model_class": "VLA", "techniques": ["layer pruning of action expert", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "DeMaVLA builds its action expert by pruning every other transformer layer to reduce cost, and trains on large real folding data with DAgger.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2501.09747"}
{"id": "2606.03847", "title": "Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot Policies", "authors": ["Xiangdong Feng", "Yuxuan Cheng", "Chen Shi", "Boyao Han", "Yuxuan Yan", "Yitong Hong", "Zhuotao Tian", "Li Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.03847", "pdf_url": "https://arxiv.org/pdf/2606.03847", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["DVAC", "denoising variance", "adaptive chunking", "replanning"], "model_class": "diffusion/flow policy", "techniques": ["denoising-variance signal", "test-time horizon"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboTwin", "CALVIN", "real"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Executes only the low-variance prefix of a flow chunk (variance of clean-action estimates across final denoising steps) and replans before uncertain actions are committed.", "relevance": 3, "relevance_reason": "Uncertainty-driven commitment of open-loop actions.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"action chunk\" execution horizon adaptive"}
{"id": "seeed-gr00t-n17-tensorrt-agx-orin", "title": "Deploy Full-Weight GR00T N1.7 with TensorRT on JetPack 7.2 and Jetson AGX Orin", "authors": ["Seeed Studio"], "year": 2026, "venue": "Seeed Studio wiki", "type": "blog", "url": "https://wiki.seeedstudio.com/deploy_full_weight_gr00t_n1.7_tensorrt_jetpack7.2_agx_orin/", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["Jetson AGX Orin", "TensorRT", "GR00T"], "model_class": "VLA", "techniques": ["TensorRT"], "base_model": "GR00T N1.7", "params": "", "hardware": "Jetson AGX Orin", "reported_metric": "0.2755 s warm inference per 16-action chunk (~3.63 chunks/s)", "metric_measures": "open-loop chunk rate", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Step-by-step guide that builds TensorRT engines for all seven GR00T N1.7 components on Jetson AGX Orin (JetPack 7.2) and validates offline action generation from a LeRobot dataset.", "relevance": 3, "relevance_reason": "verified 2026-09-30 by orchestrator: page fetched; reports 0.2755 s warm inference per 16-action chunk (batch 1) on AGX Orin, offline (no robot in the loop).", "foundational_technique": false, "verified": true, "found_via": "web: Jetson Thor GR00T N1 deployment TensorRT"}
{"id": "2603.11400", "title": "Deployment-Time Reliability of Learned Robot Policies", "authors": ["Christopher Agia"], "year": 2026, "venue": "Stanford PhD dissertation 2026", "type": "tech-report", "url": "https://arxiv.org/abs/2603.11400", "pdf_url": "https://arxiv.org/pdf/2603.11400", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["deployment-time reliability", "runtime monitoring", "influence functions"], "model_class": "other", "techniques": ["runtime monitoring", "data attribution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Dissertation collecting runtime monitoring, data-attribution and deployment-time mechanisms to improve reliability of learned manipulation policies.", "relevance": 3, "relevance_reason": "Consolidated reference on runtime monitoring.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: runtime monitoring robot policy"}
{"id": "2608.07361", "title": "Depth-Wise Probing and Pruning of the Planning Token in a Driving Vision-Language-Action Model", "authors": ["Harisankar Babu", "Benjamin Coors", "Christopher Lang", "Hendrik Berkemeyer", "Tamim Asfour", "Simon Foell"], "year": 2026, "venue": "ECCV 2026 Workshop (DriveX)", "type": "paper", "url": "https://arxiv.org/abs/2608.07361", "pdf_url": "https://arxiv.org/pdf/2608.07361", "code_url": "", "clusters": ["C2"], "keywords": ["logit lens", "planning token", "depth pruning", "driving VLA"], "model_class": "VLA", "techniques": ["layer pruning", "probing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Probes a driving VLA's planning token across depth and finds intent decodable early while trajectory quality needs depth, informing how far layers can be pruned.", "relevance": 3, "relevance_reason": "Depth-pruning analysis for driving VLAs.", "foundational_technique": false, "verified": true, "found_via": "search: pruning vision-language-action"}
{"id": "2512.01467", "title": "Differentiable Weightless Controllers: Learning Logic Circuits for Continuous Control", "authors": ["Fabian Kresse", "Christoph H. Lampert"], "year": 2026, "venue": "ICML 2026 (accepted)", "type": "paper", "url": "https://arxiv.org/abs/2512.01467", "pdf_url": "https://arxiv.org/pdf/2512.01467", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["logic circuits", "FPGA", "weightless networks", "nanojoule"], "model_class": "other", "techniques": ["logic-circuit policies", "extreme low-precision"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MuJoCo"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns control policies as differentiable logic circuits that compile to FPGA with single-cycle latency, competitive with full-precision or quantized nets.", "relevance": 3, "relevance_reason": "Extreme-compression alternative to binary networks for control.", "foundational_technique": false, "verified": true, "found_via": "search: quantization humanoid policy"}
{"id": "2609.28161", "title": "Dissecting Advantage-Guided Post-Training for Vision-Language-Action Policies", "authors": ["Jiahang Cao", "Hanye Zhao", "Hang Lai", "Shenyu Zhang", "Xiaoshen Han", "Xinghang Li", "Futeng Liu", "Wanli Peng", "Heyun Wang", "Yunhong Wang", "Jason Li", "Yong Yu", "Weinan Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28161", "pdf_url": "https://arxiv.org/pdf/2609.28161", "code_url": "", "clusters": ["C3"], "keywords": ["advantage-guided post-training", "offline evaluation"], "model_class": "VLA", "techniques": ["advantage-weighted post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Controlled dissection of advantage-guided VLA post-training design choices with cheap staged offline evaluation.", "relevance": 3, "relevance_reason": "Post-training recipe study minimizing robot evals.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.28107", "title": "Distillation for Efficient Multitask Manipulation Policies via Conditional Flow Matching", "authors": ["Shreya Deshmukh", "Imen Mahdi", "Nick Heppert", "Abhinav Valada"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28107", "pdf_url": "https://arxiv.org/pdf/2609.28107", "code_url": "", "clusters": ["C2"], "keywords": ["multi-task distillation", "velocity field", "CFM"], "model_class": "diffusion/flow policy", "techniques": ["policy distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RLBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills several single-task flow-matching experts into one fixed-size multi-task policy by matching their velocity fields.", "relevance": 3, "relevance_reason": "Multi-expert consolidation at fixed model size.", "foundational_technique": false, "verified": true, "found_via": "search: knowledge distillation robot manipulation policy"}
{"id": "2609.29669", "title": "Do World Models Make Better Robots? A Survey of Evaluation Benchmarks for Predictive Embodied Intelligence", "authors": ["Gaytri Jena", "Kapil Wanaskar", "Vinija Jain", "Aman Chadha", "Vasu Sharma", "Amitava Das"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.29669", "pdf_url": "https://arxiv.org/pdf/2609.29669", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["survey", "evaluation"], "model_class": "world model", "techniques": ["survey", "evaluation benchmarks"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Surveys evaluation benchmarks for predictive embodied models and asks whether world modeling yields measurable closed-loop gains over direct VLA policies.", "relevance": 3, "relevance_reason": "relevant: benchmarking methodology for world models vs VLAs", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2609.01453", "title": "Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds", "authors": ["Clinton Enwerem", "John S. Baras", "C. Belta"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.01453", "pdf_url": "https://arxiv.org/pdf/2609.01453", "code_url": "https://github.com/coenwerem/parcelstow", "clusters": ["C4", "C7"], "keywords": ["temporal robustness", "execution speed"], "model_class": "other", "techniques": ["execution-speed robustness evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Evaluates how imitation-learned dexterous policies hold up as task execution is sped up, comparing learner and expert under matched conditions and speedup factors.", "relevance": 3, "relevance_reason": "relevant: robustness to timing / execution speed", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.07089", "title": "Dreaming when Necessary: Advancing World Action Models with Adaptive Multi-Modal Reasoning", "authors": ["Yinzhou Tang", "Jingbo Xu", "Yu Shang", "Zihao Song", "Chen Gao", "Wei Wu", "Yong Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.07089", "pdf_url": "https://arxiv.org/pdf/2606.07089", "code_url": "", "clusters": ["C8"], "keywords": ["adaptive reasoning", "dynamic router", "dream when necessary"], "model_class": "world-action model", "techniques": ["router triggering textual or visual reasoning on demand"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "WAM with a lightweight router that invokes visual imagination only for fine-grained phases and textual reasoning at task transitions.", "relevance": 3, "relevance_reason": "Conditional imagination compute.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.23863", "title": "DreamLedger: Where to Refuse World-Model Imagination Using Execution-Settled Credit", "authors": ["Xianyao Li", "Ruitong Tian", "Rui Min", "Fang Xu", "Eric Jing Du"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.23863", "pdf_url": "https://arxiv.org/pdf/2608.23863", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["refusal of imagination", "execution-settled credit", "reliability ledger"], "model_class": "world model", "techniques": ["registering predictions as claims settled by outcomes", "history-based credit estimation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["navigation", "manipulation", "Franka replay"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Keeps a ledger of past world-model predictions and their executed outcomes to decide when a new prediction should not be trusted, lowering failures per consumed prediction.", "relevance": 3, "relevance_reason": "Monitoring trustworthiness of world-model imagination.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.16860", "title": "DreamPlan: Efficient Reinforcement Fine-Tuning of Vision-Language Planners via Video World Models", "authors": ["E. Jia", "Wei-Duo Yuan", "Tianheng Shi", "V. Guizilini", "Jiageng Mao", "Yue Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.16860", "pdf_url": "https://arxiv.org/pdf/2603.16860", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["RL fine-tuning", "video world model"], "model_class": "other", "techniques": ["RL in video world model", "ORPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "DreamPlan trains an action-conditioned video world model from zero-shot VLM exploration data and then RL-fine-tunes the VLM planner entirely inside that model, avoiding costly physical rollouts.", "relevance": 3, "relevance_reason": "relevant: sample-efficient adaptation via world-model imagination", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2306.03310"}
{"id": "2605.07727", "title": "Drifting Field Policy: A One-Step Generative Policy via Wasserstein Gradient Flow", "authors": ["Juil Koo", "Mingue Park", "Jiwon Choi", "Yun-Hong Min", "Minhyuk Sung"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.07727", "pdf_url": "https://arxiv.org/pdf/2605.07727", "code_url": "", "clusters": ["C1"], "keywords": ["Wasserstein gradient flow", "RL"], "model_class": "diffusion/flow policy", "techniques": ["one-step generation", "drifting model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Robomimic", "OGBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "One-step drifting-based generative policy trained as a Wasserstein gradient flow toward high-value actions.", "relevance": 3, "relevance_reason": "One-step policy inference", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2608.29583", "title": "Drive the Thoughts: Runtime Monitoring of VLA Reasoning-Trajectory Consistency", "authors": ["Tian Yu", "Lu Feng", "Sebastian Elbaum"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29583", "pdf_url": "https://arxiv.org/pdf/2608.29583", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["driving VLA", "chain-of-thought monitoring", "Alpamayo", "runtime monitoring", "reasoning-trajectory consistency"], "model_class": "VLA", "techniques": ["CoT-trajectory consistency monitoring", "runtime consistency monitor"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DriveAlignBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Assesses whether chain-of-thought produced by an autonomous-driving VLA can serve as a runtime specification to catch inconsistent or unsafe trajectories.", "relevance": 3, "relevance_reason": "Runtime monitoring of VLA outputs (driving).", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action; venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2604.01765", "title": "DriveDreamer-Policy: A Geometry-Grounded World-Action Model for Unified Generation and Planning", "authors": ["Yang Zhou", "Xiaofeng Wang", "Hao Shao", "Letian Wang", "Guosheng Zhao", "Jiangnan Shao", "Jiagang Zhu", "Tingdong Yu", "Zheng Zhu", "Guan Huang", "Steven L. Waslander"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.01765", "pdf_url": "https://arxiv.org/pdf/2604.01765", "code_url": "", "clusters": ["C8"], "keywords": ["driving"], "model_class": "world-action model", "techniques": ["world-action model", "geometry grounding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Geometry-grounded driving world-action model that jointly generates future world and plans with attention to latency.", "relevance": 3, "relevance_reason": "Driving WAM with deployment focus.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2605.14598", "title": "DSSP: Diffusion State Space Policy with Full-History Encoding", "authors": ["Zhiyuan Guan", "Jianshu Hu", "Han Fang", "Yunpeng Jiang", "Yize Huang", "Shujia Li", "Xiao Li", "Yutong Ban"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.14598", "pdf_url": "https://arxiv.org/pdf/2605.14598", "code_url": "", "clusters": ["C1"], "keywords": ["SSM diffusion policy", "full-history"], "model_class": "diffusion/flow policy", "techniques": ["state space model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Diffusion policy with an SSM history encoder and SSM denoiser that conditions on full history with small model size and memory.", "relevance": 3, "relevance_reason": "SSM diffusion policy.", "foundational_technique": false, "verified": true, "found_via": "search query: state space model imitation learning"}
{"id": "2603.22280", "title": "DualCoT-VLA: Visual-Linguistic Chain of Thought via Parallel Reasoning for Vision-Language-Action Models", "authors": ["Zhide Zhong", "Junfeng Li", "Junjie He", "Haodong Yan", "Xin Gong", "Guanyi Zhao", "Yingjie Cai", "Jiantao Gao", "Xu Yan", "Bingbing Liu", "Yingcong Chen", "Liuqing Yang", "Haoang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.22280", "pdf_url": "https://arxiv.org/pdf/2603.22280", "code_url": "", "clusters": ["C1"], "keywords": ["parallel CoT", "visual and linguistic reasoning", "query tokens"], "model_class": "VLA", "techniques": ["parallel latent reasoning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces autoregressive CoT with two sets of learnable query tokens for visual and linguistic reasoning computed in a single forward pass.", "relevance": 3, "relevance_reason": "Low-latency reasoning VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: reasoning VLA latency"}
{"id": "2609.10506", "title": "DUET-DINO: Simultaneous Cross-View World Modeling for Latent Planning in Robot Manipulation", "authors": ["Nisarga Nilavadi", "Ralf Römer", "Moritz Reuss", "Michael Krawez", "Tobias Jülg", "Angela P. Schoellig", "Rudolf Lioutikov", "Wolfram Burgard"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.10506", "pdf_url": "https://arxiv.org/pdf/2609.10506", "code_url": "", "clusters": ["C8"], "keywords": ["cross-view latent world model", "7-DoF latent planning", "DINOv3"], "model_class": "world model", "techniques": ["joint side+wrist view latent prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DROID", "RoboArena", "real reach/grasp"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Joint side- and wrist-camera latent world model that makes zero-shot latent planning reliable for full 7-DoF control, and finds V-JEPA 2 wrist predictions underestimate fine action effects.", "relevance": 3, "relevance_reason": "Latent planning on real robots (V-JEPA 2-AC follow-up).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2604.22152", "title": "dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model", "authors": ["Yaxuan Li", "Zhongyi Zhou", "Ye Chen", "Yaokai Xue", "Yichen Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.22152", "pdf_url": "https://arxiv.org/pdf/2604.22152", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["policy evaluation", "world model"], "model_class": "world model", "techniques": ["discrete diffusion world model", "policy evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "dWorldEval uses a discrete-diffusion world model over unified vision-language-action tokens as a scalable proxy for evaluating robot policies across many tasks and environments.", "relevance": 3, "relevance_reason": "relevant: benchmarking methodology via world-model evaluation", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310; cited by 2606.19531"}
{"id": "2606.11408", "title": "Dynamic Execution Horizon Prediction for Chunk-based Robot Policies", "authors": ["Yuchi Zhao", "Miroslav Bogdanovic", "Arjun Sohal", "Liyu Tao", "Kourosh Darvish", "Alán Aspuru-Guzik", "Florian Shkurti", "Animesh Garg"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.11408", "pdf_url": "https://arxiv.org/pdf/2606.11408", "code_url": "", "clusters": ["C2", "C6", "C7"], "keywords": ["execution horizon", "online RL", "black-box chunk policy"], "model_class": "diffusion/flow policy", "techniques": ["RL-trained horizon predictor"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains a lightweight branch with online RL to choose how many actions of each chunk to execute, keeping the chunk policy frozen; predicts short horizons in fine-grained phases.", "relevance": 3, "relevance_reason": "Adaptive open-loop commitment.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"open-loop\" action chunk reactive closed-loop"}
{"id": "2601.22153", "title": "DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation", "authors": ["Xie, Haozhe", "Wen, Beichen", "Zheng, Jiarui", "Chen, Zhaoxi", "Hong, Fangzhou", "Diao, Haiwen", "Liu, Ziwei"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.22153", "pdf_url": "https://arxiv.org/pdf/2601.22153", "code_url": "", "clusters": ["C2", "C4"], "keywords": ["DynamicVLA", "DOM benchmark", "dynamic objects", "continuous inference"], "model_class": "VLA", "techniques": ["continuous inference", "latent-aware action streaming", "compact conv encoder"], "base_model": "", "params": "0.4B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DOM", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Compact VLA with overlapped inference/execution plus the Dynamic Object Manipulation benchmark (200K synthetic episodes, 2K real) for moving-object tasks.", "relevance": 3, "relevance_reason": "Dynamic benchmark where latency matters.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: benchmark dynamic manipulation moving objects VLA reaction latency"}
{"id": "2606.27268", "title": "E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation", "authors": ["Wenming Ye", "Pei-Yan Li", "Ting-Yu Yuan", "Yuan Xu", "Xiang Wu", "Chao-Yang Zhao", "Jing Liu", "Nianfeng Liu", "Yan Huang", "Liang Wang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.27268", "pdf_url": "https://arxiv.org/pdf/2606.27268", "code_url": "", "clusters": ["C1"], "keywords": ["E-TTS", "closed-loop refinement"], "model_class": "VLA", "techniques": ["test-time scaling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Closed-loop test-time scaling framework for VLAs that feeds back evaluations into sampling for better inference efficiency.", "relevance": 3, "relevance_reason": "Compute-efficient test-time scaling.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310; cited by 2609.09250"}
{"id": "2609.27547", "title": "EBRL: Asynchronous Embodied RL by Multi-Grained Resource Management", "authors": ["Liang Mi", "Wei-Jun Wang", "Bo-Wen Gao", "Tian-Ze Yu", "Zi-Xu Hao", "Han Xiao", "Xin Ding", "Ming-Zhe Huang", "Xin He", "Lu Shi", "Hao Wu", "Hai-Peng Dai", "Gui-Hai Chen", "Yun-Xin Liu", "Ting Cao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.27547", "pdf_url": "https://arxiv.org/pdf/2609.27547", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["embodied RL system", "throughput"], "model_class": "other", "techniques": ["asynchronous pipelining", "resource management"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.30-3.47x end-to-end rollout throughput; 2.5x training convergence", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Asynchronous embodied-RL training system that overlaps rollout and training and pools CPU/GPU resources to reduce waste.", "relevance": 3, "relevance_reason": "Systems efficiency for RL fine-tuning of embodied policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "s2-ecg-rtc-2026", "title": "ECG-RTC: Executable-Cursor-Guided Local Real-Time Chunking for Smooth Execution-Side Handoff of Action-Chunking Flow Policies", "authors": ["Chu-Tian Qian", "Hui Xie", "Wen-Kun Wang"], "year": 2026, "venue": "ISCER 2026", "type": "paper", "url": "https://www.semanticscholar.org/paper/79ac1df3e8bc7230dee71f992118efcee8f2852d", "pdf_url": "", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["RTC extension", "handoff seam", "execution-side correction"], "model_class": "diffusion/flow policy", "techniques": ["handoff point selection", "local continuity correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Client-side extension to RTC that selects the handoff point via an executable cursor and corrects local continuity around the seam when a new chunk arrives.", "relevance": 3, "relevance_reason": "RTC execution-side refinement.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "10.1109/tccn.2026.3658762", "title": "Edge General Intelligence Through World Models, Large Language Models, and Agentic AI: Fundamentals, Solutions, and Challenges", "authors": ["Chang-Yuan Zhao", "Guang-Yuan Liu", "Rui-Chen Zhang", "Yin-Qiu Liu", "Jia-Cheng Wang", "Jiawen Kang", "D. Niyato", "Zan Li", "Xuemin Shen", "Zhu Han", "Su-Mei Sun", "Chau Yuen", "Dong In Kim"], "year": 2026, "venue": "IEEE Transactions on Cognitive Communications and Networking", "type": "paper", "url": "https://doi.org/10.1109/TCCN.2026.3658762", "pdf_url": "", "code_url": "", "clusters": ["C6", "C8"], "keywords": ["edge intelligence", "survey", "world models"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey positioning world models, LLMs and agentic AI for edge systems under latency, energy and privacy constraints.", "relevance": 3, "relevance_reason": "Edge deployment survey involving world models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2501.03575, 2506.01392"}
{"id": "2606.00552", "title": "Edge-Based QoS-Aware Adaptive Task Placement: A Closed-Loop Control in Multi-Robot Systems", "authors": ["Thien Tran", "Jonathan Kua", "Thuong Hoang", "Minh Tran", "Honghao Lyu", "Jiong Jin"], "year": 2026, "venue": "IEEE INDIN 2026", "type": "paper", "url": "https://arxiv.org/abs/2606.00552", "pdf_url": "https://arxiv.org/pdf/2606.00552", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["adaptive task placement", "multi-robot edge", "tail latency", "visual servoing testbed"], "model_class": "other", "techniques": ["QoS-aware adaptive placement"], "base_model": "", "params": "", "hardware": "Raspberry Pi edge nodes", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Testbed showing static edge offloading of a camera-to-manipulator pipeline amplifies tail latency and deadline misses, while QoS-aware adaptive placement keeps them low.", "relevance": 3, "relevance_reason": "Empirical tail-latency evidence for adaptive offloading.", "foundational_technique": false, "verified": true, "found_via": "snowball from 1902.05703"}
{"id": "2608.17774", "title": "Edge-Native Embodied Intelligence for Action-Aware Wireless Edge Networks", "authors": ["Yiru Wang", "Chuanao Jiang", "Jiahui Cui", "Zide Fan", "Lei Wang", "Zehui Xiong", "Dong In Kim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.17774", "pdf_url": "https://arxiv.org/pdf/2608.17774", "code_url": "", "clusters": ["C6"], "keywords": ["edge-native embodied intelligence", "6G", "confidence-aware assistance", "embodied federated learning"], "model_class": "other", "techniques": ["confidence-aware edge offloading", "edge generative adaptation", "federated learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Framework article proposing a 6G edge-embodiment loop where local agents request edge assistance when action-conditioned confidence drops, and physical experience feeds federated edge-model updates.", "relevance": 3, "relevance_reason": "Confidence-triggered edge offloading for embodied agents.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "10.1109/lra.2026.3682534", "title": "Efficient Continual Imitation Learning With Online Meta-Adapters", "authors": ["Ruiqi Zhu", "Endong Sun", "Guanhe Huang", "O. Çeliktutan"], "year": 2026, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://doi.org/10.1109/LRA.2026.3682534", "pdf_url": "", "code_url": "", "clusters": ["C3"], "keywords": ["continual imitation learning", "PEFT"], "model_class": "other", "techniques": ["adapters", "meta-learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Meta-learned adapters for continual imitation learning that transfer priors across tasks to speed adaptation.", "relevance": 3, "relevance_reason": "Parameter-efficient continual adaptation", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2603.23227", "title": "Efficient Hybrid SE(3)-Equivariant Visuomotor Flow Policy via Spherical Harmonics for Robot Manipulation", "authors": ["Qinglun Zhang", "Shen Cheng", "Tian Dan", "Haoqiang Fan", "Guang-Hui Liu", "Shuai-Cheng Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.23227", "pdf_url": "https://arxiv.org/pdf/2603.23227", "code_url": "", "clusters": ["C3", "C9"], "keywords": ["equivariant flow policy", "rectified flow", "spherical harmonics"], "model_class": "diffusion/flow policy", "techniques": ["rectified flow", "efficient equivariance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "E3Flow combines fast rectified-flow sampling with efficient SE(3)-equivariant features for visuomotor policies.", "relevance": 3, "relevance_reason": "Fast-sampling equivariant policy.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2608.09762", "title": "Efficient Real-World Online Reinforcement Learning for Robot Manipulation via Centralized Training and Critic Decomposition", "authors": ["Changhao Li", "Yifang Zhang", "Heng Zhang", "Davide Torielli", "Damiano Gasperini", "Arturo Laurenzi", "L. Muratore", "Arash Ajoudani", "Nikos G. Tsagarakis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09762", "pdf_url": "https://arxiv.org/pdf/2608.09762", "code_url": "", "clusters": ["C3"], "keywords": ["real-world RL", "sample efficiency"], "model_class": "other", "techniques": ["CTDE", "critic decomposition"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Improves sample efficiency of real-world online RL for manipulation with centralized training of multiple actors sharing a decomposed task/grasp critic.", "relevance": 3, "relevance_reason": "relevant: sample-efficient real-world RL", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2501.16664, 2505.18719"}
{"id": "2605.08799", "title": "ElasticFlow: One-Step Physics-Consistent Policy with Elastic Time Horizons for Language-Guided Manipulation", "authors": ["Kewei Chen", "Yayu Long", "Shuai Li", "Mingsheng Shang"], "year": 2026, "venue": "Findings of ACL 2026", "type": "paper", "url": "https://arxiv.org/abs/2605.08799", "pdf_url": "https://arxiv.org/pdf/2605.08799", "code_url": "", "clusters": ["C1"], "keywords": ["mean flow", "one-step", "elastic horizons"], "model_class": "VLA", "techniques": ["mean flow", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1-NFE, ~71 Hz", "metric_measures": "throughput", "benchmarks": ["LIBERO", "CALVIN", "RoboTwin"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distillation-free one-step policy based on average-velocity modeling with horizon encoding for varying control granularity.", "relevance": 3, "relevance_reason": "One-step language-conditioned policy.", "foundational_technique": false, "verified": true, "found_via": "search query: consistency policy"}
{"id": "2605.04709", "title": "ELVIS: Ensemble-Calibrated Latent Imagination for Long-Horizon Visual MPC", "authors": ["Yurui Du", "Pinhao Song", "Yutong Hu", "Renaud Detry"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.04709", "pdf_url": "https://arxiv.org/pdf/2605.04709", "code_url": "", "clusters": ["C8"], "keywords": ["Gaussian-mixture MPPI", "ensemble critics", "long-horizon latent MPC"], "model_class": "world model", "techniques": ["RSSM latent planning", "GMM-MPPI", "UCB-gated lambda-return"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Makes long-horizon Dreamer-style latent MPC reliable with multi-modal MPPI and uncertainty-gated returns that limit compounding model error.", "relevance": 3, "relevance_reason": "Uncertainty-aware latent planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "github-laelazorana-embodied-efficiency", "title": "embodied-efficiency: measuring what makes a VLA fast enough to run on the robot", "authors": ["LaelaZorana"], "year": 2026, "venue": "GitHub", "type": "repo", "url": "https://github.com/LaelaZorana/embodied-efficiency", "pdf_url": "", "code_url": "https://github.com/LaelaZorana/embodied-efficiency", "clusters": ["C1", "C4", "C5"], "keywords": ["CUDA graphs", "weight-only quantization", "batch-1 latency", "speculative decoding", "runtime supervisor"], "model_class": "VLA", "techniques": ["CUDA graph capture", "INT8/INT4 weight-only quantization", "action chunking", "speculative decoding"], "base_model": "", "params": "", "hardware": "NVIDIA T4, L4", "reported_metric": "CUDA-graph sampler 5.9x over eager (4.82 to 0.82 ms/step on T4); torchao/Marlin int4 1.2-1.6x slower than bf16 at batch 1 on L4", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Independent measurement repo showing CUDA-graph capture gives large batch-1 speedups while weight-only low-bit quantization gave no batch-1 latency gain (even slowdowns) on L4.", "relevance": 3, "relevance_reason": "Practitioner evidence that technique benefits depend on hardware/kernels; not peer-reviewed.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: embodied-efficiency-bench"}
{"id": "2607.20771", "title": "Emergent Compositional Skills in Mixture-of-Experts VLAs", "authors": ["Shlok Shah", "Rhiaan Jhaveri", "Tharun Kumar Tiruppali Kalidoss", "Chirayu Nimonkar", "Ishaan Javali", "Dhruv Shah", "RH Jhaveri"], "year": 2026, "venue": "ICML 2026 Workshop on Compositional Learning", "type": "preprint", "url": "https://arxiv.org/abs/2607.20771", "pdf_url": "https://arxiv.org/pdf/2607.20771", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["mixture of experts", "action head", "compositional skills", "mixture-of-experts", "compositional primitives"], "model_class": "VLA", "techniques": ["mixture of experts", "mixture-of-experts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains a VLA with a simple MoE action head and finds experts specialize into reusable low-level primitives while matching a monolithic head.", "relevance": 3, "relevance_reason": "MoE for VLAs is a CFP topic, though focus is interpretability not speed.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA mixture-of-experts; author sweep: Dhruv Shah"}
{"id": "2603.16218", "title": "Enabling Dynamic Tracking in Vision-Language-Action Models via Time-Discrete and Time-Continuous Velocity Feedforward", "authors": ["Johannes Hechtl", "Philipp S. Schmitt", "G. von Wichert", "Wolfram Burgard"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.16218", "pdf_url": "https://arxiv.org/pdf/2603.16218", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["dynamic tracking", "compliance"], "model_class": "VLA", "techniques": ["velocity feedforward", "B-spline actions"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Extracts velocity feedforward from chunked VLA outputs to enable compliant, high-frequency tracking on industrial robots.", "relevance": 3, "relevance_reason": "Addresses low-frequency VLA output vs high-rate control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2405.05941"}
{"id": "2603.25766", "title": "ETA-VLA: Efficient Token Adaptation via Temporal Fusion and Intra-LLM Sparsification for Vision-Language-Action Models", "authors": ["Yiru Wang", "Anqing Jiang", "Shuo Wang", "Yuwen Heng", "Zichong Gu", "Hao Sun"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.25766", "pdf_url": "https://arxiv.org/pdf/2603.25766", "code_url": "", "clusters": ["C2"], "keywords": ["ETA-VLA"], "model_class": "VLA", "techniques": ["token pruning", "temporal fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "85% visual tokens pruned, 61% fewer FLOPs, 94% accuracy retained", "metric_measures": "other", "benchmarks": ["NAVSIM v2"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA with temporal fusion and intra-LLM sparsification that prunes 85% of visual tokens.", "relevance": 3, "relevance_reason": "Visual token pruning in driving VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2607.02646", "title": "EVA-Client: A Unified Data Collection, Inference, and Deployment Framework for Embodied Policies on Real Robots", "authors": ["He Yang", "Yang Yi", "Liyao Wang", "Lin-Qing Zhong", "Donglin Yang", "Rui-Pu Wu", "Zitong Bai", "Fengjiao Chen", "Manyuan Zhang", "Lin-Jiang Huang", "Si Liu"], "year": 2026, "venue": "arXiv", "type": "repo", "url": "https://arxiv.org/abs/2607.02646", "pdf_url": "https://arxiv.org/pdf/2607.02646", "code_url": "", "clusters": ["C4", "C6", "C7"], "keywords": ["deployment", "async inference", "RTC"], "model_class": "other", "techniques": ["deployment framework", "asynchronous inference", "real-time chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "EVA-Client is an open-source layer between a policy server and robot hardware that unifies deployment, evaluation and data collection and exposes sync/async execution, temporal ensembling and Real-Time Chunking behind one config.", "relevance": 3, "relevance_reason": "relevant: real-robot deployment and real-time inference strategies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2601.18723", "title": "Eval-Actions: Fine-Grained Execution Quality Evaluation for Robotic Manipulation", "authors": ["Meng-Yuan Liu", "Ju-Yi Sheng", "Pei-Ming Li", "Zi-Yi Wang", "Tian Xu", "Tian-Tian Xu", "Hong Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.18723", "pdf_url": "https://arxiv.org/pdf/2601.18723", "code_url": "", "clusters": ["C4"], "keywords": ["evaluation", "execution quality"], "model_class": "other", "techniques": ["execution-quality evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Eval-Actions is a diagnostic methodology and real-robot benchmark that scores fine-grained execution quality of manipulation policies beyond binary success.", "relevance": 3, "relevance_reason": "relevant: evaluation methodology", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2303.04137; cited by 2609.03611"}
{"id": "2605.14950", "title": "Evo-Depth: A Lightweight Depth-Enhanced Vision-Language-Action Model", "authors": ["Tao Lin", "Yuxin Du", "Jiting Liu", "Nuobei Zhu", "Yunhe Li", "Yuqian Fu", "Yinxinyu Chen", "Hongyi Cai", "Zewei Ye", "Bing Cheng", "Kai Ye", "Yiran Mao", "Yilei Zhong", "MingKang Dong", "Junchi Yan", "Gen Li", "Bo Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.14950", "pdf_url": "https://arxiv.org/pdf/2605.14950", "code_url": "", "clusters": ["C1"], "keywords": ["lightweight VLA", "implicit depth"], "model_class": "VLA", "techniques": ["small backbone"], "base_model": "", "params": "0.9B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "0.9B VLA that adds compact implicit depth features from RGB, reporting lowest memory and highest inference speed among compared real-world baselines.", "relevance": 3, "relevance_reason": "Lightweight VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: lightweight VLA"}
{"id": "2607.20293", "title": "Evolving Cache Schedules for Fast Diffusion Policy Inference", "authors": ["Siying Wang", "Kangye Ji", "Di Wang", "Fei Cheng"], "year": 2026, "venue": "PRCV 2026", "type": "paper", "url": "https://arxiv.org/abs/2607.20293", "pdf_url": "https://arxiv.org/pdf/2607.20293", "code_url": "", "clusters": ["C1"], "keywords": ["cache schedule", "evolutionary search", "training-free"], "model_class": "diffusion/flow policy", "techniques": ["feature caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Searches offline for block-by-timestep cache-refresh schedules for diffusion policies, reusing redundant transformer activations without training.", "relevance": 3, "relevance_reason": "Feature caching for diffusion policies.", "foundational_technique": false, "verified": true, "found_via": "search query: efficient diffusion policy action chunk"}
{"id": "2606.20537", "title": "Execution-State Capsules: Graph-Bound Execution-State Checkpoint and Restore for Low-Latency, Small-Batch, On-Device Physical-AI Serving", "authors": ["Liang Su"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20537", "pdf_url": "https://arxiv.org/pdf/2606.20537", "code_url": "", "clusters": ["C5"], "keywords": ["FlashRT", "execution-state checkpoint", "CUDA graphs", "small-batch on-device serving"], "model_class": "other", "techniques": ["graph-bound execution-state capsules", "static buffers"], "base_model": "", "params": "", "hardware": "RTX 5090", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "FlashRT's author introduces checkpoint/restore of complete execution state (KV, recurrent, conv) for low-latency batch-1 on-device physical-AI serving including robot policies.", "relevance": 3, "relevance_reason": "Runtime design for small-batch on-device physical AI.", "foundational_technique": false, "verified": true, "found_via": "arXiv: Jetson AND (robot policy OR manipulation OR VLA)"}
{"id": "2607.12931", "title": "ExToken: Structured Exploration for Efficient Vision-Language-Action Reinforcement Fine-tuning", "authors": ["Yilun Kong", "Yunpeng Qing", "Guozheng Ma", "Haoyu Wang", "Li Shen", "Zhi Hou", "Dacheng Tao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.12931", "pdf_url": "https://arxiv.org/pdf/2607.12931", "code_url": "", "clusters": ["C3"], "keywords": ["exploration tokens", "sample efficiency"], "model_class": "VLA", "techniques": ["online RL", "behavior priors"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Conditions VLAs on discrete behavior tokens during rollouts to diversify exploration and improve RL sample efficiency.", "relevance": 3, "relevance_reason": "Sample-efficient RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2604.02408", "title": "F2F-AP: Flow-to-Future Asynchronous Policy for Real-time Dynamic Manipulation", "authors": ["Haoyu Wei", "Xiuwei Xu", "Ziyang Cheng", "Hang Yin", "Angyuan Ma", "Bingyao Yu", "Jie Zhou", "Jiwen Lu", "Hao-Yu Wei", "Xiu-Wei Xu", "Zi-Yang Cheng", "Ang-Yuan Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.02408", "pdf_url": "https://arxiv.org/pdf/2604.02408", "code_url": "", "clusters": ["C1", "C2", "C6", "C7"], "keywords": ["asynchronous policy", "object flow", "future observation synthesis", "dynamic manipulation", "F2F-AP"], "model_class": "diffusion/flow policy", "techniques": ["asynchronous inference", "future observation prediction", "predicted future observations", "predicted object flow", "contrastive alignment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Synthesizes future observations from predicted object flow so an asynchronous policy acts on where the scene will be, improving dynamic manipulation under latency.", "relevance": 3, "relevance_reason": "Latency compensation for async policies.", "foundational_technique": false, "verified": true, "found_via": "search query: asynchronous inference robot; snowball from 2506.07339; snowball: citations of 2506.07339 (S2)"}
{"id": "2607.28596", "title": "FA-RDP: A Frequency-Adaptive Reactive Diffusion Policy for Contact-Rich Manipulation", "authors": ["Lifeng Zhuo", "Wendi Chen", "Han Xue", "Shi-Run Tang", "Jun Lv", "Ce-Wu Lu", "Chuan Wen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.28596", "pdf_url": "https://arxiv.org/pdf/2607.28596", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["reactive control", "adaptive frequency"], "model_class": "diffusion/flow policy", "techniques": ["frequency-adaptive sampling", "consistency distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "FA-RDP switches between multi-step low-frequency sampling before contact and one-step high-frequency sampling after contact, with manifold consistency distillation for fast reactive visual-force control.", "relevance": 3, "relevance_reason": "relevant: adaptive inference frequency and one-step sampling", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.16697", "title": "FabriMAE I Trust Myself? Self-Evaluating VLA Action Generation with Markov Attention Entropy", "authors": ["Aniri", "Chen Yilin", "Jinhe Bi", "Junfei Guo", "Donglai Ran", "Xu Bian", "Zengjie Jin", "Yujun Wang", "Yijun Tian", "Volker Tresp", "Fei Shen", "Tat-Seng Chua", "Yunpu Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.16697", "pdf_url": "https://arxiv.org/pdf/2608.16697", "code_url": "", "clusters": ["C7"], "keywords": ["LIBERO-Reflect"], "model_class": "VLA", "techniques": ["attention entropy", "self-evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Reflect"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses internal visual attention entropy, formalized as a conditional generative Markov chain, as an architecture-aware reliability score for VLA actions; also used for test-time selection.", "relevance": 3, "relevance_reason": "Cheap internal self-evaluation signal for triggering interventions.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=100]"}
{"id": "2607.08575", "title": "FabriVLA: A Lightweight Vision-Language-Action Model with Conformal Action Chunk Uncertainty", "authors": ["Shiyuan Yang", "Borong Zhang", "Jizheng Zhang", "Zhijia Tao", "Junfei Guo", "Donglai Ran", "Xu Bian", "Qingbiao Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.08575", "pdf_url": "https://arxiv.org/pdf/2607.08575", "code_url": "", "clusters": ["C1", "C7", "C8"], "keywords": ["lightweight VLA", "conformal uncertainty", "layer fusion", "FabriVLA", "conformal action chunk calibration"], "model_class": "VLA", "techniques": ["small backbone", "flow matching action head", "joint conformal chunk calibration", "residual scale head"], "base_model": "", "params": "0.88B", "hardware": "", "reported_metric": "90.0% avg success on MT50; 3.3x tighter conformal radius", "metric_measures": "other", "benchmarks": ["Meta-World MT50", "LIBERO-Safety"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lightweight 0.88B VLA plus a post-hoc conformal method giving calibrated uncertainty sets over the executed action prefix from a single query.", "relevance": 3, "relevance_reason": "Calibrated chunk uncertainty for risk ranking.", "foundational_technique": false, "verified": true, "found_via": "search query: lightweight VLA; arXiv search: \"uncertainty\" \"vision-language-action\" calibration"}
{"id": "2609.34175", "title": "FailPatch: Failure Residual Patching for Vision-Language-Action Models", "authors": ["Peng Yu", "Jiacheng Wang", "Ziheng Zhang", "Xuchong Zhang", "Baoting Li", "Zhuoyuan Yu", "Yuxiang Chen", "Tiancai Wang", "Hongbin Sun"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34175", "pdf_url": "https://arxiv.org/pdf/2609.34175", "code_url": "", "clusters": ["C3"], "keywords": ["residual expert bank", "failure patching", "frozen VLA"], "model_class": "VLA", "techniques": ["residual experts", "frozen backbone"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Adds null-gated residual experts in the action hidden space of a frozen VLA, trained from success demos plus failure trajectories.", "relevance": 3, "relevance_reason": "Lightweight patching instead of full retraining.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "10.1109/lra.2026.3678124", "title": "Failure Detection With Zero-Shot Error Correction in Robotic Manipulation", "authors": ["Haojie Luo", "Xizhou Bu", "Hongbo Wang", "Shijie Guo", "Junxiu Liu", "Wei Li"], "year": 2026, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://doi.org/10.1109/LRA.2026.3678124", "pdf_url": "", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "error correction"], "model_class": "diffusion/flow policy", "techniques": ["lightweight monitor + VLM on demand"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.1x faster detection vs VLM-only", "metric_measures": "relative latency", "benchmarks": ["RLBench"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight dynamics model flags failures and triggers a VLM only when needed.", "relevance": 3, "relevance_reason": "Latency-aware failure detection.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2604.13788", "title": "Failure Identification in Imitation Learning Via Statistical and Semantic Filtering", "authors": ["Quentin Rolland", "Fabrice Mayran de Chamisso", "Jean-Baptiste Mouret"], "year": 2026, "venue": "ICRA 2026", "type": "preprint", "url": "https://arxiv.org/abs/2604.13788", "pdf_url": "https://arxiv.org/pdf/2604.13788", "code_url": "", "clusters": ["C7"], "keywords": ["failure identification", "imitation learning", "failure detection", "BotFails"], "model_class": "other", "techniques": ["anomaly detection", "conformal thresholds", "VLM filtering", "conformal prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Policy-independent failure detector using optimal-transport anomaly scores against nominal demonstrations, conformal spatio-temporal thresholds and VLM semantic filtering of benign anomalies.", "relevance": 3, "relevance_reason": "Calibrated runtime failure detection for IL policies.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:conformal AND (\"policy failure\" OR \"failure prediction\" OR \"failure detection\") AND robot*; snowball round 2: cited by 2606.21386, 2609.03611"}
{"id": "10.1109/cisr70722.2026.11680207", "title": "FAITH-Manip: A Benchmark for Offline-to-Closed-Loop Evaluation of Imitation-Learning Robotic Manipulation Policies", "authors": ["Qiming Bao", "Qi-Tong Cao", "Ran Zhao", "Han Dong", "Zhi-Sheng Lu"], "year": 2026, "venue": "2026 3rd International Conference on Intelligent Systems and Robotics (CISR)", "type": "paper", "url": "https://doi.org/10.1109/CISR70722.2026.11680207", "pdf_url": "", "code_url": "", "clusters": ["C4"], "keywords": ["offline metrics", "closed-loop evaluation", "benchmark"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Push-T", "ALOHA", "robomimic"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Benchmark measuring how faithfully offline proxies (validation loss, open-loop error) predict closed-loop success across DP, ACT and VLA policies.", "relevance": 3, "relevance_reason": "Benchmarking methodology for robot policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.19475", "title": "FASA: Feedback-Aware Sampling Adaptation for Efficient Diffusion-Based VLA Models", "authors": ["Yuchen Han", "Jianhan Wu", "Xiaoyang Qu", "Lingwei Kong", "Shiyi Li", "Jianzong Wang"], "year": 2026, "venue": "NAS 2026", "type": "paper", "url": "https://arxiv.org/abs/2609.19475", "pdf_url": "https://arxiv.org/pdf/2609.19475", "code_url": "", "clusters": ["C1"], "keywords": ["adaptive denoising steps", "feedback-aware"], "model_class": "VLA", "techniques": ["adaptive sampling steps"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.45x speedup", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts the number of diffusion sampling steps at runtime using visual and gripper-force feedback and proprioception.", "relevance": 3, "relevance_reason": "Dynamic inference for diffusion VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference"}
{"id": "10.1109/sii64115.2026.11404587", "title": "Fast Action Generation via Knowledge Distillation with Flow Matching for Social Navigation", "authors": ["Yuki Tomita", "Kohei Matsumoto", "Yuki Hyodo", "Kazuto Nakashima", "Ryo Kurazume"], "year": 2026, "venue": "IEEE/SICE SII 2026", "type": "paper", "url": "https://doi.org/10.1109/SII64115.2026.11404587", "pdf_url": "", "code_url": "", "clusters": ["C2"], "keywords": ["social navigation", "diffusion RL", "flow matching student"], "model_class": "diffusion/flow policy", "techniques": ["knowledge distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills a slow conditional diffusion navigation policy into a fast flow-matching-based generator for real-time social navigation.", "relevance": 3, "relevance_reason": "Diffusion-policy distillation for real-time mobile robots.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2405.07503"}
{"id": "2605.23163", "title": "Fast-dDrive: Efficient Block-Diffusion VLM for Autonomous Driving", "authors": ["Kewei Zhang", "Jin Wang", "Sensen Gao", "Chengyue Wu", "Yulong Cao", "Songyang Han", "Boris Ivanovic", "Langechuan Liu", "Marco Pavone", "Song Han", "Daquan Zhou", "Enze Xie"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.23163", "pdf_url": "https://arxiv.org/pdf/2605.23163", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["driving VLA"], "model_class": "VLA", "techniques": ["block diffusion", "speculative decoding", "KV cache reuse"], "base_model": "", "params": "", "hardware": "", "reported_metric": "12x throughput vs AR baseline (with SGLang)", "metric_measures": "throughput", "benchmarks": ["WOD-E2E", "nuScenes"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Block-diffusion driving VLA with causal ordering across semantic sections, scaffold speculative decoding and shared-prefix KV-cache forking for cheap test-time scaling.", "relevance": 3, "relevance_reason": "Throughput-oriented driving VLA decoding.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]"}
{"id": "2606.02307", "title": "FATE-VLA:Failue-aware test generation for vision-language-action models", "authors": ["Arusa Kanwal", "Pablo Valle", "Shaukat Ali", "Aitor Arrieta"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.02307", "pdf_url": "https://arxiv.org/pdf/2606.02307", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["test generation", "failure discovery"], "model_class": "VLA", "techniques": ["failure-aware test generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "FATE-VLA reframes VLA evaluation as active failure discovery, generating test scenes targeted at clustered failures that static random benchmarks miss.", "relevance": 3, "relevance_reason": "relevant: evaluation methodology / safety testing", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2607.29235", "title": "FBFM: A Training-Free Asynchronous Feedback Mechanism for Flow-Matching in World-Action Models Execution", "authors": ["Peize Li", "Ruimeng Zhang", "Ru Zhang", "Cong Huang", "Kai Chen", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.29235", "pdf_url": "https://arxiv.org/pdf/2607.29235", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["feedback flow matching", "world-action model", "intra-chunk re-grounding", "re-grounding", "training-free"], "model_class": "world-action model", "techniques": ["masked pseudoinverse velocity correction", "training-free", "masked pseudoinverse correction of velocity field", "cross-chunk observation feedback"], "base_model": "DreamZero, LingBot-VA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboTwin2.0", "RoboTwin 2.0", "real robot diagnostics"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Training-free correction inside flow-matching sampling that re-grounds a world-action model on the latest executed actions and observed frames, giving asynchronous feedback finer than chunk boundaries.", "relevance": 3, "relevance_reason": "Improves reliability of async WAM rollouts.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"action chunk\" asynchronous inference; search query"}
{"id": "2605.15705", "title": "Feedback World Model Enables Precise Guidance of Diffusion Policy", "authors": ["Tuo An", "Jindou Jia", "Gen Li", "Jingliang Li", "Chuhao Zhou", "Pengfei Liu", "Bofan Lyu", "Jiaqi Bai", "Xinying Guo", "Geng Li", "Jianfei Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.15705", "pdf_url": "https://arxiv.org/pdf/2605.15705", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["feedback world model"], "model_class": "world model", "techniques": ["latent observer", "action-aware guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus", "Robomimic", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Maintains a lightweight feedback state updated from observed outcomes to correct world-model predictions online, used for guiding diffusion policies.", "relevance": 3, "relevance_reason": "Closed-loop correction of world-model predictions during execution.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "2606.09311", "title": "FF-JEPA: Long-Horizon Planning in World Models with Latent Planners", "authors": ["Sergi Masip", "Jonathan Swinnen", "Yutong Hu", "Renaud Detry", "Tinne Tuytelaars"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09311", "pdf_url": "https://arxiv.org/pdf/2606.09311", "code_url": "", "clusters": ["C8"], "keywords": ["latent planner", "goal-free planning", "hierarchical JEPA"], "model_class": "world model", "techniques": ["action-free subgoal predictor + action-conditioned forward model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["PushT"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Adds an action-free latent subgoal predictor to a JEPA world model, replacing expensive long-horizon CEM with a sequence of short optimizations and removing the need for goal images.", "relevance": 3, "relevance_reason": "Cheaper long-horizon latent planning (preliminary).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2607.29596", "title": "FibVLA: An Efficient Temporal Vision-Language-Action Model with Fibonacci Sampling", "authors": ["Li Lin", "Wujun Xu", "Weiwei Meng", "Kaiwen Xia", "Kang Hao Cheong", "Shuai Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.29596", "pdf_url": "https://arxiv.org/pdf/2607.29596", "code_url": "", "clusters": ["C1"], "keywords": ["temporal context", "logarithmic sampling", "flow matching"], "model_class": "VLA", "techniques": ["history subsampling", "flow matching action expert"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Samples history frames and states at Fibonacci/log-spaced intervals so a VLA can use long temporal context without a proportional cost increase.", "relevance": 3, "relevance_reason": "Efficient long-context VLA design.", "foundational_technique": false, "verified": true, "found_via": "search query: vision-language-action efficient"}
{"id": "2609.18651", "title": "FIERCE: From Generalist Robot Policies to Fast Specialists via Progress-Failure Feedback", "authors": ["Run-Jia Tan", "Yuang Tu", "Yu-Jie Yan", "Lan Yu", "Xue-Song Tian", "Chen Lv"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18651", "pdf_url": "https://arxiv.org/pdf/2609.18651", "code_url": "https://github.com/ar-mine/FIERCE", "clusters": ["C2", "C3"], "keywords": ["generalist-to-specialist", "compact specialist"], "model_class": "other", "techniques": ["RL refinement", "progress-failure evaluator"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Refines compact specialist policies from a generalist initialization with RL driven by a learned progress/failure evaluator; only the specialist is deployed.", "relevance": 3, "relevance_reason": "Distilling generalist competence into cheap-to-deploy specialists.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2306.03310"}
{"id": "10.1109/ichora69329.2026.11537253", "title": "Fine-Tuning and Deployment of a Vision-Language-Action Based Robotic Foundation Model for Open-Source Robot Arms", "authors": ["Mücahit Emre Kabaoğlu", "M. H. B. Ucar"], "year": 2026, "venue": "2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA)", "type": "paper", "url": "https://doi.org/10.1109/ICHORA69329.2026.11537253", "pdf_url": "", "code_url": "", "clusters": ["C3", "C5"], "keywords": ["deployment", "ROS2", "low-cost arm"], "model_class": "VLA", "techniques": ["fine-tuning"], "base_model": "GR00T 1.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "End-to-end infrastructure for fine-tuning and deploying GR00T on a low-cost open-source arm.", "relevance": 3, "relevance_reason": "Low-cost deployment and efficient fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2503.14734"}
{"id": "aws-pi0-finetuning-sagemaker-hyperpod-blog-2026", "title": "Fine-Tuning π0 (Pi-Zero) for robotic manipulation on Amazon SageMaker HyperPod EKS", "authors": ["Amin Dashti", "Eric Saleh"], "year": 2026, "venue": "blog (AWS)", "type": "blog", "url": "https://aws.amazon.com/blogs/physical-ai/fine-tuning-%CF%800-pi-zero-for-robotic-manipulation-on-amazon-sagemaker-hyperpod-eks/", "pdf_url": "", "code_url": "", "clusters": ["C3", "C4", "C9"], "keywords": ["pi0 fine-tuning", "FSDP", "single ODE step", "SageMaker HyperPod"], "model_class": "VLA", "techniques": ["full fine-tuning with FSDP", "flow step reduction (1 ODE step)"], "base_model": "pi0", "params": "", "hardware": "8x H100 (ml.p5.48xlarge)", "reported_metric": "~199 ms inference with 1 ODE step (~half of 10-step); ~6 h fine-tuning on DROID/LIBERO-10", "metric_measures": "other", "benchmarks": ["DROID", "LIBERO-10 (offline MSE)"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "AWS walkthrough of full pi0 fine-tuning on 8 H100s, finding a single flow-integration step suffices after fine-tuning and halves inference latency.", "relevance": 3, "relevance_reason": "Cloud-provider technical post on VLA fine-tuning and flow-step reduction; uses GPUs rather than Trainium.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: AWS Neuron"}
{"id": "2608.12932", "title": "FlashDrive: Flash Vision-Language-Action Inference for Autonomous Driving", "authors": ["Zekai Li", "Yihao Liang", "Hong-Fei Zhang", "Jian Chen", "Yesheng Liang", "Zhijian Liu", "Hongfei Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.12932", "pdf_url": "https://arxiv.org/pdf/2608.12932", "code_url": "", "clusters": ["C1", "C2", "C5", "C9"], "keywords": ["driving VLA", "KV reuse", "speculative reasoning", "adaptive denoising", "algorithm-system co-design", "streaming KV reuse", "diffusion drafter", "adaptive step caching", "prefill", "visual encoding redundancy", "cascade of bottlenecks"], "model_class": "VLA", "techniques": ["KV-cache reuse", "adaptive denoising", "streaming KV-cache reuse", "speculative decoding with diffusion drafter", "adaptive flow-step caching", "frame-overlap reuse", "prefill acceleration", "decoding acceleration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Algorithm-system co-design (MIT Han Lab) attacking four cascaded bottlenecks of driving VLAs: frame-overlap encoding, prefill recompute, serial reasoning tokens and uniform denoising.", "relevance": 3, "relevance_reason": "Systems co-design for VLA real-time inference (driving).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.07530; arXiv: CUDA graph AND (robot OR VLA OR policy); venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2609.17008", "title": "FlexEE: Self-Speculative and KV-Compatible Early Exiting for Offloading-Aware LLM Inference", "authors": ["Qi-Hu Xie", "Zi-Wei Li", "Yi Kang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.17008", "pdf_url": "https://arxiv.org/pdf/2609.17008", "code_url": "", "clusters": ["C1", "C6"], "keywords": ["offloading", "LLM inference"], "model_class": "other", "techniques": ["early exit", "self-speculative decoding", "KV compatibility"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Early-exit framework compatible with KV caching and self-speculation for offloading-based LLM inference.", "relevance": 3, "relevance_reason": "Foundational early-exit technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2210.17323"}
{"id": "2609.35138", "title": "FlexiWorld: Learning and Planning via Flexible Action Chunks Across Multiple Time Scales", "authors": ["Shi-Du Ren", "Qi Gu", "Zheng-Hao Ni", "Jun Sun", "Jia-Qi Wang", "Damien Scieur", "Yun-Ze Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.35138", "pdf_url": "https://arxiv.org/pdf/2609.35138", "code_url": "", "clusters": ["C8"], "keywords": ["JEPA", "latent planning"], "model_class": "world model", "techniques": ["JEPA", "variable-length action chunks", "CEM planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~1.3x faster planning with longer chunks", "metric_measures": "relative latency", "benchmarks": ["PushT"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "FlexiWorld is a JEPA latent world model trained with variable-length action chunks and mixed goal spans; it can plan with longer chunks to speed up planning without retraining.", "relevance": 3, "relevance_reason": "relevant: efficient latent world-model planning", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2310.16828"}
{"id": "2606.06461", "title": "Flow-based Policy Adaptation without Policy Updates", "authors": ["Luzhe Sun", "Jingtian Ji", "Haoran Chen", "Jiawei Zhou", "Matthew R. Walter"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.06461", "pdf_url": "https://arxiv.org/pdf/2606.06461", "code_url": "", "clusters": ["C3"], "keywords": ["GLOVES", "action transport", "no policy updates"], "model_class": "VLA", "techniques": ["flow-based action correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Learned flow transports non-expert actions toward the expert distribution, adapting behavior without updating the base policy.", "relevance": 3, "relevance_reason": "Adaptation without base policy fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2602.22056", "title": "FlowCorrect: Efficient Interactive Correction of Generative Flow Policies for Robotic Manipulation", "authors": ["Edgar Welte", "Yitian Shi", "Rosa Wolf", "Maximillian Gilles", "Rania Rayyes"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.22056", "pdf_url": "https://arxiv.org/pdf/2602.22056", "code_url": "", "clusters": ["C3"], "keywords": ["interactive correction", "VR nudges", "deployment-time adaptation"], "model_class": "VLA", "techniques": ["local policy correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Adapts flow policies at deployment from sparse human pose nudges without retraining the backbone.", "relevance": 3, "relevance_reason": "Cheap deployment-time adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.08877", "title": "FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space", "authors": ["Michael Murray", "Daphne Chen", "Simran Bagaria", "Dean Fortier", "T. Hellebrekers", "Galen Mullins", "H. Gajarla", "Oier Mees", "Maya Cakmak", "A. Kolobov"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.08877", "pdf_url": "https://arxiv.org/pdf/2607.08877", "code_url": "", "clusters": ["C3"], "keywords": ["human-in-the-loop", "frozen policy adaptation"], "model_class": "diffusion/flow policy", "techniques": ["latent steering", "action inversion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adapts frozen flow/diffusion policies by inverting human corrections into noise and training a lightweight latent policy to steer the base model.", "relevance": 3, "relevance_reason": "Compute- and sample-efficient adaptation of generative robot policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2606.22303", "title": "FlowDPG: Deterministic Policy Gradient on Flow Matching Policies for Real-World Manipulation", "authors": ["Kexin Shi", "Junyao Shi", "Poorvi Hebbar", "Zhuolun Zhao", "Tarun Amarnath", "Yifan Su", "Shikhar Bahl", "Deepak Pathak"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.22303", "pdf_url": "https://arxiv.org/pdf/2606.22303", "code_url": "", "clusters": ["C3"], "keywords": ["RL fine-tuning", "flow matching"], "model_class": "diffusion/flow policy", "techniques": ["deterministic policy gradient", "BPTT-free distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "FlowDPG distills critic gradients into a flow-matching policy's velocity field at training time, avoiding costly backpropagation through the ODE for real-world RL.", "relevance": 3, "relevance_reason": "relevant: computationally efficient RL fine-tuning of flow policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2606.08015"}
{"id": "2606.05468", "title": "FlowPRO: Reward-Free Reinforced Fine-Tuning of Flow-Matching VLAs via Proximalized Preference Optimization", "authors": ["Yihao Wu", "He Zhang", "Junbo Tan", "Xueqian Wang", "Zhengyou Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.05468", "pdf_url": "https://arxiv.org/pdf/2606.05468", "code_url": "", "clusters": ["C3"], "keywords": ["preference optimization", "flow DPO", "reward-free", "intervention-rollback"], "model_class": "VLA", "techniques": ["offline preference optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real bimanual robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Reward-free offline fine-tuning of flow-matching VLAs with a proximal preference objective using paired trajectories from teleop interventions.", "relevance": 3, "relevance_reason": "Post-training method avoiding reward/critic engineering.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025"}
{"id": "2609.17210", "title": "FluxVLA Engine: A One-Stop VLA Engineering Platform for Embodied Intelligence", "authors": ["Yinhao Li", "Weixin Mao", "Zihan Lan", "Jikun Rong", "Qirui Hu", "Yiming Zhang", "Weipeng Deng", "Bowen Shen", "Minzhao Zhu", "Yiming Mao", "Yan Yang", "Chenguang Cui", "Hongyuan Chen", "Xu Huang", "Zheyi Zhao", "Pinxi Shen", "Bozhen He", "Zhen Fu", "Yifan Wang", "Zexin Zhang", "Ang Gao", "Haoyu Chen", "Chengqi Shi", "Hua Chen", "Yin-Hao Li", "Wei-Xin Mao", "Zi-Han Lan", "Ji-Kun Rong", "Qi-Rui Hu", "Yi-Ming Zhang", "Wei-Peng Deng", "Bo-Wen Shen", "Min-Zhao Zhu", "Yi-Ming Mao", "Chen-Guang Cui", "Hong-Yuan Chen", "Zhe-Yi Zhao", "Pin-Xi Shen", "Bo-Zhen He", "Yi-Fan Wang", "Ze-Xin Zhang", "An-Hong-Mei Gao", "Hao-Yun Chen", "Cheng-Qi Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.17210", "pdf_url": "https://arxiv.org/pdf/2609.17210", "code_url": "https://github.com/FluxVLA/FluxVLA", "clusters": ["C3", "C4", "C5", "C6", "C9"], "keywords": ["FluxVLA", "engineering platform", "data-to-deployment", "real-time chunking", "inference backends", "deployment", "remote GPU serving", "RTC", "VLA engineering platform"], "model_class": "VLA", "techniques": ["training/inference platform", "real-time chunking", "accelerated inference backends", "remote GPU serving", "remote serving"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Configuration-driven platform standardizing data, training, sim evaluation, optimized inference and robot interfaces for VLAs/WAMs, integrating RTC with accelerated backends for reproducible deployment.", "relevance": 3, "relevance_reason": "Infrastructure for reproducible end-to-end evaluation and deployment.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"offline reinforcement learning\"; search query: VLA acceleration evaluation; web: VLA accelerator FPGA ASIC design; search query"}
{"id": "2607.18231", "title": "FM-VLA: Force-based Memory for Vision-Language-Action Models in Contact-Rich Manipulation", "authors": ["Ruicheng Li", "Qi-Xiu Li", "Ruichun Ma", "Yu Deng", "Lin Luo", "Zhi-Ying Du", "Jianfeng Xiang", "Hui-Zhi Liang", "Ruicheng Wang", "Jiao-Long Yang", "Bai-Ning Guo"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.18231", "pdf_url": "https://arxiv.org/pdf/2607.18231", "code_url": "", "clusters": ["C1"], "keywords": ["contact-rich", "memory"], "model_class": "VLA", "techniques": ["compact force memory tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "FM-VLA encodes force history into compact tokens as a cheap alternative to image-history memory for non-Markovian contact tasks.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2411.19650"}
{"id": "2606.20867", "title": "FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation", "authors": ["Duc Minh Nguyen", "Nghiem Tuong Diep", "Binh Gia Nguyen", "Trong-Bao Ho", "Doanh Le", "Tan Q. Nguyen", "Thien-Loc Ha", "Nhiem Tran", "Bao Thach", "Nhat X. Tran", "Tuan A. Tran", "Artur Habuda", "Philip Lund Møller", "Tran Nguyen Le", "Daniel Sonntag", "Matthias Niepert", "Khoa D. Doan", "Vu Duong", "Hung Ngo", "Minh N. Vu", "Duy M. H. Nguyen", "An Thai Le", "Ngo Anh Vien"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20867", "pdf_url": "https://arxiv.org/pdf/2606.20867", "code_url": "", "clusters": ["C3"], "keywords": ["few-shot adaptation", "future conditioning", "data efficiency"], "model_class": "VLA", "techniques": ["auxiliary future prediction", "action-free co-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "95.7% LIBERO with 20 demos", "metric_measures": "other", "benchmarks": ["LIBERO", "RoboCasa", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Adds latent future-interaction prediction during adaptation to make VLAs data-efficient in few-shot imitation.", "relevance": 3, "relevance_reason": "Few-shot VLA adaptation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: few-shot adaptation vision-language-action model new embodiment efficient"}
{"id": "2609.21228", "title": "FOCAL-VLA: Subtask-Guided Geometry Distillation and Implicit World Modeling for Vision-Language-Action Models", "authors": ["Zhi-Yuan Gao", "Di Wen", "Yan Zhan", "Mohammad Khoshnazar", "Jeroen Schäfer", "Kun-Yu Peng", "Michael Beetz"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21228", "pdf_url": "https://arxiv.org/pdf/2609.21228", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["FOCAL-VLA", "implicit world model"], "model_class": "VLA", "techniques": ["geometry distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills geometry and future-dynamics cues into VLA representations so heavy 3D models are not run at inference.", "relevance": 3, "relevance_reason": "Avoids expensive auxiliary models at inference.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2606.26006", "title": "FORCE: Efficient VLA Reinforcement Fine-Tuning via Value-Calibrated Warm-up and Self-Distillation", "authors": ["Shuyi Zhang", "Yunfan Lou", "Hongyang Cheng", "Yichen Guo", "Chuyao Fu", "Yaoxu Lyu", "Xiaojie Zhang", "Haoran Li", "Pengwei Wang", "Zhongyuan Wang", "Shanghang Zhang", "Hong-Yang Cheng", "Yi-Chen Guo", "Peng-Wei Wang", "Zhong-Yuan Wang", "Shang-Hang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.26006", "pdf_url": "https://arxiv.org/pdf/2606.26006", "code_url": "", "clusters": ["C3"], "keywords": ["VLA RL", "value calibration", "sample efficiency"], "model_class": "VLA", "techniques": ["RL fine-tuning", "Q-filtering", "self-distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "32.5% faster training", "metric_measures": "other", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Three-stage VLA RL fine-tuning: value-calibrated warm-up, Q-filtered updates on own and expert actions, then self-distillation, reducing initial unlearning and sample cost.", "relevance": 3, "relevance_reason": "Sample-efficient VLA RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (interrupt* OR \"human intervention\" OR \"mid-execution\"); snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2606.27079", "title": "ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models", "authors": ["Mingyang Lyu", "Yinqian Sun", "Yiyang Jia", "Sicheng Shen", "Moquan Sha", "Huangrui Li", "Feifei Zhao", "Yi Zeng"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.27079", "pdf_url": "https://arxiv.org/pdf/2606.27079", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["ForesightSafety-VLA", "safety benchmark", "13-category taxonomy"], "model_class": "VLA", "techniques": ["benchmark"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diagnostic benchmark making safety the primary evaluation target for VLAs across physical, instruction-side and perception-side safety categories.", "relevance": 3, "relevance_reason": "Safety evaluation of VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" \"vision-language-action\""}
{"id": "2608.14022", "title": "ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models", "authors": ["Xinye Li", "Lingshuai Lin", "Lei Wang", "Liuzhou Zhang", "Jialin Cui", "Qingshan Li", "Guanchu Wang", "Qingbin Liu", "Xi Chen", "Jiang Bian", "Wai Lam"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.14022", "pdf_url": "https://arxiv.org/pdf/2608.14022", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["few-step causal video world model", "consistency distillation", "interactive world model"], "model_class": "world model", "techniques": ["teacher-forced causal training", "causal consistency distillation", "on-policy distribution matching", "1/2/4-step students"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Minecraft", "FPS gameplay"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Four-stage recipe converting a bidirectional action-conditioned video generator into 1-, 2- or 4-step causal world models with accurate control response, plus optional replay-time refinement.", "relevance": 3, "relevance_reason": "General few-step action-conditioned world-model recipe (games, transferable to robots).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.06971", "title": "Formal Methods in Robot Policy Learning and Verification: A Survey on Current Techniques and Future Directions", "authors": ["Anastasios Manganaris", "Vittorio Giammarino", "A. H. Qureshi", "Suresh Jagannathan"], "year": 2026, "venue": "TMLR 2026", "type": "paper", "url": "https://arxiv.org/abs/2602.06971", "pdf_url": "https://arxiv.org/pdf/2602.06971", "code_url": "", "clusters": ["C7"], "keywords": ["formal methods", "policy verification", "survey"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of formal specification, synthesis and verification methods applied to learned robot policies.", "relevance": 3, "relevance_reason": "Certification of learned controllers.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2410.04640 (S2)"}
{"id": "2609.37433", "title": "FP2: Equipping Robotic Foundation Models with Force Control", "authors": ["Hong-Jie Fang", "Shi-Run Tang", "Jun-Jian Hu", "Shi-Dong Zhang", "Derek Zhang", "Lin-Hao Chen", "De-Hai Li", "Ming-Yu Mei", "Wan-Xi Liu", "Ce-Wu Lu", "Shi-Quan Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.37433", "pdf_url": "https://arxiv.org/pdf/2609.37433", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["force control", "high-frequency policy", "foundation policy interface"], "model_class": "VLA", "techniques": ["dual-rate control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight interface where a task-adapted robot foundation model sets task actions while a high-frequency force policy regulates contact.", "relevance": 3, "relevance_reason": "Fast-slow split around a foundation policy", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2605.13757", "title": "FrameSkip: Learning from Fewer but More Informative Frames in VLA Training", "authors": ["Bin Yu", "Shijie Lian", "Xiao-Peng Lin", "Zhao-Long Shen", "Yu-Liang Wei", "Changti Wu", "Hang Yuan", "Hai-Shan Liu", "Bailing Wang", "Cong Huang", "Kai Chen"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.13757", "pdf_url": "https://arxiv.org/pdf/2605.13757", "code_url": "", "clusters": ["C3"], "keywords": ["training data efficiency"], "model_class": "VLA", "techniques": ["frame selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa-GR1", "SimplerEnv", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "FrameSkip selects informative frames in the dataloader so VLAs train on a compressed trajectory view without architecture changes.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2603.09415", "title": "From Flow to One Step: Real-Time Multi-Modal Trajectory Policies via Implicit Maximum Likelihood Estimation-based Distribution Distillation", "authors": ["Ju Dong", "Li-Ding Zhang", "Lei Zhang", "Yuan-You Fu", "Kaixin Bai", "Zoltán-Csaba Márton", "Zhen-Shan Bing", "Zhaopeng Chen", "Alois Knoll", "Jian-Wei Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.09415", "pdf_url": "https://arxiv.org/pdf/2603.09415", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["IMLE", "distribution distillation", "Chamfer"], "model_class": "diffusion/flow policy", "techniques": ["distribution distillation", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills a conditional flow-matching expert into a single-step student with IMLE and a set-level Chamfer loss to preserve modes.", "relevance": 3, "relevance_reason": "Teacher-student step compression.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2410.21257"}
{"id": "2609.15382", "title": "From Prediction to Decision: World-Model-Guided Action Selection for Continuous Pile Excavation", "authors": ["Ai-Ling Zhang", "Fan-Chang Gao", "Song Zhang", "Kawa Leong", "Zi-Yu Wu", "Ya-Fei Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.15382", "pdf_url": "https://arxiv.org/pdf/2609.15382", "code_url": "", "clusters": ["C5"], "keywords": ["world-action model", "Jetson AGX Orin", "TensorRT", "excavation", "field robot"], "model_class": "world-action model", "techniques": ["TensorRT mixed precision", "candidate ranking"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin", "reported_metric": "5 candidates processed in 72.4 ms (ROS2/TensorRT)", "metric_measures": "open-loop chunk rate", "benchmarks": ["full-size wheel loader"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Deploys a world-action model that ranks candidate scoops for autonomous wheel-loader excavation onboard a Jetson AGX Orin via TensorRT.", "relevance": 3, "relevance_reason": "Onboard TensorRT deployment of a world-action model on field hardware.", "foundational_technique": false, "verified": true, "found_via": "web: world model inference Jetson onboard"}
{"id": "2609.21788", "title": "From Pretraining to Proficiency: Real-World Subtask RL for Long-Horizon Manipulation with Minimal Human Intervention", "authors": ["Sichang Su", "B. Yang", "Zhi-Yun Deng", "Bo-Yuan Liang", "Yip-Fun Yeung", "Ze-Lin Wang", "Ling-Feng Sun"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21788", "pdf_url": "https://arxiv.org/pdf/2609.21788", "code_url": "", "clusters": ["C3"], "keywords": ["real-world RL fine-tuning", "minimal intervention"], "model_class": "other", "techniques": ["residual RL", "subtask-targeted practice"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Real-world residual RL that concentrates practice on bottleneck subtasks of a frozen foundation policy, needing only tens of minutes of rollouts.", "relevance": 3, "relevance_reason": "Sample-efficient real-world RL fine-tuning of foundation policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2406.09246, 2501.16664"}
{"id": "2609.19413", "title": "From Rollout to Reset: A Graph-Based Harness for Autonomous Long-Horizon Manipulation Evaluation", "authors": ["Jing Jiang", "Yue Yang", "Xinkai Jiang", "Gedas Bertasius", "Daniel J. Szafir", "Rudolf Lioutikov"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19413", "pdf_url": "https://arxiv.org/pdf/2609.19413", "code_url": "", "clusters": ["C4"], "keywords": ["HALTER", "autonomous reset", "long-horizon evaluation", "scene graph", "LLM scoring"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "Franka arm", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Extends autonomous real-robot evaluation to long-horizon tasks by planning resets over a library of learned atomic reset skills and LLM-based scoring on a scene graph.", "relevance": 3, "relevance_reason": "Scalable reproducible real-robot evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query: robot learning reproducibility"}
{"id": "2605.28816", "title": "Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players", "authors": ["Fang-Fu Liu", "Kaiming He", "Tianchang Shen", "Tianshi Cao", "Sanja Fidler", "Yue-Qi Duan", "Jun Gao", "Igor Gilitschenski", "Zian Wang", "Xuanchi Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.28816", "pdf_url": "https://arxiv.org/pdf/2605.28816", "code_url": "", "clusters": ["C8"], "keywords": ["multi-agent world model", "real-time rollout"], "model_class": "world model", "techniques": ["distillation", "causal student", "KV caching", "sparse attention"], "base_model": "", "params": "", "hardware": "", "reported_metric": "24 FPS action-responsive generation", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Multi-agent interactive world model with linear-cost hub attention, distilled into a causal KV-cached student for real-time generation.", "relevance": 3, "relevance_reason": "Efficient/real-time world model", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2501.03575"}
{"id": "2608.25659", "title": "GaussianDream++: Efficient 3D Gaussian World Modeling for Robotic Manipulation", "authors": ["Yuqing Jiang", "Zijian Zhang", "Weitao Zhou", "Jiawei Wang", "Junjie He", "Lei Yang", "Haifang Qing", "Si Liu", "Ding Zhao", "Ping Luo", "Haibao Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.25659", "pdf_url": "https://arxiv.org/pdf/2608.25659", "code_url": "", "clusters": ["C8"], "keywords": ["3D Gaussian world model", "compact", "VLA supervision"], "model_class": "world model", "techniques": ["compact Gaussian prediction prefix", "training-time 3D supervision"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Compact redesign of GaussianDream that keeps 3D Gaussian future-prediction supervision for VLAs while cutting deployment cost.", "relevance": 3, "relevance_reason": "Efficient 3D world-model supervision.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2608.24959", "title": "GaussVLA: Geometry-Aware Spatial Reasoning for Vision-Language-Action Model", "authors": ["MD SELIM SAROWAR", "Md Tanvir Islam", "Sungho Kim", "Sangtae Ahn"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.24959", "pdf_url": "https://arxiv.org/pdf/2608.24959", "code_url": "", "clusters": ["C1"], "keywords": ["small VLA", "3D reasoning", "parameter efficiency"], "model_class": "VLA", "techniques": ["compact Gaussian tokens", "Mamba backbone"], "base_model": "", "params": "200M", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Mamba-based VLA that lifts semantic and depth features into compact 3D Gaussian tokens for geometry-aware manipulation with a small parameter count.", "relevance": 3, "relevance_reason": "Parameter-efficient small VLA.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.13429", "title": "Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment", "authors": ["Dwip Dalal", "Shivansh Patel", "Chahit Jain", "Jeonghwan Kim", "Utkarsh Mishra", "Yuchen Song", "Alex Baratian", "Hyeonjeong Ha", "Heng Ji", "Svetlana Lazebnik", "Unnat Jain"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.13429", "pdf_url": "https://arxiv.org/pdf/2607.13429", "code_url": "", "clusters": ["C3"], "keywords": ["representation anchoring", "language-action alignment", "forgetting"], "model_class": "VLA", "techniques": ["distillation regularization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Adds layer-wise anchoring to a frozen VLM and language-action alignment to prevent BC fine-tuning from eroding representations.", "relevance": 3, "relevance_reason": "Replacement for costly web co-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.20776", "title": "GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies", "authors": ["Xin Chen", "Sen Chen", "Yujuan Ding", "Jian Liu", "Guoqing Wang", "Wei Ye", "Heng Tao Shen", "Yi Bin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20776", "pdf_url": "https://arxiv.org/pdf/2609.20776", "code_url": "", "clusters": ["C1"], "keywords": ["adaptive chunking", "flow matching", "uncertainty"], "model_class": "VLA", "techniques": ["adaptive action chunking", "training-free"], "base_model": "GR00T N1.5, pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Pro", "RoboCasa365", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses geometric variation of flow-matching denoising trajectories as an uncertainty proxy to pick the execution horizon from one generation.", "relevance": 3, "relevance_reason": "Adaptive chunking.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking"}
{"id": "2606.17046", "title": "Geometric Action Model for Robot Policy Learning", "authors": ["Jisang Han", "Seonghu Jeon", "Jaewoo Jung", "René Zurbrügg", "Honggyu An", "Tifanny Portela", "Marco Hutter", "Marc Pollefeys", "Seungryong Kim", "Sung‐Jin Hong"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.17046", "pdf_url": "https://arxiv.org/pdf/2606.17046", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["geometric foundation model", "latent prediction"], "model_class": "world-action model", "techniques": ["backbone sharing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Splits a pretrained geometric foundation model so one backbone encodes observations, forecasts future latent tokens and decodes actions; reported as faster and lighter than foundation-scale baselines.", "relevance": 3, "relevance_reason": "Claims speed/size advantages over large VLA/WAM baselines.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310; cited by 2609.33575"}
{"id": "2605.08638", "title": "Geometry Guided Self-Consistency for Physical AI", "authors": ["Yinwei Dai", "Zhuofu Chen", "Lijie Yang", "Ravi Netravali"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.08638", "pdf_url": "https://arxiv.org/pdf/2605.08638", "code_url": "https://github.com/dywsjtu/keystone", "clusters": ["C1", "C7"], "keywords": ["KeyStone"], "model_class": "VLA", "techniques": ["self-consistency", "parallel sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to +13.3% success at negligible latency overhead", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Samples K action chunks in parallel from shared context and returns the medoid of the largest cluster; exploits memory-bandwidth-bound diffusion to add no wall-clock latency.", "relevance": 3, "relevance_reason": "Judge-free chunk selection at negligible latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=100]"}
{"id": "2609.07006", "title": "GIFT: Goal-Injected Fine-Tuning for Efficient Manipulation Policy Adaptation", "authors": ["Xiaoyuan Fang", "Shuo Feng", "Yuxuan Wang", "Enhua Cheng", "Peng Zhou", "Piji Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.07006", "pdf_url": "https://arxiv.org/pdf/2609.07006", "code_url": "", "clusters": ["C3"], "keywords": ["goal image conditioning", "zero-initialized convolution"], "model_class": "VLA", "techniques": ["lightweight fine-tuning", "zero-init adapter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Injects generated goal-image features into pretrained VLAs through a zero-initialized conv adapter for cheap, stable fine-tuning.", "relevance": 3, "relevance_reason": "Lightweight fine-tuning adding a modality without full retraining.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.02642", "title": "GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation", "authors": ["GigaWorld Team", "Angyuan Ma", "Boyuan Wang", "Bohan Li", "Chaojun Ni", "Guo Li", "Guan Huang", "Guosheng Zhao", "Hao Li", "Hengtao Li", "Jingyu Liu", "Jiwen Lu", "Qiuping Deng", "Tingdong Yu", "Xuancheng Xu", "Xinyu Zhou", "Xiuwei Xu", "Xinze Chen", "Xiaofeng Wang", "Xiaoyu Tian", "Yang Wang", "Yifan Chang", "Yukun Zhou", "Yun Ye", "Zhenyu Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.02642", "pdf_url": "https://arxiv.org/pdf/2607.02642", "code_url": "", "clusters": ["C3", "C4"], "keywords": ["GigaWorld-1", "WMBench", "world model as evaluator"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["WMBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Systematic study of what makes world models reliable policy evaluators, with a benchmark of 7 video world models and 324k rollouts paired with real executions.", "relevance": 3, "relevance_reason": "World-model-based evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "search query: policy evaluation real robot scalable"}
{"id": "2602.20200", "title": "Global Prior Meets Local Consistency: Dual-Memory Augmented Vision-Language-Action Model for Efficient Robotic Manipulation", "authors": ["Zaijing Li", "Bing Hu", "Rui Shao", "Gongwei Chen", "Dongmei Jiang", "Pengwei Xie", "Jianye Hao", "Liqiang Nie"], "year": 2026, "venue": "CVPR 2026", "type": "paper", "url": "https://arxiv.org/abs/2602.20200", "pdf_url": "https://arxiv.org/pdf/2602.20200", "code_url": "", "clusters": ["C1"], "keywords": ["retrieved prior", "reduced NFE", "dual memory"], "model_class": "VLA", "techniques": ["retrieval-augmented initialization", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "CALVIN"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Starts denoising from task priors retrieved from similar trajectories rather than Gaussian noise, cutting NFE, plus a local consistency memory for smoothness.", "relevance": 3, "relevance_reason": "Fewer denoising steps via priors.", "foundational_technique": false, "verified": true, "found_via": "search query: denoising steps VLA"}
{"id": "2603.14245", "title": "GoldenStart: Q-Guided Priors and Entropy Control for Distilling Flow Policies", "authors": ["He Zhang", "Ying Sun", "Hui Xiong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.14245", "pdf_url": "https://arxiv.org/pdf/2603.14245", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["flow policy distillation", "RL"], "model_class": "diffusion/flow policy", "techniques": ["one-step distillation", "Q-guided prior"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Designs Q-guided initial noise and entropy control for one-step distilled flow policies in RL.", "relevance": 3, "relevance_reason": "Distillation of flow policies for fast inference", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2607.13059", "title": "GPUSimBench: Towards Scalable and Reliable GPU-Accelerated Simulators in Embodied AI", "authors": ["Huzhenyu Zhang", "Shenghai Yuan", "Wenrui Yan", "Li Ma", "Hengjie Li", "Jingcheng Pang", "Dmitry Yudin"], "year": 2026, "venue": "IROS 2026", "type": "benchmark", "url": "https://arxiv.org/abs/2607.13059", "pdf_url": "https://arxiv.org/pdf/2607.13059", "code_url": "", "clusters": ["C4"], "keywords": ["GPUSimBench", "GPU simulation", "determinism", "Isaac Lab", "Genesis"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["GPUSimBench"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Benchmarks GPU simulators for throughput/memory scaling, physical fidelity and run-to-run non-determinism, showing batched GPU execution undermines reproducibility.", "relevance": 3, "relevance_reason": "Reproducibility of sim-based evaluation infrastructure.", "foundational_technique": false, "verified": true, "found_via": "search query: embodied AI benchmark efficiency"}
{"id": "bair-blog-grasp-2026", "title": "Gradient-based Planning for World Models at Longer Horizons (BAIR Blog)", "authors": ["Michael Psenka", "Mike Rabbat", "Aditi Krishnapriyan", "Yann LeCun", "Amir Bar"], "year": 2026, "venue": "blog (BAIR)", "type": "blog", "url": "https://bair.berkeley.edu/blog/2026/04/20/grasp/", "pdf_url": "", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["world model planning", "GRASP"], "model_class": "world model", "techniques": ["parallel gradient-based planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "Push-T H=40 59.0% success, 8.5 s median; H=80 10.4%, 58.9 s", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Blog explainer for GRASP, showing faster and more successful long-horizon world-model planning than CEM/GD/LatCo.", "relevance": 3, "relevance_reason": "Blog companion to 2602.00475.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: BAIR blog"}
{"id": "2608.27079", "title": "GRAFT: Grounded and Efficient Online Reinforcement Adaptation for Fine-Grained Robot Manipulation", "authors": ["Yibo Qiu", "Haoliang Ye", "Shu'ang Sun", "Zan Huang", "Ronald X Xu", "Mingzhai Sun"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.27079", "pdf_url": "https://arxiv.org/pdf/2608.27079", "code_url": "", "clusters": ["C3"], "keywords": ["online adaptation", "visual anchors", "biomedical manipulation"], "model_class": "VLA", "techniques": ["online RL", "single-step action generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Efficient online VLA adaptation with learned view-specific visual anchors and single-step action generation to cut inference/update cost.", "relevance": 3, "relevance_reason": "Addresses compute cost of online adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.13695", "title": "GROOVE: Geometry-Guided Reduction of Operational-Space Jerk in VLA Execution", "authors": ["Sang-Ho Yun", "Minsoo Kim", "Minwoo Cho", "Hwanjo Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.13695", "pdf_url": "https://arxiv.org/pdf/2609.13695", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["GROOVE", "jerk reduction", "QP regulator", "end-effector"], "model_class": "VLA", "techniques": ["QP-based online jerk regulation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "-33.02% translational, -43.42% rotational EEF jerk", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Online QP regulator reducing within-chunk and boundary jerk of VLA end-effector commands while bounding deviation from the raw plan, with no extra VLA inference.", "relevance": 3, "relevance_reason": "Actuator-safe execution of chunks.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2608.04510", "title": "GUARD: Grounding Uncertainty and Ablation-Based Risk Detection for Diffusion-Based VLAs", "authors": ["Suhas Hegde", "Jitendra Yasaswi Bharadwaj Katta"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.04510", "pdf_url": "https://arxiv.org/pdf/2608.04510", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["failure detection", "grounding uncertainty", "diffusion VLA"], "model_class": "VLA", "techniques": ["ablation-based test-time risk detection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "GUARD detects weakly grounded (risky) diffusion-VLA actions at test time by ablating inputs, without modifying the policy.", "relevance": 3, "relevance_reason": "Test-time failure monitoring.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: arXiv API VLA AND (safety filter OR CBF OR fallback)"}
{"id": "2607.02092", "title": "Guided Action Flow: Q-Guided Inference for Flow-Matching Vision-Language-Action Policies", "authors": ["Liuhaichen Yang", "Zhuang Jiang", "Chenchao Sheng", "Zezhi Tang", "Liu Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.02092", "pdf_url": "https://arxiv.org/pdf/2607.02092", "code_url": "", "clusters": ["C3"], "keywords": ["Q-guided flow", "inference-time guidance", "Q-learning"], "model_class": "VLA", "techniques": ["critic guidance", "IQL", "critic-guided sampling", "frozen policy"], "base_model": "SmolVLA", "params": "", "hardware": "", "reported_metric": "success 47.5% -> 85.0% on real task", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Guided Action Flow steers a frozen SmolVLA's flow sampling with gradients from an offline-trained action-chunk critic, improving a real task without fine-tuning the policy.", "relevance": 3, "relevance_reason": "Cheap deployment-time adaptation from few rollouts.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=0]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2605.10051", "title": "Guided Streaming Stochastic Interpolant Policy", "authors": ["Pu-Ming Jiang", "Mei-Yi Wang", "Kelvin Lin", "Ce Hao", "Harold Soh"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2605.10051", "pdf_url": "https://arxiv.org/pdf/2605.10051", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["reactivity", "stochastic interpolants"], "model_class": "diffusion/flow policy", "techniques": ["streaming policy", "inference-time guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Derives optimal guidance for streaming stochastic-interpolant policies, enabling reactive real-time steering unlike chunk-based policies.", "relevance": 3, "relevance_reason": "Low-latency reactive generative control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.16837", "title": "HAF: Adapting Generalist VLAs to Humanoid Whole-Body Loco-manipulation via Hierarchical Action Flow and Spectral Latent RL", "authors": ["Langzhe Gu", "Chengkai Hou", "Meng Li", "Xinhua Wang", "Jiaming Liu", "Xinyuan Lv", "Bowei Zhang", "Shuanghao Bai", "Guangrun Li", "Jingyang He", "Gaole Dai", "Ziluo Ding", "Zhiyuan Xu", "Kuan Cheng", "Jian Tang", "Zhengping Che", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.16837", "pdf_url": "https://arxiv.org/pdf/2608.16837", "code_url": "", "clusters": ["C3"], "keywords": ["humanoid loco-manipulation", "spectral latent RL", "hierarchical action flow"], "model_class": "VLA", "techniques": ["latent-space RL", "hierarchical adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Adapts generalist VLAs to humanoid whole-body control and refines them with latent RL instead of tuning the large backbone.", "relevance": 3, "relevance_reason": "Avoids tuning large backbone during real RL.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: ti:VLA AND ti:RL"}
{"id": "2607.04265", "title": "HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models", "authors": ["An-Gen Ye", "Wei-Jie Ke", "Xiao-Feng Wang", "Xinze Chen", "Chaojun Ni", "Guosheng Zhao", "Boyuan Wang", "Zheng Zhu", "Jun-Jie Xie", "Da-Peng Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.04265", "pdf_url": "https://arxiv.org/pdf/2607.04265", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["world-action model adaptation", "precision insertion"], "model_class": "world-action model", "techniques": ["lightweight actor-critic adapter", "online RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Online RL with a lightweight adapter reading world-action-model latents for fast real-world adaptation on precision tasks.", "relevance": 3, "relevance_reason": "Efficient online adaptation of WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2607.13960"}
{"id": "2609.09941", "title": "HaWMPO: Hallucination-Aware World Model-based Policy Optimization for Generalist Robot Policy", "authors": ["Zengjue Chen", "Peidong Liu", "Jiawei Li", "Qi Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.09941", "pdf_url": "https://arxiv.org/pdf/2609.09941", "code_url": "", "clusters": ["C8"], "keywords": ["hallucination-aware", "world-model RL", "VLA post-training"], "model_class": "world model", "techniques": ["hallucination detection in imagined rollouts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "World-model RL pipeline for VLA post-training that detects and discounts hallucinated long-horizon transitions to avoid misleading policy updates.", "relevance": 3, "relevance_reason": "Reliability of imagined rollouts.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2604.12447", "title": "HazardArena: Evaluating Semantic Safety in Vision-Language-Action Models", "authors": ["Zixin Chen", "Yifeng Gao", "Li Wang", "Yun-Han Zhao", "Yi Liu", "Jiayu Li", "Xiang Zheng", "Zuxuan Wu", "Cong Wang", "Xing-guan Ma", "Yu-Gang Jiang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2604.12447", "pdf_url": "https://arxiv.org/pdf/2604.12447", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["HazardArena", "semantic safety"], "model_class": "VLA", "techniques": ["safety benchmark"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmark stress-testing whether VLAs recognize when otherwise valid actions become hazardous.", "relevance": 3, "relevance_reason": "Safety evaluation for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310; cited by 2604.23775"}
{"id": "2607.09776", "title": "HELP: Human-Efficient Large-Scale Robot Post-Training with Rollout Segmentation", "authors": ["Shaopeng Zhai", "Qi Zhang", "Tianyi Zhang", "Haoran Zhang", "Fuxian Huang", "Zhanhui Lin", "Zijun Xu", "Weinan Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.09776", "pdf_url": "https://arxiv.org/pdf/2607.09776", "code_url": "", "clusters": ["C3"], "keywords": ["human efficiency", "fleet post-training", "interventions"], "model_class": "VLA", "techniques": ["interactive post-training", "human-in-the-loop"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Post-training pipeline where two specialized operators supervise twelve robots, maximizing policy improvement per unit of human labor.", "relevance": 3, "relevance_reason": "Human/labor cost of iterative VLA post-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.17043", "title": "Hierarchical Advantage Weighting for Online RL Fine-Tuning of VLAs from Sparse Episode Outcomes", "authors": ["Tongyan Fang", "Siyuan Huang", "Naiyu Fang", "Ganlong Zhao", "Zhongjin Luo", "Jianbo Liu", "Xiaogang Wang", "Ying Dong", "Hongsheng Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.17043", "pdf_url": "https://arxiv.org/pdf/2606.17043", "code_url": "", "clusters": ["C3"], "keywords": ["advantage weighting", "sparse outcomes", "human intervention", "real-world RL"], "model_class": "VLA", "techniques": ["advantage-weighted BC", "offline-to-online RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real bimanual contact-rich tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Advantage-weighted BC with separate viability/efficiency critics and intervention-aware credit assignment for real-world VLA RL fine-tuning.", "relevance": 3, "relevance_reason": "Real-world RL post-training method.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025"}
{"id": "2603.27281", "title": "HiFlow: Tokenization-Free Scale-Wise Autoregressive Policy Learning via Flow Matching", "authors": ["Daichi Yashima", "Koki Seno", "Shuhei Kurita", "Yusuke Oda", "Komei Sugiura"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.27281", "pdf_url": "https://arxiv.org/pdf/2603.27281", "code_url": "", "clusters": ["C1"], "keywords": ["tokenization-free", "scale-wise"], "model_class": "diffusion/flow policy", "techniques": ["coarse-to-fine autoregression", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MimicGen", "RoboTwin 2.0"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Coarse-to-fine autoregressive flow policy on continuous actions that avoids discrete tokenizers.", "relevance": 3, "relevance_reason": "Efficient AR action generation design.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.03865", "title": "High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching", "authors": ["Yuran Chen", "Xin Cai", "Zhong-Lin Gong", "Yang Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.03865", "pdf_url": "https://arxiv.org/pdf/2607.03865", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["one-step", "recursive correction", "frequency consistency"], "model_class": "diffusion/flow policy", "techniques": ["one-step distillation", "consistency training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin", "RoboTwin 2.0", "Adroit", "DexArt", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Improves one-step flow policies with recursive correction, spectral consistency across timesteps and contrastive flow matching to avoid mode averaging.", "relevance": 3, "relevance_reason": "High-fidelity one-step policy compression.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2410.21257"}
{"id": "2609.20659", "title": "HIL-UMI: Bringing Human-in-the-Loop Post-Training of Vision-Language-Action Models to Universal Manipulation Interface", "authors": ["Zimu Han", "Yiming Zeng", "Jiyao Zhang", "Zihao Zhao", "Yuanfei Wang", "Yixiang Jin", "Shiqi Li", "Shuangben Chen", "Wei Huang", "Ruodai Li", "Hui Shen", "Hao Dong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20659", "pdf_url": "https://arxiv.org/pdf/2609.20659", "code_url": "", "clusters": ["C3"], "keywords": ["UMI", "robot-free post-training", "human-in-the-loop"], "model_class": "VLA", "techniques": ["interactive post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Queries the policy during handheld UMI demonstrations to collect targeted corrections for VLA post-training without a robot.", "relevance": 3, "relevance_reason": "Cuts robot time for post-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.10363", "title": "HiMem-WAM: Hierarchical Memory-Gated World Action Models for Robotic Manipulation", "authors": ["Xiao-Quan Sun", "Ruijiang Zhang", "Chenyi Cao", "Yihang Sun", "Jia-Hui Chen", "Ze-Tian Xu", "Bo-Yuan Chen", "Haijier Chen", "Zhen Yang", "Jia-Run Zhu", "Yi-Jun Hong", "Jing-Zhe Xu", "Jin-Li Pang", "Ming-Qi Yuan", "Jia-Yu Chen"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.10363", "pdf_url": "https://arxiv.org/pdf/2606.10363", "code_url": "", "clusters": ["C8"], "keywords": ["hierarchical latent actions", "memory gate"], "model_class": "world-action model", "techniques": ["latent memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus", "RMBench", "real robot"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "WAM with hierarchical latent actions and a boundary-triggered compact memory, performing inference without generating future video or optical flow at test time.", "relevance": 3, "relevance_reason": "Avoids test-time video generation in WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2609.36518"}
{"id": "2608.13678", "title": "hint$^2$: Hierarchical World Models for Inference-Time Temporal Logic Guidance", "authors": ["Moritz Zoellner", "Anastasios Manganaris", "Ahmed H. Qureshi", "Rohan Paleja"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.13678", "pdf_url": "https://arxiv.org/pdf/2608.13678", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["LTL", "action chunk", "world model"], "model_class": "diffusion/flow policy", "techniques": ["temporal logic guidance", "hierarchical world models", "inference-time guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses a two-level hierarchy of world models to turn long-horizon LTL specifications into inference-time guidance for short-horizon chunked policies that replan in closed loop.", "relevance": 3, "relevance_reason": "Shows how temporal-logic constraints can be imposed on chunk-level generation, a route to specification-aware chunk execution.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunk*\" AND (safety OR safe)"}
{"id": "2604.06067", "title": "HiPolicy: Hierarchical Multi-Frequency Action Chunking for Policy Learning", "authors": ["Ji-Yao Zhang", "Zimu Han", "Jun-Han Wang", "Xionghao Wu", "Shih-Ming Lin", "Jinzhou Li", "Hongwei Fan", "Rui-Hai Wu", "Dongjiang Li", "Hao Dong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.06067", "pdf_url": "https://arxiv.org/pdf/2604.06067", "code_url": "", "clusters": ["C1"], "keywords": ["hierarchical chunking", "execution efficiency"], "model_class": "diffusion/flow policy", "techniques": ["multi-frequency action chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Jointly predicts action chunks at multiple frequencies and adaptively chooses execution granularity by uncertainty.", "relevance": 3, "relevance_reason": "Improves execution efficiency of chunked policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.30828", "title": "HIRE: History-Conditioned Interaction Reasoning and High-Rate Execution for Visually Aliased Precision Manipulation", "authors": ["Rong-Ji Li", "Wen-Hao He", "Ce-Wu Lu", "Xing-Yu Chen", "Xu-Yao Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.30828", "pdf_url": "https://arxiv.org/pdf/2609.30828", "code_url": "", "clusters": ["C1"], "keywords": ["high-rate execution", "history", "precision manipulation"], "model_class": "other", "techniques": ["dual-rate reasoning/execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Couples history-conditioned interaction reasoning with a high-rate execution loop for precision manipulation under visual aliasing.", "relevance": 3, "relevance_reason": "Cross-rate reasoning/execution design", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.06480", "title": "History-Conditioned Spatio-Temporal Visual Token Pruning for Efficient Vision-Language Navigation", "authors": ["Qitong Wang", "Yijun Liang", "Ming Li", "Tianyi Zhou", "C. Rasmussen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.06480", "pdf_url": "https://arxiv.org/pdf/2603.06480", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["token pruning", "spatio-temporal", "VLN"], "model_class": "VLA", "techniques": ["training-free visual token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free spatio-temporal token pruning conditioned on history for VLA-based navigation.", "relevance": 3, "relevance_reason": "Token pruning for embodied VLAs.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2602.12062", "title": "HoloBrain-0 Technical Report", "authors": ["Xuewu Lin", "Tianwei Lin", "Yun Du", "Hongyu Xie", "Yiwei Jin", "Jiawei Li", "Shijie Wu", "Qingze Wang", "Mengdi Li", "Mengao Zhao", "Ziang Li", "Chaodong Huang", "Hongzhe Bi", "Lichao Huang", "Zhizhong Su"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2602.12062", "pdf_url": "https://arxiv.org/pdf/2602.12062", "code_url": "", "clusters": ["C1"], "keywords": ["HoloBrain-0", "RoboOrchard"], "model_class": "VLA", "techniques": ["embodiment priors", "small variant"], "base_model": "", "params": "0.2B (efficient variant)", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "LIBERO", "GenieSim", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VLA framework with embodiment priors (camera params, URDF) plus an open full-stack infrastructure; a 0.2B variant rivals larger baselines for low-latency on-device use.", "relevance": 3, "relevance_reason": "Small on-device-capable VLA variant with deployment stack.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]"}
{"id": "2606.15032", "title": "How Should World Models Be Evaluated for Embodied Decision-Making? A Decision-Making-Centric Position", "authors": ["Yang Yu", "Shiyuan Zhang", "Yi-Fei Sheng", "Hao-Xiang Ren", "Hao-Chu Lin"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.15032", "pdf_url": "https://arxiv.org/pdf/2606.15032", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["world model evaluation", "position"], "model_class": "world model", "techniques": ["evaluation methodology"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Position paper arguing world models for embodied decision-making should be judged on interventional reasoning, policy evaluation and planning rather than visual realism.", "relevance": 3, "relevance_reason": "World-model evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104; cited by 2609.09250"}
{"id": "2604.21192", "title": "How VLAs (Really) Work In Open-World Environments", "authors": ["Amir Rasouli", "Yangzheng Wu", "Zhiyuan Li", "Rui Heng Yang", "Xuan Zhao", "Charles Eret", "Sajjad Pakdamansavoji"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.21192", "pdf_url": "https://arxiv.org/pdf/2604.21192", "code_url": "", "clusters": ["C4"], "keywords": ["BEHAVIOR-1K", "evaluation protocol", "safety violations", "reproducibility"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["BEHAVIOR-1K"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Analyzes B1K challenge policies and argues final-state success metrics overstate progress; proposes protocols capturing consistency and safety violations.", "relevance": 3, "relevance_reason": "Critique of success-only metrics.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2609.35709", "title": "Humanoid Loco-Manipulation With Discrete VLA Model", "authors": ["Wenxin Shao", "Siqi Chai", "Kun Li", "Kerou Zhang", "Xinzhou Jiang", "Wei Xu", "Qiang Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.35709", "pdf_url": "https://arxiv.org/pdf/2609.35709", "code_url": "", "clusters": ["C1"], "keywords": ["humanoid", "discrete VLA", "grouped discrete diffusion decoding"], "model_class": "VLA", "techniques": ["action tokenization", "discrete diffusion decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Discrete humanoid loco-manipulation VLA with body-part tokenizers decoded via grouped discrete diffusion rather than autoregression for real-time control.", "relevance": 3, "relevance_reason": "Parallel decoding for humanoid VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: discrete diffusion VLA"}
{"id": "2608.28995", "title": "Hydra: A Navigation World Action Model with Discrete Latent Planning and Continuous Flow-Matching Execution", "authors": ["Mohammad Nazeri", "Alexandyr Card", "S. Huber", "Anuj Pokhrel", "Yu-Jun Wang", "Ruben Hammele", "Daeun Song", "Sören Pirk", "Xuesu Xiao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.28995", "pdf_url": "https://arxiv.org/pdf/2608.28995", "code_url": "", "clusters": ["C8"], "keywords": ["navigation", "latent planning", "real-time planning"], "model_class": "world-action model", "techniques": ["vector quantization", "discrete latent planning", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Hydra is a navigation world-action model that plans directly in a shared discrete latent space (avoiding costly decoding of candidates) and executes via flow matching.", "relevance": 3, "relevance_reason": "relevant: planning in latent space to enable real-time world-model planning", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2608.25053", "title": "Hydra: Phase-Aware Workload Characterization of LLM Inference across Edge SoC Generations, Backends, and Quantization Levels", "authors": ["Amir Taherin", "Sana Taghipour Anvari", "Charles Amante", "Yi-Xiao Chen", "Ruben Noroian", "Z. Feric", "N. Agostini", "Pu Zhao", "José Cano", "Bin Ren", "Yan-Zhi Wang", "David R. Kaeli"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.25053", "pdf_url": "https://arxiv.org/pdf/2608.25053", "code_url": "", "clusters": ["C5", "C9"], "keywords": ["edge SoC", "AGX Xavier/Orin/Thor", "prefill vs decode", "telemetry", "llama.cpp"], "model_class": "other", "techniques": ["phase-aware workload characterization"], "base_model": "", "params": "", "hardware": "Jetson AGX Xavier, AGX Orin, AGX Thor", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Phase-aware LLM characterization framework across three Jetson generations, backends and quantization formats, showing quantization reduces memory traffic and energy but not power monotonically.", "relevance": 3, "relevance_reason": "Edge SoC characterization directly relevant to VLA backbones though not robotics-specific.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2509.11480 (citations)"}
{"id": "2609.07581", "title": "ICI-VLA: In-Context Imitation with Spatiotemporally Aligned Demonstrations for Vision-Language-Action Models", "authors": ["Songhua Yang", "Ziyu Liu", "Xuetao Li", "Ruqi Xiao", "Kangxin Zhu", "Miao Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.07581", "pdf_url": "https://arxiv.org/pdf/2609.07581", "code_url": "", "clusters": ["C3"], "keywords": ["in-context imitation", "retrieval", "test-time adaptation"], "model_class": "VLA", "techniques": ["in-context learning", "retrieval"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Few-shot test-time adaptation of a text-action VLA by retrieving aligned micro-demonstrations as context, no gradient updates.", "relevance": 3, "relevance_reason": "Gradient-free adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"few-shot\""}
{"id": "2602.20715", "title": "IG-RFT: An Interaction-Guided RL Framework for VLA Models in Long-Horizon Robotic Manipulation", "authors": ["Zhian Su", "Weijie Kong", "Haonan Dong", "Huixu Dong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.20715", "pdf_url": "https://arxiv.org/pdf/2602.20715", "code_url": "", "clusters": ["C3"], "keywords": ["IG-RFT", "advantage weighted regression", "real-world RL"], "model_class": "VLA", "techniques": ["AWR", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Interaction-guided AWR system for sample-efficient real-world RL fine-tuning of flow VLAs on long-horizon tasks.", "relevance": 3, "relevance_reason": "Sample-efficient real-world post-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.24033", "title": "Imagine-RL: Residual-Confidence-Guided Cross-Attention for World-Model-Augmented VLA Reinforcement Learning", "authors": ["Kejia Hu", "Wentong Zhai", "Bo Zhao", "Shuai Liang", "Ke-Jia Hu", "Wen Zhai", "Bo-Tong Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24033", "pdf_url": "https://arxiv.org/pdf/2609.24033", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["VLA RL", "contact-rich", "RL post-training", "frozen VLA"], "model_class": "VLA", "techniques": ["noise-space RL", "latent world model", "critic", "frozen VLA"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Imagine-RL improves a frozen VLA via noise-space RL with a critic that uses a frozen visual-torque latent world model to imagine consequences, needing only about 100 real RL trajectories.", "relevance": 3, "relevance_reason": "relevant: sample-efficient adaptation of frozen VLAs", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]; snowball round 2: cites 2405.12213, 2406.09246"}
{"id": "2603.13335", "title": "Information-Theoretic Constraints for Continual Vision-Language-Action Alignment", "authors": ["Libang Zhao", "Qixin Zeng", "Hongyin Zhang", "Donglin Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.13335", "pdf_url": "https://arxiv.org/pdf/2603.13335", "code_url": "", "clusters": ["C3"], "keywords": ["continual learning", "Info-VLA", "mutual information"], "model_class": "VLA", "techniques": ["continual learning", "contrastive replay"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Continual VLA learning that preserves cross-modal alignment via replay-anchor contrastive learning and mutual information constraints.", "relevance": 3, "relevance_reason": "Continual adaptation of VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: continual learning vision-language-action model catastrophic forgetting arXiv"}
{"id": "2607.04988", "title": "InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization", "authors": ["Haoxiang Ma", "Junhao Cai", "Xiaoxu Xu", "Hao Li", "Yuyin Yang", "Yang Tian", "Jiafei Cao", "Hongrui Zhu", "Zherui Qiu", "Zhaxizhuoma", "Yuqiang Yang", "Jiaqi Peng", "Xueyuan Wei", "Yangkun Zhu", "Jiahao Jiang", "Xing Gao", "Hanqing Wang", "Feng Yuan", "Kailin Li", "Xueyue Zhu", "Tai Wang", "Yan Ding", "Jiangmiao Pang", "Jia Zeng", "Jingjing Zhang", "Bowen Zhou", "Yao Mu", "Chunhua Shen", "Weinan Zhang", "Hao-Xiang Ma", "Jun-Hao Cai", "Yu-Yin Yang", "Jia-Fei Cao", "Hong-Rui Zhu", "Yu-Qiang Yang", "Xue-Yuan Wei", "Yang-Kun Zhu", "Han-Qing Wang", "Xue-Yue Zhu", "Jiang-Miao Pang", "Jing-Jing Zhang", "Bo-Wen Zhou", "Wei-Nan Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.04988", "pdf_url": "https://arxiv.org/pdf/2607.04988", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["InternVLA-A1.5", "unified VLA", "world-model priors"], "model_class": "VLA", "techniques": ["latent foresight tokens", "branch dropped at inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills video-generator dynamics priors into a few latent foresight tokens during training and discards the video branch at inference to keep real-time control.", "relevance": 3, "relevance_reason": "World-model benefits without inference-time video generation cost.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight); snowball round 2: cited by 2609.30247, 2609.36471"}
{"id": "2606.19194", "title": "Invertible Neural Network Adapter for One-Step Flow Matching in Robot Manipulation", "authors": ["Yu Zhang", "Kangyi Ji", "Yongxiang Zou", "Rongtao Xu", "Feng Zheng", "Long Cheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.19194", "pdf_url": "https://arxiv.org/pdf/2606.19194", "code_url": "", "clusters": ["C1"], "keywords": ["invertible adapter", "one-step flow"], "model_class": "VLA", "techniques": ["one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "VLA inference latency 110 ms -> 61 ms", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Invertible-network adapter constrains flow generation to a latent space where a single denoising step suffices, applied to VLAs.", "relevance": 3, "relevance_reason": "One-step VLA action head.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching robot policy efficient"}
{"id": "2601.21506", "title": "IROS: A Dual-Process Architecture for Real-Time VLM-Based Indoor Navigation", "authors": ["Joonhee Lee", "Hyunseung Shin", "Jeonggil Ko"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.21506", "pdf_url": "https://arxiv.org/pdf/2601.21506", "code_url": "", "clusters": ["C1", "C6", "C9"], "keywords": ["dual process", "indoor navigation", "on-device VLM", "dual-process", "VLM"], "model_class": "other", "techniques": ["dual-system architecture", "gated VLM invocation", "fast/slow dual process"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Navigation system that runs fast reflexive modules and invokes a compact VLM only when needed on low-cost on-device hardware.", "relevance": 3, "relevance_reason": "Dual-process on-device design.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2603.06954", "title": "Is Your Safe Controller Actually Safe? A Critical Review of CBF Tautologies and Hidden Assumptions", "authors": ["Taekyung Kim"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2603.06954", "pdf_url": "https://arxiv.org/pdf/2603.06954", "code_url": "", "clusters": ["C7"], "keywords": ["CBF pitfalls", "input constraints", "valid vs candidate CBF", "tutorial"], "model_class": "other", "techniques": ["critical review"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Tutorial showing many claimed CBF-safe robot controllers rely on hidden assumptions (e.g., passively safe dynamics, unbounded inputs) and distinguishing valid from candidate CBFs.", "relevance": 3, "relevance_reason": "Caveats for CBF filters on learned policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"control barrier function\" learned policy robot manipulation"}
{"id": "2605.13013", "title": "JEDI: Joint Embedding Diffusion World Model for Online Model-Based Reinforcement Learning", "authors": ["Jing Yu Lim", "Rushi Shah", "Zarif Ikram", "Samson Yu", "Haozhe Ma", "Tze-Yun Leong", "Dianbo Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.13013", "pdf_url": "https://arxiv.org/pdf/2605.13013", "code_url": "", "clusters": ["C8"], "keywords": ["joint embedding diffusion", "latent diffusion world model", "MBRL"], "model_class": "world model", "techniques": ["end-to-end JEPA + diffusion latent dynamics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Online MBRL with a diffusion world model operating in an end-to-end learned JEPA latent space, avoiding the cost of pixel diffusion.", "relevance": 3, "relevance_reason": "Efficient latent diffusion world model.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.09630", "title": "JEPA Policy: Diffusion-Free Imitation Learning via Paired Action and Future Representation Prediction", "authors": ["Jie Xu", "Kang-Jin Yu", "Zi-Yi Jin", "Jun-Jie Gao", "Li-Qing Chen", "Yi-Xiang Li", "Shuai Tian", "Zhong-Pu Xia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.09630", "pdf_url": "https://arxiv.org/pdf/2609.09630", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["JEPA", "low latency"], "model_class": "diffusion/flow policy", "techniques": ["JEPA", "diffusion-free"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+0.29 ms model latency over MIP baseline", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "JEPA Policy jointly predicts action chunks and future representations in a shared Transformer with two forward passes, providing low-latency imitation without iterative sampling.", "relevance": 3, "relevance_reason": "relevant: low-latency diffusion-free policy", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2605.26636", "title": "JetViT: Efficient High-Resolution Vision Transformer with Post-Training Attention Search", "authors": ["Dongyun Zou", "Zhuoyang Zhang", "Junyu Chen", "Wenkun He", "Qinhe Peng", "Hanrong Ye", "Yao Lu", "Hongxu Yin", "Yu Wang", "Song Han", "Han Cai"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.26636", "pdf_url": "https://arxiv.org/pdf/2605.26636", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["hybrid attention ViT", "post-training attention search", "high-resolution vision encoder"], "model_class": "other", "techniques": ["replace full attention with linear/window attention post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "JetViT converts pretrained full-attention ViTs into hybrid linear/window-attention models via post-training search, keeping accuracy at much higher high-res throughput.", "relevance": 3, "relevance_reason": "Efficient vision encoders applicable to VLA backbones; organizer Zhuoyang Zhang.", "foundational_technique": true, "verified": true, "found_via": "author sweep: Zhuoyang Zhang"}
{"id": "2607.16074", "title": "JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models", "authors": ["Haoran Sun", "Wentao Zhang", "Junyang Hua", "Hedan Yang", "Yongjian Guo", "Yifei Zhang", "Xiaolong Xiang", "Mingxi Luo", "Jing Long", "Chen Zhao", "Chen Zhou", "Wanting Xu", "Qiming Yang", "Hui Zhang", "Song Wang", "Xiaodong Bai", "Shuai Di", "Xu Chu", "Xiaotie Deng", "Yicheng Gong", "Junwu Xiong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.16074", "pdf_url": "https://arxiv.org/pdf/2607.16074", "code_url": "", "clusters": ["C3"], "keywords": ["multi-tenant", "post-training service", "shared base models"], "model_class": "VLA", "techniques": ["multi-tenant serving", "SFT and RL service"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Service architecture for multi-tenant VLA SFT/RL/evaluation that shares resident base models across tenants to cut cost for bursty workloads.", "relevance": 3, "relevance_reason": "Infrastructure cost of post-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2608.21247", "title": "Just Noticeable Difference Modeling for Token Compression in Vision-Language-Action Models", "authors": ["Zhuoyuan Li", "Rui Zhao", "Jin Wang", "Hanwei Zhu", "Cong Zhang", "Giuseppe Valenzise", "Weisi Lin", "Kin-Man Lam"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.21247", "pdf_url": "https://arxiv.org/pdf/2608.21247", "code_url": "", "clusters": ["C1"], "keywords": ["just noticeable difference", "token compression", "KV reuse"], "model_class": "VLA", "techniques": ["token pruning", "KV cache reuse"], "base_model": "OpenVLA, OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Scores tokens by how much they can change before the VLA's action response deviates, and uses this as a criterion for pruning or stale-KV reuse.", "relevance": 3, "relevance_reason": "Compression criterion for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "2606.16533", "title": "Kairos: A Regret-Aware Native World-Action Model Stack for Physical AI", "authors": ["Kairos Team", "Fei Wang", "Shan You", "Qiming Zhang", "Tao Huang", "Zuoyi Fu", "Zhisheng Zheng", "Yunlong Xi", "Feng Lv", "Xiaoming Wu", "Zeyu Liu", "Cong Wan", "Pu Li", "Ruiqing Yang", "Xiaoou Li", "Wei Wang", "Kangkang Zhu", "Yuwei Zhang", "Shi Fu", "Zheng Zhang", "Xiaoning Wu", "Xuzeng Fan", "Dacheng Tao", "Xiaogang Wang"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2606.16533", "pdf_url": "https://arxiv.org/pdf/2606.16533", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["Kairos"], "model_class": "world-action model", "techniques": ["linear temporal attention", "system co-design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "World-action model stack emphasizing control-relevant state, hybrid linear temporal attention and deployment-aware co-design for latency and memory.", "relevance": 3, "relevance_reason": "Treats latency/memory as first-order WAM constraints.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2603.01581", "title": "KERV: Kinematic-Rectified Speculative Decoding for Embodied VLA Models", "authors": ["Zihao Zheng", "Zhihao Mao", "Maoliang Li", "Jiayu Chen", "Xinhao Sun", "Zhaobo Zhang", "Donggang Cao", "Hong Mei", "Xiang Chen"], "year": 2026, "venue": "DAC 2026", "type": "preprint", "url": "https://arxiv.org/abs/2603.01581", "pdf_url": "https://arxiv.org/pdf/2603.01581", "code_url": "", "clusters": ["C1", "C2", "C5"], "keywords": ["speculative decoding", "Kalman filter", "kinematics"], "model_class": "VLA", "techniques": ["speculative decoding", "kinematic-rectified speculative decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "27-37% acceleration | 27%-37% acceleration with nearly no success-rate loss", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses a kinematic Kalman-filter prediction to correct draft errors and adapt acceptance thresholds in VLA speculative decoding.", "relevance": 3, "relevance_reason": "Primarily C2; addresses memory-bound AR decoding. Author list incomplete (from search snippet).", "foundational_technique": false, "verified": true, "found_via": "search query: speculative decoding VLA; snowball from 2608.04428 (references)"}
{"id": "2609.00908", "title": "Knowing When to Stop: Adaptive Action Chunking via Internal Cross-Attention Dynamics in VLAs", "authors": ["Runze Xu", "Xiaolong Shan", "Shuang Dai", "Yu Wang", "Jincheng Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.00908", "pdf_url": "https://arxiv.org/pdf/2609.00908", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["adaptive chunking", "cross-attention entropy", "adaptive action chunking", "training-free"], "model_class": "VLA", "techniques": ["adaptive action chunking", "training-free", "attention-entropy truncation"], "base_model": "pi0.5, X-VLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "LIBERO", "real robot", "real"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Truncates executed chunks when action-to-observation cross-attention entropy plateaus, using attention already computed by the action expert.", "relevance": 3, "relevance_reason": "Adaptive open-loop horizon.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; arXiv search: \"action chunk\" execution horizon adaptive"}
{"id": "2603.13781", "title": "KoopmanFlow: Spectrally Decoupled Generative Control Policy via Koopman Structural Bias", "authors": ["Cheng-Si Yao", "Ge Wang", "Kai Kang", "Shen Yan", "Jia-Hao Yang", "Fan Feng", "Honghao Cai", "Xian-Xian Zeng", "Rong-Jun Chen", "Yi-Ming Zhao", "Ya-Tong Han", "Xi Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.13781", "pdf_url": "https://arxiv.org/pdf/2603.13781", "code_url": "", "clusters": ["C1"], "keywords": ["real-time generative policy", "parameter efficiency"], "model_class": "diffusion/flow policy", "techniques": ["single-step consistency", "flow matching", "Koopman prior"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Splits generation into a single-step consistency branch for slow motion and a flow-matching branch for high-frequency corrections within real-time budgets.", "relevance": 3, "relevance_reason": "Few-step generative policy designed for real-time receding-horizon control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2606.01151", "title": "Lagrangian Perturbation Diffusion Steering: Latent Reinforcement Learning for Generative Policies", "authors": ["Hikmet Simsir", "Ozgur S. Oguz"], "year": 2026, "venue": "ICML 2026", "type": "preprint", "url": "https://arxiv.org/abs/2606.01151", "pdf_url": "https://arxiv.org/pdf/2606.01151", "code_url": "", "clusters": ["C3"], "keywords": ["latent RL", "diffusion steering", "lightweight adaptation"], "model_class": "diffusion/flow policy", "techniques": ["noise-space steering", "trust region", "latent noise-space RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboMimic", "Gym", "Adroit", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns a compact noise-space perturbation for a frozen generative policy with a Lagrangian trust region, improving value while staying near the latent prior.", "relevance": 3, "relevance_reason": "Lightweight RL adaptation of diffusion policies.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]; snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2602.10556", "title": "LAP: Language-Action Pre-Training Enables Zero-shot Cross-Embodiment Transfer", "authors": ["Lihan Zha", "Asher J. Hancock", "Mingtong Zhang", "Tenny Yin", "Yixuan Huang", "Dhruv Shah", "Allen Z. Ren", "Anirudha Majumdar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.10556", "pdf_url": "https://arxiv.org/pdf/2602.10556", "code_url": "", "clusters": ["C3"], "keywords": ["LAP", "language actions", "zero-shot cross-embodiment"], "model_class": "VLA", "techniques": ["actions as language"], "base_model": "", "params": "3B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["multiple real robots"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Encodes actions as natural language to get zero-shot transfer to new embodiments, sidestepping per-embodiment fine-tuning cost.", "relevance": 3, "relevance_reason": "Reduces cross-embodiment adaptation cost.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: knowledge insulation co-training VLA training efficiency pi0.5 arXiv"}
{"id": "2603.08476", "title": "LAR-MoE: Latent-Aligned Routing for Mixture of Experts in Robotic Imitation Learning", "authors": ["Ariel Rodríguez", "Chenpan Li", "Lorenzo Mazza", "Rayan Younis", "Ortrun Hellig", "S. Bodenstedt", "Martin Wagner", "Stefanie Speidel"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.08476", "pdf_url": "https://arxiv.org/pdf/2603.08476", "code_url": "", "clusters": ["C1"], "keywords": ["LAR-MoE"], "model_class": "diffusion/flow policy", "techniques": ["mixture of experts"], "base_model": "", "params": "150M", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MoE imitation policy with latent-aligned routing that avoids expert collapse while staying parameter efficient.", "relevance": 3, "relevance_reason": "Parameter-efficient MoE policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.05468", "title": "Learning 4D Geometric Priors for Inference-Efficient World Action Models", "authors": ["Jianjun Zhang", "Jian Zhu", "Taiyi Su", "Chong Ma", "Zitai Huang", "Yi Xu", "Hanli Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.05468", "pdf_url": "https://arxiv.org/pdf/2607.05468", "code_url": "", "clusters": ["C8"], "keywords": ["4D geometric priors", "multi-expert co-training", "lightweight inference graph"], "model_class": "world-action model", "techniques": ["training-only 4D expert from frozen VGGT", "decayed read-mask attention"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Injects 4D geometry priors into a WAM through a training-only expert, preserving the original lightweight inference graph.", "relevance": 3, "relevance_reason": "Geometry without inference cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2605.24931", "title": "Learning High-Frequency Continuous Action Chunks in Latent Space", "authors": ["Kunyun Wang", "Yuhang Zheng", "Yupeng Zheng", "Jieru Zhao", "Wenchao Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24931", "pdf_url": "https://arxiv.org/pdf/2605.24931", "code_url": "https://github.com/tars-robotics/RTR", "clusters": ["C1"], "keywords": ["latent action chunks", "60 Hz", "reuse-then-refine"], "model_class": "diffusion/flow policy", "techniques": ["latent action space", "asynchronous inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns high-frequency (~60 Hz) chunks in a VAE latent space and refines reused chunks for continuity under async inference.", "relevance": 3, "relevance_reason": "High-frequency chunk generation.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking"}
{"id": "2606.15685", "title": "Learning New Tasks via Reusable Skills: Skill-Compositional Experts for Embodied Continual Learning", "authors": ["Shuai-Ke Zhang", "Shaokun Wang", "Hao Tang", "Jianlong Wu", "Liqiang Nie"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.15685", "pdf_url": "https://arxiv.org/pdf/2606.15685", "code_url": "", "clusters": ["C3"], "keywords": ["embodied continual learning"], "model_class": "other", "techniques": ["skill-compositional experts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Continual learning via reusable skill experts to add tasks without forgetting under closed-loop control.", "relevance": 3, "relevance_reason": "Continual adaptation (C3).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2310.05905"}
{"id": "2609.35003", "title": "Learning to Act under Visual Interruptions with Vision-Language-Action Models", "authors": ["Mingle Jiang", "Rui Xu", "Yunke Wang", "Chang Xu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.35003", "pdf_url": "https://arxiv.org/pdf/2609.35003", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["MAIL-Bench", "camera dropout", "visual interruptions", "visual interruption", "camera loss", "robustness"], "model_class": "VLA", "techniques": ["missing-view training", "optical-flow / world-model view imputation"], "base_model": "pi0.5, GR00T N1.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MAIL-Bench"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Benchmark and method for VLAs when camera streams drop mid-task; trains for missing views and imputes frames with extrapolation or a world model, withdrawing predictions when unreliable.", "relevance": 3, "relevance_reason": "Robustness to sensing failures relevant to real-time deployment.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation world model; arXiv search: \"streaming\" flow policy robot action"}
{"id": "2606.12475", "title": "Learning to Assist: Collaborative VLAs for Implicit Human-Robot Collaboration", "authors": ["Le Xu", "Le-Tian Li", "Alex Cuellar", "Michael Hagenow"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.12475", "pdf_url": "https://arxiv.org/pdf/2606.12475", "code_url": "", "clusters": ["C7"], "keywords": ["human-robot collaboration", "action chunk leakage"], "model_class": "VLA", "techniques": ["inference-time steering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Shows action-chunked VLAs can act prematurely in collaboration because chunks cross task transitions, and proposes an inference-time steering fix.", "relevance": 3, "relevance_reason": "Execution-horizon/chunking failure mode analysis.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2605.07079", "title": "Learning Visual Feature-Based World Models via Residual Latent Action", "authors": ["Xinyu Zhang", "Zhengtong Xu", "Yutian Tao", "Yeping Wang", "Yu She", "Abdeslam Boularias", "Zheng-Tong Xu", "Yu Tao", "Ye Wang", "Yuqin She"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.07079", "pdf_url": "https://arxiv.org/pdf/2605.07079", "code_url": "", "clusters": ["C8"], "keywords": ["RLA-WM", "feature world model", "latent action"], "model_class": "world model", "techniques": ["residual latent actions", "feature-space world model", "residual latent action", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RLA-WM predicts DINO-feature residual latent actions via flow matching, giving a feature-space world model much faster than video diffusion that supports policy learning and RL inside the model.", "relevance": 3, "relevance_reason": "Cheap world model alternative to video diffusion.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=100]; snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2605.00416", "title": "Learning While Deploying: Fleet-Scale Reinforcement Learning for Generalist Robot Policies", "authors": ["Yi Wang", "Xinchen Li", "Pengwei Xie", "Pu Yang", "Buqing Nie", "Yunuo Cai", "Qinglin Zhang", "Chendi Qu", "Jeffrey Wu", "Jianheng Song", "Xinlin Ren", "Jingshun Huang", "Mingjie Pan", "Siyuan Feng", "Zhi Chen", "Jianlan Luo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.00416", "pdf_url": "https://arxiv.org/pdf/2605.00416", "code_url": "", "clusters": ["C3"], "keywords": ["fleet learning", "offline-to-online RL", "continual post-training"], "model_class": "VLA", "techniques": ["offline-to-online RL", "distributional value learning", "adjoint matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["16 dual-arm robot fleet"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Fleet-scale continual RL post-training of a flow VLA from autonomous rollouts and interventions across 16 robots.", "relevance": 3, "relevance_reason": "Deployment-time continual adaptation at scale.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv"}
{"id": "2607.19399", "title": "Leveraging Offline Supervision for Efficient and Generalizable Reinforcement Learning in Large-Scale Vision-Language-Action Models", "authors": ["Dmitriy Poyarkov", "Aleksei Staroverov", "Aleksandr I. Panov", "D.I. Poyarkov", "A. Staroverov", "A. Panov"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.19399", "pdf_url": "https://arxiv.org/pdf/2607.19399", "code_url": "", "clusters": ["C3"], "keywords": ["VLA RL", "OOD", "RL training efficiency"], "model_class": "VLA", "techniques": ["offline-regularized RL", "offline-guided RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "about half the training budget of standard RL", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Studies hybrid offline-online RL for large VLAs, regularizing online RL with offline data or reference policies to combine OOD robustness with efficiency.", "relevance": 3, "relevance_reason": "Efficiency of VLA RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]; snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2603.28301", "title": "LIBERO-Para: A Diagnostic Benchmark and Metrics for Paraphrase Robustness in VLA Models", "authors": ["Chanyoung Kim", "Minwoo Kim", "Minseok Kang", "Hyunwoo Kim", "Dahuin Jung"], "year": 2026, "venue": "EMNLP 2026", "type": "benchmark", "url": "https://arxiv.org/abs/2603.28301", "pdf_url": "https://arxiv.org/pdf/2603.28301", "code_url": "https://github.com/cau-hai-lab/LIBERO-Para", "clusters": ["C4"], "keywords": ["LIBERO-Para", "paraphrase robustness", "language generalization"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Para"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "LIBERO variant that paraphrases action verbs and object references; seven VLAs (0.6B-7.5B) lose 22-52 points, mostly from object synonym changes, and a difficulty-weighted metric (PRIDE) is proposed.", "relevance": 3, "relevance_reason": "Robustness variant of LIBERO; relevant to whether compressed VLAs retain language grounding.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2609.05178", "title": "LIBERO-RECOVER: Beyond Task Success Towards Failure Recovery in Robotic Manipulation Models", "authors": ["Lin Liu", "Zhicheng Bao", "Lu Zhang", "Ziying Song", "Wu Yang", "Yuzheng Zhuang", "Shuai Tao", "Wulong Liu", "Caiyan Jia", "Huchuan Lu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.05178", "pdf_url": "https://arxiv.org/pdf/2609.05178", "code_url": "", "clusters": ["C4"], "keywords": ["LIBERO-Recover", "failure recovery"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Recover"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Benchmark of 1000+ failure states harvested from real policy failures on LIBERO across four recovery levels, measuring whether VLAs/WAMs can recover rather than just succeed from clean starts.", "relevance": 3, "relevance_reason": "Extends LIBERO evaluation beyond nominal success; recovery is a key deployability metric.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2606.23686", "title": "LIBERO-Safety: A Comprehensive Benchmark for Physical and Semantic Safety in Vision-Language-Action Models", "authors": ["Rongxu Cui", "Zongzheng Zhang", "Jingrui Pang", "Haohan Chi", "Jinbang Guo", "Saining Zhang", "Shaoxuan Xie", "Xin Jin", "Yao Mu", "Jiaolong Yang", "Guocai Yao", "Xianyuan Zhan", "Ya-Qin Zhang", "Hao Zhao"], "year": 2026, "venue": "ECCV 2026", "type": "benchmark", "url": "https://arxiv.org/abs/2606.23686", "pdf_url": "https://arxiv.org/pdf/2606.23686", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["LIBERO-Safety", "safety benchmark", "collision"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Safety"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Parametric LIBERO-based benchmark for physical and semantic safety with ~19.7k collision-free demos; evaluation of eight VLAs finds a tension between generalization and safe execution.", "relevance": 3, "relevance_reason": "Adds safety axis to LIBERO evaluation; relevant for deployability metrics.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2602.06556", "title": "LIBERO-X: Robustness Litmus for Vision-Language-Action Models", "authors": ["Guodong Wang", "Chenkai Zhang", "Qingjie Liu", "Jinjin Zhang", "Jiancheng Cai", "Junjie Liu", "Xinmin Liu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2602.06556", "pdf_url": "https://arxiv.org/pdf/2602.06556", "code_url": "", "clusters": ["C4"], "keywords": ["LIBERO-X", "hierarchical evaluation", "robustness"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-X"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Rethinks LIBERO with a progressive-difficulty protocol over spatial, object and instruction generalization plus a more diverse teleoperated training set; VLAs drop sharply under cumulative perturbations.", "relevance": 3, "relevance_reason": "LIBERO robustness variant aimed at more reliable VLA benchmarking.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2604.24921", "title": "Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System", "authors": ["Yifei Wei", "Linqing Zhong", "Yi Liu", "Yuxiang Lu", "Xindong He", "Maoqing Yao", "Guanghui Ren"], "year": 2026, "venue": "ACL 2026", "type": "paper", "url": "https://arxiv.org/abs/2604.24921", "pdf_url": "https://arxiv.org/pdf/2604.24921", "code_url": "", "clusters": ["C1"], "keywords": ["coarse-to-fine dual system", "hybrid action space", "asynchronous"], "model_class": "VLA", "techniques": ["dual-system architecture", "asynchronous execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Splits actions into discrete coarse directional tokens from a planner and continuous refinements from a fast refiner, run asynchronously.", "relevance": 3, "relevance_reason": "Dual-system VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "2603.10929", "title": "Lifelong Imitation Learning with Multimodal Latent Replay and Incremental Adjustment", "authors": ["Fan-Qi Yu", "Matteo Tiezzi", "Tommaso Apicella", "Cigdem Beyan", "Vittorio Murino"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.10929", "pdf_url": "https://arxiv.org/pdf/2603.10929", "code_url": "", "clusters": ["C3"], "keywords": ["lifelong imitation learning", "memory constraints"], "model_class": "other", "techniques": ["latent replay", "incremental feature adjustment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lifelong IL that replays compact multimodal latents instead of raw data, under memory and data constraints.", "relevance": 3, "relevance_reason": "Memory-efficient continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2601.09512"}
{"id": "2604.26182", "title": "Lifting Embodied World Models for Planning and Control", "authors": ["Alex N. Wang", "Trevor Darrell", "P. Izmailov", "Yutong Bai", "Amir Bar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.26182", "pdf_url": "https://arxiv.org/pdf/2604.26182", "code_url": "", "clusters": ["C8"], "keywords": ["world-model planning", "CEM"], "model_class": "world model", "techniques": ["lifted action space", "lightweight low-level policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Composes a frozen world model with a lightweight policy mapping high-level to low-level actions, making planning cheaper in high-dimensional action spaces.", "relevance": 3, "relevance_reason": "Reduces world-model planning cost.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2604.03208"}
{"id": "2605.31158", "title": "Light Interaction: Training-Free Inference Acceleration for Interactive Video World Models", "authors": ["Jiacheng Lu", "Haoyi Zhu", "Sipei Yi", "Enze Xie", "Yu Li", "Cheng Zhuo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.31158", "pdf_url": "https://arxiv.org/pdf/2605.31158", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["training-free acceleration", "adaptive computation", "interactive world model"], "model_class": "world model", "techniques": ["trajectory-dependent memory discarding", "adaptive temporal context", "early-step output reuse on revisits"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Training-free inference acceleration for interactive world models that adapts memory, context and denoising reuse to the camera trajectory.", "relevance": 3, "relevance_reason": "General world-model inference acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.30935", "title": "LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation", "authors": ["Shao-An Wang", "Ao-Cheng Luo", "Fei Huang", "Jing-Yi Xu", "Xiao-Yang Wang", "Yue-Yue Wang", "Qian-Li Ma", "Fan Yang", "Ran Mei", "Jia Wei", "Jiang-Peng Hu", "Xu-Hao Liu", "Hong-Ming Chen", "Yuan-Bin Shao", "Yiyang Lin", "Zi-Liang Li", "Liang Pan", "Xin-Hang Liu", "Yuntao Ma", "Tingxiang Fan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.30935", "pdf_url": "https://arxiv.org/pdf/2608.30935", "code_url": "", "clusters": ["C1"], "keywords": ["compact navigation model"], "model_class": "VLA", "techniques": ["RVQ action tokenizer", "visual history compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "LightNav-0 is a compact VLM-based generalist navigation model using a unified pointing interface and compressed visual history.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2502.19645"}
{"id": "2606.26183", "title": "LiMoDE: Rethinking Lifelong Robot Manipulation from a Mixture-of-Dynamic-Experts Perspective", "authors": ["Zhihao Gu", "Lin Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.26183", "pdf_url": "https://arxiv.org/pdf/2606.26183", "code_url": "", "clusters": ["C3"], "keywords": ["LiMoDE", "mixture of dynamic experts", "lifelong"], "model_class": "other", "techniques": ["MoE", "PEFT adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Dynamic MoE pretraining then lightweight task adaptation for lifelong manipulation.", "relevance": 3, "relevance_reason": "Parameter-efficient lifelong adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"robot\" AND abs:\"parameter-efficient fine-tuning\""}
{"id": "2606.27871", "title": "LocalNav: Distilling Frontier VLMs and Embodied RL for On-Device Object Goal Navigation", "authors": ["N. Baumann", "Liam Boyle", "Pu Deng", "Edoardo Ghignone", "Bo-Yang Sun", "Marc Pollefeys", "Luca Benini", "Michele Magno"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.27871", "pdf_url": "https://arxiv.org/pdf/2606.27871", "code_url": "", "clusters": ["C1", "C2", "C5", "C6"], "keywords": ["ObjectNav", "on-device VLM", "Jetson Orin", "distillation"], "model_class": "other", "techniques": ["knowledge distillation", "distillation from frontier VLM", "RL output-length regularization", "quantization"], "base_model": "Qwen3.5-4B", "params": "4B", "hardware": "Jetson Orin", "reported_metric": "82.8% cumulative reduction in inference latency", "metric_measures": "relative latency", "benchmarks": ["HM3D OVON"], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Distills frontier-model navigation reasoning into a 4B local VLM and trains it to emit shorter outputs, cutting latency for on-device ObjectNav on Jetson Orin.", "relevance": 3, "relevance_reason": "Large-to-small distillation for onboard embodied reasoning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246; snowball from 2604.24447 (citations)"}
{"id": "2607.03751", "title": "Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA Models", "authors": ["Xinyi Xie", "Zican Hu", "Zhanyu Liu", "Yicheng Dong", "Wenhao Wu", "Zhenhong Sun", "Haoran Li", "Chunlin Chen", "Zhi Wang", "Pichao Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.03751", "pdf_url": "https://arxiv.org/pdf/2607.03751", "code_url": "", "clusters": ["C3"], "keywords": ["frozen VLA", "test-time scaling"], "model_class": "VLA", "techniques": ["tree-search distillation", "Q-value reranking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "9B+SVA beats 27B VLA by 7 pts at 27% lower inference latency", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills MCTS-explored returns into a lightweight Q-model that picks among frozen-VLA candidate actions at deployment; a 9B VLA beats a 27B one at lower latency.", "relevance": 3, "relevance_reason": "Test-time evaluation as a cheaper alternative to larger models.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]"}
{"id": "2606.12688", "title": "M*: A Modular, Extensible, Serving System for Multimodal Models", "authors": ["Atindra Jha", "Naomi Sagan", "Keisuke Kamahori", "Irmak Sivgin", "Rohan Sanda", "Steven Gao", "Mark Horowitz", "Luke Zettlemoyer", "Olivia Hsu", "Jure Leskovec", "Baris Kasikci", "Stephanie Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.12688", "pdf_url": "https://arxiv.org/pdf/2606.12688", "code_url": "https://github.com/mstar-project/mstar", "clusters": ["C5", "C6"], "keywords": ["composite model serving", "dataflow graph", "VLA serving", "world model", "M*", "serving"], "model_class": "other", "techniques": ["disaggregated component placement", "model serving", "dataflow graph"], "base_model": "", "params": "", "hardware": "", "reported_metric": "20% lower latency than vLLM-Omni on BAGEL; up to 2.7x throughput on Qwen3-Omni TTS; outperforms V-JEPA 2-AC rollout baseline for robot planning", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Serving system representing composite models (VLM + action/flow heads + world models) as dataflow graphs for flexible cluster placement, including robot planning workloads.", "relevance": 3, "relevance_reason": "Cloud-side serving infrastructure for VLA/world-model fleets.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q2: arXiv abs:\"vision-language-action\" AND abs:serving; recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]"}
{"id": "2602.02212", "title": "MAIN-VLA: Modeling Abstraction of Intention and eNvironment for Vision-Language-Action Models", "authors": ["Zheyuan Zhou", "Liang Du", "Zi-Xun Sun", "Xiaoyun Zhou", "R. Ye", "Qihao Chen", "Yin-Da Chen", "Le Qiu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.02212", "pdf_url": "https://arxiv.org/pdf/2602.02212", "code_url": "", "clusters": ["C1"], "keywords": ["token pruning", "games"], "model_class": "VLA", "techniques": ["parameter-free token pruning", "semantic abstraction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MAIN-VLA abstracts instructions and visual streams into compact semantic primitives, and the resulting attention concentration enables parameter-free token pruning in fast-paced game environments.", "relevance": 3, "relevance_reason": "relevant: token pruning for real-time VLA decision-making", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2608.15748", "title": "Making two action heads agree: coordination mechanisms and a runtime collapse certificate for flow-matching policies", "authors": ["Jinhui Sun", "Wei Zhou", "Bowen Yang", "Xin-Wei Xiao", "Li Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15748", "pdf_url": "https://arxiv.org/pdf/2608.15748", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["runtime collapse certificate", "dual action heads", "joint/EE residual"], "model_class": "diffusion/flow policy", "techniques": ["dual-representation decoding residual"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses the residual between joint-space and end-effector-space decodings of a dual-head flow policy as a physically interpretable runtime signal, studying how to coordinate branches to avoid false alarms.", "relevance": 3, "relevance_reason": "Runtime self-check for flow policies.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2510.09459 (S2)"}
{"id": "2603.28545", "title": "ManipArena: Comprehensive Real-world Evaluation of Reasoning-Oriented Generalist Robot Manipulation", "authors": ["Yu Sun", "Meng Cao", "Yang Ping", "Kaidong Zhang", "Qingxuan Chen", "Rongtao Xu", "Liangwang Ruan", "Xuecheng Chen", "Dongxiu Liu", "Yunxiao Yan", "Zunnan Xu", "Runze Xu", "Charles Yang", "Peilun Zhang", "Xiaofan Li", "Ruyi Gan", "Liang Ma", "Yuehao Yin", "Jincheng Yu", "Lufang Chen", "Yuxin Liang", "Peng Zhai", "Hao Wang", "Ivan Laptev", "Ian Reid", "Qian Wang", "Xiaodan Liang", "Kai-Dong Zhang", "Qing Chen", "Rong-Tao Xu", "Liang-Wang Ruan", "Xue-Cheng Chen", "Dong-Xiu Liu", "Yu-Yan Yan", "Run-Ze Xu", "Pei-Lun Zhang", "Xiao-Fan Li", "Yue Yin", "Lu Chen", "Yu-Xin Liang", "Ian D. Reid"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2603.28545", "pdf_url": "https://arxiv.org/pdf/2603.28545", "code_url": "", "clusters": ["C4"], "keywords": ["ManipArena", "real-robot evaluation", "standardized protocol"], "model_class": "other", "techniques": ["benchmarking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Standardized real-robot evaluation framework (20 tasks, 10,812 trajectories, ~188 robot hours) with matched conditions and paired real-to-sim scenes for VLAs and WAMs.", "relevance": 3, "relevance_reason": "Real-world benchmark that accounts for latency and hardware effects.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=100]; snowball round 2: cites 2306.03310, 2504.16054"}
{"id": "2603.04363", "title": "ManipulationNet: An Infrastructure for Benchmarking Real-World Robot Manipulation with Physical Skill Challenges and Embodied Multimodal Reasoning", "authors": ["Yi-Ting Chen", "Kenny Kimble", "Edward H. Adelson", "T. Asfour", "Podshara Chanrungmaneekul", "Sachin Chitta", "Yash Chitambar", "Zi-Yang Chen", "Ken Goldberg", "Danica Kragic", "Hui Li", "Xiang Li", "Yunzhu Li", "Aaron Prather", "Nancy S. Pollard", "Máximo A. Roa-Garzón", "Robert W. Seney", "Shuo Sha", "Shih-Ting Wang", "Yu Xiang", "Kai-Feng Zhang", "Yu-Ke Zhu", "Kaiyu Hang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2603.04363", "pdf_url": "https://arxiv.org/pdf/2603.04363", "code_url": "", "clusters": ["C4"], "keywords": ["real-world benchmark", "reproducibility"], "model_class": "other", "techniques": ["real-world benchmark infrastructure"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "ManipulationNet is a global infrastructure of standardized hardware kits and a unified client for reproducible, distributed real-world manipulation benchmarking.", "relevance": 3, "relevance_reason": "relevant: real-robot benchmarking protocol", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310; cited by 2605.20774, 2607.04434"}
{"id": "2604.08995", "title": "Matrix-Game 3.0: Real-Time and Streaming Interactive World Model with Long-Horizon Memory", "authors": ["Zile Wang", "Zexiang Liu", "Jiaxing Li", "Kaichen Huang", "Baixin Xu", "Fei Kang", "Mengyin An", "Peiyu Wang", "Biao Jiang", "Yichen Wei", "Yidan Xietian", "Jiangbo Pei", "Liang Hu", "Boyi Jiang", "Hua Xue", "Zidong Wang", "Haofeng Sun", "Wei Li", "Wanli Ouyang", "Xianglong He", "Yang Liu", "Yangguang Li", "Yahui Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.08995", "pdf_url": "https://arxiv.org/pdf/2604.08995", "code_url": "", "clusters": ["C8"], "keywords": ["Matrix-Game 3.0", "720p real-time", "long-horizon memory"], "model_class": "world model", "techniques": ["residual self-correction training", "camera-aware memory retrieval", "few-step AR distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Memory-augmented interactive world model targeting 720p real-time long-form generation, adding self-correction training and camera-aware memory to the Matrix-Game 2.0 recipe.", "relevance": 3, "relevance_reason": "State of the art in real-time interactive world models with memory.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.13810", "title": "Mean Flow Policy with Instantaneous Velocity Constraint for One-step Action Generation", "authors": ["Guojian Zhan", "Letian Tao", "Pengcheng Wang", "Yixiao Wang", "Yiheng Li", "Yuxin Chen", "Hongyang Li", "Masayoshi Tomizuka", "Shengbo Eben Li"], "year": 2026, "venue": "ICLR 2026", "type": "paper", "url": "https://arxiv.org/abs/2602.13810", "pdf_url": "https://arxiv.org/pdf/2602.13810", "code_url": "", "clusters": ["C1"], "keywords": ["mean velocity policy", "one-step", "instantaneous velocity constraint"], "model_class": "diffusion/flow policy", "techniques": ["mean flow", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Robomimic", "OGBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "One-step mean-velocity policy with an instantaneous-velocity boundary constraint that keeps it expressive, improving training and inference speed.", "relevance": 3, "relevance_reason": "One-step policy class.", "foundational_technique": false, "verified": true, "found_via": "search query: mean flow policy"}
{"id": "2603.03596", "title": "MEM: Multi-Scale Embodied Memory for Vision Language Action Models", "authors": ["Marcel Torne", "Karl Pertsch", "H. Walke", "Kyle Vedder", "Suraj Nair", "Brian Ichter", "Allen Z. Ren", "Hao-Huan Wang", "Jiaming Tang", "Kyle Stachowicz", "Karan Dhabalia", "M. Equi", "Q. Vuong", "J. Springenberg", "Sergey Levine", "Chelsea Finn", "Danny Driess"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.03596", "pdf_url": "https://arxiv.org/pdf/2603.03596", "code_url": "", "clusters": ["C1"], "keywords": ["long-horizon memory", "video encoder"], "model_class": "VLA", "techniques": ["memory compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Gives robot policies multi-scale memory: compressed short-term video memory plus text-based long-term memory, enabling tasks lasting up to ~15 minutes.", "relevance": 3, "relevance_reason": "Compressed memory representation to keep long-context control tractable.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2603.14371"}
{"id": "2609.28256", "title": "MemBodied: Recurrent Associative Memory for Vision-Language-Action Models", "authors": ["Tej Deep Pala", "Navonil Majumder", "Bryce Goh", "Raphael Yee", "Jianfei Yang", "Liming Chen", "Soujanya Poria", "Jian-Fei Yang", "Li-Ming Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28256", "pdf_url": "https://arxiv.org/pdf/2609.28256", "code_url": "", "clusters": ["C1"], "keywords": ["memory", "VLA", "episodic memory", "constant context"], "model_class": "VLA", "techniques": ["recurrent associative memory", "fixed-size recurrent memory"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RMBench", "LIBERO-Long"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fixed-size episodic memory (associative state plus episode anchor) for VLAs replacing growing observation context.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=0]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2608.26545", "title": "Memory Anchors for Continual Robot Learning", "authors": ["Maximilian Du", "Zhanyi Sun", "Chen Xu", "Paarth Shah", "Masha Itkina", "Shuran Song"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.26545", "pdf_url": "https://arxiv.org/pdf/2608.26545", "code_url": "", "clusters": ["C3"], "keywords": ["memory anchors", "replay selection", "continual learning"], "model_class": "VLA", "techniques": ["experience replay selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Identifies a small set of replay samples (memory anchors) that prevent forgetting, enabling smaller replay buffers.", "relevance": 3, "relevance_reason": "Cheaper continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.18144", "title": "Memory as a Wasting Asset: Pricing Flash Endurance for Embodied Agents, and the Limits of Doing So", "authors": ["J. Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18144", "pdf_url": "https://arxiv.org/pdf/2606.18144", "code_url": "", "clusters": ["C5"], "keywords": ["flash endurance", "edge robots"], "model_class": "other", "techniques": ["memory placement", "wear-aware pricing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Prices flash write endurance for on-board robot memory and derives cost-optimal RAM/NVM/cloud placement.", "relevance": 3, "relevance_reason": "Hardware-aware resource analysis for edge robots.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2506.01844"}
{"id": "2609.11561", "title": "Memory as Plans: World-Action Modeling with Memory-Grounded Planning", "authors": ["Si-Zhe Zhao", "Hao-Zhe Xie", "Wei-Yu Zhao", "Chen-Chu Zhang", "Huan Wang", "Chen-Yang Wang", "Qinglin Liu", "Sheng-Ping Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.11561", "pdf_url": "https://arxiv.org/pdf/2609.11561", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["world-action model", "long-horizon memory"], "model_class": "world-action model", "techniques": ["fixed-length executor context", "KV caching", "memory-grounded planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RMBench"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Converts episodic memory into compact plans so the world-action executor keeps a fixed context length with KV caching, holding inference latency constant as history grows.", "relevance": 3, "relevance_reason": "Latency-bounded memory design for world-action models.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2406.09246"}
{"id": "2606.25700", "title": "Memory-Efficient Policy Libraries with Low-Rank Adaptation in Reinforcement Learning", "authors": ["Samuel Valland Lyngset", "Tor Viljen Raanaas", "Gard Sveipe", "Eirik Møller Nilsen", "Jim Torresen", "Kai Olav Ellefsen", "Tobias Lømo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.25700", "pdf_url": "https://arxiv.org/pdf/2606.25700", "code_url": "", "clusters": ["C3"], "keywords": ["policy library", "LoRA", "PPO"], "model_class": "other", "techniques": ["LoRA"], "base_model": "", "params": "", "hardware": "", "reported_metric": "20-160x lower memory than full fine-tuning; 90-95% storage saving", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "LoRA for a library of RL specialist policies cuts memory 20-160x vs full fine-tuning.", "relevance": 3, "relevance_reason": "Memory-efficient multi-task policy adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"robot\" AND abs:\"parameter-efficient fine-tuning\""}
{"id": "2606.15869", "title": "Metis: A Generalizable and Efficient World-Action Model for Autonomous Driving and Urban Navigation", "authors": ["Jingyu Li", "Zhe Liu", "Dongnan Hu", "Junjie Wu", "Zipei Ma", "Wenxiao Wu", "Chao Han", "Zhihui Hao", "Zhikang Liu", "Kun Zhan", "Jiankang Deng", "Xiatian Zhu", "Li Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.15869", "pdf_url": "https://arxiv.org/pdf/2606.15869", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["driving WAM", "asymmetric attention mask", "video bypass", "efficient"], "model_class": "world-action model", "techniques": ["MoT with separate video/action experts", "asymmetric mask for action-only inference", "avoid test-time future prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM navhard/navtest", "CityWalker", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Driving/navigation WAM whose action expert can skip explicit video generation at inference thanks to an asymmetric attention mask used during joint training.", "relevance": 3, "relevance_reason": "Efficient WAM for driving/navigation.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2605.12624", "title": "MindVLA-U1: VLA Beats VA with Unified Streaming Architecture for Autonomous Driving", "authors": ["Yuzhou Huang", "Benjin Zhu", "Hengtong Lu", "Victor Shea-Jay Huang", "Haiming Zhang", "Wei Chen", "Jifeng Dai", "Yang-Jiang Xie", "Hongsheng Li"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.12624", "pdf_url": "https://arxiv.org/pdf/2605.12624", "code_url": "", "clusters": ["C1"], "keywords": ["MindVLA-U1", "autonomous driving"], "model_class": "VLA", "techniques": ["streaming architecture", "few-step diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "16 FPS", "metric_measures": "throughput", "benchmarks": ["WOD-E2E"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Streaming driving VLA processing frames incrementally with learned memory and 2 diffusion steps, matching action-only model throughput.", "relevance": 3, "relevance_reason": "Real-time embodied VLA design.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2605.30263", "title": "minWM: A Full-Stack Open-Source Framework for Real-Time Interactive Video World Models", "authors": ["Min Zhao", "Hongzhou Zhu", "Bokai Yan", "Zihan Zhou", "Yimin Chen", "Honglie Wang", "Wenqiang Sun", "Zhengwei Fang", "Zizheng Xun", "Zihao Li", "Kaiwen Zheng", "Guande He", "Xiao Yang", "Chongxuan Li", "Fan Bao", "Jun Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30263", "pdf_url": "https://arxiv.org/pdf/2605.30263", "code_url": "", "clusters": ["C8"], "keywords": ["open-source framework", "causal forcing", "few-step AR distillation", "streaming"], "model_class": "world model", "techniques": ["AR diffusion training", "causal ODE/consistency distillation", "asymmetric DMD", "streaming inference"], "base_model": "Wan2.1-T2V-1.3B; HY1.5-TI2V-8B", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "End-to-end open pipeline that turns bidirectional video diffusion models into camera-controllable few-step autoregressive world models for low-latency rollout.", "relevance": 3, "relevance_reason": "Tooling for real-time interactive world models (non-robot actions).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.04657", "title": "MobileWAM: Bridging World Action Models to Mobile Manipulation with Chain-of-Foresight", "authors": ["Ze-Hua Fan", "Junjie He", "Wen-Xuan Song", "Xi Wang", "Wenqi Lyu", "Lin Zhao", "Fuhao Li", "Zi-Han You", "Yifei Yang", "Kai-Ming Xu", "Qihui Jiang", "Yue Jiang", "Hao-Ang Li", "Cheng Chi", "Feng Gao", "Bailin Li", "Yan Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.04657", "pdf_url": "https://arxiv.org/pdf/2608.04657", "code_url": "", "clusters": ["C8"], "keywords": ["MobileWAM", "mobile manipulation"], "model_class": "world-action model", "techniques": ["MoT", "training-only foresight"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill-HAB", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Mobile-manipulation WAM whose foresight chain is used only as training signal, so deployment runs at policy-level cost without video generation.", "relevance": 3, "relevance_reason": "Removes video-generation cost at inference.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2509.20869", "title": "Model-Based Reinforcement Learning under Random Observation Delays", "authors": ["Armin Karamzade", "Kyungmin Kim", "JB Lanier", "Davide Corsi", "Roy Fox"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.20869", "pdf_url": "https://arxiv.org/pdf/2509.20869", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["random observation delays", "model-based RL", "Dreamer", "out-of-sequence observations"], "model_class": "world model", "techniques": ["delay-aware belief filtering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Handles randomly delayed, out-of-order observations in model-based RL via a filtering process that updates the belief as observations arrive, applied to Dreamer.", "relevance": 3, "relevance_reason": "Delay-aware world-model control.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"random delays\" reinforcement learning"}
{"id": "2606.10825", "title": "MODIP: Efficient Model-Based Optimization for Diffusion Policies", "authors": ["Zakariae El Asri", "Philippe Gratias-Quiquandon", "Nicolas Thome", "Olivier Sigaud"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.10825", "pdf_url": "https://arxiv.org/pdf/2606.10825", "code_url": "", "clusters": ["C8"], "keywords": ["diffusion policy fine-tuning", "world model MPC", "terminal value"], "model_class": "world model", "techniques": ["MPC in world model generates BC targets", "terminal state value for cheaper planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["D4RL", "RoboMimic"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Fine-tunes diffusion policies with trajectories planned in a world model, using a terminal state value to cut MPC inference cost.", "relevance": 3, "relevance_reason": "Efficient world-model planning to improve policies.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.04056", "title": "Modular Safety Guardrails Are Necessary for Foundation-Model-Enabled Robots in the Real World", "authors": ["Joonkyung Kim", "Wenxi Chen", "Davood Soleymanzadeh", "Yi Ding", "Xiangbo Gao", "Zhengzhong Tu", "Ruqi Zhang", "Fan Fei", "Sushant Veer", "Yiwei Lyu", "Minghui Zheng", "Yan Gu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.04056", "pdf_url": "https://arxiv.org/pdf/2602.04056", "code_url": "", "clusters": ["C7"], "keywords": ["modular safety guardrails", "monitoring", "intervention", "position paper"], "model_class": "other", "techniques": ["position/survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Position paper arguing foundation-model robots need modular monitoring and intervention guardrails spanning action, decision and human-centered safety.", "relevance": 3, "relevance_reason": "Framework for runtime safety layers.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: safe robot foundation models"}
{"id": "2603.15265", "title": "MoE-ACT: Scaling Multi-Task Bimanual Manipulation with Sparse Task-Conditioned Mixture-of-Experts Transformers", "authors": ["Kang-Jun Guo", "Hai-Chao Liu", "Yan-Ji Sun", "Ru-Han Zhao", "Jin-Ni Zhou", "Jun Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.15265", "pdf_url": "https://arxiv.org/pdf/2603.15265", "code_url": "", "clusters": ["C1"], "keywords": ["MoE", "ACT", "parameter efficiency"], "model_class": "other", "techniques": ["mixture of experts"], "base_model": "", "params": "195M activated", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Sparse MoE extension of ACT for multi-task bimanual manipulation that beats a much larger VLA with far fewer active parameters.", "relevance": 3, "relevance_reason": "Parameter-efficient deployable policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2602.13764", "title": "MOTIF: Learning Action Motifs for Few-shot Cross-Embodiment Transfer", "authors": ["Heng Zhi", "Wentao Tan", "Lei Zhu", "Fengling Li", "Jingjing Li", "Guoli Yang", "Heng Tao Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.13764", "pdf_url": "https://arxiv.org/pdf/2602.13764", "code_url": "", "clusters": ["C3"], "keywords": ["MOTIF", "action motifs", "cross-embodiment few-shot"], "model_class": "VLA", "techniques": ["vector quantization", "few-shot transfer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Learns embodiment-agnostic action motifs for few-shot cross-embodiment VLA transfer.", "relevance": 3, "relevance_reason": "Cheap cross-embodiment adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"few-shot\""}
{"id": "2606.06245", "title": "MPCoT: Reward-Guided Multi-Path Latent Reasoning for Test-Time Scalable Vision-Language-Action", "authors": ["Boyang Zhang", "Lianlei Shan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.06245", "pdf_url": "https://arxiv.org/pdf/2606.06245", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["latent reasoning", "test-time scaling", "configurable depth"], "model_class": "VLA", "techniques": ["weight-tied latent reasoning", "multi-path aggregation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MPCoT refines multiple latent hypotheses for K weight-tied steps, providing configurable test-time compute without token-level CoT latency.", "relevance": 3, "relevance_reason": "Adjustable compute for VLA reasoning.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2607.07403", "title": "Multi-Agent Robotic Control with Onboard Vision-Language Models", "authors": ["Kajetan Rachwał", "Maciej Majek", "Bartłomiej Boczek", "Jakub Matejczyk", "Dominik Matejkowski", "Adam Dąbrowski", "Tim Seyde", "Alexander Amini", "Maria Ganzha"], "year": 2026, "venue": "PAAMS 2026", "type": "paper", "url": "https://arxiv.org/abs/2607.07403", "pdf_url": "https://arxiv.org/pdf/2607.07403", "code_url": "", "clusters": ["C6"], "keywords": ["onboard VLMs", "multi-agent system", "no external compute"], "model_class": "other", "techniques": ["onboard compact VLMs", "multi-agent orchestration"], "base_model": "compact VLMs 3-20B", "params": "", "hardware": "AMD Ryzen AI mini PC", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Fully onboard multi-agent system of compact VLMs controlling a warehouse mobile manipulator, argued as a cost-efficient alternative to cloud-dependent deployment.", "relevance": 3, "relevance_reason": "Onboard-vs-cloud cost argument.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud AND abs:vision-language-action"}
{"id": "2606.21372", "title": "NAC: Neural Action Codec for Vision-Language-Action Models", "authors": ["Ahad Jawaid", "Yu Xiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.21372", "pdf_url": "https://arxiv.org/pdf/2606.21372", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["action tokenizer", "neural codec", "compression"], "model_class": "VLA", "techniques": ["convolutional neural action codec"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "NAC borrows neural audio codec designs for action tokenization, balancing compression, latency and downstream performance.", "relevance": 3, "relevance_reason": "Token-efficient action representation.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2605.11665", "title": "Nautilus: From One Prompt to Plug-and-Play Robot Learning", "authors": ["Yufeng Jin", "Jianfei Guo", "Xiao Jia", "Yu Deng", "Ze-Chu Li", "Hange Liu", "Wei-Ran Liao", "Vignesh Prasad", "M. Franzius", "Gerhard Neumann", "Georgia Chalvatzaki"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.11665", "pdf_url": "https://arxiv.org/pdf/2605.11665", "code_url": "", "clusters": ["C4"], "keywords": ["reproducibility", "evaluation harness"], "model_class": "other", "techniques": ["agentic tooling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open-source agentic harness that turns a single prompt into reproducible evaluation, fine-tuning and deployment pipelines across policies, benchmarks and robots via typed interfaces.", "relevance": 3, "relevance_reason": "Composability/benchmarking infrastructure for cross-policy evaluation (C4).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2607.14643", "title": "NavCMPO: Critic-Guided MeanFlow Policy Optimization for Adaptive Navigation", "authors": ["Junjie An", "Yi Wu", "Xiao Liu", "Yiqun Zhou", "Yue-Chen Wu", "Xiaoqing Guan", "You Wang", "Guang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.14643", "pdf_url": "https://arxiv.org/pdf/2607.14643", "code_url": "", "clusters": ["C1"], "keywords": ["navigation", "critic guidance"], "model_class": "diffusion/flow policy", "techniques": ["few-step MeanFlow", "PPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "inference latency reduced from 85 ms to 60 ms", "metric_measures": "other", "benchmarks": ["InternVLA-N1", "Unitree Go2"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Few-step MeanFlow navigation policy with critic-guided refinement and PPO adaptation.", "relevance": 3, "relevance_reason": "Reduces diffusion-policy latency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2409.00588"}
{"id": "2608.10824", "title": "Neural Introspection Gating for Adaptive KV-Cache Reuse in Vision-Language-Action Models", "authors": ["Zhijie Wu", "Kento Kawaharazuka", "Kei Okada"], "year": 2026, "venue": "IROS 2026", "type": "paper", "url": "https://arxiv.org/abs/2608.10824", "pdf_url": "https://arxiv.org/pdf/2608.10824", "code_url": "", "clusters": ["C1"], "keywords": ["KV cache reuse", "confidence gating"], "model_class": "VLA", "techniques": ["KV cache reuse"], "base_model": "OpenVLA, OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Extends VLA-Cache by invalidating the cache whenever the top-2 action-token logit margin drops, recovering accuracy lost by blind caching while keeping most savings.", "relevance": 3, "relevance_reason": "KV reuse for VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "2608.15002", "title": "NPU Offloading of a Frozen Visual Encoder for Robot Policy Training", "authors": ["Hyojun Yun", "Seungjae Won", "Hyungpil Moon"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15002", "pdf_url": "https://arxiv.org/pdf/2608.15002", "code_url": "", "clusters": ["C4", "C5", "C9"], "keywords": ["NPU offloading", "energy per sample", "training", "Mobilint Aries2", "training energy", "INT8", "frozen encoder", "energy"], "model_class": "other", "techniques": ["INT8 NPU offload", "heterogeneous GPU+NPU training", "INT8 encoder offload", "A8W8 INT8 encoder"], "base_model": "", "params": "", "hardware": "Mobilint Aries2 NPU + RTX 5060 Ti", "reported_metric": "energy per sample -17.1% to -27.9%; time per sample +15.2% to +37.7%", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Runs a frozen visual encoder in INT8 on a low-power NPU while training the action expert on GPU, trading longer training for lower energy and slightly lower success.", "relevance": 3, "relevance_reason": "Heterogeneous accelerator use for robot policy workloads.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy energy; web: Qualcomm Snapdragon NPU on-device robot policy VLA; venue sweep: arXiv API robot AND (Trainium OR Inferentia OR NPU OR TPU)"}
{"id": "2606.03159", "title": "NVIDIA OmniDreams: Real-Time Generative World Model for Closed-Loop Autonomous Vehicle Simulation", "authors": ["Aarti Basant", "Amlan Kar", "Despoina Paschalidou", "Fangyin Wei", "Francesco Ferroni", "Guillermo Garcia Cobo", "Haithem Turki", "Huan Ling", "Jaewoo Seo", "James Lucas", "Jay Zhangjie Wu", "Jialiang Wang", "Jonathan Lorraine", "Jun Gao", "Kai He", "Katarina Tothova", "Kevin Xie", "Michal Tyszkiewicz", "Qi Wu", "Riccardo de Lutio", "Ruilong Li", "Sanja Fidler", "Seung Wook Kim", "Tianchang Shen", "Tianshi Cao", "Tobias Pfaff", "William Lew", "Xindi Wu", "Xuanchi Ren", "Yifan Lu", "Yuxuan Zhang", "Zan Gojcic", "Zian Wang"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2606.03159", "pdf_url": "https://arxiv.org/pdf/2606.03159", "code_url": "https://github.com/nv-tlabs/omni-dreams", "clusters": ["C8"], "keywords": ["OmniDreams", "closed-loop simulation"], "model_class": "world model", "techniques": ["autoregressive video generation", "post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Real-time action-conditioned driving world model post-trained from Cosmos for closed-loop policy evaluation with Alpamayo 1.", "relevance": 3, "relevance_reason": "Real-time generative simulator for policy evaluation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2609.24155", "title": "Object-Centric Conditioning for Visuomotor Flow Matching", "authors": ["Ji-Jie Li", "Xu Yang", "Jun-Hong Zou", "Chun-Ming Zhao", "Chao-Yang Zhao", "Zhen Lei", "Xiang-Yu Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24155", "pdf_url": "https://arxiv.org/pdf/2609.24155", "code_url": "", "clusters": ["C1"], "keywords": ["object-centric conditioning", "low-step inference"], "model_class": "diffusion/flow policy", "techniques": ["action-to-action flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adds object-centric conditioning to action-to-action flow matching while keeping its low-step inference efficiency.", "relevance": 3, "relevance_reason": "Preserves fast few-step flow inference", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.18953", "title": "Object-Centric Residual RL for Zero-Shot Sim-to-Real VLA Enhancement", "authors": ["Kinam Kim", "Namiko Saito", "Heecheol Kim", "Katsushi Ikeuchi", "Jaegul Choo", "Yasuyuki Matsushita"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18953", "pdf_url": "https://arxiv.org/pdf/2606.18953", "code_url": "", "clusters": ["C3"], "keywords": ["object-centric residual RL", "sim-to-real", "frozen VLA"], "model_class": "VLA", "techniques": ["residual RL", "object-pose observations"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Sim-trained object-centric residual RL corrects a frozen VLA's actions and transfers zero-shot to the real world.", "relevance": 3, "relevance_reason": "Avoids costly real RL; frozen base.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: ti:VLA AND ti:RL"}
{"id": "2605.03065", "title": "OGPO: Sample Efficient Full-Finetuning of Generative Control Policies", "authors": ["Sarvesh Patil", "Mitsuhiko Nakamoto", "Manan Agarwal", "Shashwat Saxena", "Jesse Zhang", "Giri Anantharaman", "Cleah Winston", "Chaoyi Pan", "Douglas Chen", "Nai-Chieh Huang", "Zeynep Temel", "Oliver Kroemer", "Sergey Levine", "Abhishek Gupta", "Hongkai Dai", "Paarth Shah", "Max Simchowitz"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.03065", "pdf_url": "https://arxiv.org/pdf/2605.03065", "code_url": "", "clusters": ["C3"], "keywords": ["OGPO", "off-policy", "full fine-tuning"], "model_class": "diffusion/flow policy", "techniques": ["off-policy RL", "modified PPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["manipulation simulation"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Off-policy critics plus PPO-style gradients through the full generative process for sample-efficient fine-tuning of diffusion/flow policies.", "relevance": 3, "relevance_reason": "Sample-efficient RL fine-tuning of generative policies.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: diffusion steering latent space RL DSRL fine-tune pi0 noise policy"}
{"id": "2602.23843", "title": "OmniXtreme: Breaking the Generality Barrier in High-Dynamic Humanoid Control", "authors": ["Yun-Sheng Wang", "Shao-Hang Zhu", "Peiyuan Zhi", "Yuhan Li", "Jiaxi Li", "Yong-Lu Li", "Yuchen Xiao", "Xingxing Wang", "Baoxiong Jia", "Si-Yuan Huang"], "year": 2026, "venue": "arXiv (Robotics)", "type": "preprint", "url": "https://arxiv.org/abs/2602.23843", "pdf_url": "https://arxiv.org/pdf/2602.23843", "code_url": "", "clusters": ["C5"], "keywords": ["Unitree G1", "Orin NX", "TensorRT", "flow matching", "high-dynamic humanoid"], "model_class": "diffusion/flow policy", "techniques": ["TensorRT full-pipeline", "flow-matching policy + residual policy"], "base_model": "", "params": "", "hardware": "Unitree G1 onboard Jetson Orin NX", "reported_metric": "~10 ms end-to-end inference; 50 Hz control", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "High-capacity flow-matching humanoid tracking policy whose full pipeline runs in TensorRT on the G1's onboard Orin NX at 50 Hz.", "relevance": 3, "relevance_reason": "Concrete onboard Orin NX numbers for a generative humanoid policy.", "foundational_technique": false, "verified": true, "found_via": "web: Jetson Orin quadruped/humanoid onboard inference"}
{"id": "2601.06748", "title": "On-the-Fly VLA Adaptation via Test-Time Reinforcement Learning", "authors": ["Changyu Liu", "Yiyang Liu", "Taowen Wang", "Qiao Zhuang", "James Chenhao Liang", "Wenhao Yang", "Renjing Xu", "Qifan Wang", "Dongfang Liu", "Cheng Han", "Chang-Yu Liu", "Tao-Wen Wang", "J. Liang", "Ren-Jing Xu", "Qi-Fan Wang"], "year": 2026, "venue": "Annual Meeting of the Association for Computational Linguistics", "type": "preprint", "url": "https://arxiv.org/abs/2601.06748", "pdf_url": "https://arxiv.org/pdf/2601.06748", "code_url": "", "clusters": ["C3"], "keywords": ["on-the-fly adaptation", "test-time adaptation", "RL"], "model_class": "VLA", "techniques": ["test-time RL", "progress rewards"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts VLAs on the fly at deployment with test-time reinforcement learning, avoiding separate fine-tuning phases or human interventions.", "relevance": 3, "relevance_reason": "Deployment-time adaptation without separate fine-tuning phases.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (interrupt* OR \"human intervention\" OR \"mid-execution\"); snowball round 2: cites 2405.12213, 2410.00425; cited by 2605.11750"}
{"id": "2605.01069", "title": "Online Safety Filter for Deformable Object Manipulation with Horizon Agnostic Neural Operators", "authors": ["Jiaxing Li", "Hanjiang Hu", "Zhuoyuan Wang", "Yorie Nakahira", "Changliu Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.01069", "pdf_url": "https://arxiv.org/pdf/2605.01069", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["online safety filter", "neural operator", "boundary CBF", "deformable objects"], "model_class": "other", "techniques": ["neural operator dynamics", "CBF-QP filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Real-time safety filter minimally modifying any nominal policy for deformable manipulation, using horizon-agnostic neural operators and a boundary CBF QP.", "relevance": 3, "relevance_reason": "Lightweight runtime safety filter for learned policies.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2606.31993", "title": "OopsieVerse: A Safety Benchmark with Damage-Aware Simulation for Robot Manipulation", "authors": ["Arnav Balaji", "Arpit Bahety", "Sriniket Ambatipudi", "Daniel Lam", "Junhong Xu", "Roberto Martín-Martín"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2606.31993", "pdf_url": "https://arxiv.org/pdf/2606.31993", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["OopsieVerse", "damage-aware"], "model_class": "VLA", "techniques": ["safety benchmark"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Damage-aware simulation benchmark for household manipulation used to evaluate and improve VLA physical safety.", "relevance": 3, "relevance_reason": "Safety evaluation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.15185, 2410.24164"}
{"id": "2606.29936", "title": "OpenSPM: An Environment-Transferable Robotic Key Spatial Pose Memory and Closed-Loop High-Frequency Flow-Matching Action Generation Model", "authors": ["Iok Tong Lei", "Qingchen Xie", "Yi-Fan Wang", "Yap Ying Jie", "Zhidong Deng"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.29936", "pdf_url": "https://arxiv.org/pdf/2606.29936", "code_url": "", "clusters": ["C1"], "keywords": ["high-frequency action generation"], "model_class": "diffusion/flow policy", "techniques": ["lightweight flow matching", "spatial pose memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1033.3 Hz equivalent control frequency", "metric_measures": "other", "benchmarks": ["LIBERO-Goal"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "OpenSPM combines retrieved object-centric pose memory with a lightweight flow-matching generator for high-frequency tabletop control.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "github-omega1029-openvla-oft-quantization", "title": "openvla-oft-quantization: closed-loop quantization of OpenVLA-OFT on LIBERO", "authors": ["Omega1029"], "year": 2026, "venue": "GitHub (IROS 2026 ORS workshop abstract)", "type": "repo", "url": "https://github.com/Omega1029/openvla-oft-quantization", "pdf_url": "", "code_url": "https://github.com/Omega1029/openvla-oft-quantization", "clusters": ["C2", "C4"], "keywords": ["OpenVLA-OFT", "W4A4", "rotation", "LIBERO", "Jetson AGX Orin"], "model_class": "VLA", "techniques": ["W4A4 fake quantization", "orthogonal rotation (Hadamard/DCT/PCA)", "AWQ", "mixed precision"], "base_model": "OpenVLA-OFT", "params": "7B", "hardware": "A100; Jetson AGX Orin", "reported_metric": "W4A4 lossless on LIBERO with backbone rotation; 0% success without rotation (simulated quantization, no memory/latency gain)", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Scores OpenVLA-OFT PTQ variants by LIBERO closed-loop rollouts and finds that uniform W4A4 works only if the language backbone is rotated first; the choice of rotation barely matters.", "relevance": 3, "relevance_reason": "LIBERO-level evidence for 4-bit OpenVLA-class viability; fake quant, so deployment gains unmeasured.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub repo search openvla+jetson"}
{"id": "github-openvla-serve-roofline", "title": "openvla-serve: low-latency serving, profiling, and kernel study for OpenVLA-7B", "authors": ["subhashkotaru"], "year": 2026, "venue": "GitHub", "type": "repo", "url": "https://github.com/subhashkotaru/openvla-serve", "pdf_url": "", "code_url": "https://github.com/subhashkotaru/openvla-serve", "clusters": ["C4"], "keywords": ["Nsight", "GEMV", "memory-bandwidth floor", "torch.compile", "CUDA graphs"], "model_class": "VLA", "techniques": ["torch.compile", "CUDA graphs", "Triton fused add+RMSNorm"], "base_model": "OpenVLA-7B", "params": "7B", "hardware": "NVIDIA L40S", "reported_metric": "bf16 peak 14.76 GB, 5.19 actions/s; decode 16.33 ms/token achieved vs 15.6 ms/token HBM floor (13.48 GB/864 GB/s); torch.compile+CUDA graphs 190 -> 177 ms (~7%)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Independent profiling showing OpenVLA-7B decode sits at ~95% of the memory-bandwidth floor on an L40S, so kernel fusion and CUDA graphs only trim overhead (~7%).", "relevance": 3, "relevance_reason": "Worked roofline/bandwidth-floor analysis of a VLA with measured numbers.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub repo search openvla+quantization"}
{"id": "intel-oep-pi05-core-ultra", "title": "Optimizing pi0.5 Vision-Language-Action Robotic Model on Intel Core Ultra Series 3 Processor", "authors": ["Asaad F Said", "Radwan Ibrahim"], "year": 2026, "venue": "Intel Open Edge Platform documentation", "type": "blog", "url": "https://docs.openedgeplatform.intel.com/2026.1/OEP-articles/publications/optimizing-pi0.5-lva-model.html", "pdf_url": "", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["OpenVINO", "heterogeneous CPU-GPU-NPU", "ISO-TDP comparison", "perf per watt"], "model_class": "VLA", "techniques": ["OpenVINO", "heterogeneous execution", "KV-cache handoff"], "base_model": "pi0.5 (DROID)", "params": "", "hardware": "Intel Core Ultra X7 358H; Jetson Thor; Jetson AGX Orin", "reported_metric": "294 ms at 40 W (Core Ultra) vs 287 ms Thor, 1.53 s AGX Orin at 40 W; up to 2.6x perf/W vs AGX Orin (3 cams, 10 denoising steps)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Vendor tech note mapping pi0.5 across iGPU and NPU with OpenVINO and comparing latency at equal TDP to Jetson devices, with the input configuration stated.", "relevance": 3, "relevance_reason": "ISO-power hardware-normalized comparison example; vendor source, author list partial.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: VLA deployment edge devices benchmark"}
{"id": "2607.21670", "title": "Ordered Action Tokens for Visuomotor Policy Learning", "authors": ["Chaoqi Liu", "Yue Zhao", "Hao-Nan Chen", "Xiao-Shen Han", "Jia-Wei Gao", "E. Adeli", "Yi-Lun Du"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.21670", "pdf_url": "https://arxiv.org/pdf/2607.21670", "code_url": "", "clusters": ["C1"], "keywords": ["action tokenization", "anytime inference"], "model_class": "other", "techniques": ["learned action tokenizer", "finite scalar quantization", "ordered tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learned action tokenizer with ordered, prefix-decodable tokens giving an anytime trade-off between inference cost and action fidelity.", "relevance": 3, "relevance_reason": "Action tokenization with explicit compute/fidelity trade-off.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2605.09656", "title": "ORICF -- Open Robotics Inference and Control Framework", "authors": ["Andrés Meseguer Valenzuela", "Luís Miguel Bartolín Arnau"], "year": 2026, "venue": "ICRA 2026 Workshop RoSE", "type": "paper", "url": "https://arxiv.org/abs/2605.09656", "pdf_url": "https://arxiv.org/pdf/2605.09656", "code_url": "", "clusters": ["C6"], "keywords": ["inference pipeline framework", "edge offloading", "ROS 2"], "model_class": "other", "techniques": ["declarative pipelines", "edge offloading"], "base_model": "ASR + LLM + CNN detector", "params": "", "hardware": "", "reported_metric": "up to 83.16% lower robot-side compute utilization and 65.8% lower estimated energy vs onboard", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "YAML-configured ROS 2 framework composing multimodal inference pipelines that can be moved from the robot to nearby edge computers without code changes.", "relevance": 3, "relevance_reason": "Edge offloading tooling with energy measurements.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2604.08266", "title": "Orion-Lite: Distilling LLM Reasoning into Efficient Vision-Only Driving Models", "authors": ["Jing Gu", "Niccolò Cavagnero", "Gijs Dubbelman"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.08266", "pdf_url": "https://arxiv.org/pdf/2604.08266", "code_url": "", "clusters": ["C2"], "keywords": ["vision-only driving"], "model_class": "other", "techniques": ["knowledge distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Bench2Drive"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills an LLM-based driving VLA teacher into a vision-only student using latent-feature distillation and trajectory supervision; the student surpasses the teacher in closed loop.", "relevance": 3, "relevance_reason": "Distillation removing LLM cost from driving policies.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2608.19589", "title": "OrthoSkillVLA: Continual Skill Learning via Gradient-Informed Skill Subspace Adaptation", "authors": ["Jiaqi Wang", "Zhou Fang", "Qiongfeng Shi", "Yi Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.19589", "pdf_url": "https://arxiv.org/pdf/2608.19589", "code_url": "", "clusters": ["C3"], "keywords": ["OrthoSkillVLA", "orthogonal subspace", "continual skills"], "model_class": "VLA", "techniques": ["subspace-constrained fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Component-specific gradient-informed subspace constraints for continual VLA skill learning without growing inference footprint.", "relevance": 3, "relevance_reason": "Continual adaptation with fixed model size.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"catastrophic forgetting\""}
{"id": "2606.04463", "title": "OSCAR: Omni-Embodiment Skeleton-Conditioned World Action Model for Robotics", "authors": ["Zhuo-Yuan Wu", "Jun Gao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.04463", "pdf_url": "https://arxiv.org/pdf/2606.04463", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["policy evaluation", "skeleton conditioning"], "model_class": "world-action model", "techniques": [], "base_model": "Cosmos-Predict2.5-2B", "params": "", "hardware": "single GH200", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboArena"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Skeleton-conditioned video world model fine-tuned from Cosmos-Predict2.5-2B on a single GPU for cross-embodiment virtual policy evaluation.", "relevance": 3, "relevance_reason": "World model for policy evaluation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941, 2410.24164"}
{"id": "2606.13856", "title": "Output-Level Regularization Eliminates the Seed Lottery in Single-GPU VLA Fine-Tuning", "authors": ["Jeffrin Sam", "Dzmitry Tsetserukou", "Sam, Jeffrin", "Tsetserukou, Dzmitry"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.13856", "pdf_url": "https://arxiv.org/pdf/2606.13856", "code_url": "", "clusters": ["C3", "C4"], "keywords": ["seed lottery", "output collapse", "single-GPU fine-tuning", "seed variance", "LIBERO-Object", "reproducibility"], "model_class": "VLA", "techniques": ["output-level regularization"], "base_model": "VLA-JEPA", "params": "", "hardware": "", "reported_metric": "LIBERO-Object success spans 65.2%-94.2% across 13 seeds (std 7.5%)", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Documents large seed-to-seed variance in single-GPU VLA fine-tuning on LIBERO (one in thirteen runs collapses) and proposes output-level regularizers; author notes the draft is still being revised.", "relevance": 3, "relevance_reason": "Seed variance must be accounted for when claiming small accuracy deltas from efficiency methods.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations); WebSearch: reproducibility crisis robot learning seeds variance VLA"}
{"id": "2605.11459", "title": "Overcoming Dynamics-Blindness: Training-Free Pace-and-Path Correction for VLA Models", "authors": ["Yan-Yan Zhang", "Chao-Da Song", "Vikash Singh", "Xinpeng Li", "Kai Ye", "Zhenfang Hu", "Zhongzhu Pu", "Yuhui Yin", "Vipin Chaudhary"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.11459", "pdf_url": "https://arxiv.org/pdf/2605.11459", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["pace-and-path correction", "dynamics-blindness", "training-free", "closed-form"], "model_class": "VLA", "techniques": ["closed-form inference-time correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free closed-form operator wrapping chunked VLAs that corrects pace and path for moving targets, addressing single-frame VLAs being blind to dynamics.", "relevance": 3, "relevance_reason": "Reactivity in non-stationary scenes.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2607.09590", "title": "PAC-ACT: Post-training Actor-Critic for Action Chunking Transformers", "authors": ["Yujie Pang", "Zudong Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.09590", "pdf_url": "https://arxiv.org/pdf/2607.09590", "code_url": "", "clusters": ["C3"], "keywords": ["ACT", "industrial contact"], "model_class": "other", "techniques": ["actor-critic post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RL post-training of Action Chunking Transformers for precision contact tasks while retaining low latency and memory.", "relevance": 3, "relevance_reason": "Efficient RL fine-tuning of lightweight policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.00537", "title": "PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking", "authors": ["Junnan Nie", "Jiayi Li", "Chenghao Liu", "Junyi Lao", "Jiachen Zhang", "Tianle Zhang", "Liang Lin", "Songfang Huang", "Nie, Junnan", "Li, Jiayi", "Liu, Chenghao", "Lao, Junyi", "Zhang, Jiachen", "Zhang, Tianle", "Lin, Liang", "Huang, Songfang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.00537", "pdf_url": "https://arxiv.org/pdf/2606.00537", "code_url": "", "clusters": ["C1", "C2", "C4"], "keywords": ["execution horizon", "phase-aware", "PACE", "non-monotonic success"], "model_class": "VLA", "techniques": ["adaptive action chunking", "training-free", "phase-aware chunk execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "ALOHA", "Franka real robot", "simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Shows success is task-dependent and non-monotonic in execution horizon, so a fixed horizon is unreliable; picks replanning points at low-speed transitions.", "relevance": 3, "relevance_reason": "Execution horizon is a hidden confound in reported control rates.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking; WebSearch: RoboChallenge Table30 (result list)"}
{"id": "2605.09860", "title": "PACE: Policy-Native Adaptive Decision Timing for Long-Horizon Reasoning", "authors": ["Chen Li", "Zhan-Tao Yang", "Fangyi Chen", "Hanwen Zhang", "Anudeepsekhar Bolimera", "Marios Savvides"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.09860", "pdf_url": "https://arxiv.org/pdf/2605.09860", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["PACE", "replanning"], "model_class": "other", "techniques": ["adaptive execution horizon"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns, from within the policy, how many actions to execute open-loop before replanning, trading replanning cost against error accumulation.", "relevance": 3, "relevance_reason": "Adaptive chunk execution timing.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2606.00515", "title": "PaCo-VLA: Passivity-Shielded Compliance Prior for Contact-Rich Vision-Language-Action Manipulation", "authors": ["Haofan Cao", "Zhao-Yang Li", "Zhi-Chao You", "Liang Guo", "Tianrui Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.00515", "pdf_url": "https://arxiv.org/pdf/2606.00515", "code_url": "", "clusters": ["C7"], "keywords": ["contact-rich safety", "low-rate VLA outputs"], "model_class": "VLA", "techniques": ["passivity shield", "energy tank"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Treats low-rate VLA outputs as compliance proposals checked by a high-frequency passivity shield that blocks stale or invalid predictions.", "relevance": 3, "relevance_reason": "Safety under low VLA output rates (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.08414", "title": "PACT: Self-Evolving Physical Safety Alignment for Diffusion Policies in Embodied Manipulation", "authors": ["Lingxuan Wu", "Zijian Zhu", "Lizhong Wang", "Chengyang Ying", "Hua-Yu Chen", "Xiao Yang", "Fang-Ming Liu", "Junyou Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08414", "pdf_url": "https://arxiv.org/pdf/2606.08414", "code_url": "", "clusters": ["C7"], "keywords": ["safety alignment", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["constraint distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Post-training framework that projects diffusion policies onto constraint-feasible regions by distilling constraint gradients.", "relevance": 3, "relevance_reason": "Safety of diffusion policies (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.17924", "title": "PearlVLA: Progressive Embodied Action-Plan Refinement in Latent Space", "authors": ["Bochen Yang", "Lianlei Shan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.17924", "pdf_url": "https://arxiv.org/pdf/2606.17924", "code_url": "", "clusters": ["C1"], "keywords": ["PearlVLA", "latent deliberation", "low latency"], "model_class": "VLA", "techniques": ["latent reasoning", "latent world model", "latent plan refinement", "parallel decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Moves VLA deliberation into VLM latent space, iteratively refining a latent plan with a frozen lightweight latent world model instead of textual or pixel reasoning.", "relevance": 3, "relevance_reason": "Latent reasoning to avoid explicit-reasoning latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]; snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2609.12036", "title": "Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence", "authors": ["Shilong Zou", "Shi-Lin Zhang", "Yingji Zhang", "Yu-Han Huang", "Yi Zhang", "Ze-Yuan Ding", "Han-Ning Dong", "Junwei Liao", "Yong Dai", "Jian Tang", "Xiao-Zhu Ju"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.12036", "pdf_url": "https://arxiv.org/pdf/2609.12036", "code_url": "", "clusters": ["C8"], "keywords": ["world model simulator", "policy evaluation"], "model_class": "world model", "techniques": ["few-step distillation", "sparse MoE", "causal adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "5.67x speedup over 35-step model", "metric_measures": "relative latency", "benchmarks": ["AgiBotWorld", "RoboMIND", "RoboTwin"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Pelican-Sim is a cross-embodiment world-model simulator distilled into a four-step autoregressive rollout model for faster generation.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2606.20754", "title": "Perturbation-Based Epistemic Uncertainty for Failure Detection in Vision-Language-Action Models", "authors": ["Yousung Lee", "Dongsoo Har"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.20754", "pdf_url": "https://arxiv.org/pdf/2606.20754", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["epistemic uncertainty", "low-rank weight perturbation", "failure detection"], "model_class": "VLA", "techniques": ["low-rank weight perturbation ensembles"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Estimates VLA epistemic uncertainty from disagreement among actions under random low-rank weight perturbations for training-free failure detection.", "relevance": 3, "relevance_reason": "Uncertainty-based runtime monitoring (extra compute cost).", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2603.08342", "title": "PhaForce: Phase-Scheduled Visual-Force Policy Learning with Slow Planning and Fast Correction for Contact-Rich Manipulation", "authors": ["Mingxin Wang", "Zhirun Yue", "Renhao Lu", "Yizhe Li", "Zihan Wang", "Guoping Pan", "Kangkang Dong", "Jun Cheng", "Yi Cheng", "Houde Liu", "Zi-Han Wang", "Junlei Cheng", "Hou-De Liu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.08342", "pdf_url": "https://arxiv.org/pdf/2603.08342", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["visual-force policy", "chunk planning", "PhaForce", "force feedback"], "model_class": "diffusion/flow policy", "techniques": ["slow-fast control", "residual correction", "force fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Couples a low-rate diffusion chunk planner with a high-rate force-driven residual corrector to overcome latency-limited update rates in contact-rich tasks.", "relevance": 3, "relevance_reason": "Fast correction layered under slow chunk planning.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*); snowball round 2: cites 2303.04137; cited by 2608.15938"}
{"id": "2606.27146", "title": "PhysReflect-VLA: Physical Feasibility and Self-Reflective Regulation for Reliable Vision-Language-Action Policies", "authors": ["Jiayu Yang", "Tao Yang", "Weijun Li", "Xiang Chang", "Fei Chao", "Changjing Shang", "Qiang Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.27146", "pdf_url": "https://arxiv.org/pdf/2606.27146", "code_url": "", "clusters": ["C7"], "keywords": ["execution-time reliability"], "model_class": "VLA", "techniques": ["feasibility checking", "self-reflection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a feasibility operator, transition-coherence check and LLM reflection module around a VLA to catch physically infeasible actions and generate corrective guidance online.", "relevance": 3, "relevance_reason": "Execution-time feasibility checking of VLA actions.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (\"failure recovery\" OR \"error recovery\" OR \"self-correct*\")"}
{"id": "2606.13886", "title": "PhysVLA: Towards Physically-Grounded VLA for Embodied Robotic Manipulation", "authors": ["N. Chandra", "Shriram Damodaran", "Lin Wang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.13886", "pdf_url": "https://arxiv.org/pdf/2606.13886", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["PhysVLA", "trajectory smoothing"], "model_class": "VLA", "techniques": ["inference-time wrapper"], "base_model": "", "params": "", "hardware": "", "reported_metric": "<1 ms overhead per control step", "metric_measures": "other", "benchmarks": ["LIBERO-Spatial", "Robosuite"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Plug-in inference-time wrapper for any frozen VLA that improves physical stability and success with sub-millisecond per-step overhead.", "relevance": 3, "relevance_reason": "Low-overhead runtime stabilization.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2304.13705"}
{"id": "2601.03782", "title": "PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation", "authors": ["Wenlong Huang", "Yu-Wei Chao", "Arsalan Mousavian", "Ming-Yu Liu", "Dieter Fox", "Kaichun Mo", "Li Fei-Fei", "Wen-Bing Huang", "A. Mousavian", "Fei-Fei Li"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2601.03782", "pdf_url": "https://arxiv.org/pdf/2601.03782", "code_url": "", "clusters": ["C8"], "keywords": ["PointWorld", "3D world model"], "model_class": "world model", "techniques": ["3D point flow prediction", "3D point-flow world model", "MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "0.1 s inference", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Large pretrained 3D world model predicting per-point 3D flow from RGB-D and actions, fast enough to sit inside an MPC loop for zero-shot real manipulation.", "relevance": 3, "relevance_reason": "Real-time world-model inference used for planning.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight); snowball round 2: cites 2303.04137; cited by 2602.15922"}
{"id": "2606.29908", "title": "Pondering the Way: Spatial-perceiving World Action Model for Embodied Navigation", "authors": ["Hong Chen", "Daqi Liu", "Ze-Han Zhang", "Haiguang Wang", "Tianhao Lu", "Longfei Yan", "Haiyang Sun", "Fang Li", "Hong-Bin Xie", "Bing Wang", "Guang Chen", "Hang-Jun Ye", "Yihua Tan"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.29908", "pdf_url": "https://arxiv.org/pdf/2606.29908", "code_url": "", "clusters": ["C8"], "keywords": ["navigation", "SWAM"], "model_class": "world-action model", "techniques": ["single-pass generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Navigation world-action model generating RGB-D foresight and trajectories in one pass, outperforming two-stage planners in inference efficiency.", "relevance": 3, "relevance_reason": "Efficient WAM inference.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2608.24115", "title": "PonderPounce: A Pretrained MLLM as an Episode Context Engine for Robot Control", "authors": ["Suhwan Choi", "Jaeyoon Jung", "Sungkyung Kim", "Yunsung Lee", "Youngjae Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.24115", "pdf_url": "https://arxiv.org/pdf/2608.24115", "code_url": "", "clusters": ["C1", "C6"], "keywords": ["PonderPounce", "System 1/System 2", "episode memory", "MLLM"], "model_class": "VLA", "techniques": ["asynchronous dual system", "episode memory", "asynchronous dual-system"], "base_model": "", "params": "", "hardware": "H100", "reported_metric": "p50 78 ms cognition refresh, 25 ms action-model call | p50 latency 78 ms (cognition refresh), 25 ms (action model call)", "metric_measures": "other", "benchmarks": ["RoboMME", "RoboCasa", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Asynchronous System-2 MLLM episode-context engine feeding a System-1 action model that conditions on the newest cognition and its age.", "relevance": 3, "relevance_reason": "Asynchronous fast-slow VLA with reported per-call latencies", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"action chunk*\") AND (asynchronous OR \"inference delay\" OR \"latency compensation\"); snowball round 2: cites 2402.15391, 2410.24164"}
{"id": "github-openvla-issue-324-fp8-nvfp4-libero", "title": "Post-Training Quantization (FP8/NVFP4) on OpenVLA with LIBERO Benchmark (openvla/openvla issue #324)", "authors": ["jtao34"], "year": 2026, "venue": "GitHub issue", "type": "repo", "url": "https://github.com/openvla/openvla/issues/324", "pdf_url": "", "code_url": "", "clusters": ["C2", "C4"], "keywords": ["OpenVLA", "ModelOpt", "FP8", "NVFP4", "LIBERO"], "model_class": "VLA", "techniques": ["PTQ FP8", "PTQ NVFP4 (fake-quant, dequantized to BF16 at runtime)"], "base_model": "openvla-7b-finetuned-libero-spatial / libero-10", "params": "7B", "hardware": "", "reported_metric": "model size 7.68 GB (FP8) / 4.44 GB (NVFP4); 'no significant drop' in LIBERO-Spatial/LIBERO-10 success (500 episodes each; numbers only in an image)", "metric_measures": "relative memory", "benchmarks": ["LIBERO-Spatial", "LIBERO-10"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Community report applying ModelOpt FP8 and NVFP4 PTQ to linear layers of LIBERO-finetuned OpenVLA and seeing no significant LIBERO success drop, though weights are dequantized so no speed/memory gain is realized at runtime.", "relevance": 3, "relevance_reason": "Direct LIBERO evidence for 4-bit-class OpenVLA accuracy; simulated quantization only, numbers not machine-readable.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub issues repo:openvla/openvla quantization"}
{"id": "2604.01860", "title": "Posterior Optimization with Clipped Objective for Bridging Efficiency and Stability in Generative Policy Learning", "authors": ["Yuhui Chen", "Haoran Li", "Zhennan Jiang", "Yu-Xin Qin", "Yu Wan", "Weiheng Liu", "Dong-Bin Zhao"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2604.01860", "pdf_url": "https://arxiv.org/pdf/2604.01860", "code_url": "", "clusters": ["C3"], "keywords": ["POCO", "offline-to-online"], "model_class": "VLA", "techniques": ["RL fine-tuning", "posterior optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "EM-style RL fine-tuning for chunked generative/VLA policies that distills a reward-weighted posterior, aiming at sample efficiency and stability.", "relevance": 3, "relevance_reason": "Efficient, stable RL fine-tuning of generative robot policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.04968", "title": "Potential-Guided Flow Matching for Vision-Language-Action Policy Improvement", "authors": ["Yunpeng Mei", "Jiakai He", "Hongjie Cao", "Chenyu Wang", "Xiaowen Zhu", "Yihan Zhou", "Jiamin Wang", "Chenbo Xin", "Peng Cheng", "Yuxuan Yang", "Yijie Wang", "Xinhu Zheng", "Gao Huang", "Jie Chen", "Gang Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.04968", "pdf_url": "https://arxiv.org/pdf/2606.04968", "code_url": "", "clusters": ["C3"], "keywords": ["ForesightFlow", "success potential", "critic-free improvement"], "model_class": "VLA", "techniques": ["self-guided flow matching", "best-of-K"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Flow policy that also generates success-potential trajectories, enabling policy improvement from mixed-quality data without a separate critic.", "relevance": 3, "relevance_reason": "Avoids large critic in post-training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"offline reinforcement learning\""}
{"id": "2603.10980", "title": "PPGuide: Steering Diffusion Policies with Performance Predictive Guidance", "authors": ["Zixing Wang", "Devesh K. Jha", "Ahmed H. Qureshi", "Diego Romeres", "A. H. Qureshi"], "year": 2026, "venue": "ICRA 2026", "type": "preprint", "url": "https://arxiv.org/abs/2603.10980", "pdf_url": "https://arxiv.org/pdf/2603.10980", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["PPGuide"], "model_class": "diffusion/flow policy", "techniques": ["classifier guidance", "multiple instance learning", "failure avoidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Robomimic", "MimicGen"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Self-supervised performance predictor gives real-time gradients steering a pretrained diffusion policy away from failure modes.", "relevance": 3, "relevance_reason": "Lightweight inference-time guidance.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight); snowball round 2: cites 2303.04137; cited by 2609.21220"}
{"id": "2605.08879", "title": "Preserving Foundational Capabilities in Flow-Matching VLAs through Conservative SFT", "authors": ["Tianyi Zhang", "Shaopeng Zhai", "Haoran Zhang", "Fuxian Huang", "Qi Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.08879", "pdf_url": "https://arxiv.org/pdf/2605.08879", "code_url": "", "clusters": ["C3"], "keywords": ["ConSFT", "conservative SFT", "trust-region"], "model_class": "VLA", "techniques": ["confidence-scaled loss"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Confidence-scaled SFT objective limiting parameter disruption when fine-tuning flow VLAs, with no replay data or extra modules.", "relevance": 3, "relevance_reason": "Cheap forgetting-mitigating fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"catastrophic forgetting\""}
{"id": "2603.03818", "title": "Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning", "authors": ["Huihan Liu", "Changyeon Kim", "Bo Liu", "Minghuan Liu", "Yuke Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.03818", "pdf_url": "https://arxiv.org/pdf/2603.03818", "code_url": "", "clusters": ["C3"], "keywords": ["continual learning", "experience replay", "forgetting"], "model_class": "VLA", "techniques": ["experience replay"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows large pretrained VLAs resist forgetting and need only small replay buffers for continual skill acquisition.", "relevance": 3, "relevance_reason": "Informs cheap continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: continual learning vision-language-action model catastrophic forgetting arXiv"}
{"id": "2602.21684", "title": "Primary-Fine Decoupling for Action Generation in Robotic Imitation", "authors": ["Xiao-Han Lei", "Min Wang", "Weng Zhou", "Xing-Yu Lu", "Hou-Qiang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.21684", "pdf_url": "https://arxiv.org/pdf/2602.21684", "code_url": "", "clusters": ["C1"], "keywords": ["multi-modal actions", "reactive control"], "model_class": "diffusion/flow policy", "techniques": ["MeanFlow", "mode decoupling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Adroit", "DexArt", "MetaWorld"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-stage policy that picks a discrete primary mode and then generates fine continuous actions with a MeanFlow policy.", "relevance": 3, "relevance_reason": "Few-step policy enabling reactive closed-loop control", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2605.30695", "title": "Primitive Subspaces Mediate Few-Shot Transfer in VLAs", "authors": ["Anya Singh", "Cabrel Happi", "Jai Relan", "Varun Nair", "Vidyut Baradwaj"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30695", "pdf_url": "https://arxiv.org/pdf/2605.30695", "code_url": "", "clusters": ["C3"], "keywords": ["primitive subspaces", "few-shot transfer", "no weight updates"], "model_class": "VLA", "techniques": ["LoRA", "primitive-segmented training", "in-context transfer"], "base_model": "OpenVLA; pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["REASSEMBLE", "LIBERO-Long"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Primitive-segmented LoRA training yields sub-skill subspaces enabling few-shot transfer to held-out tasks without weight updates.", "relevance": 3, "relevance_reason": "Avoids per-task fine-tuning cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:LoRA"}
{"id": "2609.33765", "title": "Principal Steering Subspaces for Online Adaptation of Frozen Generative Robot Policies", "authors": ["Jialeng Ni", "Nathan Zhao", "Kunpeng Song", "Jia-Leng Ni", "Kun-Peng Song"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33765", "pdf_url": "https://arxiv.org/pdf/2609.33765", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["PSS", "Principal Steering Subspaces"], "model_class": "diffusion/flow policy", "techniques": ["latent-noise RL", "subspace control", "latent steering", "online adaptation", "async deployment"], "base_model": "", "params": "3B (frozen VLA)", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboMimic", "humanoid"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adapts frozen diffusion/flow policies online by RL in a low-dimensional steering subspace of the latent noise, deployed on a frozen 3B VLA with latency-aware asynchronous execution.", "relevance": 3, "relevance_reason": "Efficient online adaptation of frozen generators.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"action chunk*\") AND (asynchronous OR \"inference delay\" OR \"latency compensation\"); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.22879", "title": "Prioritized Rollouts for Efficient World Model-based Vision-Language-Action Policy Optimization", "authors": ["Yifei Sheng", "Haoxiang Ren", "Zhilong Zhang", "Haonan Wang", "Runjie Xu", "Yihao Sun", "Nan Tang", "Zhichao Wu", "Lei Yuan", "Haoxin Lin", "Yang Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22879", "pdf_url": "https://arxiv.org/pdf/2609.22879", "code_url": "", "clusters": ["C8"], "keywords": ["prioritized rollouts", "policy uncertainty", "MBRL for VLA"], "model_class": "world model", "techniques": ["uncertainty-prioritized world-model rollouts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Spends world-model rollout compute on states where the VLA policy is uncertain, reducing the cost of model-based RL fine-tuning.", "relevance": 3, "relevance_reason": "Rollout budget efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2605.10925", "title": "PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models", "authors": ["Xinyu Guo", "Bin Xie", "Wei Chai", "Xianchi Deng", "Tiancai Wang", "Zhengxing Wu", "Xingyu Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.10925", "pdf_url": "https://arxiv.org/pdf/2605.10925", "code_url": "", "clusters": ["C3"], "keywords": ["prior preservation", "frozen prior expert", "adaptation expert"], "model_class": "VLA", "techniques": ["partial fine-tuning", "expert queries"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Keeps a frozen prior expert and trains an adaptation expert querying scene and motor priors, updating only ~25% of parameters.", "relevance": 3, "relevance_reason": "Parameter-efficient prior-preserving adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2608.14284", "title": "PRM-as-a-Judge 1.5: A Toolkit for Robot Process Assessment", "authors": ["Yu-Yang Liu", "Yan Shen", "Rui Chen", "Ji-Fan Zhao", "Yuxuan Tian", "Yi-Chi Zhang", "Tian-Feng Long", "Zi Yin", "Yi-Pu Wang", "Zi-Heng Qin", "Wen-Xing Tan", "Yang Shi", "Mingyu Cao", "Run-Ze Xiao", "Ziqian Wang", "Zhi-Xin Yin", "Shiwei Chu", "Yi-Fan Zhang", "Yao Mu", "Yu-Heng Ji", "Yi-Hao Wang", "Jun Yan", "Zhong-Yuan Wang", "Peng-Wei Wang", "Xiaolong Zheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.14284", "pdf_url": "https://arxiv.org/pdf/2608.14284", "code_url": "", "clusters": ["C4"], "keywords": ["process reward model", "evaluation toolkit"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Toolkit that converts rollout videos into dense progress curves for fine-grained robot process assessment beyond binary success.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2603.21669", "title": "PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditing", "authors": ["Yu-Heng Ji", "Yu-Yang Liu", "Hua-Jie Tan", "Xuchuan Huang", "Fan-Ding Huang", "Yi-Jie Xu", "Cheng Chi", "Yu-Ting Zhao", "Huaihai Lyu", "P. Co", "Mingyu Cao", "Qiongyu Zhang", "Zhe Li", "En-Shen Zhou", "Peng-Wei Wang", "Zhong-Yuan Wang", "Shang-Hang Zhang", "Xiaolong Zheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.21669", "pdf_url": "https://arxiv.org/pdf/2603.21669", "code_url": "", "clusters": ["C4"], "keywords": ["dense evaluation", "robot auditing"], "model_class": "other", "techniques": ["process reward models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Evaluation paradigm using process reward models to score progress, efficiency and stability of robot rollouts from video instead of binary success.", "relevance": 3, "relevance_reason": "Benchmarking methodology beyond success rate.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2609.03611, 2609.09250"}
{"id": "2606.09740", "title": "ProbeAct: Probe-Guided Training-Free Failure Recovery in Vision-Language-Action Models", "authors": ["Fan Zhang", "Seongbin Park", "Baharan Mirzasoleiman", "Shariar Talebi", "Nader Sehatbakhsh", "S. Talebi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09740", "pdf_url": "https://arxiv.org/pdf/2606.09740", "code_url": "", "clusters": ["C7"], "keywords": ["runtime intervention", "VLA", "training-free failure recovery", "safety net"], "model_class": "VLA", "techniques": ["hidden-state probing", "state machine", "failure recovery", "hidden-state probes", "CBF filter"], "base_model": "OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-plus"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free runtime layer that probes VLA hidden states for object positions and runs a kinematic state machine to detect and recover from grasp/place failures.", "relevance": 3, "relevance_reason": "Runtime failure detection and recovery for frozen VLAs.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\"); snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2604.26508", "title": "Progressive Semantic Communication for Efficient Edge-Cloud Vision-Language Models", "authors": ["Cyril Shih-Huan Hsu", "Wig Yuan-Cheng Cheng", "Chrysa Papagianni"], "year": 2026, "venue": "IEEE GLOBECOM 2026", "type": "paper", "url": "https://arxiv.org/abs/2604.26508", "pdf_url": "https://arxiv.org/pdf/2604.26508", "code_url": "", "clusters": ["C6"], "keywords": ["progressive semantic communication", "edge-cloud VLM", "visual token compression"], "model_class": "other", "techniques": ["meta autoencoder token compression", "progressive transmission"], "base_model": "", "params": "", "hardware": "NXP i.MX95 edge + GPU server", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Compresses VLM visual tokens into progressively refinable codes so an embedded device can send as much detail as bandwidth allows to a cloud VLM.", "relevance": 3, "relevance_reason": "Bandwidth-adaptive edge-cloud VLM inference transferable to VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.21429", "title": "Provably Safe Generative Sampling with Constricting Barrier Functions", "authors": ["Darshan Gadginmath", "Ahmed Allibhoy", "Fabio Pasqualetti"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.21429", "pdf_url": "https://arxiv.org/pdf/2602.21429", "code_url": "", "clusters": ["C7"], "keywords": ["constricting barrier functions", "safe generative sampling", "flow models"], "model_class": "other", "techniques": ["time-varying safety tube during sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Online shield for pretrained diffusion/flow models that enforces a safety tube relaxed at the noise distribution and tightening to the safe set, giving formal constraint satisfaction.", "relevance": 3, "relevance_reason": "Provable constraints for generative policy outputs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"control barrier function\" learned policy robot manipulation"}
{"id": "2605.21072", "title": "Q-ARVD: Quantizing Autoregressive Video Diffusion Models", "authors": ["Siao Tang", "Xinyin Ma", "Gongfan Fang", "Xingyi Yang", "Xinchao Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.21072", "pdf_url": "https://arxiv.org/pdf/2605.21072", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["quantization", "autoregressive video diffusion"], "model_class": "world model", "techniques": ["frame-sensitivity-aware PTQ", "outlier-aware weight quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Characterizes why standard DiT quantization fails for autoregressive video diffusion (skewed frame sensitivity, heterogeneous outliers) and proposes a tailored PTQ scheme.", "relevance": 3, "relevance_reason": "Quantization for the backbone class used by streaming world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.17163", "title": "Q-Learning With World Models", "authors": ["Perry Dong", "Yueru Jia", "Chelsea Finn", "Dorsa Sadigh"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.17163", "pdf_url": "https://arxiv.org/pdf/2608.17163", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["model-based RL"], "model_class": "world model", "techniques": ["Q-learning", "world model search"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses world models for test-time search over imagined trajectories on top of standard online Q-learning instead of training policies on imagined rollouts.", "relevance": 3, "relevance_reason": "Sample-efficient RL with world models, framed for VLA fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]"}
{"id": "2606.14801", "title": "QPILOTS: Efficient Test-Time Q-Steering for Flow Policies", "authors": ["Yifan Ruan", "Chenyang Cao", "Andreas Burger", "Ali Pesaranghader", "Kaveh Kamali", "Jaehong Kim", "Nandita Vijaykumar", "Alan Aspuru-Guzik", "Igor Gilitschenski", "Nicholas Rhinehart", "N. Vijaykumar", "Alán Aspuru-Guzik"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.14801", "pdf_url": "https://arxiv.org/pdf/2606.14801", "code_url": "", "clusters": ["C3"], "keywords": ["QPILOTS", "offline-to-online RL"], "model_class": "diffusion/flow policy", "techniques": ["Q-guided denoising", "test-time steering", "inference-time Q-steering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "90% avg success over 50 offline-to-online tasks", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Inference-time Q-steering of flow policies that computes critic gradients on projected clean actions at each denoising step, leaving the policy unmodified; also steers a frozen VLA.", "relevance": 3, "relevance_reason": "Efficient adaptation without re-training the policy.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=0]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2602.02958", "title": "Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization", "authors": ["Haocheng Xi", "Shuo Yang", "Yilong Zhao", "Muyang Li", "Han Cai", "Xingyang Li", "Yujun Lin", "Zhuoyang Zhang", "Jintao Zhang", "Xiuyu Li", "Zhiying Xu", "Jun Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.02958", "pdf_url": "https://arxiv.org/pdf/2602.02958", "code_url": "", "clusters": ["C2", "C7", "C9"], "keywords": ["KV cache quantization", "autoregressive video diffusion", "2-bit"], "model_class": "world model", "techniques": ["training-free 2-bit KV-cache quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Quant VideoGen quantizes the growing KV cache of autoregressive video diffusion models to ~2 bits, cutting memory and extending consistent long-horizon generation.", "relevance": 3, "relevance_reason": "Memory efficiency for autoregressive video world models; organizer Zhuoyang Zhang.", "foundational_technique": true, "verified": true, "found_via": "author sweep: Zhuoyang Zhang"}
{"id": "2609.34061", "title": "Quantile Head for Vision-Language-Action Models", "authors": ["Xuan Wang", "Yi-Nan Wu", "Hao-Ran Duan", "Jun-Gong Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34061", "pdf_url": "https://arxiv.org/pdf/2609.34061", "code_url": "https://github.com/xwangrs/Quantile-Head-for-VLA", "clusters": ["C1"], "keywords": ["action head", "one forward pass"], "model_class": "VLA", "techniques": ["quantile regression head"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus", "LIBERO-Pro", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Quantile Head unifies regression and flow matching into a quantile objective, predicting ordered action quantiles in one forward pass instead of costly iterative flow sampling.", "relevance": 3, "relevance_reason": "relevant: one-pass action head avoids iterative sampling", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "sail-blog-rnb-encore-2026", "title": "R&B-EnCoRe: Self-Improving Pretraining of Embodied Reasoning Vision-Language-Action Models (SAIL Blog)", "authors": ["Milan Ganai", "Katie Luo", "Jonas Frey", "Clark Barrett", "Marco Pavone"], "year": 2026, "venue": "blog (Stanford AI Lab)", "type": "blog", "url": "https://ai.stanford.edu/blog/rnb-encore/", "pdf_url": "", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["concise reasoning", "VLA"], "model_class": "VLA", "techniques": ["selective reasoning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "reasoning tokens ~halved; WidowX step ~5 s → ~3 s", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Blog explainer: selective reasoning beats exhaustive templates, shortening VLA reasoning and latency across 1B–30B models.", "relevance": 3, "relevance_reason": "Blog companion to 2602.08167.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Stanford AI Lab blog"}
{"id": "2603.18202", "title": "R2-Dreamer: Redundancy-Reduced World Models without Decoders or Augmentation", "authors": ["Naoki Morihira", "Amal Nahar", "Kartik Bharadwaj", "Yasuhiro Kato", "Akinobu Hayashi", "Tatsuya Harada"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.18202", "pdf_url": "https://arxiv.org/pdf/2603.18202", "code_url": "https://github.com/NM512/r2dreamer", "clusters": ["C8"], "keywords": ["decoder-free Dreamer", "redundancy reduction", "Barlow Twins"], "model_class": "world model", "techniques": ["decoder-free latent world model", "redundancy-reduction regularizer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "trains 1.59x faster than DreamerV3", "metric_measures": "relative latency", "benchmarks": ["DMC", "Meta-World", "DMC-Subtle"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Removes the pixel decoder from Dreamer and prevents collapse with a Barlow-Twins-style objective, matching DreamerV3/TD-MPC2 while training ~1.6x faster.", "relevance": 3, "relevance_reason": "Cheaper latent world models (DreamerV3 follow-up).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2301.04104"}
{"id": "2609.03276", "title": "R2S-Eval: Robot Evaluation with Real-to-Sim Calibration via Vision-Language Models", "authors": ["Yidi Wang", "Feixiang Ruan", "Ruoqu Chen", "Jie Yin", "Yang Yu", "Mengdi Xu", "Kaifeng Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.03276", "pdf_url": "https://arxiv.org/pdf/2609.03276", "code_url": "", "clusters": ["C4"], "keywords": ["R2S-Eval", "real-to-sim", "VLM judge", "pairwise preference"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Combines real-to-sim calibrated rollouts with VLM pairwise preference judging to produce stable policy rankings with fewer hardware trials.", "relevance": 3, "relevance_reason": "Cheaper, more stable real-world-proxy evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query: real-world robot policy evaluation protocol statistical"}
{"id": "2608.25585", "title": "RA-VLA: Retrieval-Augmented VLA for Test-Time Adaptation", "authors": ["Sanghwan Jang", "Minjin Jeon", "Minsoo Kim", "Seongjin Choi", "Dongha Kim", "Hwanjo Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.25585", "pdf_url": "https://arxiv.org/pdf/2608.25585", "code_url": "", "clusters": ["C3"], "keywords": ["retrieval-augmented", "in-context imitation", "test-time adaptation"], "model_class": "VLA", "techniques": ["retrieval augmentation", "in-context learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Retrieval-augmented VLA that adapts to new tasks from retrieved expert context without gradient updates.", "relevance": 3, "relevance_reason": "Training-free adaptation alternative to fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2603.28522", "title": "RAD-LAD: Rule and Language Grounded Autonomous Driving in Real-Time", "authors": ["Anurag Ghosh", "Srinivasa Narasimhan", "M. Chandraker", "F. Pittaluga"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.28522", "pdf_url": "https://arxiv.org/pdf/2603.28522", "code_url": "", "clusters": ["C1"], "keywords": ["language-action planner", "autonomous driving", "real-time"], "model_class": "VLA", "techniques": ["interruptible architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~20 Hz plan-only, ~10 Hz with reasoning; ~3x lower latency than prior driving LMs", "metric_measures": "other", "benchmarks": ["nuPlan", "InterPlan"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Real-time language-action driving planner that outputs a plan in one pass or with reasoning text, plus a rule-based planner for hybrid use.", "relevance": 3, "relevance_reason": "Real-time language-action planning", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2504.16054"}
{"id": "2602.10980", "title": "RADAR: Benchmarking Vision-Language-Action Generalization via Real-World Dynamics, Spatial-Physical Intelligence, and Autonomous Evaluation", "authors": ["Yuhao Chen", "Zhihao Zhan", "Xiaoxin Lin", "Zijian Song", "Hao Liu", "Qinhan Lyu", "Y. Zu", "Xiao Chen", "Zhiyuan Liu", "Tao Pu", "Tianshui Chen", "Keze Wang", "Liang Lin", "Guang-Run Wang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.10980", "pdf_url": "https://arxiv.org/pdf/2602.10980", "code_url": "", "clusters": ["C4"], "keywords": ["RADAR", "real-world dynamics"], "model_class": "VLA", "techniques": ["benchmark design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmark that critiques current VLA evaluation practice and tests generalization under realistic dynamics, physics and noise.", "relevance": 3, "relevance_reason": "Benchmarking methodology.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.00425"}
{"id": "github-radixquantvla", "title": "RadixQuantVLA: unified post-training quantization and LIBERO benchmarking stack for VLAs", "authors": ["RadixRootMind"], "year": 2026, "venue": "GitHub", "type": "repo", "url": "https://github.com/RadixRootMind/RadixQuantVLA", "pdf_url": "", "code_url": "https://github.com/RadixRootMind/RadixQuantVLA", "clusters": ["C2", "C4"], "keywords": ["W4A8", "W4A4", "GPTQ", "DuQuant", "LIBERO", "pi0.5", "GR00T"], "model_class": "VLA", "techniques": ["W4A8", "W4A4 GPTQ", "DuQuant", "QVLA mixed-bit"], "base_model": "pi0.5, GR00T-N1.5, OpenVLA, OpenVLA-OFT", "params": "", "hardware": "A100 40GB", "reported_metric": "LIBERO-Object: pi0.5 W4A8 DuQuant 99.0%, pi0.5 W4A4 GPTQ 98.0%, GR00T-N1.5 W4A8 82.0%; LIBERO-Spatial OpenVLA-OFT QVLA W8 26.0% vs FP16 60.0% (as reported)", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Unified launcher for applying several PTQ recipes to VLAs and scoring them on LIBERO; reports W4 pi0.5 near full success but low baseline numbers for OpenVLA-family runs.", "relevance": 3, "relevance_reason": "Cross-model W4 LIBERO results in one harness; baselines look under-tuned, so treat as indicative.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub repo search openvla+quantization"}
{"id": "2609.17728", "title": "RAF-VLA: Representation Alignment with the Future for End-to-End Autonomous Driving", "authors": ["Dogun Kim", "Yongjae Lee", "Joonhee Lim", "Yeina Lee", "Junhyeok Park", "Moogeun Park", "Dongsuk Kum", "Y. Lee"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.17728", "pdf_url": "https://arxiv.org/pdf/2609.17728", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["driving VLA", "world model", "no future generation"], "model_class": "VLA", "techniques": ["representation alignment", "future representation alignment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~1 ms inference overhead, 3.8% training overhead", "metric_measures": "other", "benchmarks": ["NAVSIM"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Aligns driving-VLA hidden states with pretrained future-frame representations instead of generating futures, avoiding world-model inference cost.", "relevance": 3, "relevance_reason": "Gets world-model benefits with negligible inference overhead.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=0]; snowball round 2: cites 2406.09246, 2410.24164"}
{"id": "2609.18324", "title": "RAFAIL: Relationship-Aware Failure Detection for Robotic Manipulation", "authors": ["Loris Schneider", "Edgar Welte", "R. Rayyes"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18324", "pdf_url": "https://arxiv.org/pdf/2609.18324", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "no runtime VLM"], "model_class": "other", "techniques": ["relationship-level OOD detection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Detects execution failures via OOD detection on task-relevant relationships, avoiding VLM inference at runtime.", "relevance": 3, "relevance_reason": "Low-overhead runtime failure detection.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.22689, 2503.08558"}
{"id": "2602.03310", "title": "RDT2: Exploring the Scaling Limit of UMI Data Towards Zero-Shot Cross-Embodiment Generalization", "authors": ["Songming Liu", "Bangguo Li", "Kai Ma", "Lingxuan Wu", "Hengkai Tan", "Xiao Ouyang", "Hang Su", "Jun Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.03310", "pdf_url": "https://arxiv.org/pdf/2602.03310", "code_url": "", "clusters": ["C1"], "keywords": ["RDT2", "UMI"], "model_class": "VLA", "techniques": ["residual vector quantization", "flow matching", "distillation"], "base_model": "", "params": "7B VLM backbone", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "7B-VLM robot foundation model trained on 10k+ hours of UMI data with a three-stage recipe (RVQ alignment, flow matching, distillation for real-time inference) for zero-shot cross-embodiment use.", "relevance": 3, "relevance_reason": "Distillation for real-time inference in a large VLA.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot policy\" OR \"world action model\") AND (quantiz* OR \"low-bit\")"}
{"id": "2606.30362", "title": "ReactiveBFM: Reactive Closed-Loop Motion Planning Towards Universal Humanoid Whole-Body Control", "authors": ["Xiao Chen", "Weishuai Zeng", "Xiaojie Niu", "Zirui Wang", "Jianan Li", "Huayi Wang", "Furui Xu", "Jiahe Chen", "Weixiang Zhong", "Lihe Ding", "Kailin Li", "Jiangmiao Pang", "Tai Wang", "Tianfan Xue", "Jingbo Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.30362", "pdf_url": "https://arxiv.org/pdf/2606.30362", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["ReactiveBFM", "Unitree G1"], "model_class": "other", "techniques": ["asynchronous replanning", "trajectory chunking"], "base_model": "", "params": "", "hardware": "Unitree G1", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Humanoid planning-control framework with asynchronous replanning and trajectory chunking to reconcile slow autoregressive planning with fast tracking.", "relevance": 3, "relevance_reason": "Async chunked replanning for humanoids.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv humanoid AND (VLA OR \"foundation model\") AND (onboard OR \"real-time\" OR latency)"}
{"id": "2607.22997", "title": "Real2Sim2Real for Vision-Language-Action Manipulation: An AMD ROCm-Based Pipeline", "authors": ["Qing Yang", "Xun Wang", "Ziguan Wang", "Zhenjiang Li", "Hongqiang Wang", "Dongdong Weng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.22997", "pdf_url": "https://arxiv.org/pdf/2607.22997", "code_url": "", "clusters": ["C5"], "keywords": ["AMD ROCm", "Ryzen AI", "non-CUDA stack", "SmolVLA"], "model_class": "VLA", "techniques": ["alternative accelerator software stack"], "base_model": "SmolVLA", "params": "", "hardware": "AMD Radeon AI PRO R9700, Radeon PRO W7900, Ryzen AI", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot (Franka)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Demonstrates an end-to-end VLA training, simulation, and deployment pipeline running entirely on AMD GPUs and ROCm instead of CUDA.", "relevance": 3, "relevance_reason": "Heterogeneous vendor hardware for VLA deployment.", "foundational_technique": false, "verified": true, "found_via": "arXiv: vision-language-action AND Jetson/edge/onboard"}
{"id": "2603.26360", "title": "Realtime-VLA V2: Learning to Run VLAs Fast, Smooth, and Accurate", "authors": ["Chen Yang", "Yucheng Hu", "Yunchao Ma", "Yunhuan Yang", "Jing Tan", "Hao-Qiang Fan"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2603.26360", "pdf_url": "https://arxiv.org/pdf/2603.26360", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["end-to-end deployment", "execution speed", "calibration", "Dexmal"], "model_class": "VLA", "techniques": ["system-level deployment stack", "learned execution speed"], "base_model": "pi0-class VLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Follow-up to Running VLAs at Real-time Speed describing calibration, planning/control and learned speed selection so a VLA-driven lightweight arm runs near human speed.", "relevance": 3, "relevance_reason": "Full-stack deployment beyond GPU kernels.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2510.26742 (citations)"}
{"id": "2606.31160", "title": "Reasoning-aware Speculative Decoding for Efficient Vision-Language-Action Models in Autonomous Driving", "authors": ["Anh Dung Dinh", "Simon Khan", "Flora Salim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.31160", "pdf_url": "https://arxiv.org/pdf/2606.31160", "code_url": "", "clusters": ["C1"], "keywords": ["driving VLA", "reasoning latency"], "model_class": "VLA", "techniques": ["speculative decoding", "draft specialization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Speculative decoding for driving VLA chain-of-causation reasoning where a specialized routine reasoner drafts from trajectory history and the target verifies, with a modified RoPE in the draft.", "relevance": 3, "relevance_reason": "Speculative decoding tailored to VLA reasoning latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=0]"}
{"id": "2606.23617", "title": "RECALL: Recovery Experience Collection for Active Lifelong Learning in Vision-Language-Action Models", "authors": ["Ulas Berk Karli", "Tesca Fitzgerald"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.23617", "pdf_url": "https://arxiv.org/pdf/2606.23617", "code_url": "", "clusters": ["C3"], "keywords": ["active learning", "recovery data", "continual learning"], "model_class": "VLA", "techniques": ["uncertainty-guided data collection", "replay", "EWC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uncertainty-guided active collection of recovery demos makes VLA fine-tuning more data-efficient; replay needed to avoid forgetting.", "relevance": 3, "relevance_reason": "Data-efficient continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"continual learning\""}
{"id": "2604.16677", "title": "ReconVLA: An Uncertainty-Guided and Failure-Aware Vision-Language-Action Framework for Robotic Control", "authors": ["Lingling Chen", "Zongyao Lyu", "William J. Beksi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.16677", "pdf_url": "https://arxiv.org/pdf/2604.16677", "code_url": "", "clusters": ["C7"], "keywords": ["VLA uncertainty"], "model_class": "VLA", "techniques": ["conformal prediction", "failure detection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Applies conformal prediction to pretrained VLA action tokens and to robot state to give calibrated uncertainty and pre-failure outlier detection.", "relevance": 3, "relevance_reason": "Conformal uncertainty signal usable to trigger replanning.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:conformal AND (\"policy failure\" OR \"failure prediction\" OR \"failure detection\") AND robot*"}
{"id": "2607.18540", "title": "Recti-Q: Feature-Space Rectification for Out-of-Distribution-Robust Quantized Perception in Edge Robotics", "authors": ["Hamidreza Yaghoubi Araghi", "Parastoo Pilevar", "Ming C. Lin"], "year": 2026, "venue": "IROS 2026", "type": "paper", "url": "https://arxiv.org/abs/2607.18540", "pdf_url": "https://arxiv.org/pdf/2607.18540", "code_url": "", "clusters": ["C2"], "keywords": ["PTQ robustness gap", "OOD", "LoRA adapter", "edge robotics"], "model_class": "other", "techniques": ["post-training quantization", "LoRA rectification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ImageNet-C", "PACS"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows 4-bit PTQ vision backbones lose OOD robustness despite clean accuracy and recovers it with a tiny LoRA head adapter.", "relevance": 3, "relevance_reason": "Evidence that quantization harms generalization under shift for robot perception.", "foundational_technique": false, "verified": true, "found_via": "search: 1-bit robot"}
{"id": "2602.20659", "title": "Recursive Belief Vision Language Action Models", "authors": ["Vaidehi Bagaria", "Bijo Sebastian", "Nirav Kumar Patel"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.20659", "pdf_url": "https://arxiv.org/pdf/2602.20659", "code_url": "", "clusters": ["C1"], "keywords": ["partial observability"], "model_class": "VLA", "techniques": ["latent belief state", "world-model objectives"], "base_model": "pi0 (baseline)", "params": "", "hardware": "", "reported_metric": "up to 5x lower inference latency vs baselines", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Belief-centric VLA: VLM queried once per task for intent while a compact self-supervised belief state tracks progress and conditions a diffusion policy.", "relevance": 3, "relevance_reason": "Avoids repeated VLM queries and memory growth.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]"}
{"id": "2604.22591", "title": "RedVLA: Physical Red Teaming for Vision-Language-Action Models", "authors": ["Yuhao Zhang", "Bo-Rong Zhang", "Jiaming Fan", "Jiacheng Shen", "Yishuai Cai", "Yao-Dong Yang", "Jiaming Ji"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.22591", "pdf_url": "https://arxiv.org/pdf/2604.22591", "code_url": "", "clusters": ["C7"], "keywords": ["safety", "red teaming"], "model_class": "VLA", "techniques": ["red teaming"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RedVLA is a red-teaming framework that searches for physically unsafe behaviors of VLA models before deployment.", "relevance": 3, "relevance_reason": "relevant: physical safety evaluation for deployment", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2605.12334", "title": "Reinforcing VLAs in Task-Agnostic World Models", "authors": ["Yucen Wang", "Rui Yu", "Fengming Zhang", "Junjie Lu", "Xinyao Qin", "Tianxiang Zhang", "Kaixin Wang", "Li Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.12334", "pdf_url": "https://arxiv.org/pdf/2605.12334", "code_url": "", "clusters": ["C3"], "keywords": ["RAW-Dream", "task-agnostic world model", "zero-shot reward"], "model_class": "VLA", "techniques": ["world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Decouples world/reward model learning from downstream tasks so VLA RL post-training on new tasks needs no task-specific model fine-tuning.", "relevance": 3, "relevance_reason": "Reduces per-task post-training cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"world model\" AND abs:\"post-training\""}
{"id": "10.1109/med70602.2026.11598477", "title": "REPO: Reactive Policies with Observation Augmentation", "authors": ["Aksel Vaaler", "Sigmund H. Høeg", "Helle Stige", "C. Holden"], "year": 2026, "venue": "2026 34th Mediterranean Conference on Control and Automation (MED)", "type": "paper", "url": "https://doi.org/10.1109/MED70602.2026.11598477", "pdf_url": "", "code_url": "", "clusters": ["C7"], "keywords": ["reactive policy", "dynamic environments"], "model_class": "other", "techniques": ["single-step reactive actions", "observation augmentation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RePO predicts single-step actions from observation history with adversarial embedding noise, giving reactive yet temporally consistent policies under dynamic perturbations instead of open-loop chunks.", "relevance": 3, "relevance_reason": "relevant: reactivity vs open-loop chunk execution", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.05424", "title": "Research on Intra-Chip Fusion Deployment and Optimization of Embodied Intelligence Business Operator NPU", "authors": ["Yu-Chen Zhu", "Long-Xiang Yin", "Wan-Yu Wang", "Jie-Ke Lin", "Guo-Qiang Zou", "Zi-Rui Cao", "Yu-Ling Yuan", "Xiao-Lan Fan", "Li-Fen Chen", "Hao-Zai Zheng", "Qi-Zhang He", "Hong-Yu Zhou", "Chun-Hai Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.05424", "pdf_url": "https://arxiv.org/pdf/2609.05424", "code_url": "", "clusters": ["C5"], "keywords": ["Cambricon MLU370", "Phytium", "domestic NPU", "operator fusion", "inverse kinematics"], "model_class": "other", "techniques": ["on-chip fusion deployment", "custom operators"], "base_model": "", "params": "", "hardware": "Phytium FT-2000/4 + Cambricon MLU370 (vs Jetson AGX Xavier)", "reported_metric": "18.7 ms full-pipeline single-frame latency; 2.89x vs Jetson AGX Xavier", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fuses perception (ViT), motion-blur correction and IK operators onto a single Cambricon NPU card to run the whole embodied pipeline on-chip.", "relevance": 3, "relevance_reason": "Non-NVIDIA NPU deployment for embodied pipeline.", "foundational_technique": false, "verified": true, "found_via": "web: processing-in-memory OR NPU accelerator embodied AI"}
{"id": "2609.15988", "title": "ResSafe: Learning Safety Filtering with Residual Reinforcement Learning for Humanoids", "authors": ["Gechen Qu", "Tong Zhang", "Bike Zhang", "Yen-Jen Wang", "Koushil Sreenath", "Claire Tomlin", "Jason Jangho Choi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.15988", "pdf_url": "https://arxiv.org/pdf/2609.15988", "code_url": "", "clusters": ["C7"], "keywords": ["ResSafe", "residual RL safety filter", "humanoid"], "model_class": "other", "techniques": ["residual RL as implicit safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Decouples task performance and safety for humanoids by learning a residual RL policy that acts as an implicit safety filter over a nominal policy.", "relevance": 3, "relevance_reason": "Learned safety filter for humanoid control.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" learned policy reinforcement learning robot"}
{"id": "2609.33944", "title": "ReSync: Re-Aligning the Two Clocks of Asynchronous World-Action Models", "authors": ["Xi Lin", "Feihong Zhang", "Yulong Shi", "Yanghong Mei", "Zuxing Lu", "Xiaofan Zhu", "Zihao Liang", "Zhirui Gao", "Zhaowen Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33944", "pdf_url": "https://arxiv.org/pdf/2609.33944", "code_url": "", "clusters": ["C8"], "keywords": ["asynchronous denoising", "commitment-evidence gap", "test-time compute allocation"], "model_class": "world-action model", "techniques": ["hold action state while advancing video stream", "schedule-derived compute window"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Formalizes the mismatch when actions finish denoising before the video stream and gives a parameter-free rule for how much extra world computation to spend before committing an action.", "relevance": 3, "relevance_reason": "Principled compute allocation for dual-schedule WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.22663", "title": "Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline", "authors": ["Song, Wenxuan", "Chen, Jiayi", "Sun, Xiaoquan", "Lei, Huashuo", "Qin, Yikai", "Zhao, Wei", "Ding, Pengxiang", "Zhao, Han", "Wang, Tongxin", "Hou, Pengxu", "Zhong, Zhide", "Yan, Haodong", "Wang, Donglin", "Ma, Jun", "Li, Haoang"], "year": 2026, "venue": "ICRA 2026", "type": "benchmark", "url": "https://arxiv.org/abs/2602.22663", "pdf_url": "https://arxiv.org/pdf/2602.22663", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["CEBench", "practicality", "cross-embodiment", "consumer GPU", "LLaVA-VLA"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CEBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Cross-embodiment sim+real benchmark studying VLA practicality (size, pre-training, embodiments) plus a lightweight VLA deployable on consumer GPUs.", "relevance": 3, "relevance_reason": "Benchmark framed around deployability.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: OpenReview CoRL 2025 evaluation protocol VLA benchmark flaws"}
{"id": "2608.04527", "title": "Retrieve in Time, Correct in Frequency", "authors": ["Yuze Fan", "Yue Cao", "Pengjie Gao", "Haojia Gao", "Guangqiu Guo", "Ziyue Zhang", "Junbo Tan", "Bokui Chen", "Zhuo Zou", "Xueqian Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.04527", "pdf_url": "https://arxiv.org/pdf/2608.04527", "code_url": "", "clusters": ["C7"], "keywords": ["action chunk correction", "frozen VLA"], "model_class": "VLA", "techniques": ["retrieval", "frequency-domain action correction", "test-time correction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free test-time correction for frozen VLAs: aligns the execution history to stored successful rollouts and transfers only selected frequency components of their action chunks to the current proposal.", "relevance": 3, "relevance_reason": "Closed-loop correction of chunk proposals without retraining.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*)"}
{"id": "2605.10094", "title": "Retrieve-then-Steer: Online Success Memory for Test-Time Adaptation of Generative VLAs", "authors": ["Jianchao Zhao", "Huoren Yang", "Yusong Hu", "Yuyang Gao", "Qiguan Ou", "Cong Wan", "SongLin Dong", "Zhiheng Ma", "Yihong Gong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.10094", "pdf_url": "https://arxiv.org/pdf/2605.10094", "code_url": "", "clusters": ["C3"], "keywords": ["success memory", "retrieve-then-steer", "persistent deployment"], "model_class": "VLA", "techniques": ["retrieval", "test-time steering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Stores successful deployment experience and retrieves it to steer a frozen generative VLA during repeated deployment.", "relevance": 3, "relevance_reason": "Deployment-time adaptation of frozen VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"test-time adaptation\""}
{"id": "2609.23910", "title": "ReVeal: A Reconstruction-Aware Real-to-Sim Framework for VLA Policy Evaluation", "authors": ["Wang, Xinyi", "Hao, Heng", "Hu, Wenjun", "Li, Anna Enyu", "Ma, Dizhi", "Ramani, Karthik", "Moon, Hankyu", "Kwon, Yeong-Dae"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.23910", "pdf_url": "https://arxiv.org/pdf/2609.23910", "code_url": "", "clusters": ["C4"], "keywords": ["ReVeal", "reconstruction fidelity", "real-to-sim assessment", "humanoid manipulation"], "model_class": "VLA", "techniques": [], "base_model": "GR00T, SmolVLA, pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot", "reconstructed sim"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Proposes reconstruction-fidelity metrics (NVMF, APGF) and shows reconstruction fidelity orders real-sim agreement for matched VLA evaluations.", "relevance": 3, "relevance_reason": "Validity checks for real-to-sim evaluation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-to-sim Gaussian splatting policy evaluation"}
{"id": "2608.01013", "title": "RL Bootstrapping of OpenVLA-OFT for a Novel Robot Embodiment", "authors": ["Damir Nurtdinov", "Alexei Kornaev", "Alexander Maloletov"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.01013", "pdf_url": "https://arxiv.org/pdf/2608.01013", "code_url": "", "clusters": ["C3"], "keywords": ["zero-demo embodiment alignment", "PPO", "GRPO", "cable-driven robot"], "model_class": "VLA", "techniques": ["RL in simulation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Aligns OpenVLA-OFT to a cable-driven parallel robot with no demonstrations using staged PPO then GRPO in simulation.", "relevance": 3, "relevance_reason": "Cross-embodiment adaptation without demos.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.26991", "title": "RL$^2$-VLA: Adaptive RL Latent Compositional Steering with Test-Time Scaling for Vision-Language-Action Models", "authors": ["Derek Ming Siang Tan", "Shailesh Shailesh", "Srikrishna Iyer", "William Wei Jie Teo", "Yuanliang Ju", "Qiao Gu", "Guillaume Sartoretti"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.26991", "pdf_url": "https://arxiv.org/pdf/2607.26991", "code_url": "https://rl2-vla.github.io", "clusters": ["C3"], "keywords": ["RL steering"], "model_class": "VLA", "techniques": ["latent RL", "compositional flow steering", "test-time scaling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains a small offline RL policy on VLA action-expert latents and composes its flow velocity with the frozen VLA, applying steering adaptively per timestep.", "relevance": 3, "relevance_reason": "Lightweight RL steering of frozen VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]"}
{"id": "2605.03269", "title": "RLDX-1 Technical Report", "authors": ["Dongyoung Kim", "Huiwon Jang", "Myungkyu Koo", "Jinwoo Shin", "Suhyeok Jang", "Taeyoung Kim", "Beomjun Kim", "Byungjun Yoon", "Changsung Jang", "Daewon Choi", "Dongsu Han", "Donguk Lee", "Heeseung Kwon", "Hojin Jeon", "Jaehyun Kang", "Jaekyoung Bae", "Jihyuk Lee", "Jimin Lee", "John Won", "Joonwoo Ahn", "Junhyeong Park", "Junyoung Sung", "Kyungmin Lee", "Minseong Han", "Minsung Yoon", "Sejune Joo", "Seonil Son", "Seungcheol Park", "Seunggeun Cho", "Seungjun Moon", "Seungku Kim", "Yonghoon Dong", "Yongjin Cho", "Youngchan Kim", "Chang Hwan Kim", "Dohyeon Kim", "Heecheol Kim", "Heewon Lee", "Hensen Ahn", "Hyungkyu Ryu", "Hyunsoo Choi", "Hyunsoo Shin", "Jaeheon Jung", "Jaewoo Kim", "Jinwook Kim", "Joochul Chang", "Joonsoo Kim", "Junghun Park", "Jungwoo Park", "Junho Cho", "Junhyeok Park", "Junwon Lee", "Kangwook Lee", "Kwanghoon Kim", "Kyoungwhan Choe", "Manoj Bhadu", "Nayoung Oh", "Sangjun Kim", "Sangwoo Kim", "Seunghoon Shim", "Seunghyun Kim", "Seungjun Lee", "Seungyup Ka", "Sungryol Yang", "Wook Jung", "Yashu Shukla", "Yeonjae Lee", "Yeonwoo Bae"], "year": 2026, "venue": "arXiv tech report", "type": "tech-report", "url": "https://arxiv.org/abs/2605.03269", "pdf_url": "https://arxiv.org/pdf/2605.03269", "code_url": "", "clusters": ["C1", "C4", "C5"], "keywords": ["CUDA Graph", "kernel fusion", "humanoid", "p50 latency", "RLDX-1"], "model_class": "VLA", "techniques": ["static graph + CUDA Graph capture", "custom fused kernels", "multi-stream transformer", "inference optimization"], "base_model": "RLDX-1 (MSAT)", "params": "", "hardware": "RTX 5090 + Intel Core Ultra 7 265K", "reported_metric": "per-step latency 71.2 ms eager -> 43.7 ms (1.63x) all-modality; graph stage ~1.45x then fusion to 1.61-1.63x cumulative; p50 over 300 iters after 100 warmups", "metric_measures": "other", "benchmarks": ["ALLEX humanoid real tasks", "simulation benchmarks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Dexterous humanoid VLA whose report stacks graph-level (CUDA Graph) and kernel-level (fusion) optimizations, showing gains accumulate nearly multiplicatively and are robust to added modality streams.", "relevance": 3, "relevance_reason": "Clean ablation of stacked systems optimizations with p50 reporting protocol.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q1: arXiv abs:\"vision-language-action\" AND abs:\"system-level\" AND abs:optimizations; recency sweep: arXiv humanoid AND (VLA OR \"foundation model\") AND (onboard OR \"real-time\" OR latency)"}
{"id": "2606.25800", "title": "ROAD-VLA: Robust Online Adaptation via Self-Distillation for Vision-Language-Action Models", "authors": ["Kejing Wang", "Toan Nguyen", "Minh Hoang Nguyen", "Simon Khan", "Flora D. Salim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.25800", "pdf_url": "https://arxiv.org/pdf/2606.25800", "code_url": "", "clusters": ["C3"], "keywords": ["self-distillation", "online adaptation", "token-level supervision"], "model_class": "VLA", "techniques": ["advantage-guided self-distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Builds an action-space proximal teacher from advantage-perturbed logits to give dense supervision for online VLA adaptation.", "relevance": 3, "relevance_reason": "Improves sample efficiency of online adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.27881", "title": "RoboBRIDGE: A Modular Framework for Bridging Policies to Robust Real-World Robotic Agents", "authors": ["Sihyung Yoon", "Minjong Yoo", "Sanghyun Ahn", "Seojeong Choi", "Honguk Woo"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.27881", "pdf_url": "https://arxiv.org/pdf/2607.27881", "code_url": "", "clusters": ["C7"], "keywords": ["RoboBRIDGE"], "model_class": "VLA", "techniques": ["runtime monitoring", "LoRA adapters", "async perception"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Modular orchestration layer around VLAs with a failure monitor, asynchronous perception, replanning and LoRA primitive skills.", "relevance": 3, "relevance_reason": "Runtime monitoring and async execution.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2306.03310"}
{"id": "2603.04356", "title": "RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots", "authors": ["Soroush Nasiriany", "Sepehr Nasiriany", "Abhiram Maddukuri", "Yuke Zhu"], "year": 2026, "venue": "ICLR 2026", "type": "benchmark", "url": "https://arxiv.org/abs/2603.04356", "pdf_url": "https://arxiv.org/pdf/2603.04356", "code_url": "", "clusters": ["C4"], "keywords": ["RoboCasa365", "household", "mobile manipulation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa365"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "365 household tasks across 2,500 kitchens with 600h human and 1600h synthetic demos for systematic generalist-policy evaluation.", "relevance": 3, "relevance_reason": "Large-scale sim benchmark, increasingly used for VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RoboCasa VLABench RoboTwin"}
{"id": "2609.10021", "title": "RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibility", "authors": ["Runze Xu", "Yuanfan Xu", "Cuijie Xu", "Shuang Dai", "Yining Li", "Yu Wang", "Jincheng Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.10021", "pdf_url": "https://arxiv.org/pdf/2609.10021", "code_url": "", "clusters": ["C3"], "keywords": ["data curation", "gradient compatibility", "post-training data"], "model_class": "VLA", "techniques": ["data curation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Scores post-training samples by local gradient compatibility to drop corrupted robot data before VLA post-training.", "relevance": 3, "relevance_reason": "Makes post-training more data-efficient.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.32236", "title": "RoboFFT: Finetuning generative robot policy via online reinforcement learning with forward process", "authors": ["Yu Li", "Sheng-He Hu", "Yu-Han Wang", "Yao-Xiang Pu", "Hao-Tong Zhang", "Yuan-Pei Chen", "Yao-Dong Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.32236", "pdf_url": "https://arxiv.org/pdf/2609.32236", "code_url": "", "clusters": ["C3"], "keywords": ["RL fine-tuning", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["forward-process RL", "PPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RoboFFT fine-tunes diffusion/flow robot policies with online RL by building a surrogate policy ratio from forward-noising losses, reporting better stability and training efficiency.", "relevance": 3, "relevance_reason": "relevant: efficient RL fine-tuning of generative policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.34968", "title": "RoboFL: Federated Expert Assembly for World Action Models", "authors": ["Rongyu Zhang", "Ruizhi Fan", "Yunfan Lou", "Hengyu Fang", "Shenli Zheng", "Chenrui Wu", "Yili Jin", "Li Du", "Dan Wang", "Yuan Du", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34968", "pdf_url": "https://arxiv.org/pdf/2609.34968", "code_url": "", "clusters": ["C3", "C4"], "keywords": ["federated learning", "LoRA experts", "MoE aggregation"], "model_class": "world-action model", "techniques": ["federated PEFT", "LoRA", "MoE routing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Federated adaptation of world-action models where client LoRA adapters are assembled as server-side MoE experts.", "relevance": 3, "relevance_reason": "Communication-efficient distributed adaptation via PEFT.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:LoRA"}
{"id": "2603.22126", "title": "ROBOGATE: Adaptive Failure Discovery for Safe Robot Policy Deployment via Two-Stage Boundary-Focused Sampling", "authors": ["Azuki Kim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.22126", "pdf_url": "https://arxiv.org/pdf/2603.22126", "code_url": "https://github.com/liveplex-cpu/robogate", "clusters": ["C4", "C7"], "keywords": ["ROBOGATE", "safe deployment", "pre-deployment validation", "failure boundaries"], "model_class": "other", "techniques": ["failure discovery", "adaptive sampling"], "base_model": "GR00T N1.6", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Isaac Sim", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-stage boundary-focused sampling in simulation to find failure regions of manipulation policies before deployment; benchmarks eight VLAs.", "relevance": 3, "relevance_reason": "Pre-deployment falsification complements runtime certification.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2604.09860", "title": "RoboLab: A High-Fidelity Simulation Benchmark for Analysis of Task Generalist Policies", "authors": ["Yang, Jenai Xuning", "Dagli, Rishit", "Zook, Alex", "Hadfield, Hugo", "Goyal, Ankit", "Birchfield, Stan", "Ramos, Fabio", "Tremblay, Jonathan"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2604.09860", "pdf_url": "https://arxiv.org/pdf/2604.09860", "code_url": "", "clusters": ["C4"], "keywords": ["RoboLab", "saturation", "domain overlap", "generalization axes"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboLab-120"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "High-fidelity sim benchmarking framework (120 tasks, 3 competency axes, 3 difficulties) designed against saturation and train/eval overlap, with analysis of which factors drive behavior.", "relevance": 3, "relevance_reason": "Addresses benchmark saturation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-to-sim Gaussian splatting policy evaluation"}
{"id": "2609.30715", "title": "RoboMonitor: Label-Efficient Runtime Monitoring of Robot Task Execution via Predictive Representation Learning", "authors": ["Abhiroop Ajith", "Gokul Narayanan", "Kyle Coelho", "Tingji Zhao", "Yash Shahapurkar", "Brian Zhu", "Melih Erdogan", "Ted Krubasik", "Constantinos Chamzas", "Eugen Solowjow"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.30715", "pdf_url": "https://arxiv.org/pdf/2609.30715", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["RoboMonitor", "execution monitoring", "predictive representation"], "model_class": "other", "techniques": ["self-supervised pretraining", "temporal SFT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Label-efficient vision-language execution monitor pretrained on robot trajectories with predictive objectives, then fine-tuned to track phase, failure and completion.", "relevance": 3, "relevance_reason": "Runtime monitoring.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2609.28952", "title": "RoboRecover: Benchmarking Robot Policy Recovery under Execution Deviations", "authors": ["Yang Li", "Chen Zhao", "Zhuoran Wang", "Jiankang Wang", "Chao Shao", "Yihan Lin", "Haitao Shen", "Jing Zhang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.28952", "pdf_url": "https://arxiv.org/pdf/2609.28952", "code_url": "", "clusters": ["C4"], "keywords": ["RoboRecover", "recovery", "execution deviations"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin", "LIBERO"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "2,000 deviation states reconstructed by replaying action prefixes on RoboTwin and LIBERO; recovery ability is not predicted by initial-state success.", "relevance": 3, "relevance_reason": "New evaluation dimension for closed-loop robustness.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2602.11934", "title": "Robot-DIFT: Correspondence-Sensitive Diffusion Features for Contact-Rich Robot Manipulation", "authors": ["Yu Deng", "Yufeng Jin", "Xiao-Gang Jia", "Jiahong Xue", "Gerhard Neumann", "Georgia Chalvatzaki"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.11934", "pdf_url": "https://arxiv.org/pdf/2602.11934", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["Robot-DIFT", "diffusion features"], "model_class": "other", "techniques": ["distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa", "LIBERO-10", "real robot"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills a noisy diffusion feature teacher into a deterministic single-pass backbone that supplies correspondence-rich features for real-time control.", "relevance": 3, "relevance_reason": "Distillation to make diffusion features usable at control rate.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.06749", "title": "Robotic Foundation Models for Industrial Control: A Comprehensive Survey and Readiness Assessment Framework", "authors": ["David Kube", "Simon Hadwiger", "Tobias Meisen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.06749", "pdf_url": "https://arxiv.org/pdf/2603.06749", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["industrial RFMs", "readiness assessment", "edge constraints", "survey"], "model_class": "VLA", "techniques": ["survey", "assessment framework"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey and 149-criterion readiness framework rating 324 manipulation foundation models for industry, including edge-compute and real-time requirements.", "relevance": 3, "relevance_reason": "Deployability/edge constraints of RFMs in factories.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2607.20033", "title": "Robots Acquire Manipulation Skills in Seconds from a Single Human Video", "authors": ["Guang-Yan Chen", "Meiling Wang", "Te Cui", "Zi-Chen Zhou", "Qi Shao", "Shalfun Li", "Hang Su", "Roy Gan", "Hao Wang", "Meng-Yin Fu", "Yi Yang", "Yu-Feng Yue"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.20033", "pdf_url": "https://arxiv.org/pdf/2607.20033", "code_url": "", "clusters": ["C3"], "keywords": ["HOST", "human video"], "model_class": "other", "techniques": ["one-shot adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Acquires new manipulation skills from one human video in seconds without a training loop, while retaining existing skills.", "relevance": 3, "relevance_reason": "Fast adaptation without fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.15275", "title": "RoboTTT: Context Scaling for Robot Policies", "authors": ["Yunfan Jiang", "Yevgen Chebotar", "Ruijie Zheng", "Fengyuan Hu", "Yunhao Ge", "Jimmy Wu", "Tianyuan Dai", "Scott Reed", "Li Fei-Fei", "Yuke Zhu", "Linxi \"Jim\" Fan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.15275", "pdf_url": "https://arxiv.org/pdf/2607.15275", "code_url": "", "clusters": ["C3"], "keywords": ["context scaling"], "model_class": "VLA", "techniques": ["test-time training", "fast weights", "truncated BPTT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Integrates test-time-training fast-weight layers into robot foundation models to scale visuomotor context to 8K timesteps without growing inference latency.", "relevance": 3, "relevance_reason": "Long context at constant latency; on-the-fly improvement.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=0]"}
{"id": "2608.18410", "title": "Role-Conditioned Sub-Token Routing for Efficient Vision-Language-Action Policies", "authors": ["Wei Jiang", "Wei Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.18410", "pdf_url": "https://arxiv.org/pdf/2608.18410", "code_url": "", "clusters": ["C1"], "keywords": ["sub-token compression", "KV budget", "value routing"], "model_class": "VLA", "techniques": ["KV compression", "token pruning"], "base_model": "OpenVLA-OFT-7B", "params": "", "hardware": "", "reported_metric": "total KV reduced to 9.2-11.3%", "metric_measures": "relative memory", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Compresses the value vectors of retained tokens with a role- and language-conditioned router, complementing token pruning for aggressive KV reduction.", "relevance": 3, "relevance_reason": "KV compression for VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning"}
{"id": "2606.14010", "title": "RT-VLA: Real-Time Vision-Language-Action Models via Knowledge Distillation", "authors": ["Xiangyu Huang", "Zhenlin Hua", "Han Zhou", "Shounak Sural", "Ragunathan Rajkumar", "Zhen-Tao Hua", "Hanyu Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.14010", "pdf_url": "https://arxiv.org/pdf/2606.14010", "code_url": "", "clusters": ["C1", "C2", "C8", "C9"], "keywords": ["distillation", "autonomous driving VLA", "knowledge distillation", "driving VLA", "real-time"], "model_class": "VLA", "techniques": ["distillation", "knowledge distillation"], "base_model": "SimLingo", "params": "", "hardware": "", "reported_metric": "44.8x (vision-only) / 7.9x (vision+language) lower inference time", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills the SimLingo driving VLA into a compact student that keeps language reasoning off the real-time path.", "relevance": 3, "relevance_reason": "Distillation for real-time VLA (driving domain).", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2603.21635", "title": "RTD-RAX: Fast, Safe Trajectory Planning for Systems under Unknown Disturbances", "authors": ["Evanns Morales-Cuadrado", "Long Kiu Chung", "Shreyas Kousik", "Samuel Coogan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.21635", "pdf_url": "https://arxiv.org/pdf/2603.21635", "code_url": "", "clusters": ["C7"], "keywords": ["online certification", "trajectory repair"], "model_class": "other", "techniques": ["reachability-based trajectory design", "mixed monotone reachability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Runtime-assurance extension of reachability-based trajectory design using mixed-monotone reachability for fast online certification and repair of candidate trajectories under disturbances.", "relevance": 3, "relevance_reason": "Fast online certify-then-repair loop for proposed trajectories.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2605.12651", "title": "Runtime Monitoring of Perception-Based Autonomous Systems via Embedding Temporal Logic", "authors": ["Parv Kapoor", "Abigail Hammer", "Ashish Kapoor", "Karen Leung", "Eunsuk Kang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.12651", "pdf_url": "https://arxiv.org/pdf/2605.12651", "code_url": "", "clusters": ["C7"], "keywords": ["embedding temporal logic", "runtime monitoring", "perception-based systems"], "model_class": "other", "techniques": ["temporal logic over embeddings"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Defines temporal-logic predicates via distances in learned embedding spaces so perception-driven systems can be monitored without brittle symbol-grounding modules.", "relevance": 3, "relevance_reason": "Formal runtime monitoring for learned perception.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2503.08558 (S2)"}
{"id": "2609.19665", "title": "Runtime Safety Filtering for Two-Terminal Hazards in Robotic Battery Recycling", "authors": ["Yu-Xin Cao", "Wei Song", "Xiang-Lin Yang", "Fu-Sen Guo", "Lin Li", "Xiao Cheng", "Jin-Song Dong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19665", "pdf_url": "https://arxiv.org/pdf/2609.19665", "code_url": "", "clusters": ["C7"], "keywords": ["safety filter", "fallback"], "model_class": "VLA", "techniques": ["runtime safety filter"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Studies predicate, margin and fallback design of runtime safety filters around frozen OpenVLA policies.", "relevance": 3, "relevance_reason": "Runtime safety filtering.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2309.05837"}
{"id": "2607.06558", "title": "RynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation", "authors": ["Haoyu Zhao", "Xingyue Zhao", "Hangyu Li", "Biao Gong", "Kehan Li", "Siteng Huang", "Xin Li", "Deli Zhao", "Zhongyu Li", "Hao-Yu Zhao"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.06558", "pdf_url": "https://arxiv.org/pdf/2607.06558", "code_url": "https://github.com/alibaba-damo-academy/RynnWorld-Teleop", "clusters": ["C2", "C8"], "keywords": ["digital teleoperation", "video DiT"], "model_class": "world model", "techniques": ["streaming AR distillation", "video DiT", "autoregressive distillation", "single-pass generation"], "base_model": "", "params": "", "hardware": "single H100", "reported_metric": "40+ FPS generation", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Hand-pose-driven robot world model distilled via streaming autoregressive distillation to single-pass inference for real-time digital teleoperation data generation.", "relevance": 3, "relevance_reason": "World model distillation to real-time.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.30113", "title": "SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performance", "authors": ["Teng-Long Jiang", "Chun-Pu Xu", "Jiayue Kang", "Yao Mu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.30113", "pdf_url": "https://arxiv.org/pdf/2606.30113", "code_url": "", "clusters": ["C1"], "keywords": ["SA-VLA", "VQ tokenizer"], "model_class": "VLA", "techniques": ["action tokenization", "state adapter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "State-conditioned VQ action tokenizer that reduces reconstruction error of discrete action codes while keeping discrete decoding efficient.", "relevance": 3, "relevance_reason": "Action tokenization (C1).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2410.24164"}
{"id": "2604.18482", "title": "Safe Control using Learned Safety Filters and Adaptive Conformal Inference", "authors": ["Sacha Huriot", "Ihab Tabbara", "Hussein Sibai"], "year": 2026, "venue": "L4DC 2026", "type": "preprint", "url": "https://arxiv.org/abs/2604.18482", "pdf_url": "https://arxiv.org/pdf/2604.18482", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["ACoFi", "learned safety filters", "adaptive conformal inference", "HJ reachability", "adaptive conformal filtering", "learned HJ safety filter"], "model_class": "other", "techniques": ["adaptive conformal inference on learned HJ filter", "adaptive conformal inference on learned safety filter switching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Wraps learned HJ reachability safety filters with adaptive conformal inference to adjust switching thresholds online and bound safety violations from model errors.", "relevance": 3, "relevance_reason": "Calibration technique reused by world-model safety filters.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"hamilton-jacobi\" reachability learned policy safety filter; snowball from 2502.00935 (citations)"}
{"id": "2512.11908", "title": "Safe Learning for Contact-Rich Robot Tasks: A Survey from Classical Learning-Based Methods to Safe Foundation Models", "authors": ["Heng Zhang", "Rui Dai", "Gokhan Solak", "Pokuang Zhou", "Nikos Tsagarakis", "Yu She", "Arash Ajoudani"], "year": 2026, "venue": "npj Robotics", "type": "paper", "url": "https://arxiv.org/abs/2512.11908", "pdf_url": "https://arxiv.org/pdf/2512.11908", "code_url": "", "clusters": ["C7"], "keywords": ["survey", "safe learning", "contact-rich", "safe foundation models"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of safe exploration and safe execution methods for contact-rich robot learning (constrained RL, CBFs, shields, uncertainty) and how they extend to VLA/foundation models.", "relevance": 3, "relevance_reason": "Survey bridging classical safe learning and foundation models.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: safety filter learned policy robot foundation model"}
{"id": "2608.04246", "title": "SAFECAST: Robust Failure Detection for VLA Policies with Contrast-Set Training and Calibration", "authors": ["Harshitha Rajaprakash", "Aditeya Prajapati", "Rong Xue", "Abrar Anwar", "Jesse Thomason"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.04246", "pdf_url": "https://arxiv.org/pdf/2608.04246", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["SAFECAST", "contrast sets", "functional conformal prediction", "hidden-state probe"], "model_class": "VLA", "techniques": ["contrast-set augmentation", "calibration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DROID", "LIBERO"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Improves hidden-state failure probes and their conformal calibration for deployment shift by training/calibrating on visual and language contrast-set perturbations.", "relevance": 3, "relevance_reason": "Calibrated runtime failure detection.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2605.12386", "title": "SafeManip: A Property-Driven Benchmark for Temporal Safety Evaluation in Robotic Manipulation", "authors": ["Chengyue Huang", "Khang Vo Huynh", "Mellon M. Zhang", "Sebastian Elbaum", "Zsolt Kira", "Lu Feng"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2605.12386", "pdf_url": "https://arxiv.org/pdf/2605.12386", "code_url": "", "clusters": ["C4", "C7", "C8", "C9"], "keywords": ["SafeManip", "LTLf", "temporal safety", "property-driven benchmark"], "model_class": "other", "techniques": ["temporal-logic properties over rollouts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa365", "LIBERO"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "LTLf-based temporal safety properties with per-trigger normalized metrics, audited on seven robot foundation policies; success and safety are only weakly coupled.", "relevance": 3, "relevance_reason": "Complementary metric layer to success rate.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark; arXiv search: safe robot foundation models"}
{"id": "2609.21223", "title": "SafeStage: Evaluating Safety Before, During, and After Vision-Language-Conditioned Robot Manipulation", "authors": ["Jin-Zhu Luo", "Qi Zhang", "Wei Wang", "Wei Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21223", "pdf_url": "https://arxiv.org/pdf/2609.21223", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["safety evaluation", "lifecycle"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Evaluates VLA and world-model policies for safety before, during and after manipulation under a common closed-loop protocol.", "relevance": 3, "relevance_reason": "Safety benchmarking for embodied foundation policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2406.09246"}
{"id": "2604.01346", "title": "Safety, Security, and Cognitive Risks in World Models", "authors": ["Manoj Parmar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.01346", "pdf_url": "https://arxiv.org/pdf/2604.01346", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["world model risks", "security", "survey"], "model_class": "world model", "techniques": ["survey/taxonomy of safety and security risks"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey of safety, security and cognitive risks specific to world models, from data poisoning and compounding rollout errors to automation bias in human users.", "relevance": 3, "relevance_reason": "Risk framing for world-model-based safety.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.00773", "title": "SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models", "authors": ["Jialiang Fan", "Weizhe Xu", "Oleg Sokolsky", "Insup Lee", "Fanxin Kong"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.00773", "pdf_url": "https://arxiv.org/pdf/2606.00773", "code_url": "", "clusters": ["C4", "C7", "C8", "C9"], "keywords": ["SafeVLA-Bench", "signal temporal logic", "unsafe success", "success-safety gap", "STL specifications"], "model_class": "VLA", "techniques": ["STL-based post-hoc safety evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboCasa365"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Post-hoc STL-based safety layer over LIBERO and RoboCasa365 reporting Succ-But-Unsafe and a violation severity index; many successful rollouts still violate safety constraints.", "relevance": 3, "relevance_reason": "Metric beyond success rate; useful for judging accelerated policies whose trajectories change.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark; arXiv search: SafeVLA"}
{"id": "2503.03480", "title": "SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning", "authors": ["Borong Zhang", "Yuhao Zhang", "Jiaming Ji", "Yingshan Lei", "Yishuai Cai", "Josef Dai", "Yuanpei Chen", "Yaodong Yang"], "year": 2026, "venue": "NeurIPS 2025 (Spotlight)", "type": "paper", "url": "https://arxiv.org/abs/2503.03480", "pdf_url": "https://arxiv.org/pdf/2503.03480", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["SafeVLA", "constrained RL", "CMDP", "safety alignment"], "model_class": "VLA", "techniques": ["safe RL", "min-max CMDP optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "-83.58% cumulative safety cost vs SOTA, +3.85% success", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Integrated safety approach that elicits unsafe VLA behaviors and aligns the policy via constrained RL, reducing safety violations while keeping task success.", "relevance": 3, "relevance_reason": "Training-time safety for VLAs (not runtime).", "foundational_technique": false, "verified": true, "found_via": "arXiv search: SafeVLA"}
{"id": "2609.34268", "title": "SAGE: Symbolic Action-Gating and Editing for LLM Task Planners", "authors": ["Trung Minh Bui", "JongSul Moon", "YoungOuk Kim", "Quang-Ngoc Phung", "Se-Woong Jun", "Dongin Shin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34268", "pdf_url": "https://arxiv.org/pdf/2609.34268", "code_url": "https://github.com/mtbui2010/sage_release", "clusters": ["C7"], "keywords": ["SAGE", "LLM planner"], "model_class": "other", "techniques": ["symbolic gating", "runtime monitor", "local replanning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.4-3.3x fewer LLM calls for recovery", "metric_measures": "other", "benchmarks": ["AI2-THOR"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LLM task planner with a zero-token symbolic precondition gate that blocks invalid actions before execution plus local suffix repair.", "relevance": 3, "relevance_reason": "Verify-before-execute gating with negligible cost.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot policy\" OR embodied) AND (Jetson OR \"on-device\" OR onboard)"}
{"id": "2603.08124", "title": "SaiVLA-0: Cerebrum--Pons--Cerebellum Tripartite Architecture for Compute-Aware Vision-Language-Action", "authors": ["Xiang Shi", "Wenlong Huang", "Menglin Zou", "Xinhai Sun", "Xiangqun Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.08124", "pdf_url": "https://arxiv.org/pdf/2603.08124", "code_url": "", "clusters": ["C1"], "keywords": ["SaiVLA-0", "compute-aware VLA", "tripartite architecture"], "model_class": "VLA", "techniques": ["feature caching", "hierarchical architecture", "frozen backbone", "parallel categorical decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "training 7.5 h -> 4.5 h with split caching (N1.5 head-only)", "metric_measures": "other", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Concept-and-protocol paper for a compute-aware tripartite VLA with split feature caching; preliminary LIBERO evidence of reduced training time.", "relevance": 3, "relevance_reason": "Proposes a timing protocol for verifying efficiency claims.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.32108", "title": "SAMBAR: Selective Anchoring via Method of Multipliers for Balanced Knowledge Acquisition and Retention in Vision-Language-Action Models", "authors": ["Aayushi Shrivastava", "Xunlan Zhou", "Hongrui Zhao", "Ziyu Chen", "Negar Mehr"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.32108", "pdf_url": "https://arxiv.org/pdf/2609.32108", "code_url": "", "clusters": ["C3"], "keywords": ["SAMBAR", "constrained optimization", "replay-free"], "model_class": "VLA", "techniques": ["method of multipliers", "selective anchoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Replay-free continual VLA fine-tuning cast as constrained optimization with selective parameter anchoring.", "relevance": 3, "relevance_reason": "Continual adaptation without storing old data.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"continual learning\""}
{"id": "2606.03724", "title": "Same Weights, Different Robot: A Deployment Safety View of VLA Policies", "authors": ["Jianwei Tai"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.03724", "pdf_url": "https://arxiv.org/pdf/2606.03724", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["deployment safety", "action normalization"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows identical VLA weights can act differently across robots due to action unnormalization metadata, urging pre-rollout checks.", "relevance": 3, "relevance_reason": "Deployment evaluation methodology for VLAs", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2605.15178", "title": "SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer", "authors": ["Haoyi Zhu", "Haozhe Liu", "Yuyang Zhao", "Tian Ye", "Junsong Chen", "Jincheng Yu", "Tong He", "Song Han", "Enze Xie"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.15178", "pdf_url": "https://arxiv.org/pdf/2605.15178", "code_url": "", "clusters": ["C8"], "keywords": ["hybrid linear attention", "minute-scale world model", "Gated DeltaNet"], "model_class": "world model", "techniques": ["hybrid linear + softmax attention", "two-stage generation"], "base_model": "", "params": "2.6B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "2.6B open world model with hybrid linear attention for memory-efficient minute-long 720p generation at quality comparable to much larger industrial models.", "relevance": 3, "relevance_reason": "Efficient long-context world-model architecture.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.21740", "title": "Sandwich-Residuals: Parameter-Efficient Test-time Adaptation of World Models", "authors": ["Krishnam Soni", "Aditya Sehgal", "Vedant Dave", "Elmar Rueckert"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21740", "pdf_url": "https://arxiv.org/pdf/2609.21740", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["test-time adaptation", "frozen world model", "residual adapters"], "model_class": "world model", "techniques": ["small online residual corrections around frozen predictor"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["AdaJEPA"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Adapts a frozen latent world model at test time with tiny residual modules trained on its own prediction error, improving planning under shift without touching most parameters.", "relevance": 3, "relevance_reason": "Parameter-efficient adaptation of planning world models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2602.10717", "title": "Say, Dream, and Act: Learning Video World Models for Instruction-Driven Robot Manipulation", "authors": ["Songen Gu", "Yunuo Cai", "Tian-Yu Wang", "Si Wu", "Yanwei Fu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.10717", "pdf_url": "https://arxiv.org/pdf/2602.10717", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["video world model", "instruction-driven"], "model_class": "world model", "techniques": ["adversarial distillation", "few-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts a video generator as a world model, distills it adversarially for few-step generation, and trains an action model on generated plus real observations.", "relevance": 3, "relevance_reason": "Few-step distilled world model for control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.18610", "title": "SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation", "authors": ["Wei-Cheng Tseng", "Gashon Hussein", "Yuzhu Dong", "Allen Z. Ren", "L. Shi", "Xu-Dong Wang", "Sergey Levine", "Zhao-Shuo Li", "Jinwei Gu", "Florian Shkurti", "Ming-Yu Liu", "Q. Vuong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18610", "pdf_url": "https://arxiv.org/pdf/2606.18610", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["policy evaluation", "video world model"], "model_class": "world model", "techniques": ["video world model evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SC3-Eval evaluates generalist robot policies inside a self-consistent, multi-view action-conditioned video world model as a scalable alternative to real-world testing.", "relevance": 3, "relevance_reason": "relevant: benchmarking methodology", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941, 2503.00200"}
{"id": "2602.04208", "title": "SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models", "authors": ["Hyeonbeom Choi", "Daechul Ahn", "Youhan Lee", "Taewook Kang", "Seongwon Cho", "Jonghyun Choi"], "year": 2026, "venue": "ICML 2026", "type": "preprint", "url": "https://arxiv.org/abs/2602.04208", "pdf_url": "https://arxiv.org/pdf/2602.04208", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["self-uncertainty", "active inference", "test-time scaling", "efficient inference"], "model_class": "VLA", "techniques": ["uncertainty modulation", "test-time scaling", "single-pass uncertainty modulation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Training-free inference strategy that modulates perception and action by self-uncertainty in a single forward pass, replacing costly multi-pass test-time scaling.", "relevance": 3, "relevance_reason": "Uncertainty-conditioned adaptive execution at no extra passes.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]; snowball round 2: cited by 2607.00666"}
{"id": "2603.18532", "title": "Scaling Sim-to-Real VLA Reinforcement Learning with Generative 3D Worlds", "authors": ["Andrew Choi", "Xinjie Wang", "Zhizhong Su", "Wei Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.18532", "pdf_url": "https://arxiv.org/pdf/2603.18532", "code_url": "", "clusters": ["C3"], "keywords": ["sim-to-real RL", "generative 3D worlds", "scene diversity"], "model_class": "VLA", "techniques": ["RL in simulation", "generated scenes"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "RL fine-tunes VLAs across many auto-generated 3D scenes to keep generality while transferring sim-to-real.", "relevance": 3, "relevance_reason": "Scalable RL post-training without real-world cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2602.12281", "title": "Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment", "authors": ["Jacky Kwok", "Xilun Zhang", "Mengdi Xu", "Yuejiang Liu", "Azalia Mirhoseini", "Chelsea Finn", "Marco Pavone"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.12281", "pdf_url": "https://arxiv.org/pdf/2602.12281", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["test-time scaling", "VLA alignment"], "model_class": "VLA", "techniques": ["test-time verification", "contrastive verifier"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows jointly scaling rephrased instructions and sampled actions at test time recovers correct actions, and proposes a contrastive verifier (CoVer) and hierarchical verification pipeline.", "relevance": 3, "relevance_reason": "Test-time verification of VLA outputs as an alternative to more training.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunk*\" AND (uncertaint* OR \"out-of-distribution\")"}
{"id": "2608.19425", "title": "SCAPE: Scenario-Conditioned Simulation-Augmented Policy Evaluation", "authors": ["Dijie Zhu", "Seunghun Oh", "Ruopeng Huang", "Zhiyu Huang", "Jiaqi Ma", "Chen Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.19425", "pdf_url": "https://arxiv.org/pdf/2608.19425", "code_url": "", "clusters": ["C4"], "keywords": ["SCAPE", "sim-augmented evaluation", "conformal prediction", "scenario-conditioned"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["autonomous driving", "quadruped velocity tracking"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Predicts scenario-conditioned real-world performance from few paired sim/real samples plus many sim rollouts, correcting sim bias and calibrating uncertainty with conformal prediction.", "relevance": 3, "relevance_reason": "Statistical evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation statistical"}
{"id": "2607.04609", "title": "SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies", "authors": ["Dijia Zhan", "Xuemiao Xu", "Jinyi Li", "Jie Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.04609", "pdf_url": "https://arxiv.org/pdf/2607.04609", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["chunk boundary smoothing", "training-free guidance", "SEAM", "multimodal bifurcation", "velocity-guided loss steering", "training-free"], "model_class": "VLA", "techniques": ["guided sampling", "closed-form velocity correction toward previous tail"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-10"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free steering of flow VLAs toward the previous chunk’s unexecuted tail with a closed-form velocity correction, preventing cross-chunk mode switches.", "relevance": 3, "relevance_reason": "Chunk-boundary smoothness.", "foundational_technique": false, "verified": true, "found_via": "search query: denoising steps VLA; snowball: citations of 2506.07339 (S2)"}
{"id": "2602.15543", "title": "Selective Perception for Robot: Task-Aware Attention in Multimodal VLA", "authors": ["Young-Chae Son", "Jung-Woo Lee", "Yoon-Ji Choi", "Dae-Kwan Ko", "Soo-Chul Lim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.15543", "pdf_url": "https://arxiv.org/pdf/2602.15543", "code_url": "", "clusters": ["C2"], "keywords": ["multi-view VLA", "active perception"], "model_class": "VLA", "techniques": ["view routing", "conditional computation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight router predicts task relevance of each camera view from the prompt and wrist camera and attenuates computation for low-utility views.", "relevance": 3, "relevance_reason": "Dynamic multi-view input reduction.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]"}
{"id": "2609.30092", "title": "Self-Adaptive VLA for Robust Robot Deployment", "authors": ["Hongxin Zhang", "Chunru Lin", "Tsun-Hsuan Wang", "Zhenjia Xu", "Chuang Gan", "Hong-Xin Zhang", "Chun-Tse Lin", "T. Wang", "Zhen-Jia Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.30092", "pdf_url": "https://arxiv.org/pdf/2609.30092", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["hardware shift", "self-adaptation", "test-time adaptation"], "model_class": "VLA", "techniques": ["in-context adaptation", "context compression", "context encoder", "AdaLN modulation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Post-training recipe where a plug-in context encoder compresses recent rollouts so a VLA adapts to deployment hardware shifts (wear, miscalibration) in context.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (\"failure recovery\" OR \"error recovery\" OR \"self-correct*\"); snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2609.13236", "title": "Self-Evolving AI for Humanoids: Mechanisms, Safety, and Evaluation of Post-Deployment Self-Improvement", "authors": ["Loc X. Nguyen", "Avi Deb Raha", "Huy Q. Le", "Eui-Nam Huh", "Dusit Niyato", "Choong Seon Hong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.13236", "pdf_url": "https://arxiv.org/pdf/2609.13236", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["self-evolving humanoids", "safety"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of post-deployment self-evolution for humanoids, framing admissible updates as gated by world-model verification inside human oversight.", "relevance": 3, "relevance_reason": "Links continual adaptation with safety gating.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"continual learning\" OR lifelong OR \"test-time adaptation\")"}
{"id": "2608.00725", "title": "SelfWAM: A Self-Grounded Unified World Action Model for Fast Robot Control", "authors": ["Bikang Pan", "Fan Liu", "Haotao Lu", "Jingya Wang", "Ye Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00725", "pdf_url": "https://arxiv.org/pdf/2608.00725", "code_url": "", "clusters": ["C8"], "keywords": ["self-grounded WAM", "robot self-mask", "fast action-only path"], "model_class": "world-action model", "techniques": ["action-conditioned future video + self-mask prediction in training", "unchanged action-only inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Makes the training-time video branch model action-specific consequences (via robot self-masks) while keeping a fast action-only inference path.", "relevance": 3, "relevance_reason": "Better training signal for fast WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2605.23477", "title": "Semantically Structured Mixture-of-Experts for Compositional Robotic Manipulation", "authors": ["Cheng-Cheng Deng", "Guan-Qi Chen", "Yizhou Chen", "Ze-Jia Liu", "Zhiwen Ruan", "Guan-Hua Chen", "Jia Pan"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2605.23477", "pdf_url": "https://doi.org/10.15607/rss.2026.xxii.057", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["SMoDP", "skill routing"], "model_class": "diffusion/flow policy", "techniques": ["mixture of experts", "PEFT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Semantically routed MoE diffusion policy with a lightweight skill predictor, improving parameter efficiency and enabling parameter-efficient transfer.", "relevance": 3, "relevance_reason": "Sparse-activation policy efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2606.21493", "title": "Semi-Supervised Vision-Language-Action Model", "authors": ["Hongyang He", "Jiuming Liu", "Victor Sanchez"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.21493", "pdf_url": "https://arxiv.org/pdf/2606.21493", "code_url": "", "clusters": ["C3"], "keywords": ["semi-supervised", "pseudo-actions", "label efficiency", "SemiVLA"], "model_class": "VLA", "techniques": ["self-distillation", "LoRA", "semi-supervised learning"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "89.0% LIBERO with 10% labeled trajectories", "metric_measures": "other", "benchmarks": ["LIBERO", "CALVIN"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Semi-supervised VLA adaptation that pseudo-labels action-free trajectories and plugs into PEFT methods, reducing labeled-demo needs.", "relevance": 3, "relevance_reason": "Data-efficient PEFT adaptation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: parameter-efficient fine-tuning vision-language-action model arXiv"}
{"id": "10.3390/s26113541", "title": "Sensing the Action: Rethinking Sensor Modalities and Multi-Modal Fusion in Vision–Language–Action Models for Robotic Manipulation", "authors": ["B. Ko"], "year": 2026, "venue": "Italian National Conference on Sensors", "type": "paper", "url": "https://doi.org/10.3390/s26113541", "pdf_url": "https://doi.org/10.3390/s26113541", "code_url": "", "clusters": ["C4"], "keywords": ["survey", "sensor fusion", "evaluation metrics"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Sensor-centric survey of VLAs arguing for evaluation that includes latency, efficiency and safety besides success rate.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2605.01191", "title": "Sentinel-VLA: A Metacognitive VLA Model with Active Status Monitoring for Dynamic Reasoning and Error Recovery", "authors": ["Wenhao Li", "Xiu Su", "Dan Niu", "Yichao Cao", "Hongyan Xu", "Zhe Qu", "Lei Fan", "Shan You", "Chang Xu", "Wen-Hao Li", "Dantong Niu", "Hong-Yan Xu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.01191", "pdf_url": "https://arxiv.org/pdf/2605.01191", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["metacognition", "error recovery", "status monitoring"], "model_class": "VLA", "techniques": ["execution monitoring", "on-demand reasoning", "continual learning", "orthogonal continual adapters"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VLA with a sentinel module monitoring execution status that triggers reasoning or recovery only when needed, trained on auto-generated data with a continual learning scheme.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (\"failure recovery\" OR \"error recovery\" OR \"self-correct*\"); snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2602.09580", "title": "SERNF: Sample-Efficient Real-World Dexterous Policy Fine-Tuning via Action-Chunked Critics and Normalizing Flows", "authors": ["Chenyu Yang", "D. Tarasov", "Davide Liconti", "He-Hui Zheng", "Robert K. Katzschmann", "Denis Tarasov", "Romain Guntz"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://doi.org/10.48550/arXiv.2602.09580", "pdf_url": "https://arxiv.org/pdf/2602.09580", "code_url": "", "clusters": ["C3"], "keywords": ["fine-tuning", "dexterous", "sample-efficient RL fine-tuning"], "model_class": "diffusion/flow policy", "techniques": ["action-chunked critics", "normalizing flows", "chunked critics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Sample-efficient real-world fine-tuning of dexterous policies using action-chunked critics and normalizing flows (title only; no abstract available).", "relevance": 3, "relevance_reason": "relevant (from title): sample-efficient fine-tuning", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.07193", "title": "Shield-Loco: Shielding Locomotion Policies with Predictive Safety Filtering", "authors": ["Aditya Shirwatkar", "Sebastian Sanokowski", "Shishir Kolathaya", "Aaron Johnson", "Majid Khadiv"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.07193", "pdf_url": "https://arxiv.org/pdf/2606.07193", "code_url": "", "clusters": ["C7"], "keywords": ["Shield-Loco", "predictive safety filter", "locomotion", "asynchronous optimization"], "model_class": "other", "techniques": ["sampling-based predictive filtering", "asynchronous contact optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Post-hoc predictive safety filter for RL locomotion that asynchronously searches safer contact sequences with a full-physics model when collisions are predicted.", "relevance": 3, "relevance_reason": "Asynchronous safety filter for learned policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety filter\" learned policy reinforcement learning robot"}
{"id": "2609.13231", "title": "ShieldVLA: Feasibility-Aware Safety Alignment for Vision-Language-Action Models", "authors": ["Manan Tayal", "Akshay Nambi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.13231", "pdf_url": "https://arxiv.org/pdf/2609.13231", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["ShieldVLA", "HJ reachability", "safety critic", "fine-tuning"], "model_class": "VLA", "techniques": ["HJ reachability value learning", "VLM rubric safety scores"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Safety-aligned VLA fine-tuning using a model-free HJ reachability critic from vision that separates reward optimization in feasible regions from recovery near unsafe states.", "relevance": 3, "relevance_reason": "Reachability-based safety for VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"reachability\" vision-language-action safety"}
{"id": "2603.18342", "title": "Shifting Uncertainty to Critical Moments: Towards Reliable Uncertainty Quantification for VLA Model", "authors": ["Yan-Chuan Tang", "Tao-Wen Wang", "Yue Chen", "Bo-Xuan Zhang", "Qiang Guan", "Rui-Xiang Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.18342", "pdf_url": "https://arxiv.org/pdf/2603.18342", "code_url": "", "clusters": ["C7"], "keywords": ["failure prediction"], "model_class": "VLA", "techniques": ["uncertainty quantification", "sliding-window pooling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uncertainty aggregation for VLAs that preserves short risk spikes to better predict rollout failures.", "relevance": 3, "relevance_reason": "Runtime monitoring (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2310.16828"}
{"id": "2606.00090", "title": "Silent Failures in Physical AI: A Literature Review of Runtime Action Authorization for Autonomous Systems", "authors": ["Barak Or"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.00090", "pdf_url": "https://arxiv.org/pdf/2606.00090", "code_url": "", "clusters": ["C7"], "keywords": ["runtime action authorization", "silent failures", "literature review"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reviews runtime action authorization for physical AI, arguing confident but wrong actions from VLAs/world models need an authorization layer between model and actuator.", "relevance": 3, "relevance_reason": "Framing of runtime safety gating.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2601.14323", "title": "SilentDrift: Exploiting Action Chunking for Stealthy Backdoor Attacks on Vision-Language-Action Models", "authors": ["Bingxin Xu", "Yuzhang Shang", "Binghui Wang", "Emilio Ferrara"], "year": 2026, "venue": "ACL Findings 2026", "type": "paper", "url": "https://arxiv.org/abs/2601.14323", "pdf_url": "https://arxiv.org/pdf/2601.14323", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["SilentDrift", "backdoor", "intra-chunk open loop", "action chunking vulnerability"], "model_class": "VLA", "techniques": ["smooth backdoor perturbations"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows action chunking with delta poses creates an intra-chunk visual open loop that a stealthy C2-smooth backdoor can exploit to accumulate drift.", "relevance": 3, "relevance_reason": "Open-loop execution as a safety vulnerability.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety\" \"action chunking\""}
{"id": "2609.05533", "title": "SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models", "authors": ["C. Yin", "Wang Xu", "Jun-Peng Yang", "Sikyuen Tam", "Han-Yu Liu", "Yuan Yao", "Xiang-Rui Zeng", "Jun-Bo Cui", "Ye-Quan Wang", "Zhou-Ping Yin", "Yan-Kai Lin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.05533", "pdf_url": "https://arxiv.org/pdf/2609.05533", "code_url": "", "clusters": ["C1"], "keywords": ["memory", "latency"], "model_class": "VLA", "techniques": ["prefix caching", "native video context"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses the VLM backbone's native video context as memory and prefills shared history across consecutive decisions to keep latency near a single-frame VLA.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.15759", "title": "Simulation Distillation: Pretraining World Models in Simulation for Rapid Real-World Adaptation", "authors": ["Jacob Levy", "Tyler Westenbroek", "Kevin Huang", "Fernando Palafox", "Patrick Yin", "Shayegan Omidshafiei", "Dong-Ki Kim", "Abhishek Gupta", "David Fridovich-Keil"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2603.15759", "pdf_url": "https://arxiv.org/pdf/2603.15759", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["sim-to-real", "world model adaptation"], "model_class": "world model", "techniques": ["sim-pretrained world model", "latent dynamics adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "SimDist pretrains a world model in simulation and adapts only the latent dynamics with real-world prediction losses, enabling rapid real-world improvement via planning.", "relevance": 3, "relevance_reason": "relevant: efficient real-world adaptation of world models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2402.15391"}
{"id": "2602.18224", "title": "SimVLA: A Simple VLA Baseline for Robotic Manipulation", "authors": ["Yuankai Luo", "Woping Chen", "Tong Liang", "Baiqiao Wang", "Zhenguo Li", "T. Liang", "Zhen-Guo Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.18224", "pdf_url": "https://arxiv.org/pdf/2602.18224", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["baseline", "training recipe"], "model_class": "VLA", "techniques": ["small VLA", "lightweight action head"], "base_model": "", "params": "0.5B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Minimal 0.5B-parameter VLA baseline (standard VLM plus lightweight action head, standardized training) that beats multi-billion-parameter models in simulation and matches pi0.5 on real robots.", "relevance": 3, "relevance_reason": "Evidence that small VLAs can be competitive.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2608.07468", "title": "SimWAM: A Simple World Action Model for End-to-End Autonomous Driving", "authors": ["Zongchuang Zhao", "Xin Zhou", "Tianyang Xu", "Zhengyang Sun", "Kaixuan Zhou", "Yu Wu", "Honglin Li", "Dingkang Liang", "Xiang Bai", "Kai-Xuan Zhou", "Hong-Lin Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.07468", "pdf_url": "https://arxiv.org/pdf/2608.07468", "code_url": "https://github.com/H-EmbodVis/SimWAM", "clusters": ["C1", "C8"], "keywords": ["SimWAM", "autonomous driving", "latency", "world-action model"], "model_class": "world-action model", "techniques": ["training-only video supervision", "attention masking", "RL", "isolated attention mask"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM", "nuScenes", "WOD-E2E"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving WAM that uses future-video prediction only as training supervision, with an attention mask so trajectories are predicted without generating frames at inference.", "relevance": 3, "relevance_reason": "relevant: removes world-model generation cost at inference (driving)", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2410.24164, 2501.03575"}
{"id": "10.3390/ai7090334", "title": "SIRModel: Learning Spatial Intermediate Representation to Parameter-Efficiently Fine-Tune a Vision Language Model for Manipulation", "authors": ["Li Lin", "Ming-Hao Shi", "Teng-Long Wang"], "year": 2026, "venue": "Applied Informatics", "type": "paper", "url": "https://doi.org/10.3390/ai7090334", "pdf_url": "https://doi.org/10.3390/ai7090334", "code_url": "", "clusters": ["C3"], "keywords": ["Gaussian interaction regions", "hierarchical"], "model_class": "VLA", "techniques": ["parameter-efficient fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Spatial Gaussian intermediate representation enabling parameter-efficient adaptation of a VLM for long-horizon manipulation.", "relevance": 3, "relevance_reason": "PEFT for robot VLMs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2605.15536", "title": "SkiP: When to Skip and When to Refine for Efficient Robot Manipulation", "authors": ["Ming-Tong Dai", "Guanqi Peng", "Yongjie Bai", "Feng Yan", "Chunjie Chen", "Ling-Bo Liu", "Liang Lin", "Xinyu Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.15536", "pdf_url": "https://arxiv.org/pdf/2605.15536", "code_url": "", "clusters": ["C1"], "keywords": ["step skipping", "efficient manipulation"], "model_class": "other", "techniques": ["action relabeling", "skip segments"], "base_model": "", "params": "", "hardware": "", "reported_metric": "15-40% fewer executed steps", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Relabels free-space segments so the policy leaps over redundant steps and refines only key contact segments.", "relevance": 3, "relevance_reason": "Reduces executed decision steps.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.28064", "title": "SlackDrive: Reclaiming Runtime Slack for Adaptive Driving Inference", "authors": ["Xiaohuan Pei", "Hengguang Zhou", "Yuanhao Ban", "Justin Cui", "Jiaqi Feng", "Haoyu Xie", "Tao Huang", "Pichao Wang", "Yanchao Yang", "Cho-Jui Hsieh"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28064", "pdf_url": "https://arxiv.org/pdf/2609.28064", "code_url": "", "clusters": ["C5", "C7", "C9"], "keywords": ["runtime slack", "adaptive inference", "driving WAM"], "model_class": "world-action model", "techniques": ["slack-aware adaptive computation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SlackDrive exploits runtime slack to adapt compute of driving world-action models to real-time deadlines.", "relevance": 3, "relevance_reason": "Deadline-aware adaptive inference.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2608.29432", "title": "SMILE: Smooth Motion for Improved Long-Horizon VLA Execution", "authors": ["Jongwoo Park", "E-Ro Nguyen", "Kanchana Ranasinghe", "Cristina Mata", "Xiang Li", "Michael S Ryoo", "Jong Sung Park", "E. Nguyen", "Michael S. Ryoo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29432", "pdf_url": "https://arxiv.org/pdf/2608.29432", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["smooth chunks", "amortized inference", "action chunking", "smoothness", "spline"], "model_class": "VLA", "techniques": ["B-spline action representation", "longer execution horizon", "action chunking"], "base_model": "SmolVLA, Evo1, VPP, DAWN", "params": "", "hardware": "", "reported_metric": "1.1x speedup on LIBERO (SMILE-Evo1), 1.5x on CALVIN (SMILE-VPP) | 1.1x speedup (LIBERO, SMILE-Evo1); 1.5x speedup (CALVIN, SMILE-VPP)", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "CALVIN", "real robot (xArm)", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "SMILE predicts B-spline coefficients instead of raw action chunks, allowing longer smooth execution horizons that amortize VLA inference cost while reducing jitter.", "relevance": 3, "relevance_reason": "Longer smooth chunks cut calls per second while reducing jitter.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.29768", "title": "SmoothRL: Online Reinforcement Learning During Asynchronous Execution", "authors": ["Guang Gao", "Yuxuan Nong", "Baifu Huang", "Jianan Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29768", "pdf_url": "https://arxiv.org/pdf/2608.29768", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["SmoothRL", "online RL in async loop", "value gradient"], "model_class": "VLA", "techniques": ["value-gradient online RL within async inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Online RL framework that fine-tunes a pretrained chunked policy inside an asynchronous inference loop, keeping smooth real-time execution during learning.", "relevance": 3, "relevance_reason": "Learning under async timing.", "foundational_technique": false, "verified": true, "found_via": "arXiv API all:\"action chunking\" AND latency"}
{"id": "2606.26383", "title": "SOLAR: AI-Powered Speed-of-Light Performance Analysis", "authors": ["Qijing Huang", "Sana Damani", "Zhifan Ye", "Athinagoras Skiadopoulos", "Siva Kumar Sastry Hari", "Jason Clemons", "Sahil Modi", "Jingquan Wang", "Aditya Kane", "Edward C Lin", "Humphrey Shi", "Christos Kozyrakis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.26383", "pdf_url": "https://arxiv.org/pdf/2606.26383", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["speed-of-light", "roofline", "headroom analysis", "robotics workloads"], "model_class": "other", "techniques": ["analytical performance bounds"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Automatically derives validated theoretical minimum execution times from PyTorch/JAX code, evaluated on KernelBench and robotics workloads, enabling headroom and cross-platform comparisons.", "relevance": 3, "relevance_reason": "Tooling for hardware-normalized efficiency evaluation including robotics models.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q3: arXiv abs:roofline AND abs:robot"}
{"id": "2607.04739", "title": "Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity", "authors": ["Che-Sang Park", "Junsu Ha", "Jianlong Fu", "Frank C. Park"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.04739", "pdf_url": "https://arxiv.org/pdf/2607.04739", "code_url": "", "clusters": ["C1"], "keywords": ["adaptive execution horizon", "observation sensitivity"], "model_class": "diffusion/flow policy", "techniques": ["adaptive action chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Forecasts how sensitive the action distribution is to observation changes and shortens execution horizons when sensitivity is high.", "relevance": 3, "relevance_reason": "Adaptive chunking under a sampling budget.", "foundational_technique": false, "verified": true, "found_via": "search query: efficient diffusion policy action chunk"}
{"id": "2606.27807", "title": "SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks", "authors": ["Ruiqi Song", "Dujun Nie", "Siyu Teng", "Baiyong Ding", "Xiaotong Zhang", "Dong Li", "Chenming Zhang", "Yuchen Li", "Hangbin Wu", "Long Chen", "Rui-Qi Song", "Chen-Ming Zhang"], "year": 2026, "venue": "ICML 2026", "type": "preprint", "url": "https://arxiv.org/abs/2606.27807", "pdf_url": "https://arxiv.org/pdf/2606.27807", "code_url": "", "clusters": ["C1", "C3", "C5", "C9"], "keywords": ["spiking neural networks", "energy-efficient VLA", "neuromorphic", "navigation", "energy efficiency"], "model_class": "VLA", "techniques": ["spiking neural networks", "spiking vision encoder", "spiking LLM", "population-coded spiking policy", "SNN"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces the vision encoder, language model and action head of a VLA with spiking counterparts to cut inference energy for low-power navigation and control.", "relevance": 3, "relevance_reason": "Neuromorphic efficiency for embodied policies.", "foundational_technique": false, "verified": true, "found_via": "search query: navigation VLA efficient real-time; arXiv: vision-language-action AND (accelerator OR hardware OR GPU OR energy); venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2608.07712", "title": "SpikeWorld: Fast-State Adaptation for Frozen Spiking World Models", "authors": ["Ziqiao Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.07712", "pdf_url": "https://arxiv.org/pdf/2608.07712", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["spiking world model", "1.45M parameters", "frozen adaptation"], "model_class": "world model", "techniques": ["sparse spiking network", "frozen-weight external fast-state adaptation"], "base_model": "", "params": "1.45M", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Tiny 1.45M-parameter spiking world model that adapts after deployment through external residual paths without changing frozen weights.", "relevance": 3, "relevance_reason": "Ultra-small world model suited to neuromorphic/edge settings.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.09985 (citations)"}
{"id": "2605.21800", "title": "stable-worldmodel: A Platform for Reproducible World Modeling Research and Evaluation", "authors": ["Lucas Maes", "Quentin Le Lidec", "Luiz Facury", "Nassim Massaudi", "Ayushi Chaurasia", "Francesco Capuano", "Richard Gao", "Taj Gillin", "Dan Haramati", "Damien Scieur", "Yann LeCun", "Randall Balestriero"], "year": 2026, "venue": "arXiv", "type": "repo", "url": "https://arxiv.org/abs/2605.21800", "pdf_url": "https://arxiv.org/pdf/2605.21800", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["reproducible world modeling", "platform"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open platform with fast data loading, baselines, planners and controlled-variation environments for reproducible world-model evaluation.", "relevance": 3, "relevance_reason": "Benchmarking infrastructure for world models.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2609.20791", "title": "StageGuard: Learning Stage Transitions for Long-Horizon Robot Tasks via Agentic Distillation", "authors": ["Jinbang Huang", "Yuan Hu", "Zhi-Yuan Li", "Ran Qi", "Yi-Xin Xiao", "Yang-Zheng Wu", "Teng-Yue Ba", "Zhan-Guang Zhang", "Ying-Xue Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20791", "pdf_url": "https://arxiv.org/pdf/2609.20791", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["stage transition", "online monitoring"], "model_class": "other", "techniques": ["agentic distillation", "small student VLM"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["BEHAVIOR-1K"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills teacher-VLM reasoning into a lightweight student VLM for fast stage-transition decisions in hierarchical control.", "relevance": 3, "relevance_reason": "Distillation for low-latency runtime monitoring.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2406.09246"}
{"id": "2609.22684", "title": "StateMem: Single-State Residual Memory with Adaptive Inference for Vision-Language-Action Policies", "authors": ["Wenzhuo Li", "Qiongfeng Shi", "Yi Zhou", "Wen-Zhuo Li", "Q. Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22684", "pdf_url": "https://arxiv.org/pdf/2609.22684", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["StateMem", "memory", "cache reuse"], "model_class": "VLA", "techniques": ["prefix caching", "residual memory", "adaptive routing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "20.25% fewer VLM prefix refreshes; 97.6% LIBERO | 20.25% lower VLM prefix refresh rate vs full refresh", "metric_measures": "other", "benchmarks": ["LIBERO", "RoboMemArena", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Single persistent memory token updated by prediction-error low-rank residuals that also routes cached VLM prefixes, reducing prefix refreshes.", "relevance": 3, "relevance_reason": "Adaptive inference via prefix cache reuse.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=0]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.11671", "title": "StellaVLA: In-Context Structured Demonstration for Generalizable Vision-Language-Action Models", "authors": ["Siyu Xu", "Yun-Ke Wang", "Zijian Wang", "Di Zhu", "Chenghao Xia", "Chengbin Du", "Dao-Chang Liu", "Tao Huang", "Chang Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.11671", "pdf_url": "https://arxiv.org/pdf/2608.11671", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["retrieved demonstration", "test-time adaptation"], "model_class": "VLA", "techniques": ["in-context adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["VLA-Arena", "LIBERO", "LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adapts a VLA at test time from one retrieved, automatically structured demonstration instead of fine-tuning, while inference runs only the action expert to keep control latency unchanged.", "relevance": 3, "relevance_reason": "Fine-tuning-free adaptation with no added inference latency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2605.21282", "title": "Stochastic MeanFlow Policies: One-Step Generative Control with Entropic Mirror Descent", "authors": ["Ze-Yuan Wang", "Da Li", "Yulin Chen", "Yu Gong", "Yan-Ming Guo", "Ye Shi", "Liang Bai", "Tianyuan Yu", "Yanwei Fu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2605.21282", "pdf_url": "https://arxiv.org/pdf/2605.21282", "code_url": "", "clusters": ["C1"], "keywords": ["SMFP", "online RL"], "model_class": "diffusion/flow policy", "techniques": ["one-step generation", "MeanFlow"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MuJoCo"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Stochastic MeanFlow policies that support one-step action generation within off-policy RL.", "relevance": 3, "relevance_reason": "Single-step generative policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2609.26071", "title": "StrataVLA: Hierarchical and Efficient 3D Geometric Grounding for Vision-Language-Action Models", "authors": ["Jin Cui", "Zhaoyu Pu", "Botao Cai", "Jun Ye", "Xinyue Long", "Boran Zhao", "Pengju Ren", "Zhao Pu", "Bo Cai", "Xin-Yue Long", "Bo-Ran Zhao", "Peng-Ju Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.26071", "pdf_url": "https://arxiv.org/pdf/2609.26071", "code_url": "", "clusters": ["C1"], "keywords": ["3D grounding", "VLA", "efficient inference"], "model_class": "VLA", "techniques": ["feature caching", "sparse adapters", "routing", "geometry adapters", "task-aware routing", "LRU feature cache"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 88% fewer geometry-model invocations; 98.53% LIBERO", "metric_measures": "other", "benchmarks": ["LIBERO", "SimplerEnv", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Injects frozen geometry-foundation features into selected VLA layers via sparse adapters, with task-aware routing and an LRU feature cache to cut geometry-model calls.", "relevance": 3, "relevance_reason": "Caching and routing to make 3D grounding affordable at inference.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]; snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2608.26067", "title": "StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models", "authors": ["Zhe Liu", "Jinghua Hou", "Yuxiang Lu", "Zhenya Yang", "Xianzhe Fan", "Junwei Luo", "Junyi Li", "Ruihua Han", "Zhi Hou", "Hengshuang Zhao", "Zhen-Ya Yang", "Junsong Luo", "Junyong Li", "Rui-Hua Han", "Heng-Shuang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.26067", "pdf_url": "https://arxiv.org/pdf/2608.26067", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["StreamPI", "asynchronous deployment", "temporal modeling", "pi0.5"], "model_class": "VLA", "techniques": ["streaming attention", "random-interval training", "streaming inference", "causal attention across frames"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Gives single-frame VLAs streaming multi-frame context with no extra parameters and trains with random frame intervals so the policy tolerates asynchronous, timing-perturbed deployment.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot polic*\") AND (offload* OR \"edge-cloud\" OR serving OR \"inference runtime\"); snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.10504", "title": "SUREFlow: State-space Uncertainty-aware REsidual Flow Matching for Robust Robot Manipulation", "authors": ["Md Tanvir Islam", "Sai Navaneet Peddapalli", "Sangmoon Lee", "Sangtae Ahn"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.10504", "pdf_url": "https://arxiv.org/pdf/2607.10504", "code_url": "", "clusters": ["C1"], "keywords": ["uncertainty", "small model"], "model_class": "diffusion/flow policy", "techniques": ["Mamba backbone", "residual flow matching"], "base_model": "", "params": "179M", "hardware": "", "reported_metric": "179M params comparable to 3-7B VLAs on LIBERO-PRO", "metric_measures": "other", "benchmarks": ["LIBERO", "LIBERO-PRO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Mamba-based residual flow-matching policy with uncertainty-aware refinement that rivals multi-billion VLAs with 179M parameters.", "relevance": 3, "relevance_reason": "Small efficient policy", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2609.03602", "title": "SV-WAM: An Efficient Surround-View World-Action Model for End-to-End Autonomous Driving", "authors": ["Jinyang Wang", "Shiwei Li", "Junjian Wang", "Zhiqiang Deng", "Jianbin Gao", "Yihang Zhao", "Liu Liu", "Yongjia Zhao", "Jinlong Chen", "Huirui Xu", "Yifeng Pan", "Kangwei Liu", "Fan Ren", "Ji Tao", "Minghao Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.03602", "pdf_url": "https://arxiv.org/pdf/2609.03602", "code_url": "", "clusters": ["C8"], "keywords": ["surround-view WAM", "action-only planning", "driving"], "model_class": "world-action model", "techniques": ["future video as training-only supervision", "action-centered causal mask", "drivable-area regularizer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIMv2", "nuScenes"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Keeps six-camera input in a driving WAM by using future-video prediction only as training supervision and discarding the video branch at deployment.", "relevance": 3, "relevance_reason": "Same efficiency pattern for driving WAMs.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.29828", "title": "SymVD: Symmetric Vision Language Action Distillation for Robot Manipulation", "authors": ["Hyewon Choi", "Donggyu Kim", "Soojean Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29828", "pdf_url": "https://arxiv.org/pdf/2608.29828", "code_url": "", "clusters": ["C2"], "keywords": ["equivariance", "teacher VLA", "compact student", "RL"], "model_class": "VLA", "techniques": ["policy distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills a large VLA teacher into a compact equivariant actor-critic student using symmetry-aware objectives and adaptive RL/distillation weighting.", "relevance": 3, "relevance_reason": "Teacher-student VLA compression.", "foundational_technique": false, "verified": true, "found_via": "search: distillation vision-language-action"}
{"id": "table30-v2-cvprw2026", "title": "Table30 V2: Evaluating Generalized Models by Real Robots at Scale", "authors": ["Yunchao Ma", "Ze Chen", "Erjin Zhou", "Ziming Liu", "Haowei Zhang", "Kai Liu", "Haoqiang Fan"], "year": 2026, "venue": "CVPR 2026 Workshops (pp. 4461-4467)", "type": "benchmark", "url": "https://openaccess.thecvf.com/content/CVPR2026W/GigaBrainChallenge/html/Ma_Table30_V2_Evaluating_Generalized_Models_by_Real_Robots_at_Scale_CVPRW_2026_paper.html", "pdf_url": "https://openaccess.thecvf.com/content/CVPR2026W/GigaBrainChallenge/papers/Ma_Table30_V2_Evaluating_Generalized_Models_by_Real_Robots_at_Scale_CVPRW_2026_paper.pdf", "code_url": "", "clusters": ["C4"], "keywords": ["Table30 V2", "RoboChallenge", "generalization", "real-robot evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboChallenge Table30 V2", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Upgrade of RoboChallenge's Table30 real-robot benchmark focused on testing generalization of embodied models at scale.", "relevance": 3, "relevance_reason": "verified 2026-09-30 by orchestrator: CVF open-access page and abstract fetched. Real-robot benchmark infrastructure (RoboChallenge).", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RoboChallenge Table30"}
{"id": "2608.25798", "title": "TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback", "authors": ["Jianbo Zhou", "Boyuan Zhao", "Yuzheng Zhang", "Yiyang Chen", "Wenxin Chen", "Qiuyue Li", "Xiangyang Gu", "Yuhan Cao", "Xiao Xia", "Yi Yang", "Yanzhe Hu", "Zhijie Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.25798", "pdf_url": "https://arxiv.org/pdf/2608.25798", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["TacForcing", "streaming action generation", "execution-time tactile feedback"], "model_class": "VLA", "techniques": ["streaming action expert", "execution-aware tactile attention"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Streaming action expert that generates blocks sequentially and resumes from saved intermediate states with fresh tactile feedback after each block, avoiding stale tactile conditioning.", "relevance": 3, "relevance_reason": "Reactive contact handling within chunks.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: real-time chunking"}
{"id": "2609.18497", "title": "TAO-Force: Unifying Force-Aware Perception and Fast-Slow Control for Contact-Rich Manipulation", "authors": ["Bo-Han Gan", "Xuan Wen", "Yong-Shen Zhao", "Bao-Ping Cheng", "Wen-He Jia", "Ye Wang", "Gong-Xin Yao", "Han Gao", "Jing-Yao Tang", "Lei Zhao", "Jin-Ze Ge"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18497", "pdf_url": "https://arxiv.org/pdf/2609.18497", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["force-aware", "admittance control"], "model_class": "VLA", "techniques": ["fast-slow control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Injects force feedback into a frozen VLM backbone and uses a contact-gated fast admittance branch alongside a slow position branch.", "relevance": 3, "relevance_reason": "Fast-slow design for responsive control", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2406.09246"}
{"id": "2606.09499", "title": "Targeting World Models to Compromise Robot Learning Pipelines", "authors": ["Ethan Rathbun", "Ahmed Agha", "Saaduddin Mahmud", "Christopher Amato", "Alina Oprea", "Eugene Bagdasarian"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09499", "pdf_url": "https://arxiv.org/pdf/2606.09499", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["data poisoning", "world model supply chain", "unsafe policies"], "model_class": "world model", "techniques": ["world-model-targeted poisoning attack"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows poisoning a world model used for data generation or simulation can stealthily yield unsafe deployed robot policies even when the real training data is clean.", "relevance": 3, "relevance_reason": "Safety risk of world models in robot pipelines.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.27780", "title": "Task-Prototype Guided Flow Matching for Few-Shot Generalization in Vision-Language Robot Manipulation", "authors": ["Yi-Zhao Wang", "Guan-Tao Zhang", "Jing-Bo Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.27780", "pdf_url": "https://arxiv.org/pdf/2609.27780", "code_url": "", "clusters": ["C3"], "keywords": ["few-shot adaptation"], "model_class": "diffusion/flow policy", "techniques": ["prototype-guided flow matching"], "base_model": "", "params": "", "hardware": "LeRobot SO-101 platform", "reported_metric": "54.3 ms latency, 3.9 GB peak memory, 10 Hz control", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "TP-Flow turns a few demonstrations into prototype tokens guiding a flow policy for few-shot manipulation while keeping real-time execution.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.18108", "title": "Technical Report: One-Step Drifting Action Heads for GR00T N1.7", "authors": ["Xihe Shao"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2609.18108", "pdf_url": "https://arxiv.org/pdf/2609.18108", "code_url": "https://github.com/RealManShao/lerobot/tree/feat/drif-ov", "clusters": ["C1"], "keywords": ["GR00T", "action head"], "model_class": "VLA", "techniques": ["one-step generation", "drifting models", "asynchronous chunking"], "base_model": "GR00T N1.7", "params": "", "hardware": "NVIDIA A800 (training)", "reported_metric": "action-head forward ~45.3 ms -> ~5.0 ms; backbone+head ~70.0 -> ~30.6 ms", "metric_measures": "relative latency", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces GR00T N1.7 diffusion action head with a one-step drifting head plus overlap-conditioned asynchronous chunk replacement; head time drops from about 45 ms to 5 ms but LIBERO success falls, reported as a trade-off.", "relevance": 3, "relevance_reason": "Honest negative-leaning evidence on one-step heads and long open-loop chunks.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"open-loop\" AND \"vision-language-action\" AND (execution OR monitor*)"}
{"id": "2608.07314", "title": "TEMPO: Semantic-Action Decoupled RL Post-Training for Vision-Language-Action Models", "authors": ["Ziheng Liu", "Quantao Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.07314", "pdf_url": "https://arxiv.org/pdf/2608.07314", "code_url": "", "clusters": ["C3"], "keywords": ["two-timescale RL", "frozen backbone", "action expert"], "model_class": "VLA", "techniques": ["partial-parameter RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "RL post-training that freezes the VLM backbone and updates only projection and action expert at different rates.", "relevance": 3, "relevance_reason": "Restricts RL to small components.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2603.07647", "title": "TempoFit: Plug-and-Play Layer-Wise Temporal KV Memory for Long-Horizon Vision-Language-Action Manipulation", "authors": ["Jun Sun", "Bo Yang", "Jia-Hao Zhang", "Ning Ma", "Chenchen Wu", "Si-Qi Zhang", "Yi-Ou Huang", "Qiufeng Wang", "Shangsong Liang", "Yaran Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.07647", "pdf_url": "https://arxiv.org/pdf/2603.07647", "code_url": "", "clusters": ["C1"], "keywords": ["temporal memory", "training-free"], "model_class": "VLA", "techniques": ["KV cache reuse"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-LONG", "CALVIN"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Training-free retrofit that reuses layer-wise prefix K/V across timesteps as temporal memory for frozen VLAs.", "relevance": 3, "relevance_reason": "KV-cache memory at near-real-time latency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2604.20472", "title": "Temporal Difference Calibration in Sequential Tasks: Application to Vision-Language-Action Models", "authors": ["Shelly Francis-Meretzki", "Mirco Mutti", "Yaniv Romano", "Aviv Tamar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.20472", "pdf_url": "https://arxiv.org/pdf/2604.20472", "code_url": "", "clusters": ["C7"], "keywords": ["calibration", "uncertainty"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Sequential calibration of VLA success confidence using TD value estimation.", "relevance": 3, "relevance_reason": "Uncertainty/runtime monitoring.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2603.12231", "title": "Temporal Straightening for Latent Planning", "authors": ["Ying Wang", "Oumayma Bounou", "Gaoyue Zhou", "Randall Balestriero", "Tim G. J. Rudner", "Yann LeCun", "Mengye Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.12231", "pdf_url": "https://arxiv.org/pdf/2603.12231", "code_url": "https://agenticlearning.ai/temporal-straightening", "clusters": ["C8"], "keywords": ["temporal straightening", "curvature regularizer", "gradient-based planning"], "model_class": "world model", "techniques": ["curvature regularization of latent trajectories"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["goal-reaching suite"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Regularizes JEPA latent trajectories to be locally straight so Euclidean distance tracks geodesic distance, making gradient-based latent planning more stable and successful.", "relevance": 3, "relevance_reason": "Enables cheaper gradient planning in latent world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.06491", "title": "TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies", "authors": ["Dong Jing", "Jingchen Nie", "Tianqi Zhang", "Jiaqi Liu", "Huaxiu Yao", "Zhiwu Lu", "Mingyu Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.06491", "pdf_url": "https://arxiv.org/pdf/2606.06491", "code_url": "", "clusters": ["C1"], "keywords": ["execution speed control", "trajectory retiming"], "model_class": "VLA", "techniques": ["speed conditioning", "data augmentation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Conditions a VLA on target execution speed, trained via demonstration re-timing, allowing faster motion in transit and slower in contact phases.", "relevance": 3, "relevance_reason": "Execution-speed efficiency rather than compute.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "2601.15912", "title": "TeNet: Text-to-Network for Compact Policy Synthesis", "authors": ["Ariyan Bighashdel", "Kevin Sebastian Luck"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.15912", "pdf_url": "https://arxiv.org/pdf/2601.15912", "code_url": "", "clusters": ["C1"], "keywords": ["hypernetwork", "text-to-policy", "compact policy"], "model_class": "other", "techniques": ["hypernetwork"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MuJoCo", "Meta-World"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "An LLM-conditioned hypernetwork generates a compact state-based policy once per instruction, allowing high-frequency control without running the LLM.", "relevance": 3, "relevance_reason": "Hypernetwork compact policy.", "foundational_technique": false, "verified": true, "found_via": "search query: hypernetwork robot policy"}
{"id": "2609.36182", "title": "Test-Time Adaptation of Manipulation Policies Under Actuator Degradation", "authors": ["Som Sagar", "Ransalu Senanayake"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36182", "pdf_url": "https://arxiv.org/pdf/2609.36182", "code_url": "", "clusters": ["C3"], "keywords": ["telemetry", "actuator degradation", "action rectification"], "model_class": "VLA", "techniques": ["test-time adaptation", "lightweight transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Lightweight telemetry-conditioned action rectifier that adapts a frozen policy to actuator heating/voltage sag.", "relevance": 3, "relevance_reason": "Deployment-time adaptation with frozen policy.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.26865", "title": "Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral for Edge LLM Agents", "authors": ["Amirmohammad Farzaneh", "Osvaldo Simeone"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.26865", "pdf_url": "https://arxiv.org/pdf/2607.26865", "code_url": "", "clusters": ["C6"], "keywords": ["edge-cloud", "ReAct"], "model_class": "other", "techniques": ["Learn-Then-Test calibration", "cloud deferral", "early halting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "43-65% less per-episode thinking compute", "metric_measures": "other", "benchmarks": ["GSM8K", "HotpotQA", "MBPP", "household robot"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Edge LLM agent framework that halts on-device reasoning once the action stabilizes and defers uncertain actions to the cloud, both calibrated with Learn-Then-Test for finite-sample guarantees.", "relevance": 3, "relevance_reason": "Calibrated edge-cloud deferral with guarantees.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot policy\" OR embodied) AND (Jetson OR \"on-device\" OR onboard)"}
{"id": "2603.11101", "title": "Thousand-GPU Large-Scale Training and Optimization Recipe for AI-Native Cloud Embodied Intelligence Infrastructure", "authors": ["Yongjian Guo", "Yunxuan Ma", "Haoran Sun", "Zhong Guan", "Shuai Di", "Jing Long", "Wanting Xu", "Xiaodong Bai", "Wen Huang", "Yucheng Guo", "Chen Zhou", "Qiming Yang", "Mingxi Luo", "Tianyun Zhao", "Hedan Yang", "Song Wang", "Xiaomeng Tian", "Xiaolong Xiang", "Zhen Sun", "Yu Wei", "Luqiao Wang", "Yuzhen Li", "Chenfeng Gu", "Junwu Xiong", "Yicheng Gong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.11101", "pdf_url": "https://arxiv.org/pdf/2603.11101", "code_url": "", "clusters": ["C3"], "keywords": ["thousand-GPU training", "GR00T-N1.5", "LeRobot"], "model_class": "VLA", "techniques": ["distributed training", "data pipeline optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "single-round training 15 h -> 22 min (40x)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Cloud thousand-GPU training platform on LeRobot cutting GR00T-N1.5 training round from 15 h to 22 min.", "relevance": 3, "relevance_reason": "Training-cost infrastructure.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:LeRobot"}
{"id": "2606.00966", "title": "Threading Optimization for Vision-Language-Action Model Inference in Low-Cost Smart Agricultural Manipulation", "authors": ["Keith Truongcao", "Christopher Nhu", "Zi-Jian An", "Phong X. Nguyen", "Si-Wei Cai", "Lifeng Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.00966", "pdf_url": "https://arxiv.org/pdf/2606.00966", "code_url": "", "clusters": ["C5", "C6", "C9"], "keywords": ["threading", "low-cost arm", "real-time chunking implementation"], "model_class": "VLA", "techniques": ["multithreaded RTC implementation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "System-level threaded implementation of real-time action chunking for a VLA on a low-cost agricultural manipulator.", "relevance": 3, "relevance_reason": "Practical deployment of RTC on cheap hardware.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2609.33157", "title": "TimelyDAgger: Timing-Aware Expert Querying for VLA Policy Improvement", "authors": ["Zhi-Xuan Zhao", "Pei-Yan Li", "Enhao Zhang", "Yue-Ran Tao", "Hao Wang", "Cheng-Hao Yue", "Lei Lv", "Wen-Tao Zhao", "Jia-Hao Chen", "Xin Liu", "Kang-Yao Huang", "Yu Luo", "Hua-Ping Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33157", "pdf_url": "https://arxiv.org/pdf/2609.33157", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "DAgger", "takeover timing"], "model_class": "VLA", "techniques": ["feature monitoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Robot-gated DAgger that monitors internal VLA features to time expert takeovers, with metrics linking detection, timing and improvement.", "relevance": 3, "relevance_reason": "Runtime failure monitoring for VLAs", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2406.09246, 2503.08558"}
{"id": "2605.12236", "title": "TMRL: Diffusion Timestep-Modulated Pretraining Enables Exploration for Efficient Policy Finetuning", "authors": ["Matthew Hong", "Jesse Zhang", "Anusha Nagabandi", "Abhishek Gupta"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2605.12236", "pdf_url": "https://arxiv.org/pdf/2605.12236", "code_url": "", "clusters": ["C3"], "keywords": ["TMRL"], "model_class": "diffusion/flow policy", "techniques": ["RL fine-tuning", "diffusion timestep modulation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Pretrains diffusion policies with timestep-modulated exploration so subsequent RL fine-tuning is more sample efficient.", "relevance": 3, "relevance_reason": "RL fine-tuning efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2608.10860"}
{"id": "2605.06311", "title": "Toward Visually Realistic Simulation: A Benchmark for Evaluating Robot Manipulation in Simulation", "authors": ["Yixin Zhu", "Zixiong Wang", "Jian Yang", "Jin Xie", "Jingyi Yu", "Jiayuan Gu", "Beibei Wang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2605.06311", "pdf_url": "https://arxiv.org/pdf/2605.06311", "code_url": "", "clusters": ["C4"], "keywords": ["VISER", "visual realism", "PBR", "sim-real correlation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["VISER"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Isolates lighting and material effects on sim evaluation and builds a PBR-asset benchmark (1000+ assets) whose VLA scores correlate with real-world performance.", "relevance": 3, "relevance_reason": "Improves sim-as-proxy validity for VLA evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2602.19764", "title": "Towards Dexterous Embodied Manipulation via Deep Multi-Sensory Fusion and Sparse Expert Scaling", "authors": ["Yi-Rui Sun", "Guang-Yu Zhuge", "Ke-Liang Liu", "Jie Gu", "Zhi-Hao Xia", "Qionglin Ren", "Chun-Xu Tian", "Zhongxue Ga"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.19764", "pdf_url": "https://arxiv.org/pdf/2602.19764", "code_url": "", "clusters": ["C1"], "keywords": ["multi-sensory fusion", "force", "DiT"], "model_class": "diffusion/flow policy", "techniques": ["sparse MoE"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Diffusion transformer with sparse mixture-of-experts fusing RGB, depth and force while keeping latency low for real-time control.", "relevance": 3, "relevance_reason": "MoE scaling with low inference latency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2602.10503", "title": "Towards Long-Lived Robots: Continual Learning VLA Models via Reinforcement Fine-Tuning", "authors": ["Yuan Liu", "Haoran Li", "Shuai Tian", "Yuxing Qin", "Yuhui Chen", "Yupeng Zheng", "Yongzhen Huang", "Dongbin Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.10503", "pdf_url": "https://arxiv.org/pdf/2602.10503", "code_url": "", "clusters": ["C3"], "keywords": ["LifeLong-RFT", "continual learning", "process reward"], "model_class": "VLA", "techniques": ["reinforcement fine-tuning", "continual learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Chunk-level on-policy RFT with multi-dimensional process rewards for continual VLA learning without env feedback.", "relevance": 3, "relevance_reason": "Continual adaptation with less data.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2603.20607", "title": "Towards Practical World Model-based Reinforcement Learning for Vision-Language-Action Models", "authors": ["Zhilong Zhang", "Haoxiang Ren", "Yihao Sun", "Yifei Sheng", "Haonan Wang", "Haoxin Lin", "Zhichao Wu", "Pierre-Luc Bacon", "Yang Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.20607", "pdf_url": "https://arxiv.org/pdf/2603.20607", "code_url": "", "clusters": ["C3"], "keywords": ["VLA-MBPO", "model-based RL", "unified multimodal world model"], "model_class": "VLA", "techniques": ["model-based RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Practical model-based policy optimization for VLAs using a unified multimodal model as data-efficient multi-view world model.", "relevance": 3, "relevance_reason": "Avoids real-world RL cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.26121", "title": "Towards Trustworthy Embodied Intelligence: A Systems Framework and Graded Trustworthiness Levels", "authors": ["Xinyu Yang", "Tian-Xing Chen", "Honghao Su", "Minxuan Wang", "Chen Yu", "Zhang-Zheng Tu", "Yue Chen", "Yuxiao Huo", "Lingfeng Zhang", "Yan Huang", "Yan-Jiao Qin", "Shao-Long Zhu", "Qi-Wei Liang", "Hekun Tian", "Shu-Jia Liu", "Guang-Yu Chen", "Junhao Gong", "Zi-Xuan Li", "Wen-Wei Lin", "Zijian Lin", "Wen Zhu", "Erica Chen", "Yue Yuan", "Qi-Ze Yu", "Jiaqi Liang", "Hao-Wen Yan", "Hengfei Zhao", "Weijie Wan", "Zi-Qi Xiao", "Junyuan Tang", "Bai-Jun Chen", "Kaixun Lei", "Kai-Xuan Wang", "Kai-Lun Su", "Zan-Xin Chen", "Yao Mu", "Ren-Jing Xu", "Chuqiao Lyu", "Qiankai Xiong", "Ping Luo", "Wen-Bo Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.26121", "pdf_url": "https://arxiv.org/pdf/2607.26121", "code_url": "", "clusters": ["C7"], "keywords": ["trustworthiness", "runtime monitoring", "framework"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Systems framework defining layered mechanisms and graded levels for trustworthy embodied intelligence deployment.", "relevance": 3, "relevance_reason": "Deployment safety framework incl. runtime monitoring.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2306.03310"}
{"id": "2609.33125", "title": "Train Together or Merge Later? Unifying VLA Experts via a Shared Action Interface", "authors": ["Zhizhen Zhang", "Yuxia Fu", "Zijian Wang", "Helen Huang", "Yadan Luo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33125", "pdf_url": "https://arxiv.org/pdf/2609.33125", "code_url": "", "clusters": ["C3"], "keywords": ["PolicyWeave", "model merging", "shared action interface"], "model_class": "VLA", "techniques": ["LoRA", "sparse merging"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Keeps a shared action interface and restricts task adaptation to LoRA so independently trained VLA experts merge well, avoiding joint post-training.", "relevance": 3, "relevance_reason": "Merge-friendly PEFT avoids retraining.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:LoRA"}
{"id": "2603.22991", "title": "Training-Free Interaction-Aligned Visual Token Pruning for Efficient Embodied Manipulation", "authors": ["Jintao Cheng", "Weibin Li", "Haozhe Wang", "Gang Wang", "Yipu Zhang", "Xiaoyu Tang", "Jin Wu", "Xieyuanli Chen", "Yunhui Liu", "Wei Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.22991", "pdf_url": "https://arxiv.org/pdf/2603.22991", "code_url": "", "clusters": ["C1"], "keywords": ["token pruning", "dynamic budget", "semantic-motion agreement"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.54x speedup (LIBERO), 1.48x on real robot", "metric_measures": "relative latency", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Sets a per-frame token budget from agreement between instruction semantics and image motion, then selects tokens with geometric boundary correction.", "relevance": 3, "relevance_reason": "VLA token pruning.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning"}
{"id": "2608.24111", "title": "Trajectory-Level Continuous Action Representation for Robotic Manipulation", "authors": ["Tong Yang", "Jing-Kai Jia", "Yue-Cheng Xu", "Xue-Yao Chen", "Chi Zhang", "Wen-Qiang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.24111", "pdf_url": "https://arxiv.org/pdf/2608.24111", "code_url": "", "clusters": ["C1"], "keywords": ["action representation", "control frequency"], "model_class": "other", "techniques": ["continuous action tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "MimicGen", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Encodes a fixed real-time interval of trajectory into continuous latents so representation size does not grow with control rate.", "relevance": 3, "relevance_reason": "Action tokenization independent of control frequency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2606.19998", "title": "Tri-Info: Generalizable, Interpretable Failure Prediction for VLA Models via Information Theory", "authors": ["Jinghan Yang", "Yunchao Zhang", "Wang Yuan", "Haolun Wan", "Jiaming Zhang", "Zhengyang Hu", "Yanchao Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.19998", "pdf_url": "https://arxiv.org/pdf/2606.19998", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["Tri-Info", "information-theoretic failure signals"], "model_class": "VLA", "techniques": ["information-theoretic signals"], "base_model": "", "params": "", "hardware": "", "reported_metric": "83% accuracy on real-world tasks", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Derives three information-theoretic signals (action diversity, temporal consistency, state coupling) that predict VLA failures and transfer across architectures and sim-to-real without retraining.", "relevance": 3, "relevance_reason": "Training-free, transferable failure monitor.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2609.07051", "title": "TrojanWorld: Backdooring World-Model Agents via Imagination Steering", "authors": ["Wenkai Huang", "Siyuan Liang", "Gaolei Li", "Yiming Li", "Tianhao Peng", "Jianhua Li", "Dacheng Tao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.07051", "pdf_url": "https://arxiv.org/pdf/2609.07051", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["backdoor", "world-model agents", "physical trigger"], "model_class": "world model", "techniques": ["imagination-steering backdoor attack"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows pretrained world models can be backdoored so a physical object trigger steers the agent's imagination toward attacker-chosen behavior.", "relevance": 3, "relevance_reason": "Threat to imagination-based decision/safety pipelines.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.29892", "title": "Trust Your Instincts: Confidence-Driven Test-Time RL for Vision-Language-Action Models", "authors": ["Siyao Chen", "Jiakang Yuan", "Jiaxin Wang", "Tao Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.29892", "pdf_url": "https://arxiv.org/pdf/2606.29892", "code_url": "", "clusters": ["C3"], "keywords": ["test-time RL", "confidence", "self-bootstrapping"], "model_class": "VLA", "techniques": ["test-time RL", "intrinsic reward"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Test-time RL using generation confidence/similarity to expert demos as intrinsic reward, no external success signals.", "relevance": 3, "relevance_reason": "Deployment-time self-improvement.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2603.19296", "title": "TTQ: Activation-Aware Test-Time Quantization to Accelerate LLM Inference On The Fly", "authors": ["T. Koike-Akino", "Jing Liu", "Ye Wang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2603.19296", "pdf_url": "https://arxiv.org/pdf/2603.19296", "code_url": "", "clusters": ["C2"], "keywords": ["TTQ", "activation-aware"], "model_class": "other", "techniques": ["test-time quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Quantizes LLMs on the fly at inference with online activation-aware calibration.", "relevance": 3, "relevance_reason": "Generic quantization technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323, 2305.14314"}
{"id": "2604.23609", "title": "Tube Diffusion Policy: Reactive Visual-Tactile Policy Learning for Contact-rich Manipulation", "authors": ["Teng Xue", "Alberto Rigo", "Bingjian Huang", "Jiayi Shen", "Zheng-Tong Xu", "Nicholas Colonnese", "Amirhossein H. Memar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.23609", "pdf_url": "https://arxiv.org/pdf/2604.23609", "code_url": "", "clusters": ["C7"], "keywords": ["reactive control", "action chunking"], "model_class": "diffusion/flow policy", "techniques": ["tube-based feedback", "diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Push-T", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Tube Diffusion Policy learns an observation-conditioned feedback flow around nominal action chunks so the policy can react step-wise to tactile and visual changes without waiting for new chunks.", "relevance": 3, "relevance_reason": "relevant: reactivity within action chunks", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2602.09023", "title": "TwinRL: Digital Twin-Driven Reinforcement Learning for Real-World Robotic Manipulation", "authors": ["Qinwen Xu", "Jiaming Liu", "Rui Zhou", "Shaojun Shi", "Nuowei Han", "Zhuoyang Liu", "Chenyang Gu", "Shuo Gu", "Yang Yue", "Gao Huang", "Wenzhao Zheng", "Sirui Han", "Peng Jia", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.09023", "pdf_url": "https://arxiv.org/pdf/2602.09023", "code_url": "", "clusters": ["C3"], "keywords": ["digital twin", "real-world RL", "exploration"], "model_class": "VLA", "techniques": ["sim-guided RL", "RL fine-tuning", "human-in-the-loop"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Builds a smartphone-scanned digital twin to warm up and guide exploration before real-world RL fine-tuning of a VLA, improving real-interaction efficiency.", "relevance": 3, "relevance_reason": "Reduces costly real-world interaction in VLA post-training.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv"}
{"id": "2606.18043", "title": "Uncertainty Quantification for Flow-Based Vision-Language-Action Models", "authors": ["Ralf Römer", "Maximilian Seeliger", "Saida Liu", "Ben Sturgis", "Marco Bagatella", "Daniel Marta", "Andreas Krause", "Angela P. Schoellig"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18043", "pdf_url": "https://arxiv.org/pdf/2606.18043", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["velocity-field disagreement", "epistemic uncertainty", "flow VLA", "SAVE"], "model_class": "VLA", "techniques": ["small ensemble velocity disagreement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Estimates epistemic uncertainty in flow-matching VLAs from velocity-field disagreement across a small ensemble, used for failure detection and uncertainty-guided active fine-tuning.", "relevance": 3, "relevance_reason": "Uncertainty signal for runtime monitoring.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2604.26836", "title": "Uncertainty-Aware Predictive Safety Filters for Probabilistic Neural Network Dynamics", "authors": ["Bernd Frauenknecht", "Lukas Kesper", "Daniel Mayfrank", "Henrik Hose", "Sebastian Trimpe"], "year": 2026, "venue": "RLC 2026", "type": "paper", "url": "https://arxiv.org/abs/2604.26836", "pdf_url": "https://arxiv.org/pdf/2604.26836", "code_url": "", "clusters": ["C7"], "keywords": ["safe MBRL"], "model_class": "other", "techniques": ["predictive safety filter", "probabilistic ensembles", "reachable sets"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Predictive safety filter using probabilistic-ensemble NN dynamics with reachable-set formulation and a certainty constraint for rigorous safety during model-based RL.", "relevance": 3, "relevance_reason": "Learned-dynamics predictive safety filtering.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"predictive safety filter\""}
{"id": "2608.05597", "title": "Uncertainty-Aware World Model for Aerial Image-Goal Navigation", "authors": ["Deyi Zhu", "Haoyu Fan", "Yinan Zhu", "Weichen Zhang", "Shilin Ma", "Xinlei Chen", "Yansong Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.05597", "pdf_url": "https://arxiv.org/pdf/2608.05597", "code_url": "", "clusters": ["C8"], "keywords": ["aerial navigation", "uncertainty subspace", "efficient latent world model"], "model_class": "world model", "techniques": ["trajectory scoring as conditional OOD detection", "uncertainty subspace"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Efficient latent navigation world model for UAVs that represents a subspace of plausible futures and scores trajectories via OOD detection instead of a few point predictions.", "relevance": 3, "relevance_reason": "Uncertainty-aware efficient latent planning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2505.00779 (citations)"}
{"id": "2609.28838", "title": "Uncertainty-Gated Exploration Noise Suppresses Task Collapse in Online RL Fine-Tuning of a Flow-Matching Vision-Language-Action Policy", "authors": ["Mehmet Turan Yardımcı", "Yunus Emre Çoğurcu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.28838", "pdf_url": "https://arxiv.org/pdf/2609.28838", "code_url": "", "clusters": ["C3"], "keywords": ["online RL", "task collapse", "SmolVLA", "PPO", "bfloat16"], "model_class": "VLA", "techniques": ["online RL", "exploration noise control"], "base_model": "SmolVLA", "params": "450M", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-10"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Studies per-task collapse during small-compute online RL of a flow VLA and shows an uncertainty-gated exploration controller avoids it; notes bf16-without-master-copy effects.", "relevance": 3, "relevance_reason": "Low-budget RL fine-tuning pitfalls, including precision choices.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025"}
{"id": "2608.16153", "title": "Unified Condition-Action Modeling for Accurate One-Step Action Generation", "authors": ["Xinyu Zhou", "Zikun Cai", "Kuangji Zuo", "Gen Li", "Boyu Ma", "Yanshuo Lu", "Yutong Song", "Mingqi Yuan", "Jiayu Chen", "Jianfei Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.16153", "pdf_url": "https://arxiv.org/pdf/2608.16153", "code_url": "", "clusters": ["C1"], "keywords": ["UCA-Flow"], "model_class": "diffusion/flow policy", "techniques": ["one-step flow", "unified condition-action tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "45.6x and 33.4x faster than DP3 and Simple DP3", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "One-step flow policy placing conditions and actions in one token sequence processed by a unified transformer with dual-pass supervision.", "relevance": 3, "relevance_reason": "Fast accurate one-step action generation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=100]"}
{"id": "2608.17432", "title": "UniReflex: Plug-and-Play Force Control for Pretrained Generative Policies via Fast-Slow Reflex", "authors": ["Yan Huang", "Shoujie Li", "Ziwu Song", "Wenbo Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.17432", "pdf_url": "https://arxiv.org/pdf/2608.17432", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["force control", "impedance", "plug-and-play"], "model_class": "diffusion/flow policy", "techniques": ["fast-slow reflex"], "base_model": "", "params": "", "hardware": "", "reported_metric": "25-66x lower per-step backward latency vs joint training", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adds a fast reflex network driving variable impedance control to frozen generative policies, without backbone retraining.", "relevance": 3, "relevance_reason": "Fast-slow hierarchy; reported latency reduction vs joint training.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.05133", "title": "UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation", "authors": ["Mengmeng Liu", "Dian-Kun Zhang", "Jiu-Ming Liu", "Jian-Feng Cui", "Hong-Bin Xie", "Guang Chen", "Hang-Jun Ye", "Francesco Nex", "Hao Cheng", "Michael Ying Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.05133", "pdf_url": "https://arxiv.org/pdf/2607.05133", "code_url": "", "clusters": ["C8"], "keywords": ["autonomous driving", "video-action model"], "model_class": "world-action model", "techniques": ["modality masking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "4.3x speedup over joint video-action rollout", "metric_measures": "relative latency", "benchmarks": ["NAVSIM", "nuScenes", "Bench2Drive"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Single-DiT video-action model for driving that can drop future-video denoising at test time for faster trajectory-only inference.", "relevance": 3, "relevance_reason": "WAM inference speedup", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2503.00200, 2601.16163"}
{"id": "2604.18000", "title": "Unmasking the Illusion of Embodied Reasoning in Vision-Language-Action Models", "authors": ["Haiweng Xu", "Sipeng Zheng", "Haoming Luo", "Wan-Peng Zhang", "Ziheng Xi", "Zongqing Lu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2604.18000", "pdf_url": "https://arxiv.org/pdf/2604.18000", "code_url": "", "clusters": ["C4"], "keywords": ["BeTTER", "embodied reasoning"], "model_class": "VLA", "techniques": ["diagnostic benchmark"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diagnostic benchmark using causal interventions to separate reasoning failures from execution limits in VLAs.", "relevance": 3, "relevance_reason": "Evaluation methodology for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2603.10282", "title": "Update-Free On-Policy Steering via Verifiers", "authors": ["Maria Attarian", "Ian Vyse", "Jasper Gerigk", "Evgenii Opryshko", "Yifan Ruan", "Anas Almasri", "Sumeet Singh", "Yilun Du", "Igor Gilitschenski", "Claas Voelcker"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.10282", "pdf_url": "https://arxiv.org/pdf/2603.10282", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["update-free adaptation"], "model_class": "diffusion/flow policy", "techniques": ["policy steering", "learned verifier"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+49% average success over base on 5 real tasks", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Trains verifiers on initial policy rollouts to predict action success and steers a black-box diffusion policy toward likely-successful samples without updating it.", "relevance": 3, "relevance_reason": "Verifier-gated action sampling at execution time.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:verif* AND (\"diffusion policy\" OR \"flow matching policy\")"}
{"id": "2604.18463", "title": "Using large language models for embodied planning introduces systematic safety risks", "authors": ["Tao Zhang", "Kaixian Qu", "Zhibin Li", "Jiajun Wu", "Marco Hutter", "Manling Li", "Fan Shi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.18463", "pdf_url": "https://arxiv.org/pdf/2604.18463", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["safety benchmark", "LLM planners", "DESPITE"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "best planner produces dangerous plans on 28.3% of tasks", "metric_measures": "other", "benchmarks": ["DESPITE"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DESPITE benchmark (12,279 tasks) shows planning ability scales with model size while safety awareness stays flat across 23 LLMs.", "relevance": 3, "relevance_reason": "Safety evaluation for embodied planners; relevant to safety trade-offs of smaller models.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "2606.25880", "title": "USS: Unifying Spatial-Semantic Prompting for End to End Embodied Visual Tracking", "authors": ["Yu-Han Xie", "Xin-Yu Zhou", "Kuangji Zuo", "Yan-Shuo Lu", "Fengrui Huang", "Bo-Yu Ma", "Jian-Fei Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.25880", "pdf_url": "https://arxiv.org/pdf/2606.25880", "code_url": "", "clusters": ["C1"], "keywords": ["embodied visual tracking", "prompting"], "model_class": "other", "techniques": ["compact tracker"], "base_model": "", "params": "", "hardware": "", "reported_metric": "57 FPS vs 4.8-10 FPS for MLLM trackers", "metric_measures": "throughput", "benchmarks": ["EVT-Bench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "End-to-end embodied tracker accepting spatial or language prompts, running far faster than MLLM trackers.", "relevance": 3, "relevance_reason": "Real-time embodied perception-action", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2603.14482", "title": "V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning", "authors": ["Lorenzo Mur-Labadia", "Matthew Muckley", "Amir Bar", "Mido Assran", "Koustuv Sinha", "Mike Rabbat", "Yann LeCun", "Nicolas Ballas", "Adrien Bardes"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2603.14482", "pdf_url": "https://arxiv.org/pdf/2603.14482", "code_url": "", "clusters": ["C8"], "keywords": ["V-JEPA 2.1", "dense features", "robot grasping"], "model_class": "world model", "techniques": ["dense predictive loss", "deep self-supervision"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Ego4D", "EPIC-KITCHENS", "real-robot grasping"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Successor to V-JEPA 2 with dense spatially grounded features that substantially improve real-robot grasping through latent planning.", "relevance": 3, "relevance_reason": "Seed follow-up (V-JEPA line).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.09985"}
{"id": "2603.16685", "title": "vAccSOL: Efficient and Transparent AI Vision Offloading for Mobile Robots", "authors": ["Adam Zahir", "Michele Gucciardom Falk Selker", "Anastasios Nanos", "Kostis Papazafeiropoulos", "Carlos J. Bernardos", "Nicolas Weber", "Roberto Gonzalez"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.16685", "pdf_url": "https://arxiv.org/pdf/2603.16685", "code_url": "", "clusters": ["C5", "C6"], "keywords": ["vision offloading", "quadruped", "NN compiler", "transparent dispatch"], "model_class": "other", "techniques": ["neural network compilation", "transparent local/edge dispatch"], "base_model": "", "params": "", "hardware": "commercial quadruped + edge infrastructure", "reported_metric": "up to 80% lower robot-side power and up to 24x higher vision pipeline frame rate", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Compiler plus lightweight runtime that transparently runs vision models on the robot or dispatches them to nearby edge servers, cutting robot power and raising frame rates.", "relevance": 3, "relevance_reason": "Transparent edge offloading for mobile robots.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2607.07196", "title": "Validate the Dream Before You Trust Its Verdict: Admissibility for World-Model Simulators", "authors": ["Christian Oefinger", "Finn Rasmus Schäfer", "Korbinian Moller", "Mattia Piccinini", "Johannes Betz"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.07196", "pdf_url": "https://arxiv.org/pdf/2607.07196", "code_url": "", "clusters": ["C4", "C7", "C8"], "keywords": ["admissibility", "VV&A", "world model as test oracle"], "model_class": "world model", "techniques": ["verification/validation/accreditation framework for WMs"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Argues world models used as safety or success oracles must first be accredited, and proposes admissibility checks beyond visual fidelity.", "relevance": 3, "relevance_reason": "Methodology for trusting world-model safety verdicts.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.02958", "title": "ValueFormer: A Causal Transformer Value Function with Stage-Aware Labels for Semi-Autonomous Vision-Language-Action Policies", "authors": ["Inkyu Sa", "Konstantin Stulov", "Rajat Bhageria"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.02958", "pdf_url": "https://arxiv.org/pdf/2608.02958", "code_url": "", "clusters": ["C7"], "keywords": ["ValueFormer", "critic"], "model_class": "other", "techniques": ["value function", "failure detection", "bf16"], "base_model": "", "params": "", "hardware": "single GPU", "reported_metric": "2 Hz critic; 3-5x lower serving cost", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Causal transformer critic with stage-aware labels that detects failing rollouts and weights training; serving is made cheap enough to run alongside the policy.", "relevance": 3, "relevance_reason": "Runtime monitoring with measured serving cost.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.09448", "title": "VANE: Reliable Test-Time Training for Vision-Language-Action Models via Future Visual Representation Prediction", "authors": ["Hongjin Ji", "Guoyang Xia", "Luoyang Sun", "Fangxiang Feng", "Lei Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09448", "pdf_url": "https://arxiv.org/pdf/2608.09448", "code_url": "", "clusters": ["C3"], "keywords": ["reliable TTT"], "model_class": "VLA", "techniques": ["test-time training", "prompt adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SimplerEnv"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Test-time training for VLAs that isolates candidate prompt updates and commits them only if future visual consequences support them, making adaptation reversible.", "relevance": 3, "relevance_reason": "Safer online adaptation of deployed VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]"}
{"id": "2603.17652", "title": "VectorWorld: Efficient Streaming World Model via Diffusion Flow on Vector Graphs", "authors": ["Chaokang Jiang", "Desen Zhou", "Jiuming Liu", "Kevin Li Sun"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.17652", "pdf_url": "https://arxiv.org/pdf/2603.17652", "code_url": "", "clusters": ["C8"], "keywords": ["vector-graph world model", "one-step MeanFlow", "closed-loop driving simulation"], "model_class": "world model", "techniques": ["one-step masked completion (MeanFlow)", "streaming tile outpainting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Waymo Open Motion", "nuPlan"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Streaming driving world model over lane/agent vector graphs with solver-free one-step generation to meet real-time closed-loop budgets.", "relevance": 3, "relevance_reason": "Compact non-pixel world model for real-time simulation.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2605.02037", "title": "VILAS: A VLA-Integrated Low-cost Architecture with Soft Grasping for Robotic Manipulation", "authors": ["Zi-Jian An", "Hadi Khezam", "Bill Cai", "Ran Yang", "Shijie Geng", "Yiming Feng", "Yue Zheng", "Lifeng Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.02037", "pdf_url": "https://arxiv.org/pdf/2605.02037", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["low-cost platform", "deployment"], "model_class": "VLA", "techniques": ["low-cost hardware platform"], "base_model": "pi0, pi0.5, GR00T N1.6", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VILAS is a low-cost modular arm platform with a soft gripper that supports teleoperation, data collection and deployment, used to fine-tune and compare pi0, pi0.5 and GR00T N1.6 on accessible hardware.", "relevance": 3, "relevance_reason": "relevant: practical VLA deployment on low-cost hardware", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2601.14921", "title": "Vision-Language Models on the Edge for Real-Time Robotic Perception", "authors": ["Sarat Ahmad", "Maryam Hafeez", "Syed Ali Raza Zaidi"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.14921", "pdf_url": "https://arxiv.org/pdf/2601.14921", "code_url": "", "clusters": ["C6"], "keywords": ["VLM", "6G", "ORAN/MEC", "humanoid", "WebRTC"], "model_class": "other", "techniques": ["edge offloading", "MEC deployment"], "base_model": "LLaMA-3.2-11B-Vision, Qwen2-VL-2B", "params": "", "hardware": "MEC edge node vs cloud; Unitree G1", "reported_metric": "edge 5% lower end-to-end latency than cloud for 11B VLM", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Streams humanoid sensor data over WebRTC to an ORAN/MEC edge node and compares edge vs cloud VLM deployment, finding edge keeps accuracy with modestly lower latency and small models halve latency at an accuracy cost.", "relevance": 3, "relevance_reason": "Edge vs cloud VLM for robot perception over mobile networks.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.30456", "title": "Vision-Language-Action Models: Experimental Insights from a Real-World UR5 Platform", "authors": ["Mathilde Hochedel", "Marc Lalonde"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2606.30456", "pdf_url": "https://arxiv.org/pdf/2606.30456", "code_url": "", "clusters": ["C4"], "keywords": ["UR5", "OpenVLA", "deployment gap", "reproducibility"], "model_class": "VLA", "techniques": [], "base_model": "OpenVLA, OpenVLA-OFT", "params": "", "hardware": "UR5e", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Practitioner study porting OpenVLA/OFT to a UR5e; closed-loop instability stems largely from action semantics, frame conventions, temporal alignment and preprocessing rather than the model.", "relevance": 3, "relevance_reason": "Documents hidden variables that confound real-robot evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2605.30011", "title": "VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies", "authors": ["Mingjian Gao", "Wenqiao Zhang", "Yuqian Yuan", "Yang Dai", "Binhe Yu", "Zheqi Lv", "Haoyu Zheng", "Jiaqi Zhu", "Zhiqi Ge", "Zixuan Wan", "Siliang Tang", "Yueting Zhuang", "Ming Gao", "Wen-Qiao Zhang", "Yu-Qian Yuan", "Bin-He Yu", "Zhe-Qi Lv", "Haoyuan Zheng", "Jia-Qi Zhu", "Si-Liang Tang", "Yue-Ting Zhuang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30011", "pdf_url": "https://arxiv.org/pdf/2605.30011", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["visual intermediate reasoning", "selective routing", "visual reasoning", "low-latency CoT"], "model_class": "VLA", "techniques": ["compact reasoning interface", "routing", "visual intermediate reasoning instead of text CoT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Replaces text CoT with a compact visual-evidence token interface learned via selective routing, lowering reasoning latency.", "relevance": 3, "relevance_reason": "Reasoning latency reduction.", "foundational_technique": false, "verified": true, "found_via": "search query: reasoning VLA latency; venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2602.21445", "title": "VLA Knows Its Limits: Adaptive Execution Horizons for Robot Policies", "authors": ["Haoxuan Wang", "Gengyu Zhang", "Yan Yan", "Ramana Rao Kompella", "Gaowen Liu"], "year": 2026, "venue": "ECCV 2026", "type": "paper", "url": "https://arxiv.org/abs/2602.21445", "pdf_url": "https://arxiv.org/pdf/2602.21445", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["AutoHorizon", "adaptive execution horizon", "attention analysis"], "model_class": "VLA", "techniques": ["attention-based horizon estimation", "test-time"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses action self-attention in flow VLAs as a proxy for predictive limit to set the execution horizon per chunk at test time with negligible overhead.", "relevance": 3, "relevance_reason": "Adaptive open-loop horizon for reactivity.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"action chunk\" execution horizon adaptive"}
{"id": "2605.01194", "title": "VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model", "authors": ["Wenhao Li", "Xiu Su", "Yichao Cao", "Hongyan Xu", "Xiaobo Xia", "Shan You", "Yi Chen", "Chang Xu", "Wen-Hao Li", "Hong-Yan Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.01194", "pdf_url": "https://arxiv.org/pdf/2605.01194", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["uncertainty gating", "test-time compute", "critic"], "model_class": "VLA", "techniques": ["adaptive test-time compute", "pairwise critic"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-LONG"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds an uncertainty-triggered deliberation phase to VLAs that samples candidates and picks one with a relative action critic, with amortized sampling to limit cost.", "relevance": 3, "relevance_reason": "Adaptive test-time compute only when uncertain.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]; snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2609.06508", "title": "VLA-Corrector: Stage-Aware Observable State Understanding for Prompt-Based Closed-Loop Recovery of Vision-Language-Action Policies", "authors": ["Chang Song", "Bin Qian", "Yan Feng", "Zhijie Song"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.06508", "pdf_url": "https://arxiv.org/pdf/2609.06508", "code_url": "", "clusters": ["C7"], "keywords": ["stage-aware monitoring"], "model_class": "VLA", "techniques": ["learned verifier", "prompt recovery"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learned verifier over observable history estimates stage progress and execution risk and triggers prompt-based recovery of a frozen VLA.", "relevance": 3, "relevance_reason": "Chunk-level failure verification and closed-loop recovery without retraining.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (\"failure recovery\" OR \"error recovery\" OR \"self-correct*\")"}
{"id": "2604.05323", "title": "VLA-InfoEntropy: A Training-Free Vision-Attention Information Entropy Approach for Vision-Language-Action Models Inference Acceleration and Success", "authors": ["Chuhang Liu", "Yayun He", "Zuheng Kang", "Xiaoyang Qu", "Jianzong Wang"], "year": 2026, "venue": "ICME 2026", "type": "paper", "url": "https://arxiv.org/abs/2604.05323", "pdf_url": "https://arxiv.org/pdf/2604.05323", "code_url": "", "clusters": ["C1"], "keywords": ["token selection", "entropy", "training-free"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Selects visual tokens by combining image entropy and text-attention entropy with a timestep-dependent schedule to speed up VLAs.", "relevance": 3, "relevance_reason": "VLA token reduction.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference"}
{"id": "2609.21246", "title": "VLA-Scope: Shift-Aware Failure Prediction for Vision-Language-Action Models", "authors": ["Kaiwen Zhu", "Dongfang Liu", "Liangkai Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21246", "pdf_url": "https://arxiv.org/pdf/2609.21246", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["VLA-Scope", "shift-aware failure prediction", "OOD"], "model_class": "VLA", "techniques": ["OOD gate + logistic failure predictor"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "failure ROC-AUC 0.8497 after 60 actions; OOD ROC-AUC 0.9454", "metric_measures": "other", "benchmarks": ["LIBERO-Spatial"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Detects and categorizes input shifts, then combines shift type with action-prefix and progress features to update failure risk during rollouts.", "relevance": 3, "relevance_reason": "Runtime failure prediction.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: failure detection vision-language-action"}
{"id": "2605.29605", "title": "VLAConf: Calibrated Task-Success Confidence for Vision-Language-Action Models", "authors": ["Dehao Huang", "Aoxiang Gu", "Cheng-Jie Zhang", "Bo Zou", "Wenlong Dong", "Zi-Lang Cen", "Yue Wang", "Hong Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.29605", "pdf_url": "https://arxiv.org/pdf/2605.29605", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["VLAConf", "task-success confidence", "calibration", "flow-matching VLA"], "model_class": "VLA", "techniques": ["representation-level confidence", "calibration head"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-stage confidence estimator on frozen flow-VLA representations providing calibrated task-success confidence for monitoring.", "relevance": 3, "relevance_reason": "Calibrated runtime monitoring signal.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2609.30868", "title": "VLaRL: Augmenting Vision-Language-Action Models with Simulation-Trained Latent-Conditioned Residual RL", "authors": ["Namiko Saito", "Kinam Kim", "Heecheol Kim", "Katsushi Ikeuchi", "Yasuyuki Matsushita"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.30868", "pdf_url": "https://arxiv.org/pdf/2609.30868", "code_url": "", "clusters": ["C3"], "keywords": ["VLaRL", "latent-conditioned residual RL", "sim-to-real"], "model_class": "VLA", "techniques": ["residual RL", "VLA latent conditioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Sim-trained residual RL conditioned on the frozen VLA's latents transfers to real robots with no real-world RL.", "relevance": 3, "relevance_reason": "Frozen-VLA adaptation without real RL cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"residual policy\""}
{"id": "2606.03954", "title": "VLESA: Vision-Language Embodied Safety Agent for Human Activity Monitoring", "authors": ["Hanjiang Hu", "Yiyuan Pan", "Jiaxing Li", "Xusheng Luo", "Alexander Robey", "Na Li", "Yebin Wang", "Changliu Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.03954", "pdf_url": "https://arxiv.org/pdf/2606.03954", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["safety agent", "real-time interventions", "intent-conditioned safety Q-filter", "ASIMOV-2.0"], "model_class": "other", "techniques": ["goal-conditioned safety Q-filter (GRPO)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ASIMOV-2.0"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLESA monitors egocentric video, infers intent and predicts actions, and triggers real-time safety interventions via an intent-conditioned safety Q-filter.", "relevance": 3, "relevance_reason": "Real-time safety monitoring with foundation models; organizer Changliu Liu.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2606.06323", "title": "VOLT: Vision and Language Trajectory Segmentation for Faster-than-Demonstration Policies", "authors": ["Robert Ramirez Sanchez", "Daniel J. Evans", "Dylan P. Losey", "Siddarth Jain"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.06323", "pdf_url": "https://arxiv.org/pdf/2606.06323", "code_url": "", "clusters": ["C1"], "keywords": ["VOLT", "faster-than-demonstration"], "model_class": "other", "techniques": ["trajectory segmentation", "demo acceleration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses vision-language segmentation to decide which demo parts can be accelerated, training policies that run faster than humans demonstrated.", "relevance": 3, "relevance_reason": "Execution-speed improvement.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.04910", "title": "VPWEM: Non-Markovian Visuomotor Policy With Working and Episodic Memory", "authors": ["Yu-Heng Lei", "Zhi-Xuan Liang", "Hong-Yuan Zhang", "Ping Luo"], "year": 2026, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2603.04910", "pdf_url": "https://arxiv.org/pdf/2603.04910", "code_url": "", "clusters": ["C1"], "keywords": ["non-Markovian", "episodic memory"], "model_class": "diffusion/flow policy", "techniques": ["memory compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MIKASA", "MoMaRT"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diffusion policy with working and compressed episodic memory to keep near-constant compute per step.", "relevance": 3, "relevance_reason": "Constant-cost memory for real-time constraints.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2609.22606", "title": "VT-Bridge: Bridging Pretrained Foundation VLAs to VTLAs via Lightweight Residual Adaptation", "authors": ["Yansong Wu", "Tuo Yang", "Rongping Zhao", "Lingyun Chen", "Xiao Chen", "Junnan Li", "Fan Wu", "Alois Knoll"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22606", "pdf_url": "https://arxiv.org/pdf/2609.22606", "code_url": "", "clusters": ["C3"], "keywords": ["tactile", "residual adapter", "VTLA"], "model_class": "VLA", "techniques": ["residual adapter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Lightweight residual adapter turning pretrained VLAs into vision-tactile policies with up to 50 demos per task.", "relevance": 3, "relevance_reason": "Cheap modality extension by adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2608.08725", "title": "WA-SpecDec: World-Aware Speculative Decoding for Vision-Language-Action Models", "authors": ["Zikang Wen", "Yuning Zhang", "Dong Yuan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.08725", "pdf_url": "https://arxiv.org/pdf/2608.08725", "code_url": "", "clusters": ["C1"], "keywords": ["speculative decoding", "world model", "scene-aware acceptance"], "model_class": "VLA", "techniques": ["speculative decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.5x matched-success speedup over VLA speculative decoding", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Injects world-model-derived scene awareness into VLA prefill so relaxed speculative acceptance stays safe near contact.", "relevance": 3, "relevance_reason": "Speculative decoding for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: speculative decoding VLA"}
{"id": "2608.01035", "title": "WAM-Diff2: Hierarchical AR-to-Diffusion Distillation for Highly Efficient Autonomous Driving VLA", "authors": ["Zhihao Zhu", "Hanlin Shang", "Mingwang Xu", "Feipeng Cai", "Zhuolin He", "Yaoyi Li", "Jianhua Han", "Hang Xu", "Siyu Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.01035", "pdf_url": "https://arxiv.org/pdf/2608.01035", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["AR-to-diffusion", "driving VLA", "block-wise distillation"], "model_class": "VLA", "techniques": ["hierarchical distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Converts an autoregressive driving VLA into a parallel discrete-diffusion VLA via progressive block-wise and model-wise distillation.", "relevance": 3, "relevance_reason": "Distillation-based architectural conversion for speed.", "foundational_technique": false, "verified": true, "found_via": "search: distillation vision-language-action"}
{"id": "2606.04907", "title": "WAM-Nav: Asymmetric Latent World-Action Modeling for Unified Visual Navigation", "authors": ["Ning Yang", "Yan Huang", "Kaiwen Peng", "Ziheng He", "Kai Wang", "Cui Miao", "Kailin Lyu", "Guo Li", "Xiaofeng Wang", "Zheng Zhu", "Jing Liu", "Nianfeng Liu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.04907", "pdf_url": "https://arxiv.org/pdf/2606.04907", "code_url": "", "clusters": ["C8"], "keywords": ["WAM-Nav", "navigation"], "model_class": "world-action model", "techniques": ["asymmetric joint diffusion", "joint diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Latent navigation WAM generating long-horizon actions and short-horizon foresight jointly to cut latency versus autoregressive rollouts.", "relevance": 3, "relevance_reason": "Latency-conscious WAM for navigation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]; snowball round 2: cites 2505.21851; cited by 2609.33148"}
{"id": "2608.22364", "title": "WAM-OPD: On-Policy Distillation for World Action Models", "authors": ["Liuhaichen Yang", "Zhuang Jiang", "Chenchao Sheng", "Zezhi Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.22364", "pdf_url": "https://arxiv.org/pdf/2608.22364", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["on-policy distillation", "accelerated student repair", "WAM student repair"], "model_class": "world-action model", "techniques": ["on-policy distillation", "on-policy distillation from frozen teacher", "adapter fine-tuning"], "base_model": "Flash-WAM (1-step)", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "RoboTwin 2.0 (2 tasks)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Repairs capability lost by step-distilled WAM students by letting the student act and having a frozen teacher label its own histories; preliminary gains on two RoboTwin tasks.", "relevance": 3, "relevance_reason": "Addresses accuracy loss from WAM acceleration.", "foundational_technique": false, "verified": true, "found_via": "search: one-step policy distillation robot; search query"}
{"id": "2607.06988", "title": "WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time", "authors": ["Yusen Feng", "Bingchen Han", "Jiangran Lyu", "Kai Liu", "Yi-Xin Zheng", "Yuxuan Wan", "Weiheng Liu", "Sun Han", "Ruiqing Li", "Yulong Zhang", "Fang-Fu Liu", "Xuesong Shi", "Libin Liu", "Yizhou Wang", "Zhizheng Zhang", "He Wang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.06988", "pdf_url": "https://arxiv.org/pdf/2607.06988", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["WAM-TTT", "human play"], "model_class": "world-action model", "techniques": ["test-time training", "lightweight memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Steers a frozen WAM at test time by writing unlabeled human videos into a lightweight adaptive memory, without robot actions or fine-tuning.", "relevance": 3, "relevance_reason": "Efficient adaptation of WAMs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2504.16054, 2601.16163"}
{"id": "2606.14048", "title": "WAM4D: Fast 4D World Action Model via Spatial Register Tokens", "authors": ["Ying Li", "Xiaobao Wei", "Jiajun Cao", "Hao Wang", "Xiaowei Chi", "Chengyu Bai", "Qianpu Sun", "Jiajun Li", "Xiaojie Zhang", "Peidong Jia", "Jian Tang", "Sirui Han", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.14048", "pdf_url": "https://arxiv.org/pdf/2606.14048", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["4D WAM", "spatial register tokens", "geometry priors", "fast"], "model_class": "world-action model", "techniques": ["training-only geometry readout tokens", "causal mixture attention", "register branch removed at inference", "register tokens for 3D"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Transfers 3D geometric priors into a causal video-action transformer through lightweight register tokens used only during training, so inference stays as cheap as a 2D WAM.", "relevance": 3, "relevance_reason": "Adds 3D awareness without inference cost.", "foundational_technique": false, "verified": true, "found_via": "search query; venue sweep: arXiv API 'world action model' AND efficiency"}
{"id": "2605.13959", "title": "WarmPrior: Straightening Flow-Matching Policies with Temporal Priors", "authors": ["Sinjae Kang", "Chanyoung Kim", "Kaixin Wang", "Li Zhao", "Kimin Lee"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.13959", "pdf_url": "https://arxiv.org/pdf/2605.13959", "code_url": "", "clusters": ["C1"], "keywords": ["source distribution", "temporal prior", "straight paths"], "model_class": "diffusion/flow policy", "techniques": ["warm-start prior"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces the Gaussian source with a prior built from recent action history, producing straighter flow paths and higher success.", "relevance": 3, "relevance_reason": "Flow policy design relevant to few-step sampling.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching robot policy efficient"}
{"id": "2609.37107", "title": "Waypoint-1.5: A Real-Time Video World Model for Consumer Hardware", "authors": ["Rajit Rajpal", "Shahbuland Matiana", "Liew Wei Pyn", "Anmol Agarwal", "Ryan Craig", "Andrew Lapp", "Mithun Hunsur", "Sami BuGhanem", "Scottie Fox", "Aaron Sanders Carson Poole", "Irene Park", "Dave Rossi", "Spencer Frazier", "Louis Castricato"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2609.37107", "pdf_url": "https://arxiv.org/pdf/2609.37107", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["consumer hardware", "interactive world model", "control rate vs FPS"], "model_class": "world model", "techniques": ["real-time diffusion world model", "two resolution variants"], "base_model": "", "params": "", "hardware": "consumer GPUs", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Real-time interactive diffusion world model trained on 100k hours of game video that runs on consumer hardware, and usefully separates rendered FPS, latent FPS and control rate as metrics.", "relevance": 3, "relevance_reason": "Clarifies real-time world-model metrics on cheap hardware.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2604.17040", "title": "When Spike Sparsity Does Not Translate to Deployed Cost: VS-WNO on Jetson Orin Nano", "authors": ["Jason Yoo", "Shailesh Garg", "Souvik Chakraborty", "Syed Bahauddin Alam"], "year": 2026, "venue": "arXiv (ICONS 2026 submission)", "type": "preprint", "url": "https://arxiv.org/abs/2604.17040", "pdf_url": "https://arxiv.org/pdf/2604.17040", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["sparsity", "Nsight", "launch overhead", "energy per inference"], "model_class": "other", "techniques": ["spiking sparsity"], "base_model": "", "params": "", "hardware": "Jetson Orin Nano 8GB", "reported_metric": "sparse VS-WNO 59.6 ms / 228.0 mJ vs dense WNO 53.2 ms / 180.7 mJ per inference; cudaLaunchKernel 81.6% of CUDA API time", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows measured algorithmic sparsity yields no latency or energy benefit on a Jetson GPU stack because inference is launch-dominated and dense kernels ignore sparsity.", "relevance": 3, "relevance_reason": "Clear example of theoretical efficiency not transferring to measured cost on Jetson.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q3: arXiv abs:\"energy per inference\" AND abs:edge"}
{"id": "2602.22474", "title": "When to Act, Ask, or Learn: Uncertainty-Aware Policy Steering", "authors": ["Jessie Yuan", "Yilin Wu", "Andrea Bajcsy"], "year": 2026, "venue": "RSS 2026", "type": "paper", "url": "https://arxiv.org/abs/2602.22474", "pdf_url": "https://arxiv.org/pdf/2602.22474", "code_url": "", "clusters": ["C7"], "keywords": ["ask for help", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["policy steering", "uncertainty quantification", "VLM verifier"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Policy steering framework that jointly reasons about semantic and low-level action uncertainty and chooses to execute, ask a clarifying question, or request intervention.", "relevance": 3, "relevance_reason": "Decision rule for when not to execute sampled actions.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:conformal AND (\"vision-language-action\" OR \"diffusion policy\" OR \"action chunk*\" OR \"imitation learning\")"}
{"id": "2608.22421", "title": "Where World Models Break: Natural-Input Failure Discovery", "authors": ["Zhanpeng Shi", "Zi Liang", "Rong Feng", "Shiqin Tang", "Xuyang Chen", "Hongzong Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.22421", "pdf_url": "https://arxiv.org/pdf/2608.22421", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["failure discovery", "stress testing world models", "query budget"], "model_class": "world model", "techniques": ["budgeted search for catastrophic condition-action inputs"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Formulates and solves budgeted search for natural conditions and action prefixes under which a world model's predictions collapse, before those errors reach planning or safety pipelines.", "relevance": 3, "relevance_reason": "Validating world models used for safety.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2608.02547", "title": "Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control?", "authors": ["Filippo Lazzati", "Kyle Stachowicz", "William Chen", "Alberto Maria Metelli", "Andrew Wagenmaker", "Sergey Levine"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.02547", "pdf_url": "https://arxiv.org/pdf/2608.02547", "code_url": "", "clusters": ["C1"], "keywords": ["action chunking", "implicit ensembling", "delayed policies"], "model_class": "other", "techniques": ["action chunking analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Tests why chunking helps BC, finding non-Markovian expressivity and implicit ensembling rather than temporal consistency, and matches chunking with randomized-delay ensembles.", "relevance": 3, "relevance_reason": "Understanding of action chunking.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunking"}
{"id": "2609.03681", "title": "WISE: World-model-guided Imagination Scheduling for Efficient Post-training of Vision-Language-Action Models", "authors": ["Chenhao Zhang", "Hanyu Zhao", "Hang Cheng", "Tengfei Pan", "Long Zeng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.03681", "pdf_url": "https://arxiv.org/pdf/2609.03681", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["world model imagination", "scheduling", "post-training", "imagination scheduling", "VLA post-training", "bounded horizons"], "model_class": "world model", "techniques": ["world-model-based post-training", "stage-aware scheduling of imagination", "horizon bounding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Schedules world-model imagination only at execution stages where it helps and bounds horizons to reliable ranges, making world-model-based VLA post-training cheaper and more reliable.", "relevance": 3, "relevance_reason": "Efficient post-training via imagination.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations); search query"}
{"id": "2609.20820", "title": "Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision", "authors": ["Nitish Dashora", "Douglas Chen", "Idan Shenfeld", "John Marangola", "Pulkit Agrawal", "Max Simchowitz"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.20820", "pdf_url": "https://arxiv.org/pdf/2609.20820", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["lightweight memory", "workspace token"], "model_class": "other", "techniques": ["distillation of VLM reasoning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills VLM-identified task-relevant history into a lightweight latent workspace token, avoiding VLM calls in the loop.", "relevance": 3, "relevance_reason": "Removes expensive VLM queries at deployment.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2602.16182", "title": "World Model Failure Classification and Anomaly Detection for Autonomous Inspection", "authors": ["Michelle Ho", "Muhammad Fadhil Ginting", "Isaac R. Ward", "Andrzej Reinke", "Mykel J. Kochenderfer", "Ali-akbar Agha-Mohammadi", "Shayegan Omidshafiei"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.16182", "pdf_url": "https://arxiv.org/pdf/2602.16182", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["inspection robots", "failure classification", "anomaly detection", "conformal"], "model_class": "world model", "techniques": ["world model backbone on compressed video", "conformal decision functions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Policy-agnostic world-model-based monitor that labels inspection runs as success, known failure or anomaly using conformal thresholds.", "relevance": 3, "relevance_reason": "World-model runtime monitoring in the field.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.00935 (citations)"}
{"id": "2605.00080", "title": "World Model for Robot Learning: A Comprehensive Survey", "authors": ["Bo-Han Hou", "Gen Li", "Jin-Dou Jia", "Tuo An", "Xin-Ying Guo", "Si-Cong Leng", "Hao-Ran Geng", "Yanjie Ze"], "year": 2026, "venue": "IJRR (per Semantic Scholar) / arXiv", "type": "paper", "url": "https://arxiv.org/abs/2605.00080", "pdf_url": "https://arxiv.org/pdf/2605.00080", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["world models", "survey", "robot learning"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of world models from a robot-learning viewpoint: policy coupling, learned simulators for RL/evaluation, and application domains.", "relevance": 3, "relevance_reason": "Reference map for efficient world-model section; listed on Jiajun Wu's S2 profile (full author list not verified).", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "2608.09730", "title": "World Tokens: Enhancing Embodied Policies with Training-Time World Modeling", "authors": ["Qu Tang", "Benhui Zhuang", "Bo Yuan", "Xue Yu", "Long-Teng Guo", "Jun-Lan Feng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09730", "pdf_url": "https://arxiv.org/pdf/2608.09730", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["world tokens", "world adapter"], "model_class": "VLA", "techniques": ["training-time world modeling", "branch removed at deployment"], "base_model": "", "params": "2B backbone", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "SIMPLER"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses a future-video denoiser only during training to shape world tokens consumed by the action expert, then removes it so chunks are produced at VLA-level latency.", "relevance": 3, "relevance_reason": "Efficient-deployment design for world-model-enhanced VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2609.16074", "title": "World-Action Models for Robot Learning and Control: A Survey", "authors": ["Zuxing Lu", "Hongjia Zhai", "Guanzhi Wang", "Huajian Zeng", "Jiaqi Yang", "Jingyu Liu", "Lei Cheng", "Yuantai Zhang", "Yuheng Qiu", "Zezhou Cheng", "Ivan Laptev", "Danfei Xu", "Benjamin Riviere", "Giuseppe Loianno", "Eric Xing", "Xingxing Zuo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.16074", "pdf_url": "https://arxiv.org/pdf/2609.16074", "code_url": "", "clusters": ["C8"], "keywords": ["survey", "world-action models"], "model_class": "world-action model", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Robotics-focused survey organizing WAMs by representation, transition model, action interface and training pipeline, and listing efficient inference as an open challenge.", "relevance": 3, "relevance_reason": "Survey covering WAM efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.02159", "title": "World-Coherent Decoding: Self-Verifying Test-Time Planning for World Action Models", "authors": ["Chuhan Zhang", "Seiji Ito", "Kenta Hoshino", "Satoshi Ikehata", "Ikuro Sato"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.02159", "pdf_url": "https://arxiv.org/pdf/2609.02159", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["WCD"], "model_class": "world-action model", "techniques": ["self-verification", "test-time planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Samples multiple WAM rollouts and ranks them by video surprisal and action path effort, then learns an online predictor from imagination-reality mismatch.", "relevance": 3, "relevance_reason": "Pre-execution reliability estimate for WAM chunks.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2602.02454", "title": "World-Gymnast: Training Robots with Reinforcement Learning in a World Model", "authors": ["Ansh Kumar Sharma", "Yixiang Sun", "Ninghao Lu", "Yunzhe Zhang", "Jiarao Liu", "Sherry Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.02454", "pdf_url": "https://arxiv.org/pdf/2602.02454", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["World-Gymnast", "RL in world model", "VLA finetuning", "VLM reward"], "model_class": "world model", "techniques": ["world-model RL", "RL rollouts inside action-conditioned video world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Bridge (real WidowX)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "RL-finetunes a VLA entirely inside a video world model with VLM rewards, outperforming SFT by up to 18x and a software simulator by 2x on real Bridge tasks.", "relevance": 3, "relevance_reason": "World model replaces costly real interaction; rollout speed is a bottleneck.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations); search query"}
{"id": "2606.05979", "title": "World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis", "authors": ["Yi Yang", "Zhihong Liu", "Siqi Kou", "Yiyang Chen", "Yanzhe Hu", "Jianbo Zhou", "Boyuan Zhao", "Zhijie Wei", "Xiao Xia", "Xueqi Li", "Pengfei Liu", "Zhijie Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.05979", "pdf_url": "https://arxiv.org/pdf/2606.05979", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["WLA-0"], "model_class": "world-action model", "techniques": ["autoregressive backbone", "optional world prediction"], "base_model": "", "params": "2B active", "hardware": "NVIDIA RTX 5090", "reported_metric": "40 ms per inference", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "World-language-action model on an AR transformer that predicts subtasks, subgoal images and actions, with world prediction disable-able at inference; 2B active parameters.", "relevance": 3, "relevance_reason": "Reports concrete per-inference latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2602.06508", "title": "World-VLA-Loop: Closed-Loop Learning of Video World Model and VLA Policy", "authors": ["Xiaokang Liu", "Zechen Bai", "Hai Ci", "Kevin Yuchen Ma", "Mike Zheng Shou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.06508", "pdf_url": "https://arxiv.org/pdf/2602.06508", "code_url": "", "clusters": ["C3"], "keywords": ["World-VLA-Loop", "co-evolving world model and policy"], "model_class": "VLA", "techniques": ["world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Co-trains an action-conditioned video world model and VLA with near-success data to make virtual RL fine-tuning reliable.", "relevance": 3, "relevance_reason": "Cuts real-world RL cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"world model\" AND abs:\"post-training\""}
{"id": "2603.10422", "title": "World2Act: Latent Action Post-Training from World Model Dynamics", "authors": ["An Dinh Vuong", "Tuan Van Vo", "Abdullah Sohail", "Haoran Ding", "Liang Ma", "Xiaodan Liang", "Anqing Duan", "Ivan Laptev", "Ian Reid"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.10422", "pdf_url": "https://arxiv.org/pdf/2603.10422", "code_url": "", "clusters": ["C3"], "keywords": ["latent post-training", "world model dynamics", "GR00T-N1.6"], "model_class": "VLA", "techniques": ["world-model latent alignment"], "base_model": "GR00T-N1.6", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Post-trains a VLA by aligning action representations to world-model latent dynamics rather than pixel supervision.", "relevance": 3, "relevance_reason": "Efficient world-model-based post-training.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.08162", "title": "WorldAgen: Unified State-Action Prediction with Test-Time World Model Training", "authors": ["Chi Wan", "Kangrui Wang", "Yuan Si", "Pingyue Zhang", "Manling Li"], "year": 2026, "venue": "AAAI 2026", "type": "paper", "url": "https://arxiv.org/abs/2609.08162", "pdf_url": "https://arxiv.org/pdf/2609.08162", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["adaptation", "unified state-action prediction"], "model_class": "world-action model", "techniques": ["test-time training", "world model head"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shared-transformer VLA with a world-model head and agent head separated by an attention mask, adapting via test-time training on exploratory interaction data.", "relevance": 3, "relevance_reason": "Test-time adaptation using a built-in world model.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2602.08971", "title": "WorldArena: A Unified Benchmark for Evaluating Perception and Functional Utility of Embodied World Models", "authors": ["Shang, Yu", "Li, Zhuohang", "Ma, Yiding", "Su, Weikang", "Jin, Xin", "Wang, Ziyou", "Jin, Lei", "Zhang, Xin", "Tang, Yinzhou", "Su, Haisheng", "Gao, Chen", "Wu, Wei", "Liu, Xihui", "Shah, Dhruv", "Zhang, Zhaoxiang", "Chen, Zhibo", "Zhu, Jun", "Tian, Yonghong", "Chua, Tat-Seng", "Zhu, Wenwu", "Li, Yong", "Yu Shang", "Zhuohang Li", "Yiding Ma", "Weikang Su", "Xin Jin", "Ziyou Wang", "Lei Jin", "Xin Zhang", "Yinzhou Tang", "Haisheng Su", "Chen Gao", "Wei Wu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2602.08971", "pdf_url": "https://arxiv.org/pdf/2602.08971", "code_url": "", "clusters": ["C3", "C4", "C7", "C9"], "keywords": ["WorldArena", "EWMScore", "world model as policy evaluator", "world model benchmark", "functional utility"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["WorldArena"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmark scoring embodied world models on perceptual quality and functional utility (data engine, policy evaluator, planner); finds a perception-functionality gap.", "relevance": 3, "relevance_reason": "World-model benchmarking; Dhruv Shah co-author.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: WorldArena benchmark; author sweep: Dhruv Shah"}
{"id": "2609.19600", "title": "WorldContact: A Contact-Centric World Model for Scalable Robot Learning", "authors": ["Caoliwen Wang", "Mengdi Wang", "Heng Zhang", "Shixun Huang", "Siyuan Chen", "Chao Liu", "Anpei Chen", "Zhendong Wang", "Peter Yichen Chen", "Huamin Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19600", "pdf_url": "https://arxiv.org/pdf/2609.19600", "code_url": "", "clusters": ["C8"], "keywords": ["world model", "deformable manipulation"], "model_class": "world model", "techniques": ["learned simulator", "data generation"], "base_model": "", "params": "", "hardware": "single H100", "reported_metric": "10x state-rollout speedup over source simulator", "metric_measures": "throughput", "benchmarks": ["16 bag tasks", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Contact-centric learned world model for deformable bag manipulation that steps with larger time steps than the source simulator to generate VLA fine-tuning data cheaply.", "relevance": 3, "relevance_reason": "World model as a faster data generator for VLA adaptation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2609.34206", "title": "WorldGuide: Learning Success-Failure Boundaries in Latent World Models for Vision-Language-Action Policies", "authors": ["Lin Liu", "Lu Zhang", "Ziying Song", "Wu Yang", "Yuzheng Zhuang", "Yunzhi Zhuge", "Shuai Tao", "Wulong Liu", "Huchuan Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34206", "pdf_url": "https://arxiv.org/pdf/2609.34206", "code_url": "", "clusters": ["C8"], "keywords": ["success-failure boundaries", "latent world model reward", "train-only predictor"], "model_class": "world model", "techniques": ["contrastive success/failure latent pretraining", "differentiable latent reward", "predictor discarded at deployment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Learns a latent world model that separates similar-looking successes and failures and uses it as a differentiable reward during VLA training; nothing extra runs at deployment.", "relevance": 3, "relevance_reason": "World-model failure knowledge without runtime cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2603.16666 (citations)"}
{"id": "2605.22718", "title": "WorldKV: Efficient World Memory with World Retrieval and Compression", "authors": ["Jung Yi", "Minjae Kim", "Paul Hyunbin Cho", "Wooseok Jang", "Sangdoo Yun", "Seungryong Kim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.22718", "pdf_url": "https://arxiv.org/pdf/2605.22718", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["world memory", "KV retrieval", "KV compression"], "model_class": "world model", "techniques": ["retrieve evicted KV chunks by camera/action correspondence", "key-similarity token pruning"], "base_model": "Matrix-Game-2.0", "params": "", "hardware": "", "reported_metric": "halves per-chunk KV storage (2x history under fixed budget)", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Training-free persistent memory for real-time AR world models: offloads and retrieves relevant KV chunks and prunes redundant tokens, keeping consistency without linear cost growth.", "relevance": 3, "relevance_reason": "Memory-efficient long rollouts.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2609.38059", "title": "WorldLine: Action-Driven Visual Simulation for Robotic Manipulation", "authors": ["Sheng-He Zheng", "Wen-Bo Li", "Ji-Yao Zhang", "Bin Xia", "Hao-Yang Huang", "Nan Duan", "Jia-Ya Jia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.38059", "pdf_url": "https://arxiv.org/pdf/2609.38059", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["visual simulator", "policy evaluation"], "model_class": "world model", "techniques": ["few-step distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Action-driven video simulator for manipulation with few-step distillation for efficient causal rollouts used in policy evaluation and planning.", "relevance": 3, "relevance_reason": "Efficient world model for evaluation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2410.24164"}
{"id": "2607.02431", "title": "WorldSample: Closed-loop Real-robot RL with World Modelling", "authors": ["Yuquan Xue", "Le Xu", "Zeyi Liu", "Zhenyu Wu", "Zhengyi Gu", "Xinyang Song", "Bofang Jia", "Ziwei Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.02431", "pdf_url": "https://arxiv.org/pdf/2607.02431", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["world model augmentation", "real-robot RL", "policy-paced learning"], "model_class": "VLA", "techniques": ["world-model data augmentation", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "+28% success, 59% fewer training steps", "metric_measures": "other", "benchmarks": ["real robot contact-rich tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Augments costly real rollouts with world-model-generated transitions, scheduled to avoid hallucination-induced value errors.", "relevance": 3, "relevance_reason": "Cuts real-world interaction cost of RL post-training.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv"}
{"id": "2608.09298", "title": "WorldSimProbe: Diagnosing Simulator Faithfulness in Action-Conditioned World Models for Embodied Manipulation", "authors": ["P. Co", "Sichen Hu", "Chun-Xuan Jiao", "Hong-Yang Cheng", "Yulin Luo", "Yijie Xu", "Si-Xiang Chen", "Zhongxia Zhao", "Zihao Wang", "Dafeng Chi", "Pei-Dong Liu", "Yu-Tong Chen", "Henghua Liu", "Zhihao Yuan", "Huizhu Jia", "Yu-Zheng Zhuang", "Tian-Le Zhang", "Liang Lin", "Hua-Jie Tan", "Shang-Hang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09298", "pdf_url": "https://arxiv.org/pdf/2608.09298", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["world model evaluation", "simulator faithfulness"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diagnostic contract and evaluation for whether action-conditioned world models behave like faithful physical simulators.", "relevance": 3, "relevance_reason": "Benchmarking methodology for world models.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.00425"}
{"id": "2602.13977", "title": "WoVR: World Models as Reliable Simulators for Post-Training VLA Policies with RL", "authors": ["Zhennan Jiang", "Shangqing Zhou", "Yutong Jiang", "Zefang Huang", "Mingjie Wei", "Yuhui Chen", "Tianxing Zhou", "Zhen Guo", "Hao Lin", "Quanlu Zhang", "Yu Wang", "Haoran Li", "Chao Yu", "Dongbin Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.13977", "pdf_url": "https://arxiv.org/pdf/2602.13977", "code_url": "", "clusters": ["C3"], "keywords": ["WoVR", "hallucination-aware world model RL"], "model_class": "VLA", "techniques": ["world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "World-model RL post-training that explicitly handles hallucination and long-horizon error accumulation.", "relevance": 3, "relevance_reason": "Efficient RL post-training without real interaction.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2601.04137", "title": "Wow, wo, val! A Comprehensive Embodied World Model Evaluation Turing Test", "authors": ["Chunkai Fan", "Xiaowei Chi", "Xiao-Zhu Ju", "Hao Li", "Yong Bao", "Yu-Kai Wang", "Lizhang Chen", "Zhiyuan Jiang", "Kuangzhi Ge", "Ying Li", "Wei-Shi Mi", "Qingpo Wuwu", "Pei-Dong Jia", "Yulin Luo", "Kevin Zhang", "Zhiyuan Qin", "Yong Dai", "Si-Rui Han", "Yi-Ke Guo", "Shanghang Zhang", "Jian Tang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2601.04137", "pdf_url": "https://arxiv.org/pdf/2601.04137", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["WoW-World-Eval", "embodied Turing test"], "model_class": "world model", "techniques": ["evaluation protocol"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmark with 22 metrics evaluating video world models for embodied use across perception, planning, prediction, generalization and execution.", "relevance": 3, "relevance_reason": "World-model evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2402.15391"}
{"id": "2601.11269", "title": "X-Distill: Cross-Architecture Vision Distillation for Visuomotor Learning", "authors": ["Maanping Shao", "Feihong Zhang", "Gu Zhang", "Baiye Cheng", "Zhengrong Xue", "Huazhe Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.11269", "pdf_url": "https://arxiv.org/pdf/2601.11269", "code_url": "", "clusters": ["C2"], "keywords": ["cross-architecture distillation", "DINOv2 -> ResNet-18"], "model_class": "diffusion/flow policy", "techniques": ["knowledge distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["34 simulated tasks", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills DINOv2 representations into a compact ResNet-18 on ImageNet, then fine-tunes it with a diffusion policy, outperforming both originals.", "relevance": 3, "relevance_reason": "Encoder compression via distillation for policies.", "foundational_technique": false, "verified": true, "found_via": "search: knowledge distillation robot manipulation policy"}
{"id": "2606.28758", "title": "X-Mind: Efficient Visual Chain-of-Thought via Predictive World Model for End-to-End Driving", "authors": ["Bohao Zhao", "Chengrui Wei", "Guangfeng Jiang", "Ruixin Liu", "Xuejie Lv", "Liu Liang", "Sutao Deng", "Xiuyang Fan", "Pengkun Zheng", "Jinyun Zhou", "Rui Guo", "Hanpeng Liu", "Yutong Zheng", "Yi Guo", "Xinlong Zheng", "Qingyu Luo", "Zhuangzhuang Ding", "Yu Zhang", "Hang Zhang", "Xianming Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.28758", "pdf_url": "https://arxiv.org/pdf/2606.28758", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["visual CoT", "driving"], "model_class": "VLA", "techniques": ["token compression", "block diffusion", "world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA that reasons over a compact BEV sketch (12-frame rollout compressed to 96 tokens) and folds diffusion denoising into the backbone forward pass via recurrent block diffusion.", "relevance": 3, "relevance_reason": "Efficient visual chain-of-thought via token compression.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]"}
{"id": "2608.09892", "title": "XPolicyLab: A Unified Standard and Open Ecosystem for Robot Policy Evaluation and Deployment", "authors": ["XPolicyLab Community", "Tian-Xing Chen", "Yue Chen", "Nian Tian", "Zijian Cai", "Guang-Yu Chen", "Wen-Wei Lin", "Qi-Wei Liang", "Zan-Xin Chen", "Peicheng Xiang", "Kai-Lun Su", "Zi-Xuan Li", "Junyuan Tang", "Yan-Jiao Qin", "Qiang-Yu Chen", "Shao-Long Zhu", "Teng-Long Jiang", "Yi-Qing Wang", "Xiang Li", "Jiahao Zhang", "Weijie Wan", "Bai-Jun Chen", "Honghao Su", "Ke Ye", "Shu-Jia Liu", "Kai-Xuan Wang", "Haotian Liang", "Yunze Liu", "Mingleyang Li", "Yu-Ran Wang", "Bo-Yu Chen", "Hong-Zhe Bi", "Shu-He Huang", "Heng-Kai Tan", "Ji-Song Cai", "Yao Mu", "Jun Guo", "Xiao-Feng Wang", "Zhengang Zhu", "Wei Ke", "Heng-Tao Li", "Yu Tang", "Xiao-Fan Li", "Gan-Lin Yang", "Zhang-Zheng Tu", "Shuai Yang", "Wen-Xuan Song", "Pengxiang Ding", "Kai-Dong Zhang", "Yu Sun", "Jun Guo", "Tongyong Zhang", "Yi-Xin Chen", "Rongxu Cui", "Zongzheng Zhang", "Hao-Xiang Ma", "Jun-Hao Cai", "Haoyu Zhang", "Sen-Qiao Yang", "Jinhui Ye", "Pengguang Chen", "Shu Liu", "Xiu Su", "Wenhan Fang", "Wenhao Li", "Yichao Cao", "Chengyao Wang", "Qiang Chen", "Ping Luo", "Wen-Bo Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09892", "pdf_url": "https://arxiv.org/pdf/2608.09892", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["evaluation standard", "policy serving", "reproducibility"], "model_class": "other", "techniques": ["client/server inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin", "RoboDojo", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Unified schema and adapter interface with client/server separation of policy inference and environment, integrating 42 policies for standardized sim and real evaluation.", "relevance": 3, "relevance_reason": "Benchmarking/deployment infrastructure for robot policies, supports remote inference", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.13782", "title": "Your Vision-Language-Action Model Already Has Attention Heads For Path Deviation Detection", "authors": ["Jaehwan Jeong", "Evelyn Zhu", "Jinying Lin", "Emmanuel Jaimes", "Tuan-Anh Vu", "Jungseock Joo", "Sangpil Kim", "M. Khalid Jawed"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.13782", "pdf_url": "https://arxiv.org/pdf/2603.13782", "code_url": "", "clusters": ["C7"], "keywords": ["navigation VLA", "monitoring"], "model_class": "VLA", "techniques": ["attention-head probing", "deviation detection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Finds a few attention heads in frozen navigation VLAs that detect path deviation at no extra compute, enabling recovery without external critics.", "relevance": 3, "relevance_reason": "Zero-overhead internal deviation monitor.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2609.35231", "title": "Zero-Shot Reactive Obstacle Avoidance for Generative Robot Policies", "authors": ["Wei-Hang Guo", "Lydia E. Kavraki"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.35231", "pdf_url": "https://arxiv.org/pdf/2609.35231", "code_url": "", "clusters": ["C7"], "keywords": ["obstacle avoidance", "safety"], "model_class": "diffusion/flow policy", "techniques": ["SDF gradient guidance", "training-free"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "NUDGE injects signed-distance-field gradients into diffusion/flow policies (including VLAs) at inference to reactively avoid obstacles in real time without retraining.", "relevance": 3, "relevance_reason": "relevant: inference-time safety steering", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.16590", "title": "Zetta $\\zeta$: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence", "authors": ["Xin Ding", "Liang Mi", "Ming-Zhe Huang", "Zixuan Wang", "Chao Zhang", "Zi-Xu Hao", "Fu Chen", "Xiangyu Li", "Yikai Zheng", "Yao Guo", "Wei-Jun Wang", "Kun Li", "Hao Wu", "Yun-Xin Liu", "Ting Cao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.16590", "pdf_url": "https://arxiv.org/pdf/2608.16590", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["embodied harness", "closed-loop", "self-evolving", "recovery skills"], "model_class": "VLA", "techniques": ["runtime critics", "timescale separation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "11.1x inference speedup", "metric_measures": "relative latency", "benchmarks": ["LIBERO-Pro", "RoboCasa"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Closed-loop harness around a frozen VLA with three timescale-separated loops (action-frequency governance, critic/recovery proposal, gated skill updates) plus rollout infrastructure.", "relevance": 3, "relevance_reason": "Runtime governance at action frequency with reported inference speedup.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2609.00188", "title": "ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training", "authors": ["Xionghao Wu", "Yijun Yang", "Shiyang Zhou", "Haoze Sun", "Jianhui Liu", "Songsong Yu", "Jiyao Zhang", "Wenbo Li", "Bo Wang", "Guoqing Ma", "Lin Song", "Renjie Liao", "Shenghe Zheng", "Wei Tang", "Xiaojuan Qi", "Yanwei Li", "Yuan Zhang", "Zhuotao Tian", "Haoyang Huang", "Nan Duan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.00188", "pdf_url": "https://arxiv.org/pdf/2609.00188", "code_url": "", "clusters": ["C8"], "keywords": ["ZimaBlue"], "model_class": "world-action model", "techniques": ["slow-fast dual system", "video pretraining"], "base_model": "", "params": "", "hardware": "NVIDIA RTX 4090", "reported_metric": "30 Hz action prediction", "metric_measures": "open-loop chunk rate", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "WAM trained on 120k+ hours of egocentric video with an asynchronous slow-fast design: a large slow world model plus a light fast branch for action prediction.", "relevance": 3, "relevance_reason": "Slow-fast split to make a generative WAM run at 30 Hz.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2609.21712", "title": "ZYT-World: A Real-Time Controllable World Model for Closed-Loop Autonomous-Driving Simulation", "authors": ["Boni Hu", "Xiong Wei", "Haoming Huang", "Yong Huang", "Chenbo Wang", "Yi Yang", "Jiancheng Wang", "Ruicheng Zhu", "Zhimin Yang", "Guanglai Liu", "Qiaowan Jin", "Dongzhuo Wang", "Haiwei Kuang", "Jiajun Fan", "Yue Wu", "Jiaxin Wei", "Hao Sun", "Feihong Yan", "Yuyao Zhou", "Wei Bi", "Kaixuan Wang", "Zichao Guo", "Xiaozhi Chen"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2609.21712", "pdf_url": "https://arxiv.org/pdf/2609.21712", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["driving world model", "closed-loop simulation"], "model_class": "world model", "techniques": ["consistency distillation", "distribution matching distillation", "W8A8 quantization", "tiny VAE"], "base_model": "", "params": "", "hardware": "", "reported_metric": "107.7x faster than 40-step teacher (generator only); TinyVAE 59.8x faster than Wan decoder", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving world model distilled from a 40-step bidirectional teacher into a one-step streaming generator, with a 19M TinyVAE decoder and W8A8 quantization for real-time closed-loop simulation.", "relevance": 3, "relevance_reason": "World-model distillation plus quantization for real-time use.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2606.16022", "title": "λ-Reachability: Geometric-Horizon Safety Bellman Equations for Humanoid Safety", "authors": ["Rui Chen", "Shangtao Li", "Yifan Sun", "Changliu Liu"], "year": 2026, "venue": "CoRL 2026", "type": "paper", "url": "https://arxiv.org/abs/2606.16022", "pdf_url": "https://arxiv.org/pdf/2606.16022", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["Hamilton-Jacobi reachability", "safety Bellman equation", "humanoid"], "model_class": "other", "techniques": ["reachability analysis", "learned safety value function"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Scalable Hamilton–Jacobi-style safety analysis via a geometric-horizon safety Bellman equation for high-dimensional humanoids.", "relevance": 3, "relevance_reason": "Scalable safety certificates for real-time monitoring.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2605.23128", "title": "π0-EqM: Equilibrium Matching for Closed-Loop Vision-Language-Action Control", "authors": ["Huan Liu", "Cong-Sheng Xu", "Jianmin Ji", "Yao Mu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.23128", "pdf_url": "https://arxiv.org/pdf/2605.23128", "code_url": "", "clusters": ["C1"], "keywords": ["adaptive inference depth", "state-dependent compute"], "model_class": "VLA", "techniques": ["equilibrium matching decoder"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces pi0's flow-matching expert with an equilibrium-matching decoder allowing state-dependent compute and temporal reuse across control cycles.", "relevance": 3, "relevance_reason": "Inference-depth as a design knob for VLA action heads.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2608.29749"}
{"id": "2604.15483", "title": "π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities", "authors": ["Physical Intelligence"], "year": 2026, "venue": "arXiv / tech report", "type": "tech-report", "url": "https://arxiv.org/abs/2604.15483", "pdf_url": "https://arxiv.org/pdf/2604.15483", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["pi0.7", "context conditioning", "steerable VLA", "cross-embodiment"], "model_class": "VLA", "techniques": ["diverse context conditioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Physical Intelligence's pi0.7 trains with diverse context conditioning (e.g., prompts, metadata) to become steerable, with zero-shot cross-embodiment and strong out-of-box performance.", "relevance": 3, "relevance_reason": "Frontier VLA baseline; organizer Zhuoyang Zhang listed as co-author on alphaXiv.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Zhuoyang Zhang"}
{"id": "2602.09021", "title": "χ0: Resource-Aware Robust Manipulation via Taming Distributional Inconsistencies", "authors": ["Checheng Yu", "Chong-Hao Sima", "Gangcheng Jiang", "Hai Zhang", "Haoguang Mai", "Hong-Yang Li", "Huijie Wang", "Jin Chen", "Kaiyang Wu", "Li Chen", "Lirui Zhao", "Mo-Di Shi", "Ping Luo", "Qingwen Bu", "Shi-Jia Peng", "Tianyu Li", "Yi-Bo Yuan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.09021", "pdf_url": "https://arxiv.org/pdf/2602.09021", "code_url": "", "clusters": ["C3"], "keywords": ["resource-efficient training", "long-horizon garment manipulation"], "model_class": "VLA", "techniques": ["model arithmetic (weight merging)", "stage advantage", "train-deploy alignment"], "base_model": "", "params": "", "hardware": "8x A100 (training)", "reported_metric": "~20 h data, 8 A100 GPUs for training", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Resource-aware recipe (weight-space merging, stage-aware advantage, train/deploy alignment) giving robust long-horizon dual-arm garment manipulation from modest data and compute.", "relevance": 3, "relevance_reason": "Framed around resource efficiency: 20 h of data and 8 A100s to beat pi0.5.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2607.00666, 2607.27782"}
{"id": "2606.09457", "title": "ω-EVA: Envision, Verify, and Act with Latent Interactive World Models", "authors": ["Zhenguo Sun", "Yu Sun", "Hande Huang", "Alois Knoll"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09457", "pdf_url": "https://arxiv.org/pdf/2606.09457", "code_url": "", "clusters": ["C8"], "keywords": ["latent world model", "envision-verify-act"], "model_class": "world-action model", "techniques": ["latent world model", "verify-then-act"], "base_model": "", "params": "~1.2B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "omega-EVA uses a latent action-conditioned world model to imagine the consequence of a proposed action chunk and refine it, avoiding future-video generation at inference with a ~1.2B-parameter model.", "relevance": 3, "relevance_reason": "relevant: latent (not pixel) world-model feedback keeps inference compact", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2507.09144", "title": "$I^{2}$-World: Intra-Inter Tokenization for Efficient Dynamic 4D Scene Forecasting", "authors": ["Zhimin Liao", "Ping Wei", "Ruijie Zhang", "Shuaijia Chen", "Haoxuan Wang", "Ziyang Ren"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.09144", "pdf_url": "https://arxiv.org/pdf/2507.09144", "code_url": "", "clusters": ["C8"], "keywords": ["4D occupancy world model", "tokenization", "efficiency"], "model_class": "world model", "techniques": ["intra/inter-scene tokenizers", "multi-scale residual quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Occ3D-nuScenes"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Efficient occupancy world model that compresses 3D scenes with separate spatial and temporal tokenizers for fast 4D forecasting in driving.", "relevance": 3, "relevance_reason": "Token-efficient 3D world model.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2511.05936", "title": "10 Open Challenges Steering the Future of Vision-Language-Action Models", "authors": ["Soujanya Poria", "Navonil Majumder", "Chia-Yu Hung", "Amir Ali Bagherzadeh", "Chuan Li", "Kenneth Kwok", "Ziwei Wang", "Cheston Tan", "Jiajun Wu", "David Hsu"], "year": 2025, "venue": "AAAI 2026 (Senior Track)", "type": "paper", "url": "https://arxiv.org/abs/2511.05936", "pdf_url": "https://arxiv.org/pdf/2511.05936", "code_url": "", "clusters": ["C1", "C4", "C8", "C9"], "keywords": ["survey", "open challenges", "evaluation", "efficiency", "VLA challenges", "safety"], "model_class": "VLA", "techniques": ["position paper"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Position paper listing ten milestones for VLAs, including efficiency, safety and whole-body coordination, and trends toward reaching them.", "relevance": 3, "relevance_reason": "Frames efficiency/safety as open VLA challenges; co-authored by Jiajun Wu.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA acceleration evaluation; author sweep: Jiajun Wu"}
{"id": "2508.11002", "title": "3D FlowMatch Actor: Unified 3D Policy for Single- and Dual-Arm Manipulation", "authors": ["Nikolaos Gkanatsios", "Jia-He Xu", "Matthew Bronars", "A. Mousavian", "Tsung-Wei Ke", "Katerina Fragkiadaki"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.11002", "pdf_url": "https://arxiv.org/pdf/2508.11002", "code_url": "", "clusters": ["C1"], "keywords": ["3D policy", "bimanual"], "model_class": "diffusion/flow policy", "techniques": ["flow matching", "system-level optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "over 30x faster training and inference than previous 3D diffusion policies", "metric_measures": "relative latency", "benchmarks": ["PerAct2", "RLBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "3D flow-matching policy with 3D relative attention and system optimizations, much faster than prior 3D diffusion policies.", "relevance": 3, "relevance_reason": "Large training/inference speedups", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2510.16732", "title": "A Comprehensive Survey on World Models for Embodied AI", "authors": ["Xinqing Li", "Xin He", "Le Zhang", "Min Wu", "Xiaoli Li", "Yun Liu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.16732", "pdf_url": "https://arxiv.org/pdf/2510.16732", "code_url": "https://github.com/Li-Zn-H/AwesomeWorldModels", "clusters": ["C8"], "keywords": ["world models", "embodied AI"], "model_class": "world model", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of world models for embodied AI with a three-axis taxonomy, datasets/metrics and open challenges including the performance vs real-time compute trade-off.", "relevance": 3, "relevance_reason": "Broad C8 reference noting real-time efficiency gap.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "10.1109/ecti-con64996.2025.11100596", "title": "A Study on Power Consumption of a Humanoid Welcome Robot Using Battery Usage Data", "authors": ["Sejin Park", "Donghun Noh", "Jeakweon Han"], "year": 2025, "venue": "ECTI-CON 2025", "type": "paper", "url": "https://doi.org/10.1109/ecti-con64996.2025.11100596", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["humanoid", "battery", "component power breakdown", "computing load"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Measures component-level power of a battery humanoid, identifying actuation and high-performance computing as major loads, and proposes a power management strategy.", "relevance": 3, "relevance_reason": "One of few measured humanoid power breakdowns including compute.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q2: OpenAlex title_and_abstract.search: humanoid robot battery power compute"}
{"id": "2503.01238", "title": "A Taxonomy for Evaluating Generalist Robot Manipulation Policies", "authors": ["Jensen Gao", "Suneel Belkhale", "S. Dasari", "Ashwin Balakrishna", "Dhruv Shah", "D. Sadigh"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2503.01238", "pdf_url": "https://arxiv.org/pdf/2503.01238", "code_url": "", "clusters": ["C4"], "keywords": ["generalization taxonomy", "evaluation protocol"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Bridge V2", "ALOHA 2"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Proposes a fine-grained taxonomy (STAR-Gen) of generalization axes with reproducible guidelines for evaluating generalist manipulation policies, with real-world case studies.", "relevance": 3, "relevance_reason": "Benchmarking methodology for generalist robot policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2503.24278, 2506.18123; cited by 2607.00666, 2607.14439"}
{"id": "2509.15937", "title": "A Vision-Language-Action-Critic Model for Robotic Real-World Reinforcement Learning", "authors": ["Shaopeng Zhai", "Qi Zhang", "Tianyi Zhang", "Fuxian Huang", "Haoran Zhang", "Ming Zhou", "Shengzhe Zhang", "Litao Liu", "Sixu Lin", "Jiangmiao Pang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.15937", "pdf_url": "https://arxiv.org/pdf/2509.15937", "code_url": "", "clusters": ["C3"], "keywords": ["VLAC", "process reward model", "real-world RL"], "model_class": "VLA", "techniques": ["reward model", "asynchronous real-world RL", "human-in-the-loop"], "base_model": "InternVL", "params": "", "hardware": "", "reported_metric": "~30% to ~90% success within 200 real episodes", "metric_measures": "other", "benchmarks": ["4 real manipulation tasks"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Unified VLA-critic model giving dense progress rewards for sample-efficient real-world RL with graded human intervention.", "relevance": 3, "relevance_reason": "Sample-efficient real-world post-training.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv"}
{"id": "2502.20168", "title": "Accelerating Model-Based Reinforcement Learning with State-Space World Models", "authors": ["Maria Krinner", "Elie Aljalbout", "Angel Romero", "Davide Scaramuzza"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.20168", "pdf_url": "https://arxiv.org/pdf/2502.20168", "code_url": "", "clusters": ["C8"], "keywords": ["state-space world model", "Mamba-style", "MBRL training speed"], "model_class": "world model", "techniques": ["parallelizable state-space model dynamics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Uses state-space models as the world model in model-based RL to parallelize dynamics training and substantially cut wall-clock training time, validated on real quadrotor control.", "relevance": 3, "relevance_reason": "Faster world-model training for robot RL.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2507.09061", "title": "Action Chunking and Exploratory Data Collection Yield Exponential Improvements in Behavior Cloning for Continuous Control", "authors": ["Thomas T. Zhang", "Daniel Pfrommer", "Chaoyi Pan", "Nikolai Matni", "Max Simchowitz"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.09061", "pdf_url": "https://arxiv.org/pdf/2507.09061", "code_url": "", "clusters": ["C1"], "keywords": ["action chunking theory", "compounding errors"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Theory showing action chunking and exploratory data collection avoid exponential compounding errors in imitation learning.", "relevance": 3, "relevance_reason": "Theoretical grounding for action chunking.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2512.01809, 2602.04215"}
{"id": "2510.02851", "title": "Action Deviation-Aware Inference for Low-Latency Wireless Robots", "authors": ["Jeyoung Park", "Yeonsub Lim", "Seungeun Oh", "Jihong Park", "Jinho Choi", "Seong-Lyun Kim"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.02851", "pdf_url": "https://arxiv.org/pdf/2510.02851", "code_url": "", "clusters": ["C1", "C3", "C6", "C9"], "keywords": ["device-edge speculative inference", "wireless robots", "speculative decoding", "action deviation"], "model_class": "other", "techniques": ["speculative decoding", "edge-cloud collaboration", "deviation-aware speculative inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "39.2% lower end-to-end latency; ~40% fewer server calls", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "On-device draft policy sends actions for server verification only when predicted deviation is large, cutting communication and latency for networked robots.", "relevance": 3, "relevance_reason": "Speculative inference across device/edge.", "foundational_technique": false, "verified": true, "found_via": "search query: speculative decoding VLA; venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2503.18938", "title": "AdaWorld: Learning Adaptable World Models with Latent Actions", "authors": ["Shenyuan Gao", "Siyuan Zhou", "Yilun Du", "Jun Zhang", "Chuang Gan"], "year": 2025, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2503.18938", "pdf_url": "https://arxiv.org/pdf/2503.18938", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["latent action", "adaptable world model"], "model_class": "world model", "techniques": ["latent actions", "efficient adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "AdaWorld pretrains world models with self-supervised latent actions from video so they can be adapted to new environments and action spaces with few interactions and low cost.", "relevance": 3, "relevance_reason": "relevant: efficient adaptation of world models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2306.03310; cited by 2602.06949, 2606.00267"}
{"id": "2511.00088", "title": "Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail", "authors": ["Nvidia Yan Wang", "Wen-Jie Luo", "Junjie Bai", "Yulong Cao", "Tong Che", "Ke Chen", "Yuxiao Chen", "Jenna Diamond", "Yifan Ding", "Wenhao Ding", "Liang Feng", "Greg Heinrich", "Jack Huang", "Peter Karkus", "Boyi Li", "Pin-Zu Li", "Tsung-Yi Lin", "Dong-Run Liu", "Ming-Yu Liu", "Langechuan Liu", "Zhijian Liu", "Jason Lu", "Yu Mao", "Pavlo Molchanov", "Lindsey Pavao", "Zheng-Hao Peng", "Michael Ranzinger", "E. Schmerling", "Shida Shen", "Yun-Fei Shi", "Sarah Tariq", "Ran Tian", "Tilman Wekel", "Xinshuo Weng", "Tianjun Xiao", "Eric Yang", "Xiao-Dong Yang", "Yurong You", "Xiaohui Zeng", "Wenyuan Zhang", "B. Ivanovic", "Marco Pavone"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.00088", "pdf_url": "https://arxiv.org/pdf/2511.00088", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["autonomous driving", "reasoning VLA", "real-time"], "model_class": "VLA", "techniques": ["diffusion trajectory decoder"], "base_model": "", "params": "0.5B-7B", "hardware": "", "reported_metric": "99 ms latency (on-vehicle road tests)", "metric_measures": "other", "benchmarks": ["closed-loop simulation", "on-vehicle"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Driving VLA combining causal reasoning traces with a diffusion trajectory decoder, validated on-vehicle in real time.", "relevance": 3, "relevance_reason": "Reports real-time on-vehicle latency for a reasoning VLA", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2504.16054, 2505.23705"}
{"id": "2509.21006", "title": "AnywhereVLA: Language-Conditioned Exploration and Mobile Manipulation", "authors": ["Konstantin Gubernatorov", "Artem Voronov", "Roman Voronov", "Sergei Pasynkov", "Stepan Perminov", "Ziang Guo", "Dzmitry Tsetserukou"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.21006", "pdf_url": "https://arxiv.org/pdf/2509.21006", "code_url": "", "clusters": ["C5"], "keywords": ["mobile manipulation", "Jetson Orin NX", "Intel NUC", "SmolVLA", "onboard"], "model_class": "VLA", "techniques": ["heterogeneous onboard compute split", "modular classical+VLA stack"], "base_model": "SmolVLA", "params": "", "hardware": "Jetson Orin NX (perception + VLA) + Intel NUC (SLAM, control)", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot (SO-101 mobile)"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Mobile manipulation system that runs a fine-tuned SmolVLA and perception on a Jetson Orin NX and classical SLAM/exploration on an Intel NUC, fully onboard.", "relevance": 3, "relevance_reason": "Concrete heterogeneous onboard compute partition with a VLA.", "foundational_technique": false, "verified": true, "found_via": "arXiv: Jetson AND (robot policy OR manipulation OR VLA)"}
{"id": "2510.02298", "title": "ARMADA: Autonomous Online Failure Detection and Human Shared Control Empower Scalable Real-World Deployment and Adaptation", "authors": ["Wenye Yu", "Jun Lv", "Zixi Ying", "Yang Jin", "Chuan Wen", "Ce-Wu Lu"], "year": 2025, "venue": "IEEE RA-L 2025", "type": "paper", "url": "https://arxiv.org/abs/2510.02298", "pdf_url": "https://arxiv.org/pdf/2510.02298", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["ARMADA", "online failure detection", "shared control", "multi-robot deployment"], "model_class": "other", "techniques": ["autonomous failure detection", "human-in-the-loop"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Multi-robot deployment system where an online failure detector requests human shared control only when needed, enabling scalable data collection and adaptation.", "relevance": 3, "relevance_reason": "Failure detection triggering fallback at deployment.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2503.08558 (S2)"}
{"id": "2511.14148", "title": "AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models", "authors": ["Yuhua Jiang", "Shuang Cheng", "Yan Ding", "Feifei Gao", "Biqing Qi"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.14148", "pdf_url": "https://arxiv.org/pdf/2511.14148", "code_url": "https://github.com/YuhuaJiang2002/AsyncVLA", "clusters": ["C1"], "keywords": ["asynchronous flow matching", "self-correction"], "model_class": "VLA", "techniques": ["non-uniform flow schedule", "selective refinement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Generates action tokens on non-uniform timestep schedules and refines only low-confidence tokens, improving long-horizon stability.", "relevance": 3, "relevance_reason": "Flow VLA generation scheme; accuracy-focused.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "2506.09350", "title": "Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation", "authors": ["Shanchuan Lin", "Ceyuan Yang", "Hao He", "Jianwen Jiang", "Yu-Xi Ren", "X. Xia", "Yang Zhao", "Xuefeng Xiao", "Lu Jiang"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2506.09350", "pdf_url": "https://arxiv.org/pdf/2506.09350", "code_url": "", "clusters": ["C8"], "keywords": ["real-time video generation"], "model_class": "world model", "techniques": ["adversarial post-training", "one-step generation", "KV cache"], "base_model": "", "params": "8B", "hardware": "H100", "reported_metric": "24 fps streaming at 736x416 on a single H100", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Turns a pretrained video diffusion model into a one-step autoregressive real-time interactive video generator.", "relevance": 3, "relevance_reason": "Foundational real-time video/world generation technique", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2402.15391"}
{"id": "2506.13757", "title": "AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning", "authors": ["Zewei Zhou", "Tian-Hui Cai", "Seth Z. Zhao", "Yun Zhang", "Zhiyu Huang", "Bolei Zhou", "Jia-Qi Ma"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2506.13757", "pdf_url": "https://arxiv.org/pdf/2506.13757", "code_url": "", "clusters": ["C1"], "keywords": ["autonomous driving VLA", "reasoning efficiency"], "model_class": "VLA", "techniques": ["adaptive fast/slow reasoning", "GRPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["nuPlan", "nuScenes", "Waymo", "CARLA"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA with fast (trajectory-only) and slow (CoT) modes, using RL fine-tuning to skip unnecessary reasoning.", "relevance": 3, "relevance_reason": "Adaptive reasoning to cut inference cost in an embodied VLA.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2510.24795, 2602.02459"}
{"id": "2511.18151", "title": "AVERY: Intent-Driven Adaptive VLM Split Computing via Embodied Self-Awareness for Efficient Disaster Response Systems", "authors": ["Rajat Bhattacharjya", "Sing-Yao Wu", "Hyunwoo Oh", "Chaewon Nam", "Suyeon Koo", "Mohsen Imani", "Elaheh Bozorgzadeh", "Nikil Dutt"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.18151", "pdf_url": "https://arxiv.org/pdf/2511.18151", "code_url": "", "clusters": ["C6"], "keywords": ["UAV", "VLM split computing", "intent-driven", "low-bandwidth"], "model_class": "other", "techniques": ["functional dual-stream split", "adaptive depth-wise split"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Splits VLM work into a fast low-resolution context stream and slow high-fidelity insight stream, partitioning the latter across edge and cloud based on operator intent and network state.", "relevance": 3, "relevance_reason": "Adaptive split computing for embodied VLM under unstable links.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2509.04063", "title": "Balancing Signal and Variance: Adaptive Offline RL Post-Training for VLA Flow Models", "authors": ["Hongyin Zhang", "Shiyuan Zhang", "Junxi Jin", "Qixin Zeng", "Yifan Qiao", "Hongchao Lu", "Donglin Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.04063", "pdf_url": "https://arxiv.org/pdf/2509.04063", "code_url": "", "clusters": ["C3"], "keywords": ["ARFM", "offline RL", "flow matching"], "model_class": "VLA", "techniques": ["offline RL post-training", "adaptive loss scaling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Offline RL post-training for flow VLAs via adaptively scaled flow-matching loss balancing signal and variance.", "relevance": 3, "relevance_reason": "Cheap offline RL post-training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"continual learning\""}
{"id": "2503.12533", "title": "Being-0: A Humanoid Robotic Agent with Vision-Language Models and Modular Skills", "authors": ["Haoqi Yuan", "Yu Bai", "Yuhui Fu", "Bohan Zhou", "Yicheng Feng", "Xinrun Xu", "Yi Zhan", "Börje F. Karlsson", "Zongqing Lu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.12533", "pdf_url": "https://arxiv.org/pdf/2503.12533", "code_url": "", "clusters": ["C5"], "keywords": ["humanoid agent", "cloud FM + onboard VLM connector", "low-cost onboard compute"], "model_class": "other", "techniques": ["hierarchical cloud/onboard split", "lightweight onboard VLM connector"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Humanoid agent that keeps a large foundation model offboard while running a lightweight VLM connector and skill library on low-cost onboard devices for real-time operation.", "relevance": 3, "relevance_reason": "Explicit onboard/offboard compute partition on a full-size humanoid.", "foundational_technique": false, "verified": true, "found_via": "arXiv: onboard AND humanoid AND (FM OR VLA) AND (GPU OR Jetson OR compute)"}
{"id": "2505.05540", "title": "Benchmarking Vision, Language, & Action Models in Procedurally Generated, Open Ended Action Environments", "authors": ["Pranav Guruprasad", "Yang-Yue Wang", "Sudipta Chowdhury", "Harshvardhan Digvijay Sikka", "Paul Pu Liang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2505.05540", "pdf_url": "https://arxiv.org/pdf/2505.05540", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "zero-shot evaluation", "procedural environments"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Evaluates VLA and VLM models zero-shot on procedurally generated open-ended action environments to systematically characterize generalization.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2306.03310"}
{"id": "2502.11880", "title": "Bitnet.cpp: Efficient Edge Inference for Ternary LLMs", "authors": ["Jinheng Wang", "Hansong Zhou", "Ting Song", "Shijie Cao", "Yan Xia", "Ting Cao", "Jianyu Wei", "Shu-Ming Ma", "Hongyu Wang", "F. Wei"], "year": 2025, "venue": "ACL 2025", "type": "paper", "url": "https://arxiv.org/abs/2502.11880", "pdf_url": "https://arxiv.org/pdf/2502.11880", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["ternary inference", "mpGEMM", "edge CPU"], "model_class": "other", "techniques": ["ternary kernels"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 6.25x speed over full-precision baselines", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Inference library with ternary lookup-table and int2 kernels for fast lossless BitNet b1.58 inference on edge CPUs.", "relevance": 3, "relevance_reason": "Runtime enabling 1-bit VLAs on edge.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2506.07530 (references)"}
{"id": "2512.11769", "title": "BLURR: A Boosted Low-Resource Inference for Vision-Language-Action Models", "authors": ["Xiaoyu Ma", "Zhengqing Yuan", "Zheyuan Zhang", "Kaiwen Shi", "Lichao Sun", "Yanfang Ye", "Xiao-Yu Ma", "Zheng-Qing Yuan", "Zhe-Yuan Zhang", "Kai-Wen Shi", "Li-Chao Sun", "Yan-Fang Ye"], "year": 2025, "venue": "The Web Conference 2026 (demo)", "type": "preprint", "url": "https://arxiv.org/abs/2512.11769", "pdf_url": "https://arxiv.org/pdf/2512.11769", "code_url": "", "clusters": ["C1", "C2", "C5"], "keywords": ["inference wrapper", "prefix KV cache", "mixed precision", "single-step rollout", "KV cache", "commodity hardware", "pi0"], "model_class": "VLA", "techniques": ["KV caching", "mixed precision", "few-step sampling", "instruction-prefix KV cache", "mixed-precision execution", "single-step rollout"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SIMPLER", "SimplerEnv"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free wrapper for pi0 combining prefix KV caching, mixed precision and single-step rollout to make VLA control responsive on commodity hardware.", "relevance": 3, "relevance_reason": "Light systems contribution (demo paper).", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference; web: TensorRT-LLM OpenVLA acceleration kernel fusion"}
{"id": "2509.09560", "title": "Boosting Embodied AI Agents through Perception-Generation Disaggregation and Asynchronous Pipeline Execution", "authors": ["Shulai Zhang", "Ao Xu", "Quan Chen", "Han Zhao", "Weihao Cui", "Ningxin Zheng", "Haibin Lin", "Xin Liu", "Minyi Guo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.09560", "pdf_url": "https://arxiv.org/pdf/2509.09560", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["perception-generation disaggregation", "pipeline parallelism", "inference frequency"], "model_class": "other", "techniques": ["disaggregated pipelined execution", "shared public context"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.54x average throughput with 102.7% of original accuracy", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Auras splits perception and generation of embodied agents into a controlled parallel pipeline with a shared context to raise 'thinking' frequency.", "relevance": 3, "relevance_reason": "System co-design for embodied inference rate.", "foundational_technique": false, "verified": true, "found_via": "arXiv: embodied AND (accelerator OR co-design)"}
{"id": "2512.19083", "title": "CoDrone: Autonomous Drone Navigation Assisted by Edge and Cloud Foundation Models", "authors": ["Pengyu Chen", "Tao Ouyang", "Ke Luo", "Weijie Hong", "Xu Chen"], "year": 2025, "venue": "IEEE Internet of Things Journal", "type": "paper", "url": "https://arxiv.org/abs/2512.19083", "pdf_url": "https://arxiv.org/pdf/2512.19083", "code_url": "", "clusters": ["C6"], "keywords": ["UAV", "cloud-edge-end", "foundation models", "DRL scheduler"], "model_class": "other", "techniques": ["cloud-edge-end collaboration", "DRL offload scheduling"], "base_model": "Depth Anything V2 + VLM", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Drone navigation framework that calls edge-hosted depth foundation models and cloud VLM interaction only when needed, with a DRL scheduler deciding offloading.", "relevance": 3, "relevance_reason": "Cloud-edge-end offloading of foundation models for aerial robots.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2505.11123", "title": "Conditioning Matters: Training Diffusion Policies is Faster Than You Think", "authors": ["Zibin Dong", "Yicheng Liu", "Yin-Chuan Li", "Hang Zhao", "Jian-Ye Hao"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2505.11123", "pdf_url": "https://arxiv.org/pdf/2505.11123", "code_url": "", "clusters": ["C3"], "keywords": ["training efficiency", "loss collapse"], "model_class": "diffusion/flow policy", "techniques": ["condition-dependent source distribution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cocos anchors flow-matching source noise on condition semantics, preventing loss collapse and speeding diffusion-policy training.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2507.17383", "title": "Confidence Calibration in Vision-Language-Action Models", "authors": ["Thomas P. Zollo", "Richard Zemel"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2507.17383", "pdf_url": "https://arxiv.org/pdf/2507.17383", "code_url": "", "clusters": ["C7"], "keywords": ["Platt scaling", "prompt ensembles"], "model_class": "VLA", "techniques": ["confidence calibration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Studies VLA confidence calibration and proposes lightweight fixes to predict success reliably.", "relevance": 3, "relevance_reason": "Runtime reliability monitoring.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310; cited by 2604.23775"}
{"id": "2506.12544", "title": "Constrained Diffusers for Safe Planning and Control", "authors": ["Jichen Zhang", "Liqun Zhao", "Antonis Papachristodoulou", "Jack Umenberger"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.12544", "pdf_url": "https://arxiv.org/pdf/2506.12544", "code_url": "", "clusters": ["C7"], "keywords": ["constrained diffusion", "safe planning"], "model_class": "diffusion/flow policy", "techniques": ["constrained sampling", "discrete CBF"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds constraints to pretrained diffusion models via constrained Langevin sampling (projection, primal-dual, augmented Lagrangian) with discrete CBFs for online safety.", "relevance": 3, "relevance_reason": "Training-free constraint enforcement on sampled trajectories.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\")"}
{"id": "2511.18085", "title": "Continually Evolving Skill Knowledge in Vision Language Action Model", "authors": ["Yuxuan Wu", "Guangming Wang", "Zhiheng Yang", "Tianchen Deng", "Maoqing Yao", "Brian Sheil", "Hesheng Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.18085", "pdf_url": "https://arxiv.org/pdf/2511.18085", "code_url": "", "clusters": ["C3"], "keywords": ["Stellar VLA", "knowledge space", "expert routing"], "model_class": "VLA", "techniques": ["expert routing", "continual imitation learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Continual imitation learning for VLAs via a learned knowledge space and routing, without adding parameters.", "relevance": 3, "relevance_reason": "Parameter-constant continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"continual learning\""}
{"id": "2509.23823", "title": "Control Your Robot: A Unified System for Robot Control and Policy Deployment", "authors": ["Nian Tian", "Wei Ke", "Yao Mu", "Tian-Xing Chen", "Shao-Long Zhu", "Bingshan Hu"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.23823", "pdf_url": "https://arxiv.org/pdf/2509.23823", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["deployment framework", "low-latency data collection"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Control Your Robot is a modular framework unifying data collection and low-latency policy deployment across robot platforms.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2406.01586"}
{"id": "2506.16211", "title": "ControlVLA: Few-shot Object-centric Adaptation for Pre-trained Vision-Language-Action Models", "authors": ["Pu-Hao Li", "Ying-Ying Wu", "Ziheng Xi", "Wan-Lin Li", "Yuzhe Huang", "Zhi-Yuan Zhang", "Ying-Han Chen", "Jia-Nan Wang", "Song-Chun Zhu", "Teng-Yu Liu", "Si-Yuan Huang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.16211", "pdf_url": "https://arxiv.org/pdf/2506.16211", "code_url": "", "clusters": ["C3"], "keywords": ["few-shot", "fine-tuning"], "model_class": "VLA", "techniques": ["ControlNet-style adaptation", "zero-init projections"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "ControlVLA adds object-centric conditioning to a pretrained VLA through zero-initialized ControlNet-style layers, enabling few-shot (10-20 demos) real-world task adaptation.", "relevance": 3, "relevance_reason": "relevant: efficient few-shot adaptation", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2510.24795, 2511.22697"}
{"id": "2510.16240", "title": "Cosmos-Surg-DVRK: World Foundation Model-Based Automated Online Evaluation of Surgical Robot Policy Learning", "authors": ["L. Zbinden", "Nigel Nelson", "Juo-Tung Chen", "Xinhao Chen", "Ji Woong Kim", "Mahdi Azizian", "Axel Krieger", "Sean Huver"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2510.16240", "pdf_url": "https://arxiv.org/pdf/2510.16240", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["Cosmos", "surgical robotics"], "model_class": "world model", "techniques": ["world-model evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fine-tunes the Cosmos world foundation model for surgical tasks to evaluate policies automatically in simulation, correlating with real dVRK outcomes.", "relevance": 3, "relevance_reason": "Cheap world-model-based policy evaluation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941; cited by 2602.06949"}
{"id": "2503.22020", "title": "CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models", "authors": ["Qingqing Zhao", "Yao Lu", "Moo Jin Kim", "Zipeng Fu", "Zhuoyang Zhang", "Yecheng Wu", "Zhaoshuo Li", "Qianli Ma"], "year": 2025, "venue": "CVPR 2025", "type": "paper", "url": "https://arxiv.org/abs/2503.22020", "pdf_url": "https://arxiv.org/pdf/2503.22020", "code_url": "", "clusters": ["C1", "C7", "C9"], "keywords": ["visual chain-of-thought", "subgoal image prediction", "VLA"], "model_class": "VLA", "techniques": ["autoregressive image+action generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VLA that first autoregressively predicts a future subgoal image and then actions, adding explicit visual reasoning; improves over prior VLAs by ~17% in real-world tasks.", "relevance": 3, "relevance_reason": "Reasoning-heavy VLA whose latency cost motivates ForeAct/efficiency work.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Zhuoyang Zhang"}
{"id": "2509.06819", "title": "CRISP—Compliant ROS2 Controllers for Learning-Based Manipulation Policies and Teleoperation", "authors": ["Daniel San José Pro", "Oliver Hausdörfer", "Ralf Römer", "Maximilian Dösch", "Martin Schuck", "Angela P. Schoellig"], "year": 2025, "venue": "IEEE Robotics and Automation Practice", "type": "paper", "url": "https://arxiv.org/abs/2509.06819", "pdf_url": "https://doi.org/10.1109/rap.2026.3690413", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["ROS2", "controllers", "policy deployment"], "model_class": "other", "techniques": ["compliant control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lightweight C++ ROS2 compliant controllers that smoothly track low-frequency learned policy outputs in real time.", "relevance": 3, "relevance_reason": "Deployment infrastructure bridging slow policies and real-time control", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2506.19816", "title": "CronusVLA: Towards Efficient and Robust Manipulation via Multi-Frame Vision-Language-Action Modeling", "authors": ["Hao Li", "Shuai Yang", "Yilun Chen", "Xinyi Chen", "Xiaoda Yang", "Yang Tian", "Hanqing Wang", "Tai Wang", "Dahua Lin", "Feng Zhao", "Jiangmiao Pang", "Yi-Lun Chen", "Han-Qing Wang", "Da-Hua Lin", "Jiang-Miao Pang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.19816", "pdf_url": "https://arxiv.org/pdf/2506.19816", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["CronusVLA", "SimplerEnv-OR"], "model_class": "VLA", "techniques": ["feature chunking", "multi-frame", "multi-frame feature chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SimplerEnv", "SimplerEnv-OR", "real robot", "LIBERO"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Extends single-frame VLAs to multi-frame inputs by aggregating cached features instead of feeding frames through the VLM, and adds a robustness benchmark.", "relevance": 3, "relevance_reason": "Efficient multi-frame modeling avoiding VLM overhead.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]; snowball round 2: cites 2304.13705; cited by 2604.24622"}
{"id": "2510.10125", "title": "Ctrl-World: A Controllable Generative World Model for Robot Manipulation", "authors": ["Yanjiang Guo", "L. Shi", "Jianyu Chen", "Chelsea Finn"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.10125", "pdf_url": "https://arxiv.org/pdf/2510.10125", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["controllable world model", "policy evaluation in imagination", "multi-view video"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DROID"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Multi-view action-conditioned world model trained on DROID that lets generalist policies be rolled out and ranked in imagination, and used to synthesize fine-tuning data.", "relevance": 3, "relevance_reason": "World model as a cheaper evaluation/improvement substitute for real rollouts (benchmarking methodology).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2405.05941"}
{"id": "2507.02190", "title": "cVLA: Towards Efficient Camera-Space VLAs", "authors": ["Max Argus", "Jelena Bratulić", "Houman Masnavi", "M. Velikanov", "Nick Heppert", "Abhinav Valada", "Thomas Brox"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.02190", "pdf_url": "https://arxiv.org/pdf/2507.02190", "code_url": "", "clusters": ["C1"], "keywords": ["camera-space", "lightweight VLA"], "model_class": "VLA", "techniques": ["camera-space waypoint prediction", "next-token prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "cVLA predicts end-effector trajectory waypoints in image coordinates with a lightweight VLM-based design, making it cheaper to train and embodiment-agnostic, with sim-to-real transfer.", "relevance": 3, "relevance_reason": "relevant: efficient-to-train lightweight VLA", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941, 2405.12213; cited by 2510.24795"}
{"id": "2506.05064", "title": "DemoSpeedup: Accelerating Visuomotor Policies via Entropy-Guided Demonstration Acceleration", "authors": ["Lingxiao Guo", "Zhengrong Xue", "Zijing Xu", "Huazhe Xu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.05064", "pdf_url": "https://arxiv.org/pdf/2506.05064", "code_url": "", "clusters": ["C1"], "keywords": ["execution speed", "demonstration speedup"], "model_class": "diffusion/flow policy", "techniques": ["demonstration acceleration", "entropy-guided downsampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 3x faster task execution", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DemoSpeedup downsamples demonstrations more aggressively in high-entropy (casual) segments so trained policies such as ACT or Diffusion Policy execute tasks up to 3x faster.", "relevance": 3, "relevance_reason": "relevant: faster real-time execution of visuomotor policies", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2512.01031, 2602.18813"}
{"id": "2503.13217", "title": "Dense Policy: Bidirectional Autoregressive Learning of Actions", "authors": ["Yue Su", "Xinyu Zhan", "Hong-Jie Fang", "Han Xue", "Haoshu Fang", "Yong-Lu Li", "Cewu Lu", "Lixin Yang"], "year": 2025, "venue": "IEEE International Conference on Computer Vision", "type": "paper", "url": "https://arxiv.org/abs/2503.13217", "pdf_url": "https://arxiv.org/pdf/2503.13217", "code_url": "", "clusters": ["C1"], "keywords": ["Dense Policy"], "model_class": "other", "techniques": ["bidirectional autoregression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Encoder-only policy that expands actions coarse-to-fine with logarithmic-time inference.", "relevance": 3, "relevance_reason": "Lightweight fast-inference policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2505.03912"}
{"id": "2509.13574", "title": "Dense-Jump Flow Matching with Non-Uniform Time Scheduling for Robotic Policies: Mitigating Multi-Step Inference Degradation", "authors": ["Zidong Chen", "Zihao Guo", "Peng Wang", "ThankGod Itua Egbe", "Yan Lyu", "Chenghao Qian"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.13574", "pdf_url": "https://arxiv.org/pdf/2509.13574", "code_url": "", "clusters": ["C1"], "keywords": ["non-uniform time schedule", "dense jump", "few-step"], "model_class": "diffusion/flow policy", "techniques": ["flow matching", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Finds more Euler steps hurt flow policies and proposes U-shaped training time sampling plus a single jump past a cutoff at inference.", "relevance": 3, "relevance_reason": "Few-step flow sampling insight.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching robot policy efficient"}
{"id": "github-openpi-issue-386-jetson-orin", "title": "Deploying openpi0 on jetson orin encountered an error (Physical-Intelligence/openpi issue #386)", "authors": ["zhaodaxiang", "GaohaoZhou-ops", "LordLKY"], "year": 2025, "venue": "GitHub issue", "type": "repo", "url": "https://github.com/Physical-Intelligence/openpi/issues/386", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["pi0", "pi0.5", "Jetson Orin", "unified memory", "OOM", "JAX"], "model_class": "VLA", "techniques": [], "base_model": "pi0 / pi0.5 (openpi)", "params": "~3B", "hardware": "Jetson AGX Orin 64GB; Orin NX 16GB; Orin Nano Super 8GB", "reported_metric": "community: pi0.5 ~1 s per inference on AGX Orin 64GB (PyTorch); OOM on Orin NX 16GB and Orin Nano 8GB; JAX preallocation causes swapping on unified memory", "metric_measures": "other", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Community thread on running openpi on Jetson: JAX memory preallocation stalls unified-memory Orins, pi0.5 reaches ~1 s/inference on 64 GB AGX Orin, and 16 GB Orin NX runs out of memory.", "relevance": 3, "relevance_reason": "First-hand (unverified-by-paper) evidence that 3B-class VLAs do not fit Orin NX 16GB without further compression.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub issues repo:Physical-Intelligence/openpi jetson"}
{"id": "2509.14758", "title": "Designing Latent Safety Filters using Pre-Trained Vision Models", "authors": ["Ihab Tabbara", "Yuxuan Yang", "Ahmad Hamzeh", "Maxwell Astafyev", "Hussein Sibai"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.14758", "pdf_url": "https://arxiv.org/pdf/2509.14758", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["latent safety filter", "pretrained vision models", "HJ reachability"], "model_class": "world model", "techniques": ["PVR backbones for failure classifiers and HJ filters", "PVR backbones for failure classifiers, HJ filters and latent world models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Studies how frozen, fine-tuned or scratch pretrained vision encoders work as backbones for failure classifiers, HJ safety filters and latent world models, with deployment considerations.", "relevance": 3, "relevance_reason": "Design study for world-model-based safety filters.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"hamilton-jacobi\" reachability learned policy safety filter; search query"}
{"id": "2512.08991", "title": "Deterministic World Models for Closed-loop Reachability Analysis of End-to-End Vision-based Control", "authors": ["Yuang Geng", "Zhongzheng Zhang", "Chengzhen Jiang", "Yanru Li", "Xinyang Wang", "Zhuoyang Zhou", "Hoang-Dung Tran", "Ivan Ruchkin"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.08991", "pdf_url": "https://arxiv.org/pdf/2512.08991", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["deterministic world model", "closed-loop reachability", "verification", "conformal"], "model_class": "world model", "techniques": ["latent-free state-to-image decoder", "reachability with conformal inflation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CARLA braking", "control benchmarks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Replaces the stochastic camera model with a deterministic state-to-image world model so closed-loop reachability of end-to-end vision controllers can be verified, with conformal bounds transferring guarantees to the real system.", "relevance": 3, "relevance_reason": "World models for formal safety verification.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2503.00653", "title": "Discrete Codebook World Models for Continuous Control", "authors": ["Aidan Scannell", "Mohammadreza Nakhaei", "Kalle Kujanpää", "Yi Zhao", "Kevin Sebastian Luck", "Arno Solin", "Joni Pajarinen"], "year": 2025, "venue": "ICLR 2025", "type": "paper", "url": "https://arxiv.org/abs/2503.00653", "pdf_url": "https://arxiv.org/pdf/2503.00653", "code_url": "", "clusters": ["C8"], "keywords": ["discrete codebook", "latent world model", "decision-time planning"], "model_class": "world model", "techniques": ["discrete stochastic codebook latents", "MPPI planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DMControl", "Meta-World", "MyoSuite"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows discrete codebook latent states beat continuous latents for continuous control and builds a TD-MPC2-style planner on them (DCWM).", "relevance": 3, "relevance_reason": "Latent representation choice for efficient planning world models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828"}
{"id": "2503.19757", "title": "Dita: Scaling Diffusion Transformer for Generalist Vision-Language-Action Policy", "authors": ["Zhi Hou", "Tianyi Zhang", "Yuwen Xiong", "Haonan Duan", "Hengjun Pu", "Ronglei Tong", "Chengyang Zhao", "Xizhou Zhu", "Yu Qiao", "Jifeng Dai", "Yuntao Chen"], "year": 2025, "venue": "ICCV 2025", "type": "paper", "url": "https://arxiv.org/abs/2503.19757", "pdf_url": "https://arxiv.org/pdf/2503.19757", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion transformer", "in-context conditioning"], "model_class": "VLA", "techniques": ["diffusion transformer action head"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Scalable diffusion-transformer VLA that denoises action sequences with in-context conditioning on raw visual tokens; positioned as a lightweight open baseline.", "relevance": 3, "relevance_reason": "Diffusion VLA architecture.", "foundational_technique": false, "verified": true, "found_via": "search query: lightweight VLA"}
{"id": "2503.02175", "title": "DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models", "authors": ["Saeed Ranjbar Alvar", "Gursimran Singh", "Mohammad Akbari", "Yong Zhang"], "year": 2025, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2503.02175", "pdf_url": "https://arxiv.org/pdf/2503.02175", "code_url": "", "clusters": ["C1"], "keywords": ["max-min diversity", "training-free"], "model_class": "other", "techniques": ["visual token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DivPrune prunes visual tokens in LMMs by maximizing diversity of the kept set, cutting latency and memory without fine-tuning.", "relevance": 3, "relevance_reason": "generic efficiency technique connected to 6 core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135"}
{"id": "2508.03645", "title": "DiWA: Diffusion Policy Adaptation with World Models", "authors": ["Akshay L Chandra", "Iman Nematollahi", "Chen Huang", "T. Welschehold", "Wolfram Burgard", "Abhinav Valada"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.03645", "pdf_url": "https://arxiv.org/pdf/2508.03645", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["diffusion policy fine-tuning", "sample efficiency"], "model_class": "diffusion/flow policy", "techniques": ["offline RL in world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Fine-tunes diffusion policies entirely offline with RL inside a learned world model, needing orders of magnitude fewer physical interactions.", "relevance": 3, "relevance_reason": "Efficient adaptation of diffusion policies via world models.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2606.16690"}
{"id": "2512.22615", "title": "Dream-VL & Dream-VLA: Open Vision-Language and Vision-Language-Action Models with Diffusion Language Model Backbone", "authors": ["Jiacheng Ye", "Shansan Gong", "Jiahui Gao", "Junming Fan", "Shuang Wu", "Wei Bi", "H. Bai", "Lifeng Shang", "Ling-Peng Kong"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.22615", "pdf_url": "https://arxiv.org/pdf/2512.22615", "code_url": "", "clusters": ["C1"], "keywords": ["dVLA", "action chunking"], "model_class": "VLA", "techniques": ["diffusion language model backbone", "parallel decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "SimplerEnv"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Builds a VLM and VLA on a diffusion-LLM backbone whose bidirectional decoding suits parallel action-chunk generation and speeds fine-tuning convergence.", "relevance": 3, "relevance_reason": "Alternative non-autoregressive backbone enabling parallel action generation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2604.25050, 2605.13382"}
{"id": "2505.12705", "title": "DreamGen: Unlocking Generalization in Robot Learning through Video World Models", "authors": ["Joel Jang", "Seonghyeon Ye", "Zongyu Lin", "Jiannan Xiang", "Johan Bjorck", "Yu Fang", "Fengyuan Hu", "Spencer Huang", "Kaushil Kundalia", "Yen-Chen Lin", "Loic Magne", "Ajay Mandlekar", "Avnish Narayan", "You Liang Tan", "Guanzhi Wang", "Jing Wang", "Qi Wang", "Yinzhen Xu", "Xiaohui Zeng", "Kaiyuan Zheng", "Ruijie Zheng", "Ming-Yu Liu", "Luke Zettlemoyer", "Dieter Fox", "Jan Kautz", "Scott Reed", "Yuke Zhu", "Linxi Fan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.12705", "pdf_url": "https://arxiv.org/pdf/2505.12705", "code_url": "", "clusters": ["C8"], "keywords": ["DreamGen", "neural trajectories", "synthetic robot data"], "model_class": "world model", "techniques": ["fine-tuned image-to-video world model", "IDM / latent-action pseudo-labels"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["humanoid real robot (22 new behaviors)"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "NVIDIA pipeline that generates synthetic robot videos with a fine-tuned video world model and labels them with pseudo-actions, letting a humanoid learn 22 new behaviors from one-task teleop data.", "relevance": 3, "relevance_reason": "Offline world-model use (data generation) that moves compute off the robot.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2509.13774", "title": "Dual-Actor Fine-Tuning of VLA Models: A Talk-and-Tweak Human-in-the-Loop Approach", "authors": ["Piao-Piao Jin", "Qi Wang", "Guo-Kang Sun", "Ziwen Cai", "Pinjia He", "Yang-Wei You"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.13774", "pdf_url": "https://arxiv.org/pdf/2509.13774", "code_url": "", "clusters": ["C3"], "keywords": ["RL fine-tuning", "dual actor"], "model_class": "VLA", "techniques": ["human-in-the-loop RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "100% success on three tasks within 101 minutes of online fine-tuning", "metric_measures": "other", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Human-in-the-loop dual-actor RL fine-tuning of VLAs that converts corrections into language commands, reaching high success in short online training.", "relevance": 3, "relevance_reason": "Efficient real-world RL fine-tuning", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2509.25681", "title": "dVLA: Diffusion Vision-Language-Action Model with Multimodal Chain-of-Thought", "authors": ["Junjie Wen", "Minjie Zhu", "Jiaming Liu", "Zhiyuan Liu", "Yicun Yang", "Linfeng Zhang", "Shanghang Zhang", "Yichen Zhu", "Yi Xu"], "year": 2025, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2509.25681", "pdf_url": "https://arxiv.org/pdf/2509.25681", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion VLA", "multimodal CoT", "KV caching"], "model_class": "VLA", "techniques": ["discrete diffusion decoding", "KV caching", "prefix attention mask"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real Franka"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Unified diffusion VLA that jointly generates reasoning and actions, using prefix attention masking and KV caching to cut inference latency.", "relevance": 3, "relevance_reason": "Diffusion-LM VLA with acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "dyna-dyna1-2025", "title": "Dynamism v1 (DYNA-1) Model: A Breakthrough in Performance and Production-Ready Embodied AI", "authors": ["Dyna Robotics"], "year": 2025, "venue": "blog (Dyna Robotics)", "type": "blog", "url": "https://www.dyna.co/research/dyna-1", "pdf_url": "", "code_url": "", "clusters": ["C1", "C4", "C8", "C9"], "keywords": ["production foundation model", "reward model", "24-hour autonomy", "foundation reward model", "RL in deployment", "error recovery"], "model_class": "VLA", "techniques": ["foundation reward model", "autonomous data flywheel", "reward-model-in-the-loop RL", "autonomous data segmentation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "99.4% success over 24 h; ~60% of human folding speed | 99.4% success over 24 h autonomous run; 850+ napkins/24 h at ~60% human speed", "metric_measures": "throughput", "benchmarks": ["real robot (napkin folding)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Dyna's commercial foundation model for sustained dexterous autonomy (napkin folding), driven by a foundation reward model for exploration, recovery and data curation; throughput rather than compute efficiency.", "relevance": 3, "relevance_reason": "Deployment reliability/throughput metrics from a speaker's company; no latency data.", "foundational_technique": false, "verified": true, "found_via": "seed; blog sweep: Dyna"}
{"id": "2510.27545", "title": "EBT-Policy: Energy Unlocks Emergent Physical Reasoning Capabilities", "authors": ["Travis Davies", "Yiqi Huang", "Alexi Gladstone", "Yun-Xin Liu", "Xiang Chen", "Heng Ji", "Huxian Liu", "Luhui Hu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.27545", "pdf_url": "https://arxiv.org/pdf/2510.27545", "code_url": "", "clusters": ["C1"], "keywords": ["EBT-Policy", "few-step inference"], "model_class": "other", "techniques": ["energy-based transformer", "dynamic compute"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2 inference steps vs 100 for Diffusion Policy (50x fewer) on some tasks", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Energy-based transformer policy that beats diffusion policies with less training/inference compute, sometimes converging in two steps.", "relevance": 3, "relevance_reason": "Compute-efficient generative policy class.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2501.09747"}
{"id": "2508.12638", "title": "edgeVLM: Cloud-edge Collaborative Real-time VLM based on Context Transfer", "authors": ["Chen Qian", "Xinran Yu", "Zewen Huang", "Danyang Li", "Qiang Ma", "Fan Dang", "Xuan Ding", "Guangyong Shang", "Zheng Yang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.12638", "pdf_url": "https://arxiv.org/pdf/2508.12638", "code_url": "", "clusters": ["C6"], "keywords": ["context transfer", "small/large VLM", "cloud latency fluctuation"], "model_class": "other", "techniques": ["delayed large-model outputs as context for edge small model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Uses delayed cloud large-VLM outputs as historical context to guide a real-time edge small VLM, tolerating fluctuating cloud latency.", "relevance": 3, "relevance_reason": "Latency-tolerant cloud/edge collaboration pattern analogous to CloudEdgeVLA.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2503.18684", "title": "Efficient Continual Adaptation of Pretrained Robotic Policy with Online Meta-Learned Adapters", "authors": ["Ruiqi Zhu", "Endong Sun", "Guanhe Huang", "Oya Celiktutan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.18684", "pdf_url": "https://arxiv.org/pdf/2503.18684", "code_url": "", "clusters": ["C3"], "keywords": ["OMLA", "meta-learned adapters", "continual adaptation"], "model_class": "other", "techniques": ["adapters", "meta-learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Meta-learns adapter initialization online so each new task's adapter benefits from previous tasks during continual adaptation of a pretrained policy.", "relevance": 3, "relevance_reason": "Adapter-based continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: adapter tuning robot foundation model prompt tuning visuomotor policy efficient transfer"}
{"id": "2502.09829", "title": "Efficient Evaluation of Multi-Task Robot Policies With Active Experiment Selection", "authors": ["Abrar Anwar", "Rohan Gupta", "Zain Merchant", "Sayan Ghosh", "Willie Neiswanger", "Jesse Thomason"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.09829", "pdf_url": "https://arxiv.org/pdf/2502.09829", "code_url": "", "clusters": ["C4"], "keywords": ["efficient evaluation", "multi-task policies"], "model_class": "other", "techniques": ["active testing", "expected information gain"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Frames robot policy evaluation as active testing, choosing informative trials to cut evaluation cost.", "relevance": 3, "relevance_reason": "Evaluation-cost methodology (C4).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2603.13616, 2607.14439"}
{"id": "2508.03660", "title": "Efficient Morphology-Aware Policy Transfer to New Embodiments", "authors": ["Michael Przystupa", "Hongyao Tang", "Martin Jagersand", "Santiago Miret", "Mariano Phielipp", "Matthew E. Taylor", "Glen Berseth"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.03660", "pdf_url": "https://arxiv.org/pdf/2508.03660", "code_url": "", "clusters": ["C3"], "keywords": ["morphology-aware policy", "PEFT", "prefix tuning", "adapters"], "model_class": "other", "techniques": ["adapters", "prefix tuning", "partial fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulated morphologies"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Compares PEFT variants for specializing morphology-aware policies to new embodiments; tuning <1% of parameters beats zero-shot and cuts samples vs scratch.", "relevance": 3, "relevance_reason": "PEFT for cross-embodiment transfer.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: adapter tuning robot foundation model prompt tuning visuomotor policy efficient transfer"}
{"id": "2504.09927", "title": "Efficient Task-specific Conditional Diffusion Policies: Shortcut Model Acceleration and SO(3) Optimization", "authors": ["Haiyong Yu", "Yanqiong Jin", "Yonghao He", "Wei Sui"], "year": 2025, "venue": "CVPR 2025 Workshop (MEIS)", "type": "paper", "url": "https://arxiv.org/abs/2504.09927", "pdf_url": "https://arxiv.org/pdf/2504.09927", "code_url": "", "clusters": ["C1"], "keywords": ["shortcut model", "SO(3) diffusion"], "model_class": "diffusion/flow policy", "techniques": ["shortcut model", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~5x faster than DDIM Diffusion Policy", "metric_measures": "relative latency", "benchmarks": ["RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Classifier-free shortcut diffusion policy with SO(3)-aware diffusion for fast task-specific action generation.", "relevance": 3, "relevance_reason": "Shortcut model acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query: shortcut model policy robot"}
{"id": "2512.10310", "title": "Efficient-VLN: A Simple yet Strong Baseline for Efficient Vision-Language Navigation", "authors": ["Duo Zheng", "Shijia Huang", "Yan-Yang Li", "Liwei Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.10310", "pdf_url": "https://arxiv.org/pdf/2512.10310", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["efficient VLN", "history reprocessing", "training efficiency"], "model_class": "VLA", "techniques": ["memory reuse"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Efficient-VLN reduces latency from reprocessing visual history and other training/data bottlenecks in MLLM navigation agents.", "relevance": 3, "relevance_reason": "Efficient embodied navigation baseline.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "2512.02020", "title": "EfficientFlow: Efficient Equivariant Flow Policy Learning for Embodied AI", "authors": ["Jianlei Chang", "Ruofeng Mei", "Wei Ke", "Xiangyu Xu"], "year": 2025, "venue": "AAAI 2026", "type": "paper", "url": "https://arxiv.org/abs/2512.02020", "pdf_url": "https://arxiv.org/pdf/2512.02020", "code_url": "", "clusters": ["C1"], "keywords": ["equivariant flow", "acceleration regularization"], "model_class": "diffusion/flow policy", "techniques": ["equivariance", "trajectory regularization", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Equivariant flow policy with an acceleration-regularization loss that straightens trajectories for faster sampling and better data efficiency.", "relevance": 3, "relevance_reason": "Fast flow policy.", "foundational_technique": false, "verified": true, "found_via": "search query: flow matching robot policy efficient"}
{"id": "2510.11682", "title": "Ego-Vision World Model for Humanoid Contact Planning", "authors": ["Hang Liu", "Yu-Man Gao", "Sangli Teng", "Yufeng Chi", "Y. Shao", "Zhongyu Li", "Maani Ghaffari", "K. Sreenath"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.11682", "pdf_url": "https://arxiv.org/pdf/2510.11682", "code_url": "", "clusters": ["C8"], "keywords": ["humanoid", "contact planning"], "model_class": "world model", "techniques": ["latent world model", "sampling-based MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Compressed-latent world model combined with sampling MPC for real-time humanoid contact planning from depth.", "relevance": 3, "relevance_reason": "Real-time world-model planning onboard a humanoid", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2402.15391"}
{"id": "2511.01224", "title": "Embodiment Transfer Learning for Vision-Language-Action Models", "authors": ["Chengmeng Li", "Yaxin Peng"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.01224", "pdf_url": "https://arxiv.org/pdf/2511.01224", "code_url": "", "clusters": ["C3"], "keywords": ["ET-VLA", "synthetic continued pretraining", "multi-robot"], "model_class": "VLA", "techniques": ["synthetic warm-up", "fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Transfers pretrained VLAs to multi-robot embodiments with synthetic continued pretraining before target fine-tuning, cutting real demo needs.", "relevance": 3, "relevance_reason": "Cross-embodiment adaptation cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"cross-embodiment\" AND abs:\"fine-tuning\" AND abs:efficient"}
{"id": "2505.09723", "title": "EnerVerse-AC: Envisioning Embodied Environments with Action Condition", "authors": ["Yuxin Jiang", "Shengcong Chen", "Si-Yuan Huang", "Liliang Chen", "Peng-Fei Zhou", "Yue Liao", "Xindong He", "Chiming Liu", "Hongsheng Li", "Maoqing Yao", "Guanghui Ren"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.09723", "pdf_url": "https://arxiv.org/pdf/2505.09723", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["action-conditioned world model", "evaluator"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Action-conditioned multi-view world model used as data engine and policy evaluator without physical robots.", "relevance": 3, "relevance_reason": "World model for cheap evaluation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2402.15391"}
{"id": "2509.12437", "title": "Enhancing Physical Consistency in Lightweight World Models", "authors": ["Dingrui Wang", "Zhexiao Sun", "Zhouheng Li", "Cheng Wang", "Youlun Peng", "Hongyuan Ye", "Baha Zarrouki", "Wei Li", "Mattia Piccinini", "Lei Xie", "Johannes Betz"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.12437", "pdf_url": "https://arxiv.org/pdf/2509.12437", "code_url": "", "clusters": ["C5", "C8"], "keywords": ["lightweight BEV world model", "physics-informed", "edge deployment"], "model_class": "world model", "techniques": ["soft-mask training", "warm-start inference"], "base_model": "", "params": "400M", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Compact 400M physics-informed BEV world model that closes much of the gap to large world models for edge deployment in driving.", "relevance": 3, "relevance_reason": "Small world models for edge devices.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2509.18953", "title": "Eva-VLA: Evaluating Vision-Language-Action Models' Robustness Under Real-World Physical Variations", "authors": ["Han-Qing Liu", "Shouwei Ruan", "Jiahuan Long", "Junqi Wu", "Jiacheng Hou", "Hui Tang", "Ting-Song Jiang", "Wei-En Zhou", "Wenbiao Yao"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.18953", "pdf_url": "https://arxiv.org/pdf/2509.18953", "code_url": "", "clusters": ["C4"], "keywords": ["Eva-VLA", "physical variations"], "model_class": "VLA", "techniques": ["robustness evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Evaluation framework that parametrizes real-world physical variations and uses black-box optimization to find worst cases for VLAs.", "relevance": 3, "relevance_reason": "Evaluation methodology for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310; cited by 2607.29169"}
{"id": "2508.06295", "title": "Evaluating Robot Program Performance with Power Consumption Driven Metrics in Lightweight Industrial Robots", "authors": ["Heredia, Juan", "Kolvig-Raun, Emil Stubbe", "Sorensen, Sune Lundo", "Kjaergaard, Mikkel Baun"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.06295", "pdf_url": "https://arxiv.org/pdf/2508.06295", "code_url": "", "clusters": ["C4"], "keywords": ["power consumption metrics", "energy utilization coefficient", "UR5e", "industrial robots"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "UR5e", "reported_metric": "", "metric_measures": "energy", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Evaluates robot programs by the robot's electrical power profile with normalized energy-utilization, energy-conversion and reliability metrics.", "relevance": 3, "relevance_reason": "Physical energy metrics for task execution (energy per action at actuator level).", "foundational_technique": false, "verified": true, "found_via": "WebSearch: robot policy evaluation energy consumption per task"}
{"id": "2507.17049", "title": "Evaluating Uncertainty and Quality of Visual Language Action-enabled Robots", "authors": ["Pablo Valle", "Cheng-Jie Lu", "Shaukat Ali", "Aitor Arrieta"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.17049", "pdf_url": "https://arxiv.org/pdf/2507.17049", "code_url": "", "clusters": ["C4"], "keywords": ["evaluation", "uncertainty"], "model_class": "VLA", "techniques": ["uncertainty metrics", "quality metrics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proposes and validates uncertainty and execution-quality metrics for VLA-enabled robots against expert labels, arguing binary success rate is an insufficient evaluation oracle.", "relevance": 3, "relevance_reason": "relevant: evaluation methodology for VLAs", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2510.03827, 2604.23775"}
{"id": "2511.05397", "title": "EveryDayVLA: A Vision-Language-Action Model for Affordable Robotic Manipulation", "authors": ["Samarth Chopra", "Alex McMoil", "Ben Carnovale", "Evan Sokolson", "R. Kubendran", "Samuel Dickerson"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.05397", "pdf_url": "https://arxiv.org/pdf/2511.05397", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["low-cost robot", "affordable VLA"], "model_class": "VLA", "techniques": ["adaptive-horizon ensemble"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Pairs a sub-$300 6-DoF arm with a VLA outputting discrete and continuous actions, using uncertainty-triggered re-planning for safe operation.", "relevance": 3, "relevance_reason": "Low-cost deployment of robot foundation models with uncertainty-driven replanning.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2604.24622, 2605.11567"}
{"id": "2509.15880", "title": "eVGGT: An Efficient Geometry-Aware Vision Encoder for Visuomotor Policies", "authors": ["A. Vuong", "Minh Nhat Vu", "Ian D. Reid"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.15880", "pdf_url": "https://arxiv.org/pdf/2509.15880", "code_url": "", "clusters": ["C2"], "keywords": ["VGGT distillation", "geometry-aware encoder"], "model_class": "other", "techniques": ["knowledge distillation"], "base_model": "VGGT (teacher)", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Distills the heavy VGGT geometry model into a lightweight encoder usable inside ACT/Diffusion Policy with better success than 2D encoders.", "relevance": 3, "relevance_reason": "Encoder compression via distillation for policies.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2410.21257"}
{"id": "2512.14666", "title": "EVOLVE-VLA: Test-Time Training from Environment Feedback for Vision-Language-Action Models", "authors": ["Ze-Chen Bai", "Chen Gao", "M. Shou"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.14666", "pdf_url": "https://arxiv.org/pdf/2512.14666", "code_url": "", "clusters": ["C3"], "keywords": ["test-time adaptation", "few-shot"], "model_class": "VLA", "techniques": ["test-time training", "learned progress reward"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Test-time training framework letting VLAs keep adapting through environment interaction with few or no task demos, using learned progress feedback.", "relevance": 3, "relevance_reason": "Efficient deployment-time adaptation of VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2605.11750"}
{"id": "2505.09694", "title": "EWMBench: Evaluating Scene, Motion, and Semantic Quality in Embodied World Models", "authors": ["Hu Yue", "Si-Yuan Huang", "Yue Liao", "Shengcong Chen", "Peng-Fei Zhou", "Liliang Chen", "Maoqing Yao", "Guanghui Ren"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2505.09694", "pdf_url": "https://arxiv.org/pdf/2505.09694", "code_url": "https://github.com/AgibotTech/EWMBench", "clusters": ["C4", "C8"], "keywords": ["world model benchmark"], "model_class": "world model", "techniques": ["world-model evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "EWMBench evaluates embodied video world models on scene consistency, motion correctness and semantic alignment rather than generic video quality.", "relevance": 3, "relevance_reason": "relevant: benchmarking methodology for world models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2501.03575; cited by 2505.19017, 2508.05635"}
{"id": "2510.14300", "title": "Expertise need not monopolize: Action-Specialized Mixture of Experts for Vision-Language-Action Learning", "authors": ["Weijie Shen", "Yitian Liu", "Yuhao Wu", "Zhixuan Liang", "Sijia Gu", "Dehui Wang", "Tian Nian", "Lei Xu", "Yusen Qin", "Jiangmiao Pang", "Xinping Guan", "Xiaokang Yang", "Yao Mu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.14300", "pdf_url": "https://arxiv.org/pdf/2510.14300", "code_url": "", "clusters": ["C1"], "keywords": ["MoE action expert", "upcycling", "decoupled routing"], "model_class": "VLA", "techniques": ["mixture of experts", "upcycling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboTwin", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "AdaMoE upcycles a dense VLA's action-expert FFNs into sparse MoE layers with a scale adapter decoupled from the router, adding capacity at similar compute.", "relevance": 3, "relevance_reason": "Sparse MoE VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA mixture-of-experts"}
{"id": "2510.01642", "title": "FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models", "authors": ["Zi-Jun Lin", "Jiafei Duan", "Haoquan Fang", "Dieter Fox", "Ranjay Krishna", "Cheston Tan", "Bi-Han Wen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.01642", "pdf_url": "https://arxiv.org/pdf/2510.01642", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "recovery", "runtime monitoring"], "model_class": "VLA", "techniques": ["failure data generation"], "base_model": "LLaVA-OneVision-7B", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Automatically generates failure cases with executable recovery actions and trains a VLM that detects and helps VLAs recover from failures.", "relevance": 3, "relevance_reason": "Failure detection/recovery (C7 runtime monitoring).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.00425"}
{"id": "2505.21594", "title": "Fast and Cost-effective Speculative Edge-Cloud Decoding with Early Exits", "authors": ["Yeshwanth Venkatesha", "Souvik Kundu", "Priyadarshini Panda"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.21594", "pdf_url": "https://arxiv.org/pdf/2505.21594", "code_url": "", "clusters": ["C2", "C6"], "keywords": ["speculative decoding", "edge-cloud", "early exit"], "model_class": "other", "techniques": ["edge draft / cloud verify speculative decoding", "early exits"], "base_model": "Vicuna-68M draft, Llama2-7B target", "params": "", "hardware": "Jetson Nano client + A100 server", "reported_metric": "up to 35% latency reduction vs cloud autoregressive decoding", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Small draft model on the device and large target model in the cloud with early exits, so drafting overlaps verification; demonstrated with a robot use case.", "relevance": 3, "relevance_reason": "Edge-cloud speculative decoding applicable to autoregressive VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:edge-cloud AND abs:robot"}
{"id": "2507.23318", "title": "FastDriveVLA: Efficient End-to-End Driving via Plug-and-Play Reconstruction-based Token Pruning", "authors": ["Jiajun Cao", "Qizhe Zhang", "Peidong Jia", "Xuhui Zhao", "Bo Lan", "Xiaoan Zhang", "Zhuo Li", "Xiaobao Wei", "Sixiang Chen", "Liyun Li", "Xianming Liu", "Ming Lu", "Yang Wang", "Shanghang Zhang", "Jia-Jun Cao", "Qi-Zhe Zhang", "Pei-Dong Jia", "Lizhuo", "Si-Xiang Chen", "Xian-Ming Liu", "Shang-Hang Zhang"], "year": 2025, "venue": "AAAI Conference on Artificial Intelligence", "type": "paper", "url": "https://arxiv.org/abs/2507.23318", "pdf_url": "https://arxiv.org/pdf/2507.23318", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["FastDriveVLA", "autonomous driving VLA"], "model_class": "VLA", "techniques": ["token pruning", "reconstruction", "visual token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["nuScenes"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Plug-and-play MAE-style reconstruction pruner that keeps foreground visual tokens for driving VLAs, trained with an adversarial foreground/background scheme.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=100]; snowball round 2: cited by 2510.17111, 2510.24795"}
{"id": "10.1145/3761812", "title": "FC-GPU: Feedback Control GPU Scheduling for Real-time Embedded Systems", "authors": ["Srinivasan Subramaniyan", "Xiaorui Wang"], "year": 2025, "venue": "ACM TECS (2025)", "type": "paper", "url": "https://doi.org/10.1145/3761812", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["GPU scheduling", "real-time embedded"], "model_class": "other", "techniques": ["feedback control scheduling", "MIMO control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Feedback-control GPU scheduler that models contention as a MIMO system and adapts task rates from measured response times instead of relying on WCET estimates.", "relevance": 3, "relevance_reason": "Closed-loop alternative when WCET of GPU inference cannot be bounded.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2504.15517", "title": "Few-Shot Vision-Language Action-Incremental Policy Learning", "authors": ["Mingchen Song", "Xiang Deng", "Guoqiang Zhong", "Qi Lv", "Jia Wan", "Yin-Chuan Li", "Jianye Hao", "Weili Guan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2504.15517", "pdf_url": "https://arxiv.org/pdf/2504.15517", "code_url": "", "clusters": ["C3"], "keywords": ["few-shot continual learning"], "model_class": "other", "techniques": ["prompt graph evolution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Formulates few-shot action-incremental learning and proposes a task-prompt graph policy for continual learning with few demos.", "relevance": 3, "relevance_reason": "Few-shot continual adaptation (C3).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2511.22697"}
{"id": "2506.16201", "title": "FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation", "authors": ["Sen Wang", "Le Wang", "Sanpin Zhou", "Jingyi Tian", "Jiayi Li", "Haowen Sun", "Wei Tang"], "year": 2025, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2506.16201", "pdf_url": "https://arxiv.org/pdf/2506.16201", "code_url": "", "clusters": ["C1"], "keywords": ["Mamba", "flow matching"], "model_class": "diffusion/flow policy", "techniques": ["state space model", "flow matching", "region-aware perception"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RLBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "FlowRAM combines region-aware perception, Mamba state-space fusion with linear complexity, and flow matching to make high-precision manipulation policies more efficient than iterative diffusion.", "relevance": 3, "relevance_reason": "relevant: efficiency-oriented policy architecture", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2510.24795, 2512.00903"}
{"id": "2509.23655", "title": "Focusing on What Matters: Object-Agent-centric Tokenization for Vision Language Action models", "authors": ["Rokas Bendikas", "Daniel Dijkman", "Markus Peschl", "Sanjay Haresh", "Pietro Mazzaglia"], "year": 2025, "venue": "CoRL 2025", "type": "paper", "url": "https://arxiv.org/abs/2509.23655", "pdf_url": "https://arxiv.org/pdf/2509.23655", "code_url": "", "clusters": ["C2"], "keywords": ["Oat-VLA"], "model_class": "VLA", "techniques": ["object-centric tokenization"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Object-agent-centric tokenization reducing VLA visual tokens to a few, converging at least twice as fast as OpenVLA.", "relevance": 3, "relevance_reason": "Token reduction for efficient VLA training.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=100]"}
{"id": "2511.20721", "title": "Foundry: Distilling 3D Foundation Models for the Edge", "authors": ["Guillaume Letellier", "Siddharth Srivastava", "Frédéric Jurie", "Gaurav Sharma"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.20721", "pdf_url": "https://arxiv.org/pdf/2511.20721", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["distilling 3D foundation models", "edge"], "model_class": "other", "techniques": ["distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Foundry distills large self-supervised 3D foundation models into compact edge-deployable versions for robots/AR.", "relevance": 3, "relevance_reason": "Compressing perception backbones for robots.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: arXiv API robot AND cloud AND edge"}
{"id": "2509.04018", "title": "FPC-VLA: A Vision-Language-Action Framework with a Supervisor for Failure Prediction and Correction", "authors": ["Yifan Yang", "Zhi-Xiang Duan", "Tianshi Xie", "Fuyu Cao", "Pin-Xi Shen", "Peili Song", "Piao-Piao Jin", "Guo-Kang Sun", "Shaoqing Xu", "Yang-Wei You", "Jingtai Liu", "Chenyang Zhao"], "year": 2025, "venue": "Expert systems with applications", "type": "preprint", "url": "https://doi.org/10.1016/j.eswa.2026.131742", "pdf_url": "https://arxiv.org/pdf/2509.04018", "code_url": "", "clusters": ["C7"], "keywords": ["failure prediction", "correction"], "model_class": "VLA", "techniques": ["VLM supervisor", "keyframe supervisor"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SIMPLER", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Pairs a VLA with a VLM supervisor that predicts failures and issues corrections, plus a fusion module refining actions.", "relevance": 3, "relevance_reason": "Runtime failure prediction with bounded overhead", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2605.21854, 2606.06660; snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2503.20523", "title": "GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving", "authors": ["Lloyd Russell", "Anthony Hu", "Lorenzo Bertoni", "George Fedoseev", "Jamie Shotton", "Elahe Arani", "Gianluca Corrado"], "year": 2025, "venue": "arXiv (Wayve tech report)", "type": "tech-report", "url": "https://arxiv.org/abs/2503.20523", "pdf_url": "https://arxiv.org/pdf/2503.20523", "code_url": "", "clusters": ["C8"], "keywords": ["GAIA-2", "driving world model", "high-compression tokenizer"], "model_class": "world model", "techniques": ["latent diffusion", "32x spatial compression video tokenizer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["multi-camera driving generation"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Wayve's controllable multi-camera driving world model; uses an unusually aggressive 32x spatial compression tokenizer to cut token count and speed up inference.", "relevance": 3, "relevance_reason": "Tokenizer-level efficiency in a production world model.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2509.00576", "title": "Galaxea Open-World Dataset and G0 Dual-System VLA Model", "authors": ["Tao Jiang", "Tianyuan Yuan", "Yicheng Liu", "Chenhao Lu", "Jianning Cui", "Xiao Liu", "Shuiqi Cheng", "Jiyang Gao", "Huazhe Xu", "Hang Zhao"], "year": 2025, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2509.00576", "pdf_url": "https://arxiv.org/pdf/2509.00576", "code_url": "", "clusters": ["C1"], "keywords": ["dual-system", "dataset", "VLM planner + VLA executor"], "model_class": "VLA", "techniques": ["dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Galaxea's G0 couples a VLM planner with a VLA executor, trained on a large single-embodiment open-world dataset.", "relevance": 3, "relevance_reason": "Dual-system VLA from industry.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "2405.00846", "title": "Gameplay Filters: Robust Zero-Shot Safety through Adversarial Imagination", "authors": ["Duy P. Nguyen", "Kai-Chieh Hsu", "Wenhao Yu", "Jie Tan", "Jaime F. Fisac"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2405.00846", "pdf_url": "https://arxiv.org/pdf/2405.00846", "code_url": "", "clusters": ["C7"], "keywords": ["gameplay filter", "adversarial imagination", "quadruped", "full-order safety filter"], "model_class": "other", "techniques": ["game-theoretic predictive safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Predictive safety filter that continually simulates matches between a learned safety policy and a learned adversary, giving a full-order 36-D quadruped filter robust zero-shot to large perturbations.", "relevance": 3, "relevance_reason": "Scalable safety filter for learned legged control (Hsu/Fisac).", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"predictive safety filter\" learning-based"}
{"id": "2510.03342", "title": "Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer", "authors": ["Gemini Robotics Team", "Abbas Abdolmaleki", "Saminda Abeyruwan", "Joshua Ainslie", "Jean-Baptiste Alayrac", "Montserrat Gonzalez Arenas", "Ashwin Balakrishna", "Nathan Batchelor", "Alex Bewley", "Jeff Bingham", "Michael Bloesch", "Konstantinos Bousmalis", "Philemon Brakel", "Anthony Brohan", "Thomas Buschmann", "Arunkumar Byravan", "Serkan Cabi", "Ken Caluwaerts", "Federico Casarini", "Christine Chan", "Oscar Chang", "London Chappellet-Volpini", "Jose Enrique Chen", "Xi Chen", "Hao-Tien Lewis Chiang", "Krzysztof Choromanski", "Adrian Collister", "David B. D'Ambrosio", "Sudeep Dasari", "Todor Davchev", "Meet Kirankumar Dave", "Coline Devin", "Norman Di Palo", "Tianli Ding", "Carl Doersch", "Adil Dostmohamed", "Yilun Du", "Debidatta Dwibedi", "Sathish Thoppay Egambaram", "Michael Elabd", "Tom Erez", "Xiaolin Fang", "Claudio Fantacci", "Cody Fong", "Erik Frey", "Chuyuan Fu", "Ruiqi Gao", "Marissa Giustina", "Keerthana Gopalakrishnan", "Laura Graesser", "Oliver Groth", "Agrim Gupta", "Roland Hafner", "Steven Hansen", "Leonard Hasenclever", "Sam Haves", "Nicolas Heess", "Brandon Hernaez", "Alex Hofer", "Jasmine Hsu", "Lu Huang", "Sandy H. Huang", "Atil Iscen", "Mithun George Jacob", "Deepali Jain", "Sally Jesmonth", "Abhishek Jindal", "Ryan Julian", "Dmitry Kalashnikov", "M. Emre Karagozler", "Stefani Karp", "Matija Kecman", "J. Chase Kew", "Donnie Kim", "Frank Kim", "Junkyung Kim", "Thomas Kipf", "Sean Kirmani", "Ksenia Konyushkova", "Li Yang Ku", "Yuheng Kuang", "Thomas Lampe", "Antoine Laurens", "Tuan Anh Le", "Isabel Leal", "Alex X. Lee", "Tsang-Wei Edward Lee", "Guy Lever", "Jacky Liang", "Li-Heng Lin", "Fangchen Liu", "Shangbang Long", "Caden Lu", "Sharath Maddineni", "Anirudha Majumdar", "Kevis-Kokitsi Maninis", "Andrew Marmon", "Sergio Martinez", "Assaf Hurwitz Michaely", "Niko Milonopoulos", "Joss Moore", "Robert Moreno", "Michael Neunert", "Francesco Nori", "Joy Ortiz", "Kenneth Oslund", "Carolina Parada", "Emilio Parisotto", "Amaris Paryag", "Acorn Pooley", "Thomas Power", "Alessio Quaglino", "Haroon Qureshi", "Rajkumar Vasudeva Raju", "Helen Ran", "Dushyant Rao", "Kanishka Rao", "Isaac Reid", "David Rendleman", "Krista Reymann", "Miguel Rivas", "Francesco Romano", "Yulia Rubanova", "Peter Pastor Sampedro", "Pannag R Sanketi", "Dhruv Shah", "Mohit Sharma", "Kathryn Shea", "Mohit Shridhar", "Charles Shu", "Vikas Sindhwani", "Sumeet Singh", "Radu Soricut", "Rachel Sterneck", "Ian Storz", "Razvan Surdulescu", "Jie Tan", "Jonathan Tompson", "Saran Tunyasuvunakool", "Jake Varley", "Grace Vesom", "Giulia Vezzani", "Maria Bauza Villalonga", "Oriol Vinyals", "René Wagner", "Ayzaan Wahid", "Stefan Welker", "Paul Wohlhart", "Chengda Wu", "Markus Wulfmeier", "Fei Xia", "Ted Xiao", "Annie Xie", "Jinyu Xie", "Peng Xu", "Sichun Xu", "Ying Xu", "Zhuo Xu", "Jimmy Yan", "Sherry Yang", "Skye Yang", "Yuxiang Yang", "Hiu Hong Yu", "Wenhao Yu", "Wentao Yuan", "Yuan Yuan", "Jingwei Zhang", "Tingnan Zhang", "Zhiyuan Zhang", "Allan Zhou", "Guangyao Zhou", "Yuxiang Zhou"], "year": 2025, "venue": "arXiv / tech report", "type": "tech-report", "url": "https://arxiv.org/abs/2510.03342", "pdf_url": "https://arxiv.org/pdf/2510.03342", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["Gemini Robotics 1.5", "thinking", "motion transfer", "multi-embodiment VLA", "thinking VLA"], "model_class": "VLA", "techniques": ["interleaved reasoning", "motion transfer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Next-generation Gemini Robotics VLA that learns across embodiments via Motion Transfer and interleaves natural-language reasoning with actions, paired with an ER 1.5 planning model.", "relevance": 3, "relevance_reason": "Organizer-lab frontier VLA; think-before-acting raises latency questions central to the workshop.", "foundational_technique": false, "verified": true, "found_via": "search query: Gemini Robotics; author sweep: Dhruv Shah"}
{"id": "2503.08663", "title": "Generating Robot Constitutions & Benchmarks for Semantic Safety", "authors": ["Pierre Sermanet", "Anirudha Majumdar", "Alex Irpan", "Dmitry Kalashnikov", "Vikas Sindhwani"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.08663", "pdf_url": "https://arxiv.org/pdf/2503.08663", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["ASIMOV benchmark", "robot constitutions", "semantic safety"], "model_class": "VLA", "techniques": ["constitution generation", "constitutional AI"], "base_model": "", "params": "", "hardware": "", "reported_metric": "84.3% top alignment rate on ASIMOV", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Releases the ASIMOV benchmark for semantic safety of robot foundation models and auto-generates robot constitutions that let a model reject unsafe instructions/actions.", "relevance": 3, "relevance_reason": "DeepMind semantic safety layer for foundation-model robots.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: ASIMOV semantic safety robot constitutions"}
{"id": "2504.18945", "title": "Generative AI in Embodied Systems: System-Level Analysis of Performance, Efficiency and Scalability", "authors": ["Zi-Shen Wan", "Jiayi Qian", "Yu-Hang Du", "Jason Jabbour", "Yi-Lun Du", "Y. Zhao", "A. Raychowdhury", "Tushar Krishna", "V. Reddi"], "year": 2025, "venue": "IEEE ISPASS 2025", "type": "paper", "url": "https://arxiv.org/abs/2504.18945", "pdf_url": "https://arxiv.org/pdf/2504.18945", "code_url": "", "clusters": ["C5", "C9"], "keywords": ["embodied agents", "workload characterization", "LLM planning latency"], "model_class": "other", "techniques": ["system-level benchmarking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "First system-level workload characterization of LLM-based embodied agent systems, identifying planning/communication latency and prompt growth as key inefficiencies.", "relevance": 3, "relevance_reason": "Computer-architecture characterization of embodied AI workloads.", "foundational_technique": false, "verified": true, "found_via": "web: robotics computing workload characterization"}
{"id": "2510.14643", "title": "Generative Models From and for Sampling-Based MPC: A Bootstrapped Approach for Adaptive Contact-Rich Manipulation", "authors": ["Lara Brudermüller", "Brandon Hung", "Xinghao Zhu", "Jiu-Guang Wang", "Nick Hawes", "Preston Culbertson", "Simon Le Cleac'h"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2510.14643", "pdf_url": "https://arxiv.org/pdf/2510.14643", "code_url": "", "clusters": ["C3", "C7", "C9"], "keywords": ["amortized MPC", "flow matching proposals", "sampling-based MPC"], "model_class": "diffusion/flow policy", "techniques": ["flow-matching proposal distribution for MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Amortizes sampling-based MPC by training flow-matching proposal models on its own noisy control sequences, making online planning more sample-efficient.", "relevance": 3, "relevance_reason": "Efficient online planning with generative priors.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2510.07092", "title": "Generative World Modelling for Humanoids: 1X World Model Challenge Technical Report", "authors": ["Riccardo Mereu", "Aidan Scannell", "Yuxin Hou", "Yi Zhao", "Aditya Jitta", "Antonio Dominguez", "Luigi Acerbi", "Amos Storkey", "Paul Chang"], "year": 2025, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2510.07092", "pdf_url": "https://arxiv.org/pdf/2510.07092", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["1X World Model Challenge", "compression track", "humanoid video prediction"], "model_class": "world model", "techniques": ["Wan-2.2 TI2V-5B with AdaLN state conditioning + LoRA", "spatio-temporal transformer on Cosmos tokens"], "base_model": "", "params": "", "hardware": "", "reported_metric": "23.0 dB PSNR (sampling); Top-500 CE 6.6386 (compression)", "metric_measures": "other", "benchmarks": ["1X World Model Challenge"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Winning technical report for the 1X humanoid world-model challenge, covering both pixel-sampling and discrete-token compression tracks.", "relevance": 3, "relevance_reason": "Benchmark for compact (tokenized) humanoid world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "nvidia-jetson-edge-ai-getting-started-blog-2025", "title": "Getting Started with Edge AI on NVIDIA Jetson: LLMs, VLMs, and Foundation Models for Robotics", "authors": ["Chitoku Yato", "Khalil BenKhaled"], "year": 2025, "venue": "blog (NVIDIA Technical Blog)", "type": "blog", "url": "https://developer.nvidia.com/blog/getting-started-with-edge-ai-on-nvidia-jetson-llms-vlms-and-foundation-models-for-robotics/", "pdf_url": "", "code_url": "", "clusters": ["C5", "C9"], "keywords": ["Jetson", "vLLM", "TensorRT", "model sizing per device"], "model_class": "other", "techniques": ["TensorRT", "vLLM", "quantization"], "base_model": "", "params": "", "hardware": "Orin Nano Super, AGX Orin, AGX Thor", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Tutorial mapping which model sizes fit each Jetson tier and how to serve LLM/VLM/VLA models on-device with vLLM and TensorRT.", "relevance": 3, "relevance_reason": "Practical on-device deployment guidance.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: NVIDIA GEAR / Isaac"}
{"id": "nvidia-gr00t-n1.5-2025", "title": "GR00T N1.5", "authors": ["NVIDIA GEAR Lab"], "year": 2025, "venue": "tech report / project page (NVIDIA Research)", "type": "tech-report", "url": "https://research.nvidia.com/labs/gear/gr00t-n1_5/", "pdf_url": "", "code_url": "https://github.com/NVIDIA/Isaac-GR00T", "clusters": ["C1"], "keywords": ["GR00T", "frozen VLM", "Eagle 2.5"], "model_class": "VLA", "techniques": ["frozen VLM backbone", "dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "GR00T revision that keeps the Eagle 2.5 VLM frozen through pretraining and fine-tuning with a simplified adapter.", "relevance": 3, "relevance_reason": "Dual-system VLA version.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2503.14734"}
{"id": "2512.08186", "title": "Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and-Language Navigation", "authors": ["Meng Wei", "Chen-Yang Wan", "Jiaqi Peng", "Xi-Qian Yu", "Yu-Qiang Yang", "Delin Feng", "Wenzhe Cai", "Chen-Ming Zhu", "Tai Wang", "Jiang-Miao Pang", "Xi-Hui Liu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.08186", "pdf_url": "https://arxiv.org/pdf/2512.08186", "code_url": "", "clusters": ["C1"], "keywords": ["vision-language navigation", "real-time control"], "model_class": "VLA", "techniques": ["dual-system"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Dual-system VLN model: a VLM plans waypoints slowly while a lightweight DiT policy produces fast trajectories.", "relevance": 3, "relevance_reason": "Dual-system design for real-time embodied control.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2602.02459, 2606.31497"}
{"id": "2502.05485", "title": "HAMSTER: Hierarchical Action Models For Open-World Robot Manipulation", "authors": ["Y. Li", "Yuquan Deng", "Jesse Zhang", "J. Jang", "Marius Memmel", "Caelan Reed Garrett", "Fabio Ramos", "Dieter Fox", "†. AnqiLi", "†. AbhishekGupta", "Ankit Goyal", "Nvidia"], "year": 2025, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2502.05485", "pdf_url": "https://arxiv.org/pdf/2502.05485", "code_url": "", "clusters": ["C1"], "keywords": ["hierarchical VLA", "2D path", "off-domain data"], "model_class": "VLA", "techniques": ["hierarchical / dual-system"], "base_model": "VLM high-level + 3D low-level policy", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Hierarchical VLA where a fine-tuned VLM predicts a coarse 2D end-effector path that guides a smaller 3D-aware low-level control policy.", "relevance": 3, "relevance_reason": "Hierarchical/dual-system design listed under C1; offloads fine-grained control to a lightweight policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2508.09262", "title": "Harnessing Input-Adaptive Inference for Efficient VLN", "authors": ["Dongwoo Kang", "Akhil Perincherry", "Zachary Coalson", "Aiden Gabriel", "Stefan Lee", "Sanghyun Hong"], "year": 2025, "venue": "ICCV 2025", "type": "paper", "url": "https://arxiv.org/abs/2508.09262", "pdf_url": "https://arxiv.org/pdf/2508.09262", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["VLN", "input-adaptive inference", "early exit thresholds"], "model_class": "other", "techniques": ["early exit", "caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows naive input-adaptive methods hurt VLN transformers and proposes view selection, importance-based early-exit thresholds and view caching to cut compute.", "relevance": 3, "relevance_reason": "Early-exit compression for embodied navigation models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2411.02359"}
{"id": "2512.05693", "title": "HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies", "authors": ["Zhiying Du", "Bei Liu", "Yaobo Liang", "Yichao Shen", "Haidong Cao", "Xiangyu Zheng", "Zhiyuan Feng", "Zuxuan Wu", "Jiaolong Yang", "Yu-Gang Jiang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.05693", "pdf_url": "https://arxiv.org/pdf/2512.05693", "code_url": "https://github.com/ZhiyingDu/HiMoE-VLA", "clusters": ["C1"], "keywords": ["hierarchical MoE", "heterogeneous data", "action expert"], "model_class": "VLA", "techniques": ["mixture of experts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN", "LIBERO", "xArm7", "ALOHA"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Hierarchical MoE action module specializing boundary layers by action space and middle layers by observation/embodiment heterogeneity.", "relevance": 3, "relevance_reason": "MoE VLA architecture.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA mixture-of-experts"}
{"id": "2509.17141", "title": "History-Aware Visuomotor Policy Learning via Point Tracking", "authors": ["Jing-Jing Chen", "Hong-Jie Fang", "Chen-Xi Wang", "Shi-Quan Wang", "Cewu Lu"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.17141", "pdf_url": "https://arxiv.org/pdf/2509.17141", "code_url": "", "clusters": ["C1"], "keywords": ["history-aware policy"], "model_class": "other", "techniques": ["point tracking", "history encoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Encodes history via tracked points to give policies memory with high computational efficiency.", "relevance": 3, "relevance_reason": "Efficient history representation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2506.07127", "title": "Human-assisted Robotic Policy Refinement via Action Preference Optimization", "authors": ["Wenke Xia", "Yichu Yang", "Hongtao Wu", "Xiao Ma", "Tao Kong", "Di Hu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.07127", "pdf_url": "https://arxiv.org/pdf/2506.07127", "code_url": "", "clusters": ["C3"], "keywords": ["APO", "action preference optimization", "human-assisted"], "model_class": "VLA", "techniques": ["preference optimization", "human interventions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Refines deployed VLAs from human-intervention trajectories via adaptive-reweighted action preference optimization.", "relevance": 3, "relevance_reason": "Post-deployment refinement method.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"preference optimization\""}
{"id": "2506.01182", "title": "Humanoid World Models: Open World Foundation Models for Humanoid Robotics", "authors": ["Muhammad Qasim Ali", "Aditya Sridhar", "Shahbuland Matiana", "Alexander Wong", "Mohammad K. Al-Sharman"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.01182", "pdf_url": "https://arxiv.org/pdf/2506.01182", "code_url": "", "clusters": ["C8"], "keywords": ["humanoid world model", "lightweight"], "model_class": "world model", "techniques": ["parameter sharing", "masked transformers", "flow matching"], "base_model": "", "params": "", "hardware": "1-2 GPUs", "reported_metric": "33-53% model size reduction", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lightweight open humanoid world models trainable on 1-2 GPUs, with parameter-sharing that shrinks models.", "relevance": 3, "relevance_reason": "Resource-efficient world models.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2402.15391"}
{"id": "2505.21432", "title": "Hume: Introducing System-2 Thinking in Visual-Language-Action Model", "authors": ["Haoming Song", "Delin Qu", "Yuanqi Yao", "Qizhi Chen", "Qi Lv", "Yiwen Tang", "Modi Shi", "Guanghui Ren", "Maoqing Yao", "Bin Zhao", "Dong Wang", "Xuelong Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.21432", "pdf_url": "https://arxiv.org/pdf/2505.21432", "code_url": "", "clusters": ["C1"], "keywords": ["System 2 thinking", "value-guided sampling", "cascaded denoising"], "model_class": "VLA", "techniques": ["dual-system architecture", "asynchronous execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Low-frequency System 2 samples and value-ranks action candidates while a lightweight System 1 denoises them into fluid actions asynchronously.", "relevance": 3, "relevance_reason": "Dual-system VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: dual-system VLA"}
{"id": "2510.26670", "title": "Hybrid Consistency Policy: Decoupling Multi-Modal Diversity and Real-Time Efficiency in Robotic Manipulation", "authors": ["Qianyou Zhao", "Yuliang Shen", "Xuanran Zhai", "Ce Hao", "Duidi Wu", "Jin Qi", "Jie Hu", "Qiaojun Yu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.26670", "pdf_url": "https://arxiv.org/pdf/2510.26670", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["hybrid consistency", "multimodality", "one-step jump", "hybrid sampling", "consistency jump", "multi-modality"], "model_class": "diffusion/flow policy", "techniques": ["consistency distillation", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Runs a short stochastic denoising prefix then a single consistency jump, preserving multimodality while cutting latency relative to an 80-step DDPM teacher.", "relevance": 3, "relevance_reason": "Distillation trade-off between modes and speed.", "foundational_technique": false, "verified": true, "found_via": "search query: consistency policy; search: consistency distillation visuomotor policy"}
{"id": "2509.16072", "title": "I-FailSense: Towards General Robotic Failure Detection with Vision-Language Models", "authors": ["Clémence Grislain", "Hamed Rahimi", "Olivier Sigaud", "Mohamed Chetouani"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.16072", "pdf_url": "https://arxiv.org/pdf/2509.16072", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "semantic misalignment"], "model_class": "other", "techniques": ["lightweight classification heads", "ensembling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VLM-based failure detector with lightweight heads on internal layers for semantic-misalignment failures.", "relevance": 3, "relevance_reason": "Runtime failure detection (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2606.23085, 2609.03611"}
{"id": "2510.19356", "title": "Imitation Learning Policy based on Multi-Step Consistent Integration Shortcut Model", "authors": ["Yu Fang", "Xinyu Wang", "Xuehe Zhang", "Wanli Xue", "Mingwei Zhang", "Shengyong Chen", "Jie Zhao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.19356", "pdf_url": "https://arxiv.org/pdf/2510.19356", "code_url": "", "clusters": ["C1"], "keywords": ["shortcut model", "multi-step consistency", "one-step"], "model_class": "diffusion/flow policy", "techniques": ["shortcut model", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "One-step shortcut flow policy trained with multi-step consistency losses and adaptive gradient balancing.", "relevance": 3, "relevance_reason": "One-step policy.", "foundational_technique": false, "verified": true, "found_via": "search query: shortcut model policy robot"}
{"id": "2512.10946", "title": "ImplicitRDP: An End-to-End Visual-Force Diffusion Policy With Structural Slow-Fast Learning", "authors": ["Wendi Chen", "Han Xue", "Yi Wang", "Fangyuan Zhou", "Jun Lv", "Yang Jin", "Shi-Run Tang", "Chuan Wen", "Ce-Wu Lu"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2512.10946", "pdf_url": "https://arxiv.org/pdf/2512.10946", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["slow-fast", "force control"], "model_class": "diffusion/flow policy", "techniques": ["slow-fast learning", "visual-force fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "ImplicitRDP processes asynchronous slow visual and fast force tokens in one diffusion policy with causal attention so force control runs at action rate while chunks stay coherent.", "relevance": 3, "relevance_reason": "relevant: slow-fast design for reactive control", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2509.02808", "title": "Improving the Resilience of Quadrotors in Underground Environments by Combining Learning-based and Safety Controllers", "authors": ["I. R. Ward", "Mark W. Paral", "Kristopher Riordan", "M. Kochenderfer"], "year": 2025, "venue": "CoDIT 2025", "type": "paper", "url": "https://arxiv.org/abs/2509.02808", "pdf_url": "https://arxiv.org/pdf/2509.02808", "code_url": "", "clusters": ["C7"], "keywords": ["runtime monitor", "normalizing flow", "controller switching", "quadrotor", "simplex"], "model_class": "other", "techniques": ["OOD monitor", "switch to safe controller"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses a normalizing-flow OOD measure as a runtime monitor to switch a quadrotor from a learned controller to a safety controller in unseen subterranean environments.", "relevance": 3, "relevance_reason": "Classic fallback-controller architecture for learned policies.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2410.04640 (S2)"}
{"id": "2511.20714", "title": "Inferix: A Block-Diffusion based Next-Generation Inference Engine for World Simulation", "authors": ["Inferix Team", "Tianyu Feng", "Yizeng Han", "Jiahao He", "Yuanyu He", "Xi Lin", "Teng Liu", "Hanfeng Lu", "Jiasheng Tang", "Wei Wang", "Zhiyuan Wang", "Jichao Wu", "Mingyang Yang", "Yinghao Yu", "Zeyu Zhang", "Bohan Zhuang"], "year": 2025, "venue": "arXiv", "type": "repo", "url": "https://arxiv.org/abs/2511.20714", "pdf_url": "https://arxiv.org/pdf/2511.20714", "code_url": "https://github.com/alibaba-damo-academy/Inferix", "clusters": ["C5", "C8"], "keywords": ["inference engine", "block diffusion", "KV cache management"], "model_class": "other", "techniques": ["semi-autoregressive block-diffusion decoding", "LLM-style KV cache management"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Inference engine specialized for block-diffusion (semi-autoregressive) world simulators, bringing LLM-style KV cache management to interactive video world models.", "relevance": 3, "relevance_reason": "Serving infrastructure for world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2510.01389", "title": "INSIGHT: INference-time Sequence Introspection for Generating Help Triggers in Vision-Language-Action Models", "authors": ["Ulas Berk Karli", "Ziyao Shangguan", "Tesca Fitzgerald"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.01389", "pdf_url": "https://arxiv.org/pdf/2510.01389", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["INSIGHT", "help triggers", "token-level uncertainty", "pi0-FAST"], "model_class": "VLA", "techniques": ["token entropy / Dirichlet uncertainty features", "transformer classifier"], "base_model": "pi0-FAST", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns when a VLA should request help from sequences of per-token entropy, log-probability and aleatoric/epistemic estimates.", "relevance": 3, "relevance_reason": "Runtime help-request monitor.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2503.08558 (S2)"}
{"id": "isaac-gr00t-policy-server", "title": "Isaac GR00T: policy server-client (ZMQ) remote inference", "authors": ["NVIDIA"], "year": 2025, "venue": "GitHub repository docs", "type": "repo", "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/policy.md", "pdf_url": "", "code_url": "https://github.com/NVIDIA/Isaac-GR00T", "clusters": ["C6"], "keywords": ["policy server", "ZeroMQ", "remote inference", "Jetson Thor"], "model_class": "VLA", "techniques": ["remote policy server", "onboard Jetson deployment option"], "base_model": "GR00T N1.x", "params": "", "hardware": "GPU server or Jetson AGX Thor", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "GR00T ships a ZMQ policy server so the model can run on a separate GPU machine while a light robot client streams observations and receives actions, alongside Jetson Thor onboard installs.", "relevance": 3, "relevance_reason": "Standard remote-inference tooling for an open humanoid VLA.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2510.15388", "title": "Iterative Refinement of Flow Policies in Probability Space for Online Reinforcement Learning", "authors": ["Mingyang Sun", "Pengxiang Ding", "Weinan Zhang", "Donglin Wang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2510.15388", "pdf_url": "https://arxiv.org/pdf/2510.15388", "code_url": "", "clusters": ["C3"], "keywords": ["SWFP", "JKO"], "model_class": "diffusion/flow policy", "techniques": ["stepwise flow fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fine-tunes pretrained flow policies via a cascade of small flow blocks, lowering compute and memory for online RL adaptation.", "relevance": 3, "relevance_reason": "Efficient RL fine-tuning of flow policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2505.10183", "title": "KAITIAN: A Unified Communication Framework for Enabling Efficient Collaboration Across Heterogeneous Accelerators in Embodied AI Systems", "authors": ["Jieke Lin", "Wanyu Wang", "Longxiang Yin", "Yinhe Han"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.10183", "pdf_url": "https://arxiv.org/pdf/2505.10183", "code_url": "", "clusters": ["C5"], "keywords": ["heterogeneous accelerators", "GPU + Cambricon MLU", "communication library", "load balancing"], "model_class": "other", "techniques": ["unified communication abstraction", "load-adaptive scheduling"], "base_model": "", "params": "", "hardware": "NVIDIA GPUs + Cambricon MLUs", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PyTorch extension unifying vendor communication libraries so embodied-AI workloads can be split across heterogeneous GPUs and NPUs with load-adaptive scheduling.", "relevance": 3, "relevance_reason": "Heterogeneous accelerator collaboration explicitly for embodied AI.", "foundational_technique": false, "verified": true, "found_via": "arXiv: embodied AND (accelerator OR co-design)"}
{"id": "2509.21354", "title": "KV-Efficient VLA: A Method to Speed up Vision Language Models with RNN-Gated Chunked KV Cache", "authors": ["Wanshun Xu", "Long Zhuang", "Lianlei Shan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.21354", "pdf_url": "https://arxiv.org/pdf/2509.21354", "code_url": "", "clusters": ["C1"], "keywords": ["KV cache compression", "recurrent gating", "history context"], "model_class": "VLA", "techniques": ["KV cache compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.34x speedup, 1.87x less KV memory, 24.6% FLOPs saved", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Chunks the historical KV cache and uses a recurrent gate to summarize and drop low-utility history for long-horizon VLAs.", "relevance": 3, "relevance_reason": "KV memory for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA KV cache"}
{"id": "2511.17898", "title": "L1 Sample Flow for Efficient Visuomotor Learning", "authors": ["Wei Song", "Zhetao Chen", "Tao Xu", "Xianchao Zeng", "Xinyu Zhou", "Lixin Yang", "Donglin Wang", "Cewu Lu", "Yong-Lu Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.17898", "pdf_url": "https://arxiv.org/pdf/2511.17898", "code_url": "", "clusters": ["C1"], "keywords": ["few-step flow", "visuomotor policy"], "model_class": "diffusion/flow policy", "techniques": ["sample prediction", "L1 objective", "two-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2 NFEs per action sequence", "metric_measures": "other", "benchmarks": ["MimicGen", "RoboMimic", "PushT"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Reformulates flow matching as sample prediction with an L1 loss, needing only two network evaluations per action sequence.", "relevance": 3, "relevance_reason": "Few-step flow policy for faster inference and training.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2609.21216"}
{"id": "2509.20623", "title": "Latent Activation Editing: Inference-Time Refinement of Learned Policies for Safer Multirobot Navigation", "authors": ["Satyajeet Das", "Darren Chiu", "Zhehui Huang", "Lars Lindemann", "Gaurav S. Sukhatme"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.20623", "pdf_url": "https://arxiv.org/pdf/2509.20623", "code_url": "", "clusters": ["C7"], "keywords": ["LAE", "quadrotor navigation"], "model_class": "other", "techniques": ["activation editing", "runtime monitoring", "latent world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~90% fewer cumulative collisions vs unedited baseline", "metric_measures": "other", "benchmarks": ["simulation", "Crazyflie real"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Inference-time latent activation editing: an online classifier flags risky intermediate activations and an editor shifts them (via a latent collision world model) to make pretrained multirobot navigation policies more cautious.", "relevance": 3, "relevance_reason": "Runtime safety intervention inside a learned policy without retraining.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "2508.05941", "title": "Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution", "authors": ["Zhanyi Sun", "Shuran Song"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.05941", "pdf_url": "https://arxiv.org/pdf/2508.05941", "code_url": "", "clusters": ["C7"], "keywords": ["covariate shift", "OOD recovery"], "model_class": "diffusion/flow policy", "techniques": ["latent barrier", "dynamics model", "test-time steering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Treats expert latent embeddings as an implicit barrier and uses a learned dynamics model at inference to steer a diffusion policy back in-distribution.", "relevance": 3, "relevance_reason": "Latent barrier-style correction for visuomotor chunks.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\")"}
{"id": "2502.14819", "title": "Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models", "authors": ["Vlad Sobal", "Wancong Zhang", "Kyunghyun Cho", "Randall Balestriero", "Tim G. J. Rudner", "Yann LeCun"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.14819", "pdf_url": "https://arxiv.org/pdf/2502.14819", "code_url": "", "clusters": ["C8"], "keywords": ["PLDM", "reward-free offline data", "latent dynamics planning"], "model_class": "world model", "techniques": ["JEPA latent dynamics + planning vs offline RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["navigation suites"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Compares offline RL with planning in a JEPA latent dynamics model on reward-free data, finding latent planning generalizes better to new tasks and layouts.", "relevance": 3, "relevance_reason": "Evidence for latent world-model planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2505.09561", "title": "Learning Long-Context Diffusion Policies via Past-Token Prediction", "authors": ["Marcel Torne", "Andy Tang", "Yue-Jiang Liu", "Chelsea Finn"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.09561", "pdf_url": "https://arxiv.org/pdf/2505.09561", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["long-context", "training efficiency"], "model_class": "diffusion/flow policy", "techniques": ["auxiliary past-token prediction", "cached embeddings"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">10x faster policy training", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Past-token prediction regularizer and cached-embedding training for long-context diffusion policies.", "relevance": 3, "relevance_reason": "Training memory/compute savings.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2508.20840", "title": "Learning Primitive Embodied World Models: Towards Scalable Robotic Learning", "authors": ["Qiao Sun", "Liujia Yang", "Wei Tang", "Wei Huang", "Kaixin Xu", "Yong-Chao Chen", "Mingyu Liu", "Jiange Yang", "Haoyi Zhu", "Ya-Ting Wang", "Tong He", "Yi-Lun Chen", "Xili Dai", "Nanyang Ye", "Qin-Ying Gu"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2508.20840", "pdf_url": "https://arxiv.org/pdf/2508.20840", "code_url": "", "clusters": ["C8"], "keywords": ["primitive world models", "inference latency"], "model_class": "world model", "techniques": ["short-horizon primitive generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Restricts video world models to fixed short-horizon primitives, improving data efficiency and lowering inference latency, with a VLM planner for composition.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2402.15391"}
{"id": "2506.05165", "title": "LiPo: A Lightweight Post-optimization Framework for Smoothing Action Chunks Generated by Learned Policies", "authors": ["Dongwoo Son", "Suhan Park"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.05165", "pdf_url": "https://arxiv.org/pdf/2506.05165", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["chunk smoothing", "jerk minimization", "post-optimization", "LiPo"], "model_class": "other", "techniques": ["trajectory optimization", "chunk blending", "inference-aware chunk scheduling", "linear blending", "bounded trajectory optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Post-optimizes learned action chunks with overlap scheduling to hide inference delay, blending, and jerk-minimizing optimization within bounded deviation for dynamic tasks.", "relevance": 3, "relevance_reason": "Smooth execution under delay.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk inference delay; arXiv search: \"inference delay\" robot policy"}
{"id": "2509.22622", "title": "LongLive: Real-time Interactive Long Video Generation", "authors": ["Shuai Yang", "Wei Huang", "Ruihang Chu", "Yicheng Xiao", "Yuyang Zhao", "Xianbang Wang", "Mu-Yang Li", "En-Ze Xie", "Ying-Cong Chen", "Yao Lu", "Song Han", "Yu-Kang Chen"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.22622", "pdf_url": "https://arxiv.org/pdf/2509.22622", "code_url": "", "clusters": ["C8"], "keywords": ["real-time video generation"], "model_class": "world model", "techniques": ["KV re-cache", "short-window attention", "INT8"], "base_model": "", "params": "", "hardware": "NVIDIA H100", "reported_metric": "20.7 FPS", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LongLive is a causal frame-level autoregressive video generator with KV recaching and attention sinks for real-time long video.", "relevance": 3, "relevance_reason": "generic efficiency technique connected to 5 core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2506.08009"}
{"id": "2508.13009", "title": "Matrix-game 2.0: An open-source, real-time, and streaming interactive world model", "authors": ["Xianglong He", "Chunli Peng", "Zexiang Liu", "Boyang Wang", "Yifan Zhang", "Qi Cui", "Fei Kang", "Biao Jiang", "Mengyin An", "Yangyang Ren", "Baixin Xu", "Hao-Xiang Guo", "Kaixiong Gong", "Size Wu", "Wei Li", "Xuchen Song", "Yang Liu", "Yangguang Li", "Yahui Zhou"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.13009", "pdf_url": "https://arxiv.org/pdf/2508.13009", "code_url": "https://github.com/SkyworkAI/Matrix-Game", "clusters": ["C8"], "keywords": ["Matrix-Game 2.0", "real-time interactive world model", "few-step AR"], "model_class": "world model", "techniques": ["causal few-step distillation", "frame-level action injection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "25 FPS minute-level generation", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Open-source interactive world model using few-step causal autoregressive diffusion with keyboard/mouse action injection, streaming at 25 FPS.", "relevance": 3, "relevance_reason": "Open real-time interactive world model recipe (games).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2512.20861", "title": "Memory-Efficient Acceleration of Block Low-Rank Foundation Models on Resource Constrained GPUs", "authors": ["Pierre Abillama", "Changwoo Lee", "Juechu Dong", "David Blaauw", "Dennis Sylvester", "Hun-Seok Kim"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.20861", "pdf_url": "https://arxiv.org/pdf/2512.20861", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["block low-rank", "Monarch", "BLAST", "Triton kernels", "Jetson Orin Nano", "roofline"], "model_class": "other", "techniques": ["custom Triton kernels", "partial fusion", "memory layout optimization"], "base_model": "", "params": "", "hardware": "Jetson Orin Nano, A40", "reported_metric": "up to 3.76x speedup and 3x compression vs dense PyTorch", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Roofline analysis shows block-low-rank transformers become memory-bound for multi-token inference; custom Triton kernels restore speedups on Jetson Orin Nano.", "relevance": 3, "relevance_reason": "Edge-GPU kernel work for compressed foundation models.", "foundational_technique": true, "verified": true, "found_via": "arXiv: Jetson AND (robot policy OR manipulation OR VLA)"}
{"id": "2511.18810", "title": "MergeVLA: Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent", "authors": ["Yuxia Fu", "Zhizhen Zhang", "Yuqi Zhang", "Zijian Wang", "Zi Huang", "Yadan Luo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.18810", "pdf_url": "https://arxiv.org/pdf/2511.18810", "code_url": "", "clusters": ["C3"], "keywords": ["MergeVLA", "model merging", "LoRA divergence"], "model_class": "VLA", "techniques": ["model merging", "LoRA"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Diagnoses why VLA experts are non-mergeable and designs merge-friendly fine-tuning for multi-skill agents.", "relevance": 3, "relevance_reason": "Merging avoids multiple checkpoints/joint retraining.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2512.15692", "title": "mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs", "authors": ["Jonas Pai", "Liam Achenbach", "Victoriano Montesinos", "Benedek Forrai", "Oier Mees", "Elvis Nava"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.15692", "pdf_url": "https://arxiv.org/pdf/2512.15692", "code_url": "", "clusters": ["C8"], "keywords": ["video-action model", "inverse dynamics decoder", "sample efficiency"], "model_class": "world-action model", "techniques": ["pretrained video model latents", "flow-matching IDM action decoder"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Pairs an internet-scale video model with a flow-matching inverse-dynamics decoder that reads its latent video plans, giving 10x sample efficiency over VLA baselines.", "relevance": 3, "relevance_reason": "Representative inverse-dynamics WAM; relevant to compute trade-offs of video planning.", "foundational_technique": false, "verified": true, "found_via": "snowball from NVIDIA WAM blog"}
{"id": "2504.08388", "title": "MineWorld: a Real-Time and Open-Source Interactive World Model on Minecraft", "authors": ["Junliang Guo", "Yang Ye", "Tianyu He", "Hao-Yu Wu", "Yushu Jiang", "Tim Pearce", "Jiang Bian"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2504.08388", "pdf_url": "https://arxiv.org/pdf/2504.08388", "code_url": "", "clusters": ["C8"], "keywords": ["real-time world model", "parallel decoding"], "model_class": "world model", "techniques": ["parallel decoding", "autoregressive transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "4-7 frames per second", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MineWorld is an autoregressive visual-action world model for Minecraft with a parallel decoding scheme that makes interactive generation real-time.", "relevance": 3, "relevance_reason": "relevant: real-time world model via parallel decoding", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2405.15223; cited by 2505.19017, 2602.06949"}
{"id": "2512.00975", "title": "MM-ACT: Learn from Multimodal Parallel Generation to Act", "authors": ["Haotian Liang", "Xinyi Chen", "Bin Wang", "Mingkang Chen", "Yitian Liu", "Yuhao Zhang", "Zanxin Chen", "Tianshuo Yang", "Yilun Chen", "Jiangmiao Pang", "Dong Liu", "Xiaokang Yang", "Yao Mu", "Wenqi Shao", "Ping Luo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.00975", "pdf_url": "https://arxiv.org/pdf/2512.00975", "code_url": "https://github.com/HHYHRHY/MM-ACT", "clusters": ["C1"], "keywords": ["unified multimodal generation", "one-step parallel action decoding"], "model_class": "VLA", "techniques": ["parallel action decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "RoboTwin 2.0", "real Franka"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Unified text/image/action VLA using re-mask parallel decoding for text/images and one-step parallel decoding for actions.", "relevance": 3, "relevance_reason": "Parallel action decoding.", "foundational_technique": false, "verified": true, "found_via": "search query: parallel decoding action"}
{"id": "2507.01843", "title": "MoIRA: Modular Instruction Routing Architecture for Multi-Task Robotics", "authors": ["Dmytro Kuzmenko", "Nadiya Shvai"], "year": 2025, "venue": "Neurocomputing", "type": "paper", "url": "https://arxiv.org/abs/2507.01843", "pdf_url": "https://arxiv.org/pdf/2507.01843", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["modular MoE", "low-overhead inference"], "model_class": "VLA", "techniques": ["LoRA experts", "text-based routing"], "base_model": "GR00T-N1, pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "GR1"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Routes tasks to LoRA-adapted VLA experts (GR00T-N1, pi0) via zero-shot text routing for low-overhead modular deployment.", "relevance": 3, "relevance_reason": "Parameter-efficient modular experts.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2603.09298"}
{"id": "2503.08007", "title": "MoRE: Unlocking Scalability in Reinforcement Learning for Quadruped Vision-Language-Action Models", "authors": ["Han Zhao", "Wen-Xuan Song", "Dong-Lin Wang", "Xinyang Tong", "Pengxiang Ding", "Xuelian Cheng", "Zongyuan Ge"], "year": 2025, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2503.08007", "pdf_url": "https://arxiv.org/pdf/2503.08007", "code_url": "", "clusters": ["C3"], "keywords": ["quadruped VLA"], "model_class": "VLA", "techniques": ["LoRA experts", "sparse MoE", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Quadruped VLA with multiple LoRA experts as a sparse MoE, fine-tuned with RL on mixed-quality data.", "relevance": 3, "relevance_reason": "Parameter-efficient adaptation (C3).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2506.05985, 2506.13725"}
{"id": "2510.18337", "title": "MoTVLA: A Vision-Language-Action Model with Unified Fast-Slow Reasoning", "authors": ["Wenhui Huang", "Changhe Chen", "Han Qi", "Chen Lv", "Yilun Du", "Heng Yang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.18337", "pdf_url": "https://arxiv.org/pdf/2510.18337", "code_url": "", "clusters": ["C1"], "keywords": ["mixture of transformers", "fast reasoning", "domain expert"], "model_class": "VLA", "techniques": ["mixture of transformers"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Mixture-of-transformers VLA where a domain-expert transformer produces fast motion-decomposition reasoning for the action expert, avoiding slow general reasoning.", "relevance": 3, "relevance_reason": "Fast-slow reasoning VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: fast slow VLA"}
{"id": "2505.12410", "title": "MTIL: Encoding Full History with Mamba for Temporal Imitation Learning", "authors": ["Yulin Zhou", "Yuankai Lin", "Fanzhe Peng", "Jiahui Chen", "Kaiji Huang", "Hua Yang", "Zhouping Yin"], "year": 2025, "venue": "IEEE RA-L", "type": "paper", "url": "https://arxiv.org/abs/2505.12410", "pdf_url": "https://arxiv.org/pdf/2505.12410", "code_url": "", "clusters": ["C1"], "keywords": ["Mamba", "full history", "non-Markovian"], "model_class": "other", "techniques": ["state space model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ACT sim", "Robomimic", "LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses Mamba's linear-time recurrence to encode full trajectory history into a compact state for imitation policies.", "relevance": 3, "relevance_reason": "SSM for efficient history.", "foundational_technique": false, "verified": true, "found_via": "search query: Mamba robot policy"}
{"id": "2509.21073", "title": "Normalizing Flows are Capable Models for Bi-manual Visuomotor Policy", "authors": ["Jialong Li", "Simon Kristoffersson Lind", "Wenrui Xie", "Maj Stenmark", "Volker Kruger"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.21073", "pdf_url": "https://arxiv.org/pdf/2509.21073", "code_url": "", "clusters": ["C1"], "keywords": ["uncertainty-aware policy", "bimanual"], "model_class": "diffusion/flow policy", "techniques": ["normalizing flows", "single-pass sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Normalizing-flow visuomotor policy with single-pass sampling and tractable likelihoods, giving faster training and lower latency than diffusion.", "relevance": 3, "relevance_reason": "Computationally efficient generative policy class.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2510.03895", "title": "NoTVLA: Semantics-Preserving Robot Adaptation via Narrative Action Interfaces", "authors": ["Zheng Huang", "Mingyu Liu", "Xiaoyi Lin", "Muzhi Zhu", "Can-Yu Zhao", "Zong-Ze Du", "Ye Lin", "Xiaoman Li", "Yiduo Jia", "Hao Zhong", "Hao Chen", "Chunhua Shen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.03895", "pdf_url": "https://arxiv.org/pdf/2510.03895", "code_url": "", "clusters": ["C1"], "keywords": ["NoTVLA", "catastrophic forgetting"], "model_class": "VLA", "techniques": ["sparse trajectories", "temporal compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "over 10x less compute than pi0", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains VLAs on sparse, compressed end-effector trajectories instead of dense action chunks, reducing forgetting and reportedly using an order of magnitude less compute than pi0.", "relevance": 3, "relevance_reason": "Explicit compute reduction relative to pi0.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2609.30833"}
{"id": "2510.09667", "title": "OmniSAT: Compact Action Token, Faster Auto Regression", "authors": ["Huaihai Lyu", "Chaofan Chen", "Senwei Xie", "Pengwei Wang", "Xiansheng Chen", "Shanghang Zhang", "Changsheng Xu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.09667", "pdf_url": "https://arxiv.org/pdf/2510.09667", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["residual quantization", "compact action tokens", "autoregressive VLA"], "model_class": "VLA", "techniques": ["action tokenization", "compact action tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "6.8x shorter training sequences", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "B-spline normalization plus multi-stage residual quantization per action subspace to shorten AR action sequences.", "relevance": 3, "relevance_reason": "Faster AR decoding via tokenization.", "foundational_technique": false, "verified": true, "found_via": "search query: action tokenization; venue sweep: arXiv API 'vision-language-action' AND quantization"}
{"id": "2505.11917", "title": "OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning", "authors": ["Fanqi Lin", "Ruiqian Nai", "Ying-Dong Hu", "Jiacheng You", "Junming Zhao", "Yang Gao"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2505.11917", "pdf_url": "https://arxiv.org/pdf/2505.11917", "code_url": "", "clusters": ["C1"], "keywords": ["System 1/2", "reasoning on demand"], "model_class": "VLA", "techniques": ["adaptive reasoning", "dual-system"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Single VLA that switches between explicit reasoning and direct acting, reasoning only at critical moments to avoid the latency and mismatch of separate dual-system models.", "relevance": 3, "relevance_reason": "Adaptive reasoning designed partly to remove dual-system latency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2509.22195"}
{"id": "2504.18057", "title": "Opportunistic Collaborative Planning with Large Vision Model Guided Control and Joint Query-Service Optimization", "authors": ["Jiayi Chen", "Shuai Wang", "Guoliang Li", "Wei Xu", "Guangxu Zhu", "Derrick Wing Kwan Ng", "Chengzhong Xu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2504.18057", "pdf_url": "https://arxiv.org/pdf/2504.18057", "code_url": "", "clusters": ["C6"], "keywords": ["cloud large vision model", "local MPC", "query timing"], "model_class": "other", "techniques": ["cloud-guided MPC", "collaboration timing optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Local MPC navigation is guided by a cloud large vision model that is queried only when detection confidence is low, jointly optimizing when to ask the cloud.", "relevance": 3, "relevance_reason": "When-to-query-cloud policy for hybrid local/cloud control.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2501.13987", "title": "OstQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting", "authors": ["Xing Hu", "Yuan Cheng", "Da-Wei Yang", "Zhi-Xuan Chen", "Zu-Kang Xu", "Jiang-Yong Yu", "Chen Xu", "Zhihang Yuan", "Zhe Jiang", "Si-Fan Zhou"], "year": 2025, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2501.13987", "pdf_url": "https://arxiv.org/pdf/2501.13987", "code_url": "", "clusters": ["C2"], "keywords": ["orthogonal and scaling transforms", "W4A4KV4"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns orthogonal and scaling transforms to fit weight/activation distributions for better low-bit LLM PTQ.", "relevance": 3, "relevance_reason": "Foundational quantization technique", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323"}
{"id": "2506.20966", "title": "Parallels Between VLA Model Post-Training and Human Motor Learning: Progress, Challenges, and Trends", "authors": ["Tian-Yu Xiang", "Ao-Qun Jin", "Xiao-Hu Zhou", "Mei-Jiang Gui", "Xiao-Liang Xie", "Shi-Qi Liu", "Shuang-Yi Wang", "Sheng-Bin Duan", "Fu-Chao Xie", "Wen-Kai Wang", "Si-Cheng Wang", "Ling-Yun Li", "Tian Tu", "Zeng-Guang Hou"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.20966", "pdf_url": "https://arxiv.org/pdf/2506.20966", "code_url": "", "clusters": ["C3"], "keywords": ["post-training survey", "motor learning"], "model_class": "VLA", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey of VLA post-training organized by analogy to human motor learning.", "relevance": 3, "relevance_reason": "Overview of adaptation methods.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2511.01571", "title": "PixelVLA: Advancing Pixel-level Understanding in Vision-Language-Action Model", "authors": ["Wen-Qi Liang", "Gan Sun", "Yao He", "Jia-Hua Dong", "Suyan Dai", "Ivan Laptev", "Salman Khan", "Yang Cong"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2511.01571", "pdf_url": "https://arxiv.org/pdf/2511.01571", "code_url": "", "clusters": ["C1"], "keywords": ["PixelVLA", "pixel-level"], "model_class": "VLA", "techniques": ["low-cost pretraining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.5% of OpenVLA pretraining cost", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds pixel-level understanding to VLAs and reports large gains over OpenVLA at only 1.5% of its pretraining cost.", "relevance": 3, "relevance_reason": "Training-cost efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2607.14852"}
{"id": "2506.01380", "title": "Playing with Transformer at 30+ FPS via Next-Frame Diffusion", "authors": ["Xinle Cheng", "Tianyu He", "Jiayi Xu", "Junliang Guo", "Di He", "Jiang Bian"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.01380", "pdf_url": "https://arxiv.org/pdf/2506.01380", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["next-frame diffusion", "speculative sampling", "consistency distillation"], "model_class": "world model", "techniques": ["block-causal attention", "video consistency distillation", "action-speculative frame sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "30+ FPS", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Autoregressive next-frame diffusion transformer made real-time via video consistency distillation and speculative generation of several frames under the current action, discarding them if the action changes.", "relevance": 3, "relevance_reason": "Speculative decoding idea for action-conditioned world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2512.03724", "title": "PosA-VLA: Enhancing Action Generation via Pose-Conditioned Anchor Attention", "authors": ["Zi-Wen Li", "Xin Wang", "Hanlue Zhang", "Runnan Chen", "Runqi Lin", "Xiao He", "Han Huang", "Yan-Dong Guo", "Fakhri Karray", "Tong-Liang Liu", "Ming-Ming Gong"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.03724", "pdf_url": "https://arxiv.org/pdf/2512.03724", "code_url": "", "clusters": ["C1"], "keywords": ["attention anchoring", "efficient execution"], "model_class": "VLA", "techniques": ["pose-conditioned anchor attention", "lightweight architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PosA-VLA steers VLA visual attention toward task-relevant regions using pose-conditioned supervision, cutting redundant motions and keeping inference light without extra perception modules.", "relevance": 3, "relevance_reason": "relevant: time-efficient execution and lightweight inference", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "hf-blog-gr00t-n15-so101", "title": "Post-Training Isaac GR00T N1.5 for LeRobot SO-101 Arm", "authors": ["You Liang Tan", "Fengyuan Hu"], "year": 2025, "venue": "Hugging Face blog", "type": "blog", "url": "https://huggingface.co/blog/nvidia/gr00t-n1-5-so101-tuning", "pdf_url": "", "code_url": "", "clusters": ["C3"], "keywords": ["GR00T N1.5", "SO-101", "post-training tutorial", "LoRA"], "model_class": "VLA", "techniques": ["partial fine-tuning", "LoRA", "freezing diffusion model"], "base_model": "GR00T N1.5", "params": "", "hardware": "single RTX 4080-class GPU", "reported_metric": "~25GB VRAM default; ~15GB with diffusion-model tuning disabled", "metric_measures": "relative memory", "benchmarks": ["SO-101 real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "NVIDIA/HF tutorial for post-training GR00T N1.5 on 80 SO-101 episodes, with VRAM-saving options (freeze diffusion model, LoRA).", "relevance": 3, "relevance_reason": "Practical memory budget for humanoid-VLA fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: GR00T N1 post-training fine-tuning humanoid LoRA memory NVIDIA"}
{"id": "2512.16911", "title": "Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning", "authors": ["Andrew Wagenmaker", "Perry Dong", "Raymond Tsao", "Chelsea Finn", "Sergey Levine"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.16911", "pdf_url": "https://arxiv.org/pdf/2512.16911", "code_url": "", "clusters": ["C3"], "keywords": ["posterior BC", "pretraining for RL finetuning"], "model_class": "diffusion/flow policy", "techniques": ["BC pretraining design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Pretrains BC policies to model posterior uncertainty, making subsequent RL fine-tuning more sample-efficient.", "relevance": 3, "relevance_reason": "Improves RL fine-tuning efficiency via initialization.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "10.1145/3768622", "title": "Real-time, Work-conserving GPU Scheduling for Concurrent DNN Inference", "authors": ["Mingcong Han", "Rong Chen", "Weihang Shen", "Hanze Zhang", "Jinrong Yang", "Haibo Chen"], "year": 2025, "venue": "ACM TOCS (2025)", "type": "paper", "url": "https://doi.org/10.1145/3768622", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["real-time GPU scheduling", "Reef"], "model_class": "other", "techniques": ["GPU kernel preemption", "DNN serving"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reef: GPU DNN serving with microsecond-scale reset-based kernel preemption so real-time inference preempts best-effort work while keeping utilization high.", "relevance": 3, "relevance_reason": "GPU preemption mechanism enabling bounded latency for critical inference.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2511.04665", "title": "Real-to-Sim Robot Policy Evaluation with Gaussian Splatting Simulation of Soft-Body Interactions", "authors": ["Zhang, Kaifeng", "Sha, Shuo", "Jiang, Hanxiao", "Loper, Matthew", "Song, Hyunjong", "Cai, Guangyan", "Xu, Zhuo", "Hu, Xiaochen", "Zheng, Changxi", "Li, Yunzhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.04665", "pdf_url": "https://arxiv.org/pdf/2511.04665", "code_url": "", "clusters": ["C4"], "keywords": ["Gaussian splatting", "soft-body digital twin", "real-to-sim evaluation"], "model_class": "other", "techniques": [], "base_model": "ACT, Diffusion Policy, pi0, SmolVLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["deformable tasks", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Soft-body digital twins rendered with 3D Gaussian Splatting for evaluating policies on deformable tasks, correlating with real execution.", "relevance": 3, "relevance_reason": "Real-to-sim evaluation for deformables.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-to-sim Gaussian splatting policy evaluation"}
{"id": "2512.19562", "title": "REALM: A Real-to-Sim Validated Benchmark for Generalization in Robotic Manipulation", "authors": ["Martin Sedlacek", "Pavlo Yefanov", "Georgy Ponimatkin", "Jai Bardhan", "Simon Pilc", "Mederic Fourmy", "Evangelos Kazakos", "Cees G. M. Snoek", "Josef Sivic", "Vladimir Petrik"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2512.19562", "pdf_url": "https://arxiv.org/pdf/2512.19562", "code_url": "", "clusters": ["C4"], "keywords": ["REALM", "real-to-sim validated", "perturbations"], "model_class": "VLA", "techniques": [], "base_model": "pi0, pi0-FAST, GR00T N1.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["REALM"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "High-fidelity simulator with 15 perturbation factors, aligned control, and validated sim-real correlation for VLA generalization testing.", "relevance": 3, "relevance_reason": "Validated sim proxy for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA evaluation simulation real correlation"}
{"id": "2510.12710", "title": "Reflection-Based Task Adaptation for Self-Improving VLA", "authors": ["Baicheng Li", "Dong Wu", "Zi-Ke Yan", "Xinchen Liu", "Ze-Cui Zeng", "Lusong Li", "Hongbin Zha"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2510.12710", "pdf_url": "https://arxiv.org/pdf/2510.12710", "code_url": "", "clusters": ["C3"], "keywords": ["self-improvement", "efficient adaptation"], "model_class": "VLA", "techniques": ["reflective RL", "quality-guided SFT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Self-improving loop that synthesizes dense rewards from failure analysis and imitates good successes to adapt VLAs quickly in-situ.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2510.09976", "title": "Reinforcement Fine-Tuning of Flow-Matching Policies for Vision-Language-Action Models", "authors": ["Mingyang Lyu", "Yinqian Sun", "Erliang Lin", "Huangrui Li", "Ruolin Chen", "Feifei Zhao", "Yi Zeng"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.09976", "pdf_url": "https://arxiv.org/pdf/2510.09976", "code_url": "", "clusters": ["C3"], "keywords": ["flow policy optimization", "FPO", "pi0", "online RL"], "model_class": "VLA", "techniques": ["online RL", "flow-matching likelihood surrogate"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "ALOHA sim"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Makes PPO-style RL tractable for flow-matching VLAs by replacing likelihood ratios with per-sample flow-loss changes.", "relevance": 3, "relevance_reason": "RL fine-tuning for flow VLAs.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: reinforcement learning fine-tuning VLA flow matching pi0 arXiv 2025"}
{"id": "2511.20633", "title": "Reinforcing Action Policies by Prophesying", "authors": ["Jiahui Zhang", "Ze Huang", "Chun Gu", "Zipei Ma", "Li Zhang", "Zi-Pei Ma"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.20633", "pdf_url": "https://arxiv.org/pdf/2511.20633", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["ProphRL", "RL fine-tuning", "learned simulator"], "model_class": "world model", "techniques": ["world model simulator", "Flow-GRPO", "world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Pretrained action-to-video world model (Prophet) as a rollout simulator plus a stabilized Flow-GRPO variant for data- and compute-efficient VLA post-training.", "relevance": 3, "relevance_reason": "Context: world-model RL post-training (links=15).", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=100]; snowball round 2: cites 2306.03310, 2402.15391"}
{"id": "2512.24638", "title": "Resolving State Ambiguity in Robot Manipulation via Adaptive Working Memory Recoding", "authors": ["Qingda Hu", "Ziheng Qiu", "Zi-Jun Xu", "Kaizhao Zhang", "Xizhou Bu", "Zuo-Lei Sun", "Bo Zhang", "Jie-Ru Zhao", "Zhongxue Gan", "Wenchao Ding"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2512.24638", "pdf_url": "https://arxiv.org/pdf/2512.24638", "code_url": "", "clusters": ["C1"], "keywords": ["PAM", "long history"], "model_class": "diffusion/flow policy", "techniques": ["feature caching", "working memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">20 Hz inference with ~10 s history", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Visuomotor policy with cached context features as working memory, supporting a 300-frame history at high inference speed.", "relevance": 3, "relevance_reason": "Long-history policy kept real-time.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2509.21243", "title": "RetoVLA: Reusing Register Tokens for Spatial Reasoning in Vision-Language-Action Models", "authors": ["Jiyeon Koo", "Taewan Cho", "Hyunjoon Kang", "Eunseom Pyo", "Tae Gyun Oh", "Taeryang Kim", "Andrew Jaeyong Choi", "Hyun-Soo Kang", "T. Oh"], "year": 2025, "venue": "2026 IEEE International Conference on Robotics and Automation (ICRA)", "type": "preprint", "url": "https://arxiv.org/abs/2509.21243", "pdf_url": "https://arxiv.org/pdf/2509.21243", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["lightweight VLA", "spatial reasoning", "compressed VLA"], "model_class": "VLA", "techniques": ["register tokens", "register token reuse"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Reuses ViT register tokens as global spatial context injected into the action module of lightweight VLAs without adding parameters.", "relevance": 3, "relevance_reason": "Recovers spatial reasoning in small VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=100]; snowball round 2: cites 2106.09685, 2303.04137"}
{"id": "2507.21452", "title": "Retrieve-Augmented Generation for Speeding up Diffusion Policy without Additional Training", "authors": ["Sodtavilan Odonchimed", "Tatsuya Matsushima", "Simon Holk", "Yusuke Iwasawa", "Yutaka Matsuo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.21452", "pdf_url": "https://arxiv.org/pdf/2507.21452", "code_url": "", "clusters": ["C1"], "keywords": ["retrieval", "warm start", "training-free"], "model_class": "diffusion/flow policy", "techniques": ["retrieval-augmented initialization", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "at 20x acceleration, +7% accuracy over consistency policy", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Retrieves the most similar expert action from a demonstration database and starts denoising from an intermediate step.", "relevance": 3, "relevance_reason": "Training-free diffusion acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query: consistency policy"}
{"id": "2509.15965", "title": "RLinf: Flexible and Efficient Large-scale Reinforcement Learning via Macro-to-Micro Flow Transformation", "authors": ["Chao Yu", "Yuanqing Wang", "Zhen Guo", "Hao Lin", "Si Xu", "Hongzhi Zang", "Quanlu Zhang", "Yongji Wu", "Chunyang Zhu", "Junhao Hu", "Zixiao Huang", "Mingjie Wei", "Yuqing Xie", "Ke Yang", "Bo Dai", "Zhexuan Xu", "Jiakun Du", "Xiangyuan Wang", "Xu Fu", "Letong Shi", "Zhihao Liu", "Kang Chen", "Weilin Liu", "Gang Liu", "Boxun Li", "Jianlei Yang", "Zhi Yang", "Guohao Dai", "Yu Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.15965", "pdf_url": "https://arxiv.org/pdf/2509.15965", "code_url": "", "clusters": ["C3"], "keywords": ["RL system", "M2Flow", "elastic pipelining"], "model_class": "other", "techniques": ["RL training system", "scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.07x-2.43x end-to-end training throughput", "metric_measures": "throughput", "benchmarks": ["embodied RL", "reasoning RL"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "RL training system that decomposes and reschedules RL workflows for better hardware utilization, used for embodied VLA RL.", "relevance": 3, "relevance_reason": "Training-throughput infrastructure for VLA RL.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RLinf VLA reinforcement learning infrastructure embodied arXiv"}
{"id": "2502.07837", "title": "RoboBERT: An End-to-end Multimodal Robotic Manipulation Model", "authors": ["Sicheng Wang", "Sheng Liu", "Jianhua Shan", "Jianwei Zhang", "Hao Gao", "Hai-Song Han", "Yi-Peng Chen", "Kang Wei", "Chengkun Zhang", "Kai-Kit Wong", "Jie Zhao", "Lei Zhao", "Bin Fang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.07837", "pdf_url": "https://arxiv.org/pdf/2502.07837", "code_url": "", "clusters": ["C3"], "keywords": ["lightweight", "training efficiency"], "model_class": "diffusion/flow policy", "techniques": ["two-stage training", "lightweight architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RoboBERT is a lightweight end-to-end manipulation model trained in two stages (fixed instruction, then diverse phrasings) to avoid heavy fine-tuning costs while reaching strong CALVIN results.", "relevance": 3, "relevance_reason": "relevant: lower training cost, lightweight model", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2411.19650; cited by 2508.21046, 2602.00686"}
{"id": "2506.17811", "title": "RoboMonkey: Scaling Test-Time Sampling and Verification for Vision-Language-Action Models", "authors": ["Jacky Kwok", "Christopher Agia", "Rohan Sinha", "Matthew Foutter", "Shulu Li", "Ion Stoica", "Azalia Mirhoseini", "Marco Pavone"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.17811", "pdf_url": "https://arxiv.org/pdf/2506.17811", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["RoboMonkey", "test-time scaling", "sampling and verification", "inference-time scaling law"], "model_class": "VLA", "techniques": ["Gaussian perturbation sampling", "VLM action verifier"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows action error follows an inference-time scaling law in number of samples and uses sampling plus a learned verifier at deployment to boost robustness.", "relevance": 3, "relevance_reason": "Robustness vs latency tradeoff of runtime verification.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2410.04640 (S2)"}
{"id": "2512.10394", "title": "RoboNeuron: A Middle-Layer Infrastructure for Agent-Driven Orchestration in Embodied AI", "authors": ["Weifan Guan", "Qinghao Hu", "Hua-Sen Xi", "Chenxiao Zhang", "Aosheng Li", "Jian Cheng"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.10394", "pdf_url": "https://arxiv.org/pdf/2512.10394", "code_url": "", "clusters": ["C5", "C6"], "keywords": ["ROS2", "MCP", "serving stack"], "model_class": "VLA", "techniques": ["middleware"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Middleware bridging agent tool APIs and ROS2, isolating VLA backend/runtime/acceleration changes.", "relevance": 3, "relevance_reason": "Deployment/serving infrastructure.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2509.19658", "title": "RoboSSM: Scalable In-context Imitation Learning via State-Space Models", "authors": ["Youngju Yoo", "Jiaheng Hu", "Yifeng Zhu", "Bo Liu", "Qiang Liu", "Roberto Martín-Martín", "Peter Stone"], "year": 2025, "venue": "IROS 2026", "type": "paper", "url": "https://arxiv.org/abs/2509.19658", "pdf_url": "https://arxiv.org/pdf/2509.19658", "code_url": "https://github.com/youngjuY/RoboSSM", "clusters": ["C1"], "keywords": ["SSM", "in-context imitation", "Longhorn"], "model_class": "other", "techniques": ["state space model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces transformers with a Longhorn SSM for in-context imitation learning, giving linear-time inference and longer-prompt extrapolation.", "relevance": 3, "relevance_reason": "SSM policy.", "foundational_technique": false, "verified": true, "found_via": "search query: state space model imitation learning"}
{"id": "2506.18088", "title": "RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation", "authors": ["Chen, Tianxing", "Chen, Zanxin", "Chen, Baijun", "Cai, Zijian", "Liu, Yibin", "Li, Zixuan", "Liang, Qiwei", "Lin, Xianliang", "Ge, Yiheng", "Gu, Zhenyu", "Deng, Weiliang", "Guo, Yubin", "Nian, Tian", "Xie, Xuanbing", "Chen, Qiangyu", "Su, Kailun", "Xu, Tianling", "Liu, Guodong", "Hu, Mengkang", "Gao, Huan-ang", "Wang, Kaixuan", "Liang, Zhixuan", "Qin, Yusen", "Yang, Xiaokang", "Luo, Ping", "Mu, Yao"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2506.18088", "pdf_url": "https://arxiv.org/pdf/2506.18088", "code_url": "", "clusters": ["C4"], "keywords": ["RoboTwin 2.0", "bimanual", "domain randomization", "data generation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Scalable bimanual data generator and benchmark with 50 tasks, five embodiments and five-axis domain randomization with unified evaluation protocols.", "relevance": 3, "relevance_reason": "Widely used VLA sim benchmark.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: RoboCasa VLABench RoboTwin"}
{"id": "2512.08333", "title": "Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging", "authors": ["Yajat Yadav", "Zhiyuan Zhou", "Andrew Wagenmaker", "Karl Pertsch", "Sergey Levine"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.08333", "pdf_url": "https://arxiv.org/pdf/2512.08333", "code_url": "", "clusters": ["C3"], "keywords": ["weight interpolation", "model merging", "robust finetuning"], "model_class": "VLA", "techniques": ["model merging", "weight interpolation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Interpolates fine-tuned and pretrained VLA weights to add a new skill while keeping generalist abilities, at no extra inference cost.", "relevance": 3, "relevance_reason": "Cheap robust adaptation via merging.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: real-world RL fine-tuning generalist robot policy human-in-the-loop VLA arXiv"}
{"id": "2510.09036", "title": "RoDyn: Taming Interactive Robot-Dynamic 2.5D World Model for Robotic Manipulation", "authors": ["Chuanrui Zhang", "Zhengxian Wu", "Guanxing Lu", "Yan-Song Tang", "Ziwei Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.09036", "pdf_url": "https://arxiv.org/pdf/2510.09036", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["2.5D world model", "efficient latent space", "robot-dynamic tokenizer"], "model_class": "world model", "techniques": ["geometry-aware latent dynamics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RoDyn models robot-object dynamics in a compact geometry-aware 2.5D latent space rather than full 2D video.", "relevance": 3, "relevance_reason": "Efficient world model representation.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2509.25161", "title": "Rolling Forcing: Autoregressive Long Video Diffusion in Real Time", "authors": ["Kunhao Liu", "Wenbo Hu", "Jiale Xu", "Ying Shan", "Shijian Lu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.25161", "pdf_url": "https://arxiv.org/pdf/2509.25161", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["Rolling Forcing", "attention sink", "streaming long video"], "model_class": "world model", "techniques": ["joint denoising with increasing noise levels", "attention sink", "few-step distillation over long windows"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Real-time streaming long-video generation that denoises a window of frames at staggered noise levels with an attention sink, limiting error accumulation.", "relevance": 3, "relevance_reason": "Foundational streaming technique mirrored by Rolling-WAM.", "foundational_technique": true, "verified": true, "found_via": "search query"}
{"id": "2510.10975", "title": "RoVer: Robot Reward Model as Test-Time Verifier for Vision-Language-Action Model", "authors": ["Mingtong Dai", "Lingbo Liu", "Yongjie Bai", "Yang Liu", "Zhouxia Wang", "Rui SU", "Chunjie Chen", "Liang Lin", "Xinyu Wu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.10975", "pdf_url": "https://arxiv.org/pdf/2510.10975", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["RoVer"], "model_class": "VLA", "techniques": ["process reward model", "test-time verification", "feature caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Robot process reward model used as a test-time verifier that scores and refines candidate VLA actions, caching shared perception features to evaluate more candidates.", "relevance": 3, "relevance_reason": "Verifier-based action selection with amortized cost.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\")"}
{"id": "2503.06736", "title": "Safe, Task-Consistent Manipulation with Operational Space Control Barrier Functions", "authors": ["Daniel Morton", "Marco Pavone"], "year": 2025, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "type": "paper", "url": "https://arxiv.org/abs/2503.06736", "pdf_url": "https://arxiv.org/pdf/2503.06736", "code_url": "", "clusters": ["C7"], "keywords": ["safety filter", "real-time"], "model_class": "other", "techniques": ["control barrier functions", "operational space control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Operational-space CBF safety filters for manipulators that enforce many constraints cheaply while keeping task consistency.", "relevance": 3, "relevance_reason": "Real-time safety filtering (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2510.08464, 2511.06385"}
{"id": "2507.13871", "title": "Safety Certification in the Latent space using Control Barrier Functions and World Models", "authors": ["Mehul Anand", "Shishir Kolathaya"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.13871", "pdf_url": "https://arxiv.org/pdf/2507.13871", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["control barrier certificates", "latent space", "semi-supervised safety"], "model_class": "world model", "techniques": ["neural barrier + safe controller learned in world-model latent space"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Learns control barrier certificates and a safe controller in a world model's latent space with little labeled safety data.", "relevance": 3, "relevance_reason": "World-model latent safety certification.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.00935 (citations)"}
{"id": "2509.12674", "title": "Safety filtering of robotic manipulation under environment uncertainty: a computational approach", "authors": ["A. Johansson", "Daniel M. Lindmark", "Viktor Wiberg", "Martin Servin"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.12674", "pdf_url": "https://arxiv.org/pdf/2509.12674", "code_url": "", "clusters": ["C7"], "keywords": ["uncertainty", "safety filtering"], "model_class": "other", "techniques": ["simulation-based safety filter", "parallel re-evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Physics-simulation safety filter that densely rolls out nominal parameters and sparsely re-evaluates critical transitions under uncertainty.", "relevance": 3, "relevance_reason": "Safety filtering (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2309.05837"}
{"id": "2501.14009", "title": "Scalable and Interpretable Verification of Image-based Neural Network Controllers for Autonomous Vehicles", "authors": ["Aditya Parameshwaran", "Yue Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2501.14009", "pdf_url": "https://arxiv.org/pdf/2501.14009", "code_url": "", "clusters": ["C7"], "keywords": ["image-based NN controllers", "autonomous vehicles"], "model_class": "other", "techniques": ["VAE", "polytope verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Encodes images with a VAE into an interpretable latent space and builds convex polytopes there for scalable verification of image-based driving controllers.", "relevance": 3, "relevance_reason": "Latent-space verification of vision controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2511.11520", "title": "Scalable Policy Evaluation with Video World Models", "authors": ["Wei-Cheng Tseng", "Jinwei Gu", "Qinsheng Zhang", "Hanzi Mao", "Ming-Yu Liu", "Florian Shkurti", "Lin Yen-Chen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.11520", "pdf_url": "https://arxiv.org/pdf/2511.11520", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["policy evaluation", "action-conditioned video generation", "pretrained video models"], "model_class": "world model", "techniques": ["adding action conditioning to pretrained video generators"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["policy ranking vs real"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "NVIDIA study adding action conditioning to pretrained video generators for robot policy evaluation, analyzing data diversity, pretraining and failure cases.", "relevance": 3, "relevance_reason": "World models as cheap evaluators.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2509.19524", "title": "Score the Steps, Not Just the Goal: VLM-Based Subgoal Evaluation for Robotic Manipulation", "authors": ["Ramy Elmallah", "Krish Chhajer", "Chi-Guhn Lee"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.19524", "pdf_url": "https://arxiv.org/pdf/2509.19524", "code_url": "", "clusters": ["C4"], "keywords": ["StepEval", "VLM judge"], "model_class": "other", "techniques": ["subgoal evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proposes reporting per-subgoal success via a lightweight VLM-based evaluator instead of a single success rate, also tracking evaluation latency and cost.", "relevance": 3, "relevance_reason": "Evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941; cited by 2607.01088"}
{"id": "2511.08583", "title": "SeFA-Policy: Fast and Accurate Visuomotor Policy Learning with Selective Flow Alignment", "authors": ["Rong Xue", "Jiageng Mao", "Mingtong Zhang", "Yue Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.08583", "pdf_url": "https://arxiv.org/pdf/2511.08583", "code_url": "https://github.com/RongXueZoe/SeFA", "clusters": ["C1"], "keywords": ["rectified flow", "reflow", "one-step"], "model_class": "diffusion/flow policy", "techniques": ["rectified flow", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">98% lower inference latency", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Corrects observation drift introduced by reflow distillation using expert demonstrations, keeping one-step rectified-flow inference accurate.", "relevance": 3, "relevance_reason": "One-step rectified flow policy.", "foundational_technique": false, "verified": true, "found_via": "search query: rectified flow policy robot"}
{"id": "2509.19292", "title": "SOE: Sample-Efficient Robot Policy Self-Improvement via On-Manifold Exploration", "authors": ["Yang Jin", "Jun Lv", "Han Xue", "Wendi Chen", "Chuan Wen", "Ce-Wu Lu"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.19292", "pdf_url": "https://arxiv.org/pdf/2509.19292", "code_url": "", "clusters": ["C3"], "keywords": ["SOE", "self-improvement"], "model_class": "other", "techniques": ["on-manifold exploration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Constrains exploration to a learned compact latent manifold for safer, more sample-efficient policy self-improvement.", "relevance": 3, "relevance_reason": "Sample-efficient policy improvement.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2505.22179", "title": "Speculative Decoding Meets Quantization: Compatibility Evaluation and Hierarchical Framework Design", "authors": ["Yudi Zhang", "Weilin Zhao", "Xu Han", "Tiejun Zhao", "Wang Xu", "Hailong Cao", "Conghui Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.22179", "pdf_url": "https://arxiv.org/pdf/2505.22179", "code_url": "https://github.com/AI9Stars/SpecMQuant", "clusters": ["C2"], "keywords": ["EAGLE-2", "4-bit weights", "tree drafts", "compatibility"], "model_class": "other", "techniques": ["speculative decoding", "weight quantization"], "base_model": "Llama-3-70B (4-bit)", "params": "", "hardware": "A100", "reported_metric": "hierarchical scheme 2.78x speedup on W4 Llama-3-70B, 1.31x over EAGLE-2", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Finds tree-style speculative verification erodes the memory-bound gains of 4-bit weights, then fixes it with an intermediate small model turning tree drafts into sequence drafts.", "relevance": 3, "relevance_reason": "Direct evidence that speculative decoding and quantization interfere; relevant to VLA spec-decoding + quant stacks.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q1: arXiv id lookup of known speculative-decoding x quantization composition works"}
{"id": "2512.00062", "title": "SpeedAug: Policy Acceleration via Tempo-Enriched Policy and RL Fine-Tuning", "authors": ["Taewook Nam", "Junmo Cho", "Youngsoo Jang", "S. J. Hwang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2512.00062", "pdf_url": "https://arxiv.org/pdf/2512.00062", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["SpeedAug", "policy acceleration"], "model_class": "other", "techniques": ["RL fine-tuning", "speed augmentation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.8x task throughput with 16 min online interaction", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns task-optimal execution tempo by pretraining on speed-augmented demos and RL fine-tuning, speeding up real task execution.", "relevance": 3, "relevance_reason": "Policy execution speed-up.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2511.15605", "title": "SRPO: Self-Referential Policy Optimization for Vision-Language-Action Models", "authors": ["Senyu Fei", "Siyin Wang", "Li Ji", "Ao Li", "Shiduo Zhang", "Liming Liu", "Jinlong Hou", "Jingjing Gong", "Xianzhong Zhao", "Xipeng Qiu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.15605", "pdf_url": "https://arxiv.org/pdf/2511.15605", "code_url": "", "clusters": ["C3"], "keywords": ["SRPO", "self-referential", "dense rewards"], "model_class": "VLA", "techniques": ["GRPO-style RL", "latent world representations"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses the model's own successes as references to reward failed trajectories, improving VLA-RL training efficiency.", "relevance": 3, "relevance_reason": "Improves RL post-training efficiency.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"world model\" AND abs:\"post-training\""}
{"id": "2510.05057", "title": "StaMo: Unsupervised Learning of Generalizable Robot Motion from Compact State Representation", "authors": ["Mingyu Liu", "Jiuhe Shu", "Hui Chen", "Zeju Li", "Can-Yu Zhao", "Jiange Yang", "Shenyuan Gao", "Hao Chen", "Chunhua Shen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.05057", "pdf_url": "https://arxiv.org/pdf/2510.05057", "code_url": "", "clusters": ["C8"], "keywords": ["compact representation", "latent action"], "model_class": "world model", "techniques": ["compact state representation", "latent action"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "StaMo learns a two-token compressed state representation with a lightweight encoder and DiT decoder; token differences act as latent actions and improve VLAs with minimal inference overhead.", "relevance": 3, "relevance_reason": "relevant: compact state for efficient world modeling", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310; cited by 2602.06949"}
{"id": "2511.07899", "title": "Statistically Assuring Safety of Control Systems using Ensembles of Safety Filters and Conformal Prediction", "authors": ["Ihab Tabbara", "Yuxuan Yang", "Hussein Sibai"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.07899", "pdf_url": "https://arxiv.org/pdf/2511.07899", "code_url": "", "clusters": ["C7"], "keywords": ["safety filter"], "model_class": "other", "techniques": ["HJ reachability", "conformal prediction", "ensembles"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses conformal prediction and ensembles of learned HJ safety filters to give probabilistic safety guarantees despite errors in learned value functions.", "relevance": 3, "relevance_reason": "Statistical assurance for learned safety filters.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:conformal AND (ti:robot* OR ti:planning OR ti:control)"}
{"id": "2507.05240", "title": "StreamVLN: Streaming Vision-and-Language Navigation via SlowFast Context Modeling", "authors": ["Meng Wei", "Chen-Yang Wan", "Xiqian Yu", "Tai Wang", "Yu-Qiang Yang", "Xiaohan Mao", "Chen-Ming Zhu", "Wenzhe Cai", "Han-Qing Wang", "Yi-Lun Chen", "Xi-Hui Liu", "Jiang-Miao Pang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.05240", "pdf_url": "https://arxiv.org/pdf/2507.05240", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["streaming VLN", "slow-fast context", "KV cache"], "model_class": "VLA", "techniques": ["slow-fast context modeling", "token pruning of memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "StreamVLN handles continuous video for navigation with a fast sliding-window dialogue context plus a compressed slow memory for low latency.", "relevance": 3, "relevance_reason": "Efficient streaming embodied model.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'vision-language-action real-time inference latency'"}
{"id": "10.1109/robio66223.2025.11379720", "title": "Structurally-Fused Goal-Aware Network Distillation for Anomaly Detection in Diffusion Policy", "authors": ["Zherui Song", "Yifei Yang", "Zhongxiang Zhou", "Hongxiang Yu", "Xingwen Zheng", "Yue Wang", "Rong Xiong"], "year": 2025, "venue": "IEEE International Conference on Robotics and Biomimetics", "type": "paper", "url": "https://doi.org/10.1109/ROBIO66223.2025.11379720", "pdf_url": "", "code_url": "", "clusters": ["C7"], "keywords": ["anomaly detection", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["network distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Goal-aware teacher-student distillation for low-latency anomaly detection during diffusion-policy execution.", "relevance": 3, "relevance_reason": "Runtime failure detection with minimal latency", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2509.19486", "title": "Supercomputing for High-speed Avoidance and Reactive Planning in Robots", "authors": ["Kieran S. Lachmansingh", "José R. González-Estrada", "Jacob Chisholm", "Ryan E. Grant", "Matthew K. X. J. Pan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.19486", "pdf_url": "https://arxiv.org/pdf/2509.19486", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["HPC offloading", "reactive planning", "remote cluster"], "model_class": "other", "techniques": ["HPC offloading", "distributed A* (MPI)"], "base_model": "", "params": "", "hardware": "local and remote (~300 km) HPC clusters", "reported_metric": "mean planning latency 22.9 ms local / 30.0 ms remote", "metric_measures": "other", "benchmarks": ["7-DOF arm projectile avoidance"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Proof of concept showing a remote HPC cluster can plan projectile-avoidance motions for a 7-DOF arm within tens of milliseconds, motivating onboard reflexes plus offloaded bursty planning.", "relevance": 3, "relevance_reason": "Quantifies when remote compute suffices for reactive control.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2506.03574", "title": "SwitchVLA: Execution-Aware Task Switching for Vision-Language-Action Models", "authors": ["Meng Li", "Zhen Zhao", "Zhengping Che", "Fei Liao", "Kun Wu", "Zhiyuan Xu", "Pei Ren", "Zhao Jin", "Ning Liu", "Jian Tang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.03574", "pdf_url": "https://arxiv.org/pdf/2506.03574", "code_url": "", "clusters": ["C1", "C7"], "keywords": ["interruption", "task switching", "VLA"], "model_class": "VLA", "techniques": ["execution-aware conditioning", "task switching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lets a VLA switch tasks when new instructions arrive mid-execution by conditioning on contact-phase execution state instead of assuming static intent.", "relevance": 3, "relevance_reason": "Mid-execution interruption/switching of VLA behavior.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"action chunking\" AND (\"closed-loop\" OR reactiv*)"}
{"id": "2511.04421", "title": "Temporal Action Selection for Action Chunking", "authors": ["Yueyang Weng", "Xiaopeng Zhang", "Yongjin Mu", "Yingcong Zhu", "Yanjie Li", "Qi Liu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.04421", "pdf_url": "https://arxiv.org/pdf/2511.04421", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["temporal action selection", "action chunking", "reactivity vs consistency"], "model_class": "diffusion/flow policy", "techniques": ["multi-chunk action selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Caches predicted chunks from multiple steps and learns to select among them each timestep, aiming for both reactivity and decision consistency.", "relevance": 3, "relevance_reason": "Reactive chunk execution.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2503.02511", "title": "TeTRA-VPR: A Ternary Transformer Approach for Compact Visual Place Recognition", "authors": ["Oliver Grainge", "Michael Milford", "Indu Bodala", "Sarvapali D. Ramchurn", "Shoaib Ehsan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.02511", "pdf_url": "https://arxiv.org/pdf/2503.02511", "code_url": "", "clusters": ["C2"], "keywords": ["ternary ViT", "2-bit", "binary embedding", "place recognition"], "model_class": "other", "techniques": ["ternary/2-bit quantization", "progressive distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Progressively quantizes a ViT backbone to ternary/2-bit and binarizes its embedding with distillation for compact visual place recognition on drones and mobile robots.", "relevance": 3, "relevance_reason": "Ultra-low-bit transformer for robot perception.", "foundational_technique": false, "verified": true, "found_via": "search: 1-bit robot"}
{"id": "2506.08440", "title": "TGRPO :Fine-tuning Vision-Language-Action Model via Trajectory-wise Group Relative Policy Optimization", "authors": ["Zengjue Chen", "Runliang Niu", "He Kong", "Qi Wang", "Qianli Xing", "Zipei Fan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.08440", "pdf_url": "https://arxiv.org/pdf/2506.08440", "code_url": "", "clusters": ["C3"], "keywords": ["TGRPO", "trajectory-wise GRPO"], "model_class": "VLA", "techniques": ["GRPO", "online RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Trajectory-level group relative policy optimization for fine-tuning VLAs without a critic.", "relevance": 3, "relevance_reason": "Critic-free RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2506.09422", "title": "Time-Unified Diffusion Policy with Action Discrimination for Robotic Manipulation", "authors": ["Ye Niu", "Sanpin Zhou", "Yizhe Li", "De-Hao Ye", "Le Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.09422", "pdf_url": "https://arxiv.org/pdf/2506.09422", "code_url": "", "clusters": ["C1"], "keywords": ["fast denoising", "RLBench"], "model_class": "diffusion/flow policy", "techniques": ["time-unified velocity field"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RLBench"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Time-unified denoising velocity field that speeds up diffusion-policy action generation, with larger gains at fewer iterations.", "relevance": 3, "relevance_reason": "Faster diffusion-policy inference.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2505.18227", "title": "Token Reduction Should Go Beyond Efficiency in Generative Models - From Vision, Language to Multimodality", "authors": ["Zheng-Lun Kong", "Yi-Ze Li", "Fan-Hu Zeng", "Lei Xin", "Shvat Messica", "Xue Lin", "Pu Zhao", "M. Kellis", "Hao Tang", "M. Zitnik"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.18227", "pdf_url": "https://arxiv.org/pdf/2505.18227", "code_url": "", "clusters": ["C1"], "keywords": ["position paper", "multimodal"], "model_class": "other", "techniques": ["token reduction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Argues token reduction in generative models should serve goals beyond efficiency across vision, language and multimodal models.", "relevance": 3, "relevance_reason": "Foundational token-reduction perspective cited by core items.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2506.10100, 2509.12594"}
{"id": "2509.14630", "title": "Toward Embodiment Equivariant Vision-Language-Action Policy", "authors": ["Anzhe Chen", "Yifei Yang", "Zhenjie Zhu", "Kechun Xu", "Zhongxiang Zhou", "Rong Xiong", "Yue Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.14630", "pdf_url": "https://arxiv.org/pdf/2509.14630", "code_url": "https://github.com/hhcaz/e2vla", "clusters": ["C3"], "keywords": ["embodiment equivariance", "cross-embodiment", "action decoder"], "model_class": "VLA", "techniques": ["equivariant action decoder", "geometry-aware architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Designs VLA action spaces equivariant to embodiment configuration changes so fine-tuning to new robots is cheaper.", "relevance": 3, "relevance_reason": "Reduces cross-embodiment adaptation cost.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: few-shot adaptation vision-language-action model new embodiment efficient"}
{"id": "2511.14178", "title": "Towards Deploying VLA without Fine-Tuning: Plug-and-Play Inference-Time VLA Policy Steering via Embodied Evolutionary Diffusion", "authors": ["Zhuo Li", "Junjia Liu", "Zhipeng Dong", "Tao Teng", "Quentin Rouxel", "Darwin Caldwell", "Fei Chen"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.14178", "pdf_url": "https://arxiv.org/pdf/2511.14178", "code_url": "", "clusters": ["C3"], "keywords": ["VLA-Pilot", "inference-time steering", "no fine-tuning"], "model_class": "VLA", "techniques": ["inference-time steering", "evolutionary diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Plug-and-play inference-time steering for zero-shot deployment of VLAs without fine-tuning or data collection.", "relevance": 3, "relevance_reason": "Adaptation with no training cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2512.01715", "title": "Transport Discrepancy as a Reliability Signal for Vision-Language-Action Models", "authors": ["Wan-Peng Zhang", "Ye Wang", "Hao Luo", "Haoqi Yuan", "Yicheng Feng", "Chao-Yi Xu", "Sipeng Zheng", "Qin Jin", "Zongqing Lu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.01715", "pdf_url": "https://arxiv.org/pdf/2512.01715", "code_url": "", "clusters": ["C7"], "keywords": ["reliability signal", "drift detection"], "model_class": "VLA", "techniques": ["reliability gating"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses transport discrepancy between backbone and action-expert features as a per-step reliability signal and refinement gate in flow VLAs.", "relevance": 3, "relevance_reason": "Runtime reliability monitoring.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2507.01424", "title": "TriVLA: A Triple-System-Based Unified Vision-Language-Action Model with Episodic World Modeling for General Robot Control", "authors": ["Zhenyang Liu", "Yongchong Gu", "Sixiao Zheng", "Yanwei Fu", "Xiangyang Xue", "Yu-Gang Jiang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.01424", "pdf_url": "https://arxiv.org/pdf/2507.01424", "code_url": "", "clusters": ["C1"], "keywords": ["triple system", "episodic world model", "flow policy"], "model_class": "VLA", "techniques": ["multi-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~36 Hz operation", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adds a video-diffusion dynamics system to VLM + flow-policy systems for episodic world modeling, reported to run around 36 Hz.", "relevance": 3, "relevance_reason": "Multi-system VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: System 1 System 2 robot"}
{"id": "2511.05275", "title": "TwinVLA: Data-Efficient Bimanual Manipulation with Twin Single-Arm Vision-Language-Action Models", "authors": ["Hokyun Im", "Euijin Jeong", "Andrey Kolobov", "Jianlong Fu", "Youngwoon Lee"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.05275", "pdf_url": "https://arxiv.org/pdf/2511.05275", "code_url": "", "clusters": ["C3"], "keywords": ["TwinVLA", "bimanual", "modular composition"], "model_class": "VLA", "techniques": ["model composition"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Composes two pretrained single-arm VLAs into a bimanual VLA, needing far less bimanual data.", "relevance": 3, "relevance_reason": "Data-efficient embodiment adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2503.01876", "title": "Uncertainty Comes for Free: Human-in-the-Loop Policies with Diffusion Models", "authors": ["Zhan-Peng He", "Yifeng Cao", "M. Ciocarlie"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.01876", "pdf_url": "https://arxiv.org/pdf/2503.01876", "code_url": "", "clusters": ["C7"], "keywords": ["human-in-the-loop", "diffusion uncertainty", "help seeking"], "model_class": "diffusion/flow policy", "techniques": ["uncertainty from generative process"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Derives an uncertainty metric from the diffusion sampling process so policies request human help only when needed.", "relevance": 3, "relevance_reason": "Cheap uncertainty for fallback triggering.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2410.04640 (S2)"}
{"id": "2505.06111", "title": "UniVLA: Learning to Act Anywhere with Task-centric Latent Actions", "authors": ["Qingwen Bu", "Yan-Ting Yang", "Ji-Song Cai", "Shenyuan Gao", "Guanghui Ren", "Maoqing Yao", "Ping Luo", "Hong-Yang Li"], "year": 2025, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2505.06111", "pdf_url": "https://arxiv.org/pdf/2505.06111", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["latent actions", "cross-embodiment", "compute-efficient pretraining"], "model_class": "VLA", "techniques": ["latent action model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "less than 1/20 of pretraining compute and 1/10 downstream data vs OpenVLA", "metric_measures": "other", "benchmarks": ["manipulation benchmarks", "navigation benchmarks", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns task-centric latent actions from videos in DINO feature space so a generalist policy can be pretrained across embodiments and decoded cheaply to each robot.", "relevance": 3, "relevance_reason": "Heavily cited VLA that claims large pretraining-compute and data savings vs OpenVLA", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2601.00844", "title": "Value-guided action planning with JEPA world models", "authors": ["Matthieu Destrade", "Oumayma Bounou", "Quentin Le Lidec", "Jean Ponce", "Yann LeCun"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.00844", "pdf_url": "https://arxiv.org/pdf/2601.00844", "code_url": "", "clusters": ["C8"], "keywords": ["value-shaped latent space", "quasi-distance", "JEPA planning"], "model_class": "world model", "techniques": ["embedding distance approximates goal-conditioned value"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simple control tasks"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shapes JEPA embeddings so that latent distance approximates a goal-conditioned value, making simple distance-based planning far more effective.", "relevance": 3, "relevance_reason": "Improves latent planning quality cheaply.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.09985"}
{"id": "2508.01622", "title": "VFP: Variational Flow-Matching Policy for Multi-Modal Robot Manipulation", "authors": ["Xuanran Zhai", "Qian Zhao", "Qiao-Jun Yu", "Ce Hao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.01622", "pdf_url": "https://arxiv.org/pdf/2508.01622", "code_url": "", "clusters": ["C1"], "keywords": ["flow matching", "multi-modality"], "model_class": "diffusion/flow policy", "techniques": ["variational prior", "MoE decoder"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Variational flow-matching policy with OT alignment and MoE decoder capturing multi-modality while keeping fast inference and compact size.", "relevance": 3, "relevance_reason": "Efficient flow policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2601.21712"}
{"id": "2505.14357", "title": "Vid2World: Crafting Video Diffusion Models to Interactive World Models", "authors": ["Siqiao Huang", "Jialong Wu", "Qixing Zhou", "Shangchen Miao", "Mingsheng Long"], "year": 2025, "venue": "ICLR 2026", "type": "paper", "url": "https://arxiv.org/abs/2505.14357", "pdf_url": "https://arxiv.org/pdf/2505.14357", "code_url": "", "clusters": ["C8"], "keywords": ["causalization", "video diffusion to world model", "action guidance"], "model_class": "world model", "techniques": ["video diffusion causalization", "causal action guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["robot manipulation", "3D game", "navigation"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Converts pretrained bidirectional video diffusion models into autoregressive, action-controllable interactive world models by causalizing architecture and training objective.", "relevance": 3, "relevance_reason": "Standard route from video models to interactive world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2512.03044", "title": "Video2Act: A Dual-System Video Diffusion Policy with Robotic Spatio-Motional Modeling", "authors": ["Yue-Ru Jia", "Jiaming Liu", "Sheng-Bang Liu", "Rui Zhou", "Wanhe Yu", "Yu-Yang Yan", "Xiaowei Chi", "Yan-Dong Guo", "Bo-Xin Shi", "Shanghang Zhang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.03044", "pdf_url": "https://arxiv.org/pdf/2512.03044", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["dual-system", "System 1/System 2"], "model_class": "diffusion/flow policy", "techniques": ["asynchronous dual-system", "video diffusion features"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses a video diffusion model as a slow System 2 providing motion-aware conditions to a fast DiT action head that runs asynchronously.", "relevance": 3, "relevance_reason": "Asynchronous dual-system design to address VDM inference inefficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2609.28431"}
{"id": "2511.07732", "title": "ViPRA: Video Prediction for Robot Actions", "authors": ["Sandeep Routray", "Hengkai Pan", "Unnat Jain", "Shikhar Bahl", "Deepak Pathak"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2511.07732", "pdf_url": "https://arxiv.org/pdf/2511.07732", "code_url": "", "clusters": ["C1"], "keywords": ["video prediction", "high-frequency control"], "model_class": "VLA", "techniques": ["chunked flow-matching decoder", "latent actions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 22 Hz", "metric_measures": "closed-loop control", "benchmarks": ["SIMPLER", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "ViPRA pretrains a video-language model on latent actions from actionless videos and decodes them with a chunked flow-matching head for smooth high-frequency control.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2512.15258", "title": "VLA-AN: An Efficient and Onboard Vision-Language-Action Framework for Aerial Navigation in Complex Environments", "authors": ["Yuze Wu", "Mo Zhu", "Xingxing Li", "Yuheng Du", "Yuxin Fan", "Wenjun Li", "Zhichao Han", "Xin Zhou", "Fei Gao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.15258", "pdf_url": "https://arxiv.org/pdf/2512.15258", "code_url": "", "clusters": ["C1", "C3", "C5"], "keywords": ["aerial navigation", "onboard VLA", "UAV navigation", "onboard", "3D Gaussian Splatting data"], "model_class": "VLA", "techniques": ["lightweight action module", "deployment optimization", "deployment pipeline optimization", "geometric safety correction"], "base_model": "", "params": "", "hardware": "onboard UAV compute", "reported_metric": "8.3x inference throughput improvement onboard", "metric_measures": "throughput", "benchmarks": ["aerial navigation", "real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Aerial-navigation VLA with a lightweight action head, safety correction, and an optimized onboard pipeline that raises inference throughput on resource-constrained drones.", "relevance": 3, "relevance_reason": "Onboard VLA deployment on UAV with system-level optimization.", "foundational_technique": false, "verified": true, "found_via": "search query: navigation VLA efficient real-time; arXiv: vision-language-action AND Jetson/edge/onboard"}
{"id": "2512.22539", "title": "VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models", "authors": ["Borong Zhang", "Jiahao Li", "Jiachen Shen", "Yuhao Zhang", "Yishuai Cai", "Lu Liu", "Hailu Ji", "Yuanpei Chen", "Juntao Dai", "Jiaming Ji", "Yaodong Yang"], "year": 2025, "venue": "ICML 2026", "type": "benchmark", "url": "https://arxiv.org/abs/2512.22539", "pdf_url": "https://arxiv.org/pdf/2512.22539", "code_url": "", "clusters": ["C4"], "keywords": ["VLA-Arena", "difficulty levels", "perturbation probes"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["VLA-Arena"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "170 tasks over safety, distractor, extrapolation and long-horizon suites with three difficulty levels and orthogonal language/visual perturbation levels, fine-tuning only on L0.", "relevance": 3, "relevance_reason": "Structured VLA benchmark design.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA leaderboard"}
{"id": "2509.18183", "title": "VLA-LPAF: Lightweight Perspective-Adaptive Fusion for Vision-Language-Action to Enable More Unconstrained Robotic Manipulation", "authors": ["Jinyue Bian", "Zhaoxing Zhang", "Zhengyu Liang", "Shiwei Zheng", "Sheng-Tao Zhang", "Rong Shen", "Chen Yang", "Anzhou Hou"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.18183", "pdf_url": "https://arxiv.org/pdf/2509.18183", "code_url": "", "clusters": ["C3"], "keywords": ["VLA-LPAF", "multi-view"], "model_class": "VLA", "techniques": ["lightweight adapter"], "base_model": "RoboFlamingo", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lightweight module fine-tuned on single-view images to make VLAs robust to camera perspective changes.", "relevance": 3, "relevance_reason": "Cheap adaptation module.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2509.22643", "title": "VLA-Reasoner: Empowering Vision-Language-Action Models with Reasoning via Online Monte Carlo Tree Search", "authors": ["Wenkai Guo", "Guanxing Lu", "Haoyuan Deng", "Zhenyu Wu", "Yansong Tang", "Ziwei Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.22643", "pdf_url": "https://arxiv.org/pdf/2509.22643", "code_url": "", "clusters": ["C8"], "keywords": ["MCTS", "test-time scaling", "world model rollouts", "VLA"], "model_class": "world model", "techniques": ["online Monte Carlo tree search over world-model rollouts of VLA actions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Plug-in test-time planner that rolls out sampled VLA action trajectories in a world model and uses MCTS to pick foresighted actions for long-horizon tasks.", "relevance": 3, "relevance_reason": "Test-time world-model planning for VLAs (compute trade-off).", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2507.05116", "title": "VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting", "authors": ["Juyi Lin", "Amir Taherin", "Arash Akbari", "Arman Akbari", "Lei Lu", "Guangyu Chen", "Taskin Padir", "Xiaomeng Yang", "Weiwei Chen", "Yiqian Li", "Xue Lin", "David Kaeli", "Pu Zhao", "Yanzhi Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.05116", "pdf_url": "https://arxiv.org/pdf/2507.05116", "code_url": "https://github.com/LukeLIN-web/VOTE", "clusters": ["C1", "C5"], "keywords": ["fewer action tokens", "ensemble voting", "edge throughput"], "model_class": "VLA", "techniques": ["parallel action token generation", "trajectory ensemble voting"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "39x faster than OpenVLA; 46 Hz throughput on edge platforms", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fine-tunes VLAs to emit far fewer, parallel action tokens and ensembles current and past predictions, reaching 46 Hz on edge platforms.", "relevance": 3, "relevance_reason": "Edge-platform throughput numbers.", "foundational_technique": false, "verified": true, "found_via": "arXiv: vision-language-action AND Jetson/edge/onboard"}
{"id": "2507.01016", "title": "VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers", "authors": ["Yating Wang", "Haoyi Zhu", "Mingyu Liu", "Jiange Yang", "Hao-Shu Fang", "Tong He"], "year": 2025, "venue": "ICCV 2025", "type": "paper", "url": "https://arxiv.org/abs/2507.01016", "pdf_url": "https://arxiv.org/pdf/2507.01016", "code_url": "", "clusters": ["C1"], "keywords": ["VQ action tokenizer", "synthetic data scaling"], "model_class": "VLA", "techniques": ["action tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Scales a VQ action tokenizer on a large mostly synthetic trajectory corpus, speeding inference and smoothing actions of VLAs.", "relevance": 3, "relevance_reason": "Action tokenization.", "foundational_technique": false, "verified": true, "found_via": "search query: action tokenization"}
{"id": "2410.14040", "title": "WARPD: World model Assisted Reactive Policy Diffusion", "authors": ["Shashank Hegde", "Satyajeet Das", "Gautam Salhotra", "Gaurav S. Sukhatme"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.14040", "pdf_url": "https://arxiv.org/pdf/2410.14040", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["WARPD", "closed-loop policy generation", "parameter-space diffusion"], "model_class": "diffusion/flow policy", "techniques": ["hypernetwork/diffusion over policy weights"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diffuses weights of small closed-loop neural policies instead of open-loop trajectories, extending horizons with perturbation robustness and cutting inference cost.", "relevance": 3, "relevance_reason": "Alternative to open-loop chunks for reactivity at low cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"open-loop\" action chunk reactive closed-loop"}
{"id": "2511.09515", "title": "WMPO: World Model-based Policy Optimization for Vision-Language-Action Models", "authors": ["Fangqi Zhu", "Zhengyang Yan", "Zicong Hong", "Quanxin Shou", "Xiao Ma", "Song Guo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.09515", "pdf_url": "https://arxiv.org/pdf/2511.09515", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["WMPO", "pixel world model", "on-policy GRPO"], "model_class": "world model", "techniques": ["world-model RL", "GRPO", "pixel-space world model aligned with VLA features", "on-policy GRPO in imagination"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "On-policy RL for VLAs inside a pixel-space video world model, improving sample efficiency and producing emergent self-correction without real-environment interaction.", "relevance": 3, "relevance_reason": "World-model-based VLA post-training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:GRPO; search query"}
{"id": "2509.24948", "title": "World-Env: Leveraging World Model as a Virtual Environment for VLA Post-Training", "authors": ["Junjin Xiao", "Yandan Yang", "Xinyuan Chang", "Ronghan Chen", "Feng Xiong", "Mu Xu", "Wei-Shi Zheng", "Qing Zhang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.24948", "pdf_url": "https://arxiv.org/pdf/2509.24948", "code_url": "", "clusters": ["C3"], "keywords": ["world model environment", "non-resettable", "post-training"], "model_class": "VLA", "techniques": ["world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses a video world model as a virtual environment with a VLM success detector for RL post-training in data-scarce settings.", "relevance": 3, "relevance_reason": "Avoids costly real resets.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2506.00613", "title": "WorldGym: World Model as An Environment for Policy Evaluation", "authors": ["Julian Quevedo", "Ansh Kumar Sharma", "Yixiang Sun", "Varad Suryavanshi", "Percy Liang", "Sherry Yang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.00613", "pdf_url": "https://arxiv.org/pdf/2506.00613", "code_url": "", "clusters": ["C3", "C4", "C8"], "keywords": ["WorldGym", "world model evaluation", "policy ranking", "Monte Carlo rollouts", "VLM reward"], "model_class": "world model", "techniques": ["autoregressive action-conditioned video model", "VLM-based reward"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot", "VLA real-robot policies"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Autoregressive action-conditioned video model used as an evaluation environment with VLM rewards; success rates correlate with real ones and preserve ranking across policy sizes/checkpoints.", "relevance": 3, "relevance_reason": "World-model-based evaluation.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA evaluation simulation real correlation; search query"}
{"id": "2512.14614", "title": "WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling", "authors": ["Wen-Qiang Sun", "Haiyu Zhang", "Haoyuan Wang", "Jun-Tao Wu", "Ze-Han Wang", "Zhenwei Wang", "Yun-Hong Wang", "Jun Zhang", "Teng-Fei Wang", "Chun-Chao Guo"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2512.14614", "pdf_url": "https://arxiv.org/pdf/2512.14614", "code_url": "", "clusters": ["C8"], "keywords": ["real-time interactive world model", "streaming"], "model_class": "world model", "techniques": ["context forcing distillation", "memory reconstitution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "720p at 24 FPS", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "WorldPlay is a streaming video diffusion world model distilled for real-time interaction with long-term geometric consistency.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2506.08009"}
{"id": "2511.14759", "title": "π*0.6: a VLA That Learns From Experience", "authors": ["Physical Intelligence", "Ali Amin", "Raichelle Aniceto", "Ashwin Balakrishna", "Kevin Black", "Ken Conley", "Grace Connors", "James Darpinian", "Karan Dhabalia", "Jared DiCarlo", "Danny Driess", "M. Equi", "A. Esmail", "Yun-Hao Fang", "Chelsea Finn", "Catherine Glossop", "Thomas Godden", "I. Goryachev", "Lachy Groom", "H. Hancock", "Karol Hausman", "Gashon Hussein", "Brian Ichter", "S. Jakubczak", "Rowan Jen", "Tim Jones", "Ben Katz", "Liyiming Ke", "Chandra Kuchi", "Marinda Lamb", "Devin LeBlanc", "Sergey Levine", "Adrian Li-Bell", "Yao Lu", "Vishnu Mano", "Mohith Mothukuri", "Suraj Nair", "Karl Pertsch", "Allen Z. Ren", "Charvi Sharma", "L. Shi", "Jost Laura Smith", "Tobias Springenberg", "Kyle Stachowicz", "Will Stoeckle", "Alex Swerdlow", "James Tanner", "Marcel Torne", "Quan Vuong", "Anna Walling", "Hao-Huan Wang", "Blake Williams", "Sukwon Yoo", "Li-Li Yu", "Ury Zhilinsky", "Zhi-Yuan Zhou"], "year": 2025, "venue": "arXiv / tech report", "type": "tech-report", "url": "https://www.semanticscholar.org/paper/eb3723273cd9e22f2215e6aa4fc4980f1e948ead", "pdf_url": "https://arxiv.org/pdf/2511.14759", "code_url": "", "clusters": ["C3", "C4", "C9"], "keywords": ["RECAP", "advantage conditioning", "RL from deployment", "pi0.6", "Physical Intelligence", "pi*0.6"], "model_class": "VLA", "techniques": ["advantage-conditioned RL", "interventions", "RL from experience"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RECAP trains VLAs with RL via advantage conditioning on demonstrations, on-policy data and interventions, improving throughput and robustness on real tasks.", "relevance": 3, "relevance_reason": "Context: major industrial VLA (links=16). No abstract available.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Physical Intelligence; snowball round 2: cites 2409.00588, 2410.24164"}
{"id": "hf-pi0-lerobot-blog-2025", "title": "π0 and π0-FAST: Vision-Language-Action Models for General Robot Control (LeRobot port)", "authors": ["Dana Aubakirova", "Pablo Montalvo", "Mustafa Shukor", "Remi Cadene"], "year": 2025, "venue": "blog (Hugging Face)", "type": "blog", "url": "https://huggingface.co/blog/pi0", "pdf_url": "", "code_url": "https://github.com/huggingface/lerobot", "clusters": ["C1", "C5", "C9"], "keywords": ["pi0", "pi0-FAST", "FlexAttention", "PyTorch port"], "model_class": "VLA", "techniques": ["FAST tokenization", "FlexAttention for block masks"], "base_model": "pi0 / pi0-FAST", "params": "", "hardware": "", "reported_metric": "pi0-FAST ~5x faster training than diffusion VLAs (quoted); pi0 outputs 50 Hz trajectories", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Explains porting pi0/pi0-FAST to PyTorch in LeRobot, including FlexAttention for irregular prefix/action attention masks and FAST's DCT+BPE tokenization.", "relevance": 3, "relevance_reason": "Implementation details affecting VLA inference efficiency.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Hugging Face LeRobot"}
{"id": "2410.15549", "title": "A Dual Process VLA: Efficient Robotic Manipulation Leveraging VLM", "authors": ["ByungOk Han", "Jaehong Kim", "Jinhyeok Jang"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.15549", "pdf_url": "https://arxiv.org/pdf/2410.15549", "code_url": "", "clusters": ["C1"], "keywords": ["dual process", "large System 2", "small System 1"], "model_class": "VLA", "techniques": ["dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large low-frequency VLM reasoning paired with a small fast motor-control model.", "relevance": 3, "relevance_reason": "Early dual-process VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: real-time VLA inference"}
{"id": "2409.04576", "title": "ActionFlow: Equivariant, Accurate, and Efficient Policies with Spatially Symmetric Flow Matching", "authors": ["Niklas Funk", "Julen Urain", "João Carvalho", "V. Prasad", "G. Chalvatzaki", "Jan Peters"], "year": 2024, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2409.04576", "pdf_url": "https://arxiv.org/pdf/2409.04576", "code_url": "", "clusters": ["C1"], "keywords": ["ActionFlow"], "model_class": "diffusion/flow policy", "techniques": ["flow matching", "equivariance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SE(3)-equivariant flow-matching policy combining spatial symmetry with fast flow-matching inference.", "relevance": 3, "relevance_reason": "Fast-inference generative policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2503.07404"}
{"id": "2409.01083", "title": "Affordance-based Robot Manipulation with Flow Matching", "authors": ["Fan Zhang", "Michael Gienger"], "year": 2024, "venue": "Frontiers in Robotics and AI", "type": "paper", "url": "https://arxiv.org/abs/2409.01083", "pdf_url": "https://arxiv.org/pdf/2409.01083", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["affordance", "PEFT"], "model_class": "diffusion/flow policy", "techniques": ["prompt tuning", "flow matching", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~85% lower inference time (1-step flow matching vs 16-step DDIM)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Parameter-efficient prompt tuning of a frozen vision model for affordances plus a flow-matching policy that runs in one step.", "relevance": 3, "relevance_reason": "PEFT and few-step flow policy with reported inference-time savings.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2310.05905"}
{"id": "2401.17583", "title": "Agile But Safe: Learning Collision-Free High-Speed Legged Locomotion", "authors": ["Tairan He", "Chong Zhang", "Wenli Xiao", "Guanqi He", "Changliu Liu", "Guanya Shi"], "year": 2024, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2401.17583", "pdf_url": "https://arxiv.org/pdf/2401.17583", "code_url": "", "clusters": ["C7"], "keywords": ["legged locomotion", "safety", "onboard"], "model_class": "other", "techniques": ["reach-avoid value network"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Agile quadruped policy with a learned reach-avoid value network that switches to a recovery policy for collision-free high-speed motion.", "relevance": 3, "relevance_reason": "Safety under high-speed control, onboard compute.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2502.00935, 2505.00779"}
{"id": "2410.00371", "title": "AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation", "authors": ["Jiafei Duan", "Wilbert Pumacay", "Nishanth Kumar", "Yi-Ru Wang", "Shulin Tian", "Wentao Yuan", "Ranjay Krishna", "Dieter Fox", "A. Mandlekar", "Yi-Jie Guo"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2410.00371", "pdf_url": "https://arxiv.org/pdf/2410.00371", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "FailGen"], "model_class": "other", "techniques": ["VLM fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open VLM fine-tuned on procedurally generated failure trajectories to detect and explain manipulation failures.", "relevance": 3, "relevance_reason": "Runtime failure detection (C7 monitoring).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2502.01828, 2506.09937"}
{"id": "2401.12963", "title": "AutoRT: Embodied Foundation Models for Large Scale Orchestration of Robotic Agents", "authors": ["Michael Ahn", "Debidatta Dwibedi", "Chelsea Finn", "Montse Gonzalez Arenas", "Keerthana Gopalakrishnan", "Karol Hausman", "Brian Ichter", "Alex Irpan", "Nikhil Joshi", "Ryan Julian", "Sean Kirmani", "Isabel Leal", "Edward Lee", "Sergey Levine", "Yao Lu", "Isabel Leal", "Sharath Maddineni", "Kanishka Rao", "Dorsa Sadigh", "Pannag Sanketi", "Pierre Sermanet", "Quan Vuong", "Stefan Welker", "Fei Xia", "Ted Xiao", "Peng Xu", "Steve Xu", "Zhuo Xu"], "year": 2024, "venue": "ICRA 2024 VLMNM Workshop", "type": "paper", "url": "https://arxiv.org/abs/2401.12963", "pdf_url": "https://arxiv.org/pdf/2401.12963", "code_url": "", "clusters": ["C6"], "keywords": ["fleet orchestration", "foundation models", "data collection at scale"], "model_class": "other", "techniques": ["LLM/VLM fleet orchestration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["20+ robots, 77k real episodes"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "AutoRT uses VLMs and LLMs to propose and vet tasks for a fleet of 20+ robots across buildings, scaling autonomous and teleoperated data collection.", "relevance": 3, "relevance_reason": "Fleet-level orchestration with foundation models (Google DeepMind).", "foundational_technique": false, "verified": true, "found_via": "arXiv API: fleet foundation model robot"}
{"id": "2412.01818", "title": "Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning in VLMs", "authors": ["Qi-Zhe Zhang", "Aosong Cheng", "Ming Lu", "Renrui Zhang", "Zhi-Yong Zhuo", "Jia-Jun Cao", "Shaobo Guo", "Qi She", "Shang-Hang Zhang"], "year": 2024, "venue": "IEEE International Conference on Computer Vision", "type": "paper", "url": "https://arxiv.org/abs/2412.01818", "pdf_url": "https://arxiv.org/pdf/2412.01818", "code_url": "", "clusters": ["C1"], "keywords": ["VisPruner", "LVLM"], "model_class": "other", "techniques": ["visual token pruning"], "base_model": "LLaVA-1.5-7B", "params": "", "hardware": "", "reported_metric": "-91% FLOPs, -75% latency on LLaVA-1.5-7B", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Plug-and-play visual token pruning for VLMs using visual cues rather than text-visual attention.", "relevance": 3, "relevance_reason": "Generic VLM token pruning used by VLA pruning work.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135; cited by 2509.09090"}
{"id": "2411.11406", "title": "Bridging the Resource Gap: Deploying Advanced Imitation Learning Models onto Affordable Embedded Platforms", "authors": ["Haizhou Ge", "Ruixiang Wang", "Zhu-ang Xu", "Hongrui Zhu", "Ruichen Deng", "Yuhang Dong", "Zeyu Pang", "Guyue Zhou", "Junyu Zhang", "Lu Shi"], "year": 2024, "venue": "IEEE ROBIO 2024", "type": "paper", "url": "https://arxiv.org/abs/2411.11406", "pdf_url": "https://arxiv.org/pdf/2411.11406", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["embedded deployment", "temporal ensemble with dropped actions"], "model_class": "other", "techniques": ["model compression", "asynchronous execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Pipeline for migrating transformer imitation policies to edge devices via compression and an asynchronous temporal ensemble that drops stale actions.", "relevance": 3, "relevance_reason": "Edge deployment of chunked policies.", "foundational_technique": false, "verified": true, "found_via": "search query: temporal ensembling"}
{"id": "2409.15654", "title": "Cambricon-LLM: A Chiplet-Based Hybrid Architecture for On-Device Inference of 70B LLM", "authors": ["Zhongkai Yu", "Shengwen Liang", "Tianyun Ma", "Yunke Cai", "Ziyuan Nan", "Di Huang", "Xinkai Song", "Yifan Hao", "Jie Zhang", "Tian Zhi", "Yongwei Zhao", "Zidong Du", "Xing Hu", "Qi Guo", "Tianshi Chen"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2409.15654", "pdf_url": "https://arxiv.org/pdf/2409.15654", "code_url": "", "clusters": ["C5"], "keywords": ["chiplet", "NPU + NAND flash", "in-flash computing", "70B on-device", "robotics edge"], "model_class": "other", "techniques": ["hybrid NPU-flash architecture", "hardware tiling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Chiplet architecture pairing an NPU with compute-capable NAND flash to run 70B LLMs on edge devices such as robots despite low arithmetic intensity at batch 1.", "relevance": 3, "relevance_reason": "Memory-capacity/bandwidth-oriented edge accelerator explicitly citing robotics.", "foundational_technique": true, "verified": true, "found_via": "arXiv: NPU AND (robot OR VLA OR embodied)"}
{"id": "2408.08592", "title": "Case Study: Runtime Safety Verification of Neural Network Controlled System", "authors": ["Frank Yang", "Sinong Simon Zhan", "Yixuan Wang", "Chao Huang", "Qi Zhu"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2408.08592", "pdf_url": "https://arxiv.org/pdf/2408.08592", "code_url": "", "clusters": ["C7"], "keywords": ["POLAR-Express", "ROS2"], "model_class": "other", "techniques": ["runtime verification", "reachability", "controller switching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ROS2 Flatland simulation"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Case study using POLAR-Express online to verify a LiDAR NN controller on a Turtlebot and switch to an avoidance controller when verification fails.", "relevance": 3, "relevance_reason": "Runtime (online) formal verification of a neural controller with switching.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2412.04455", "title": "Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection", "authors": ["En-Shen Zhou", "Qi Su", "Cheng Chi", "Zhizheng Zhang", "Zhongyuan Wang", "Tiejun Huang", "Lu Sheng", "He Wang"], "year": 2024, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2412.04455", "pdf_url": "https://arxiv.org/pdf/2412.04455", "code_url": "", "clusters": ["C7"], "keywords": ["failure detection", "monitoring"], "model_class": "other", "techniques": ["VLM code generation", "constraint monitoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Code-as-Monitor has a VLM write code that checks spatio-temporal constraints for real-time reactive and proactive failure detection in manipulation.", "relevance": 3, "relevance_reason": "relevant: runtime failure monitoring", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2606.16690, 2606.23085"}
{"id": "2310.05921", "title": "Conformal Decision Theory: Safe Autonomous Decisions from Imperfect Predictions", "authors": ["Jordan Lekeufack", "Anastasios N. Angelopoulos", "Andrea Bajcsy", "Michael I. Jordan", "Jitendra Malik"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2310.05921", "pdf_url": "https://arxiv.org/pdf/2310.05921", "code_url": "", "clusters": ["C7"], "keywords": ["conformal decision theory", "backup policy switching", "statistical guarantees"], "model_class": "other", "techniques": ["conformal calibration of decisions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Calibrates decisions such as when to switch from a nominal to a safe backup policy directly with conformal methods, giving risk guarantees without world-model assumptions.", "relevance": 3, "relevance_reason": "Principled fallback switching.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"backup policy\" learned robot"}
{"id": "2401.06066", "title": "DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models", "authors": ["Damai Dai", "C. Deng", "Cheng-Gang Zhao", "R.-X. Xu", "Hua-Zuo Gao", "De-Li Chen", "Jia-Shi Li", "Wang-Ding Zeng", "Xing-Kai Yu", "Yu Wu", "Zhen-Da Xie", "Y.-K. Li", "Pan-Pan Huang", "Fuli Luo", "C. Ruan", "Zhi-Fang Sui", "W. Liang"], "year": 2024, "venue": "Annual Meeting of the Association for Computational Linguistics", "type": "paper", "url": "https://arxiv.org/abs/2401.06066", "pdf_url": "https://arxiv.org/pdf/2401.06066", "code_url": "", "clusters": ["C1"], "keywords": ["MoE", "expert specialization"], "model_class": "other", "techniques": ["mixture-of-experts", "fine-grained experts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MoE design with fine-grained and shared experts for better specialization at fixed compute.", "relevance": 3, "relevance_reason": "Foundational MoE technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2508.21046, 2509.12594"}
{"id": "2405.12399", "title": "Diffusion for World Modeling: Visual Details Matter in Atari", "authors": ["Eloi Alonso", "Adam Jelley", "Vincent Micheli", "Anssi Kanervisto", "Amos Storkey", "Tim Pearce", "François Fleuret"], "year": 2024, "venue": "NeurIPS 2024", "type": "paper", "url": "https://arxiv.org/abs/2405.12399", "pdf_url": "https://arxiv.org/pdf/2405.12399", "code_url": "https://github.com/eloialonso/diamond", "clusters": ["C8"], "keywords": ["DIAMOND", "diffusion world model", "Atari"], "model_class": "world model", "techniques": ["pixel-space diffusion world model", "few-step EDM sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Atari 100k", "CS:GO"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "RL agent trained entirely in a diffusion world model, showing few-step EDM sampling keeps the model fast enough for imagination while preserving visual details.", "relevance": 3, "relevance_reason": "Diffusion world models for imagination-based RL.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2408.14837", "title": "Diffusion Models Are Real-Time Game Engines", "authors": ["Dani Valevski", "Yaniv Leviathan", "Moab Arar", "Shlomi Fruchter"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2408.14837", "pdf_url": "https://arxiv.org/pdf/2408.14837", "code_url": "", "clusters": ["C8"], "keywords": ["neural game engine", "real-time"], "model_class": "world model", "techniques": ["diffusion next-frame prediction"], "base_model": "", "params": "", "hardware": "single TPU", "reported_metric": "20 FPS", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diffusion model simulating DOOM interactively in real time over long trajectories.", "relevance": 3, "relevance_reason": "Real-time interactive world model reference.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2503.00200, 2506.08009"}
{"id": "2402.03570", "title": "Diffusion World Model", "authors": ["Zihan Ding", "Amy Zhang", "Yuan A. Tian", "Qinqing Zheng"], "year": 2024, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2402.03570", "pdf_url": "https://arxiv.org/pdf/2402.03570", "code_url": "", "clusters": ["C8"], "keywords": ["offline RL", "model-based value"], "model_class": "world model", "techniques": ["multi-step diffusion dynamics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["D4RL"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diffusion World Model predicts multi-step futures in one forward pass, avoiding recursive one-step queries.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2412.03293", "title": "Diffusion-VLA: Generalizable and Interpretable Robot Foundation Model via Self-Generated Reasoning", "authors": ["Junjie Wen", "Minjie Zhu", "Yi-Chen Zhu", "Zhibin Tang", "Jinming Li", "Zhongyi Zhou", "Chengmeng Li", "Xiaoyu Liu", "Ya-Xin Peng", "Chaomin Shen", "Fei-Fei Feng"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.03293", "pdf_url": "https://arxiv.org/pdf/2412.03293", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion + autoregression", "VLA", "DiffusionVLA", "reasoning injection"], "model_class": "VLA", "techniques": ["autoregressive reasoning + diffusion head"], "base_model": "", "params": "2B-72B", "hardware": "single A6000", "reported_metric": "82 Hz (DiffusionVLA-2B)", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VLA that combines an autoregressive VLM with a diffusion action head to scale robot foundation models (title-based; no abstract available).", "relevance": 3, "relevance_reason": "Context: frequently cited hybrid VLA (links=17). No abstract available.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503; snowball round 2: cited by 2505.19017, 2510.24795"}
{"id": "2409.14719", "title": "DiSPo: Diffusion-SSM based Policy Learning for Coarse-to-Fine Action Discretization", "authors": ["Nayoung Oh", "Jaehyeong Jang", "Moonkyeong Jung", "Daehyung Park"], "year": 2024, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2409.14719", "pdf_url": "https://arxiv.org/pdf/2409.14719", "code_url": "", "clusters": ["C1"], "keywords": ["DiSPo", "coarse-to-fine"], "model_class": "diffusion/flow policy", "techniques": ["state space model", "Mamba"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diffusion-Mamba policy that varies action granularity, saving inference in less critical regions and training memory-efficiently.", "relevance": 3, "relevance_reason": "Efficient SSM-based policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2603.15046"}
{"id": "2406.01721", "title": "DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs", "authors": ["Haokun Lin", "Hao-Bo Xu", "Yi-Chen Wu", "Jing-Zhi Cui", "Yingtao Zhang", "Lin-Zhan Mou", "Linqi Song", "Zhe-Nan Sun", "Ying Wei"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2406.01721", "pdf_url": "https://arxiv.org/pdf/2406.01721", "code_url": "", "clusters": ["C2"], "keywords": ["outliers", "W4A4"], "model_class": "other", "techniques": ["post-training quantization", "rotation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Redistributes massive activation outliers with block rotations and zigzag permutation for stronger 4-bit LLM quantization.", "relevance": 3, "relevance_reason": "Foundational quantization technique connected to core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323"}
{"id": "2401.15077", "title": "EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty", "authors": ["Yuhui Li", "Fangyun Wei", "Chao Zhang", "Hong-Yang Zhang"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2401.15077", "pdf_url": "https://arxiv.org/pdf/2401.15077", "code_url": "", "clusters": ["C1"], "keywords": ["LLM decoding acceleration"], "model_class": "other", "techniques": ["speculative sampling", "feature-level drafting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Speculative sampling that drafts at the feature level to accelerate LLM decoding.", "relevance": 3, "relevance_reason": "Foundational speculative-decoding technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2507.22424, 2512.15773"}
{"id": "2408.09972", "title": "Edge-Cloud Collaborative Motion Planning for Autonomous Driving with Large Language Models", "authors": ["Jiao Chen", "Suyan Dai", "Fangfang Chen", "Zuohong Lv", "Jianhua Tang"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2408.09972", "pdf_url": "https://arxiv.org/pdf/2408.09972", "code_url": "", "clusters": ["C6"], "keywords": ["edge-cloud LLM", "autonomous driving", "drift detection"], "model_class": "other", "techniques": ["selective cloud upload", "edge small LLM + cloud GPT-4"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "EC-Drive runs small LLMs at the vehicle edge and uploads only drift-flagged data to a cloud LLM for motion planning, reducing latency and cloud traffic.", "relevance": 3, "relevance_reason": "Selective edge-to-cloud escalation for embodied planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2410.01220", "title": "Effective Tuning Strategies for Generalist Robot Manipulation Policies", "authors": ["Wen-Bo Zhang", "Yang Li", "Yanyuan Qiao", "Si-Yuan Huang", "Jia-Jun Liu", "Feras Dayoub", "Xiao Ma", "Lingqiao Liu"], "year": 2024, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2410.01220", "pdf_url": "https://arxiv.org/pdf/2410.01220", "code_url": "", "clusters": ["C3"], "keywords": ["few-shot adaptation", "generalist policies"], "model_class": "other", "techniques": ["fine-tuning strategies"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Empirical study of fine-tuning design choices for adapting generalist manipulation policies with limited samples.", "relevance": 3, "relevance_reason": "Efficient adaptation recipes (C3).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2607.00666"}
{"id": "2406.19320", "title": "Efficient World Models with Context-Aware Tokenization", "authors": ["Vincent Micheli", "Eloi Alonso", "François Fleuret"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2406.19320", "pdf_url": "https://arxiv.org/pdf/2406.19320", "code_url": "", "clusters": ["C8"], "keywords": ["transformer world model", "MBRL"], "model_class": "world model", "techniques": ["delta tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "order of magnitude faster to train than prior attention-based world models", "metric_measures": "other", "benchmarks": ["Crafter"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Delta-IRIS world model that encodes stochastic deltas between steps to cut token sequence length and training cost.", "relevance": 3, "relevance_reason": "Efficient world model", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104"}
{"id": "2412.10855", "title": "Fast and Robust Visuomotor Riemannian Flow Matching Policy", "authors": ["Haoran Ding", "Noémie Jaquier", "Jan Peters", "Leonel Rozo"], "year": 2024, "venue": "IEEE T-RO", "type": "paper", "url": "https://arxiv.org/abs/2412.10855", "pdf_url": "https://arxiv.org/pdf/2412.10855", "code_url": "", "clusters": ["C1"], "keywords": ["Riemannian flow matching", "stability"], "model_class": "diffusion/flow policy", "techniques": ["flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Flow-matching policy on Riemannian manifolds with a stable variant, faster to train and sample than diffusion or consistency policies.", "relevance": 3, "relevance_reason": "Fast flow policy.", "foundational_technique": false, "verified": true, "found_via": "search query: consistency policy"}
{"id": "2407.08608", "title": "FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision", "authors": ["Jay Shah", "Ganesh Bikshandi", "Ying Zhang", "Vijay Thakkar", "P. Ramani", "Tri Dao"], "year": 2024, "venue": "NeurIPS 2024", "type": "paper", "url": "https://arxiv.org/abs/2407.08608", "pdf_url": "https://arxiv.org/pdf/2407.08608", "code_url": "https://github.com/Dao-AILab/flash-attention", "clusters": ["C5"], "keywords": ["Hopper", "warp specialization", "TMA", "FP8 attention"], "model_class": "other", "techniques": ["asynchronous attention kernel", "FP8 block quantization"], "base_model": "", "params": "", "hardware": "H100", "reported_metric": "1.5-2.0x speedup over FlashAttention-2 on H100", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Exploits Hopper tensor-core/TMA asynchrony and FP8 with incoherent processing to accelerate attention on H100.", "relevance": 3, "relevance_reason": "Hardware-specific kernel design principle relevant to Thor/Blackwell-class edge GPUs.", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2410.09426", "title": "FlatQuant: Flatness Matters for LLM Quantization", "authors": ["Yuxuan Sun", "Rui-Kang Liu", "Hao-Li Bai", "Han Bao", "Kang Zhao", "Yue-Ning Li", "Jia-Xin Hu", "Xianzhi Yu", "Lu Hou", "Chun Yuan", "Xin Jiang", "Wu-Long Liu", "Jun Yao"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2410.09426", "pdf_url": "https://arxiv.org/pdf/2410.09426", "code_url": "", "clusters": ["C2"], "keywords": ["FlatQuant", "W4A4"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 2.3x prefill, 1.7x decoding speedup vs FP16", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Post-training LLM quantization that learns per-layer affine transforms to flatten weights/activations, with a fused kernel.", "relevance": 3, "relevance_reason": "Generic quantization technique used by robot-model compression work.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2602.02110, 2602.20309"}
{"id": "2406.05221", "title": "GCAPS: GPU Context-Aware Preemptive Priority-based Scheduling for Real-Time Tasks", "authors": ["Yidi Wang", "Cong Liu", "Daniel Wong", "Hyoseung Kim"], "year": 2024, "venue": "ECRTS 2024", "type": "paper", "url": "https://arxiv.org/abs/2406.05221", "pdf_url": "https://arxiv.org/pdf/2406.05221", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["GCAPS", "real-time GPU"], "model_class": "other", "techniques": ["GPU preemption", "response-time analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driver-level GPU context-aware preemptive priority scheduling with response-time analysis for GPU-using real-time tasks, requiring only one-line macros.", "relevance": 3, "relevance_reason": "Analyzable GPU scheduling needed to bound neural inference latency.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"real-time\" AND Jetson AND (\"response time\" OR deadline* OR schedulab*)"}
{"id": "2403.13358", "title": "GeRM: A Generalist Robotic Model with Mixture-of-experts for Quadruped Robot", "authors": ["Wenxuan Song", "Han Zhao", "Pengxiang Ding", "Can Cui", "Shangke Lyu", "Yaning Fan", "Donglin Wang"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2403.13358", "pdf_url": "https://arxiv.org/pdf/2403.13358", "code_url": "", "clusters": ["C1"], "keywords": ["mixture of experts", "quadruped", "offline RL"], "model_class": "VLA", "techniques": ["mixture of experts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["QUARD-Auto"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Quadruped VLA using an MoE transformer to raise capacity while keeping inference cheap, trained with offline RL on demonstrations and suboptimal data.", "relevance": 3, "relevance_reason": "Early MoE VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA mixture-of-experts"}
{"id": "github-openvla-issue-152-gpu-size", "title": "GPU size for deploy (openvla/openvla issue #152) and 'how much memory is needed' (#57)", "authors": ["notmyname16", "kpertsch"], "year": 2024, "venue": "GitHub issue", "type": "repo", "url": "https://github.com/openvla/openvla/issues/152", "pdf_url": "", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["OpenVLA", "memory requirement", "int4"], "model_class": "VLA", "techniques": ["int4 quantization"], "base_model": "OpenVLA-7B", "params": "7B", "hardware": "", "reported_metric": "maintainer: 16 GB GPU for bf16 load (#152); inference needs ~11 GB bf16 and ~6 GB with 4-bit (#57); verify_openvla.py comments: 8-bit ~9-10 GB, 4-bit ~6-7 GB VRAM (#145)", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "OpenVLA maintainers state the 7B policy needs a 16 GB GPU in bf16 but can drop to about 6 GB with int4, the figure most edge-deployment claims start from.", "relevance": 3, "relevance_reason": "Authoritative (maintainer) memory figures relevant to 16 GB Jetson fit.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub issues repo:openvla/openvla memory GPU"}
{"id": "2411.19309", "title": "GRAPE: Generalizing Robot Policy via Preference Alignment", "authors": ["Zijian Zhang", "Kaiyuan Zheng", "Zhaorun Chen", "Joel Jang", "Yi Li", "Siwei Han", "Chaoqi Wang", "Mingyu Ding", "Dieter Fox", "Huaxiu Yao"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2411.19309", "pdf_url": "https://arxiv.org/pdf/2411.19309", "code_url": "", "clusters": ["C3"], "keywords": ["GRAPE", "trajectory preference optimization"], "model_class": "VLA", "techniques": ["preference alignment", "TPO"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Aligns VLAs at trajectory level with preferences from success and failure trials and stage-wise cost guidance.", "relevance": 3, "relevance_reason": "Post-training alignment of VLAs.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2410.21229", "title": "HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots", "authors": ["Tairan He", "Wenli Xiao", "Toru Lin", "Zhengyi Luo", "Zhenjia Xu", "Zhenyu Jiang", "Changliu Liu", "Guanya Shi", "Xiaolong Wang", "Linxi Fan", "Yuke Zhu"], "year": 2024, "venue": "ICRA 2025", "type": "paper", "url": "https://arxiv.org/abs/2410.21229", "pdf_url": "https://arxiv.org/pdf/2410.21229", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["whole-body controller", "policy distillation", "multi-mode control", "humanoid"], "model_class": "other", "techniques": ["multi-teacher distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Distills a motion-imitation teacher into one compact whole-body controller supporting many command modes, used as low-level layer under humanoid foundation models (e.g., GR00T).", "relevance": 3, "relevance_reason": "Compact onboard low-level controller in System-1/System-2 humanoid stacks.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2405.05439", "title": "How Generalizable Is My Behavior Cloning Policy? A Statistical Approach to Trustworthy Performance Evaluation", "authors": ["Joseph A. Vincent", "Haruki Nishimura", "Masha Itkina", "Paarth Shah", "Mac Schwager", "Thomas Kollar"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2405.05439", "pdf_url": "https://arxiv.org/pdf/2405.05439", "code_url": "", "clusters": ["C4"], "keywords": ["statistical bounds", "CDF lower bound", "behavior cloning evaluation"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Provides confidence-guaranteed lower bounds on the full performance distribution of a BC policy from few rollouts.", "relevance": 3, "relevance_reason": "Statistical evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation statistical"}
{"id": "2405.14867", "title": "Improved Distribution Matching Distillation for Fast Image Synthesis", "authors": ["Tianwei Yin", "Michael Gharbi", "Taesung Park", "Richard Zhang", "Eli Shechtman", "Frédo Durand", "William T. Freeman"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2405.14867", "pdf_url": "https://arxiv.org/pdf/2405.14867", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion distillation", "few-step sampling"], "model_class": "other", "techniques": ["distribution matching distillation", "one-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Improves distribution matching distillation of diffusion models into one/few-step generators without a regression loss, adding a GAN term and multi-step support.", "relevance": 3, "relevance_reason": "Foundational few-step diffusion distillation technique that robot policies build on.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2412.09265, 2506.08009"}
{"id": "2410.23208", "title": "Kinetix: Investigating the Training of General Agents through Open-Ended Physics-Based Control Tasks", "authors": ["M. Matthews", "Michael Beukman", "Chris Lu", "J. Foerster"], "year": 2024, "venue": "ICLR 2025", "type": "benchmark", "url": "https://arxiv.org/abs/2410.23208", "pdf_url": "https://arxiv.org/pdf/2410.23208", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["Kinetix", "physics-based control tasks", "JAX"], "model_class": "other", "techniques": ["hardware-accelerated simulation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Procedurally generated 2D physics control environments with a JAX engine; its dynamic tasks became the standard simulated benchmark for inference-delay robustness (RTC and follow-ups).", "relevance": 3, "relevance_reason": "Benchmark underlying most async-execution evaluations.", "foundational_technique": false, "verified": true, "found_via": "snowball: references of 2506.07339 (S2)"}
{"id": "2402.02750", "title": "KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache", "authors": ["Zi-Rui Liu", "Jiayi Yuan", "Hongye Jin", "Shao-Chen Zhong", "Zhaozhuo Xu", "Vladimir Braverman", "Beidi Chen", "Xia Hu"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2402.02750", "pdf_url": "https://arxiv.org/pdf/2402.02750", "code_url": "", "clusters": ["C2"], "keywords": ["2-bit", "LLM serving"], "model_class": "other", "techniques": ["KV-cache quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2.6x less peak memory; 2.35-3.47x throughput", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "KIVI quantizes the KV cache to 2 bits (per-channel keys, per-token values) without tuning, reducing memory and increasing throughput.", "relevance": 3, "relevance_reason": "generic efficiency technique connected to 5 core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323"}
{"id": "2401.18079", "title": "KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization", "authors": ["Coleman Hooper", "Sehoon Kim", "Hiva Mohammadzadeh", "Michael W. Mahoney", "Y. Shao", "Kurt Keutzer", "A. Gholami"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2401.18079", "pdf_url": "https://arxiv.org/pdf/2401.18079", "code_url": "", "clusters": ["C2"], "keywords": ["long context", "LLM"], "model_class": "other", "techniques": ["KV cache quantization"], "base_model": "", "params": "", "hardware": "A100-80GB", "reported_metric": "~1.7x speedup vs fp16 matvec (LLaMA-7B)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Sub-4-bit KV cache quantization for long-context LLM inference.", "relevance": 3, "relevance_reason": "Foundational technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2404.00456, 2603.16952"}
{"id": "2404.10498", "title": "LAECIPS: Large Vision Model Assisted Adaptive Edge-Cloud Collaboration for IoT-based Embodied Intelligence System", "authors": ["Shijing Hu", "Zhihui Lu", "Xin Xu", "Ruijun Deng", "Xin Du", "Qiang Duan"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2404.10498", "pdf_url": "https://arxiv.org/pdf/2404.10498", "code_url": "", "clusters": ["C6"], "keywords": ["large vision model", "edge-cloud collaboration", "hard input mining", "embodied inspection"], "model_class": "other", "techniques": ["confidence-based cloud routing", "continual learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Routes hard or uncertain inspection inputs from lightweight edge models to a cloud large vision model while routine cases stay at the edge.", "relevance": 3, "relevance_reason": "Difficulty-aware edge/cloud routing for embodied systems.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:edge-cloud AND abs:robot"}
{"id": "2404.16710", "title": "LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding", "authors": ["Mostafa Elhoushi", "Akshat Shrivastava", "Diana Liskovich", "Basil Hosmer", "Bram Wasti", "Liangzhen Lai", "Anas Mahmoud", "Bilge Acun", "Saurabh Agarwal", "Ahmed Roman", "Ahmed A Aly", "Beidi Chen", "Carole-Jean Wu"], "year": 2024, "venue": "Annual Meeting of the Association for Computational Linguistics", "type": "paper", "url": "https://arxiv.org/abs/2404.16710", "pdf_url": "https://aclanthology.org/2024.acl-long.681.pdf", "code_url": "", "clusters": ["C1"], "keywords": ["LLM inference"], "model_class": "other", "techniques": ["early exit", "self-speculative decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 2.16x speedup (summarization)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Layer dropout and early-exit training enabling early-exit inference and self-speculative decoding.", "relevance": 3, "relevance_reason": "Foundational technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2411.02359, 2509.21930"}
{"id": "2406.08234", "title": "MaIL: Improving Imitation Learning with Mamba", "authors": ["Xiaogang Jia", "Qian Wang", "Atalay Donat", "Bowen Xing", "Ge Li", "Hongyi Zhou", "Onur Celik", "Denis Blessing", "Rudolf Lioutikov", "Gerhard Neumann"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2406.08234", "pdf_url": "https://arxiv.org/pdf/2406.08234", "code_url": "https://github.com/ALRhub/MaIL", "clusters": ["C1"], "keywords": ["Mamba", "imitation learning"], "model_class": "other", "techniques": ["state space model backbone"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Mamba-based imitation policy that outperforms transformers in low-data regimes on LIBERO.", "relevance": 3, "relevance_reason": "SSM policy architecture.", "foundational_technique": false, "verified": true, "found_via": "search query: Mamba robot policy"}
{"id": "2403.12844", "title": "MELTing point: Mobile Evaluation of Language Transformers", "authors": ["Stefanos Laskaridis", "Kleomenis Katevas", "Lorenzo Minto", "Hamed Haddadi"], "year": 2024, "venue": "MobiCom 2024", "type": "paper", "url": "https://arxiv.org/abs/2403.12844", "pdf_url": "https://arxiv.org/pdf/2403.12844", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["on-device LLM", "energy", "thermal", "Jetson", "benchmark infrastructure"], "model_class": "other", "techniques": ["quantization"], "base_model": "", "params": "", "hardware": "Android, iOS, NVIDIA Jetson", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Benchmark infrastructure measuring on-device LLM latency, memory, energy and thermal behavior across phones and Jetsons, finding inference memory-bound and sustained execution thermally limited.", "relevance": 3, "relevance_reason": "Energy/thermal measurement methodology for sustained on-device transformer inference.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q3: arXiv abs:\"large language model\" AND abs:Jetson AND abs:benchmark"}
{"id": "2402.14905", "title": "MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases", "authors": ["Zechun Liu", "Changsheng Zhao", "Forrest N. Iandola", "Chen Lai", "Yuandong Tian", "Igor Fedorov", "Yun-Yang Xiong", "Ernie Chang", "Yang-Yang Shi", "Raghuraman Krishnamoorthi", "Liangzhen Lai", "Vikas Chandra"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2402.14905", "pdf_url": "https://arxiv.org/pdf/2402.14905", "code_url": "", "clusters": ["C1", "C5"], "keywords": ["MobileLLM", "on-device"], "model_class": "other", "techniques": ["compact architecture", "weight sharing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Design study for sub-billion LLMs for on-device use, favoring deep-thin architectures and block-wise weight sharing.", "relevance": 3, "relevance_reason": "Generic small-model design for on-device use.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135; cited by 2411.02359"}
{"id": "2412.04468", "title": "NVILA: Efficient Frontier Visual Language Models", "authors": ["Zhijian Liu", "Ligeng Zhu", "Baifeng Shi", "Zhuoyang Zhang", "Yuming Lou", "Shang Yang", "Haocheng Xi", "Shiyi Cao"], "year": 2024, "venue": "CVPR 2025", "type": "paper", "url": "https://arxiv.org/abs/2412.04468", "pdf_url": "https://arxiv.org/pdf/2412.04468", "code_url": "", "clusters": ["C2", "C9"], "keywords": ["scale-then-compress", "visual token compression", "efficient VLM"], "model_class": "other", "techniques": ["visual token compression", "FP8 training", "quantized deployment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Efficient VLM family that scales spatial/temporal resolution then compresses visual tokens, optimizing cost from training to deployment; used as a backbone in several VLAs.", "relevance": 3, "relevance_reason": "Efficient VLM backbone commonly used for VLAs (foundational).", "foundational_technique": true, "verified": true, "found_via": "author sweep: Zhuoyang Zhang"}
{"id": "2405.07515", "title": "OpenBot-Fleet: A System for Collective Learning with Real Robots", "authors": ["Matthias Müller", "Samarth Brahmbhatt", "Ankur Deka", "Quentin Leboutet", "David Hafner", "Vladlen Koltun"], "year": 2024, "venue": "ICRA 2024", "type": "paper", "url": "https://arxiv.org/abs/2405.07515", "pdf_url": "https://arxiv.org/pdf/2405.07515", "code_url": "", "clusters": ["C6"], "keywords": ["fleet learning", "cloud robotics", "smartphone robots", "Firebase"], "model_class": "other", "techniques": ["fleet data upload", "cloud policy training", "policy push to fleet"], "base_model": "", "params": "", "hardware": "", "reported_metric": ">80% navigation success in unseen homes", "metric_measures": "other", "benchmarks": ["72 robots in homes"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Low-cost fleet of 72 smartphone robots that upload data to the cloud, learn navigation policies there, and receive them back, demonstrating a full cloud fleet-learning loop.", "relevance": 3, "relevance_reason": "Cost-effective fleet-cloud learning system.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2409.06078", "title": "PEERNet: An End-to-End Profiling Tool for Real-Time Networked Robotic Systems", "authors": ["Aditya Narayanan", "Pranav Kasibhatla", "Minkyu Choi", "Po-han Li", "Ruihan Zhao", "Sandeep Chinchali"], "year": 2024, "venue": "IROS 2024", "type": "paper", "url": "https://arxiv.org/abs/2409.06078", "pdf_url": "https://arxiv.org/pdf/2409.06078", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["profiling", "networked robotics", "offloading decisions"], "model_class": "other", "techniques": ["end-to-end profiling"], "base_model": "", "params": "", "hardware": "Jetson Orin, Franka Panda", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Real-time profiler for networked robot systems measuring sensors, network, model latency and devices to inform when to offload, e.g., VLM queries from a Jetson Orin.", "relevance": 3, "relevance_reason": "Measurement tooling for offloading decisions.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2410.21845", "title": "Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning", "authors": ["Jianlan Luo", "Charles Xu", "Jeffrey Wu", "Sergey Levine"], "year": 2024, "venue": "Science Robotics 2025", "type": "preprint", "url": "https://arxiv.org/abs/2410.21845", "pdf_url": "https://arxiv.org/pdf/2410.21845", "code_url": "", "clusters": ["C3"], "keywords": ["HIL-SERL", "human-in-the-loop RL", "real-world RL"], "model_class": "other", "techniques": ["off-policy RL (RLPD)", "human interventions", "pretrained visual backbone"], "base_model": "", "params": "", "hardware": "", "reported_metric": "near-perfect success within 1-2.5 h real-world training", "metric_measures": "other", "benchmarks": ["real dexterous manipulation"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Human-in-the-loop vision RL system reaching near-perfect success on dexterous tasks in 1-2.5 hours of real training.", "relevance": 3, "relevance_reason": "Reference sample-efficient real-world RL, base for VLA RL fine-tuning systems.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: HIL-SERL SERL sample-efficient real-world robot reinforcement learning fine-tuning pretrained"}
{"id": "2410.16981", "title": "Proleptic Temporal Ensemble for Improving the Speed of Robot Tasks Generated by Imitation Learning", "authors": ["Hyeonjun Park", "Daegyu Lim", "Seungyeon Kim", "Sumin Park"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.16981", "pdf_url": "https://arxiv.org/pdf/2410.16981", "code_url": "", "clusters": ["C1", "C2", "C7"], "keywords": ["temporal ensemble", "execution speed", "proleptic temporal ensemble", "ACT"], "model_class": "other", "techniques": ["temporal ensembling", "temporal ensembling variant"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 3x task execution speed vs ACT", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Temporal-ensemble variant that executes future actions early to speed up imitation-learned tasks beyond demonstrator speed without extra computation.", "relevance": 3, "relevance_reason": "Execution-speed improvement for chunked policies.", "foundational_technique": false, "verified": true, "found_via": "search query: temporal ensembling; arXiv search: temporal ensembling action chunking"}
{"id": "10.1109/access.2024.3367002", "title": "Pruning With Scaled Policy Constraints for Light-Weight Reinforcement Learning", "authors": ["Seongmin Park", "Hyungmin Kim", "Hyunhak Kim", "Jungwook Choi"], "year": 2024, "venue": "IEEE Access", "type": "paper", "url": "https://doi.org/10.1109/ACCESS.2024.3367002", "pdf_url": "", "code_url": "", "clusters": ["C2"], "keywords": ["structured pruning", "DRL", "embedded"], "model_class": "other", "techniques": ["structured pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Structured pruning of deep RL policies with scaled policy constraints to keep performance at high pruning ratios on standard hardware.", "relevance": 3, "relevance_reason": "Same group as QAIL; structured pruning of control policies.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2412.01034 (references)"}
{"id": "2410.08119", "title": "Q-VLM: Post-training Quantization for Large Vision-Language Models", "authors": ["Changyuan Wang", "Ziwei Wang", "Xiu-Wei Xu", "Yansong Tang", "Jie Zhou", "Jiwen Lu"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2410.08119", "pdf_url": "https://arxiv.org/pdf/2410.08119", "code_url": "", "clusters": ["C2"], "keywords": ["VLM quantization"], "model_class": "other", "techniques": ["post-training quantization", "cross-layer dependency"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PTQ framework for large VLMs that searches rounding strategies accounting for cross-layer dependencies.", "relevance": 3, "relevance_reason": "Foundational VLM quantization technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2602.03782, 2604.11572"}
{"id": "2405.04532", "title": "QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving", "authors": ["Yu-Jun Lin", "Haotian Tang", "Shang Yang", "Zhe-Kai Zhang", "Guang-Xuan Xiao", "Chuang Gan", "Song Han"], "year": 2024, "venue": "Conference on Machine Learning and Systems", "type": "paper", "url": "https://arxiv.org/abs/2405.04532", "pdf_url": "https://arxiv.org/pdf/2405.04532", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["LLM serving", "GPU kernels"], "model_class": "other", "techniques": ["W4A8KV4 quantization", "system co-design"], "base_model": "", "params": "", "hardware": "A100, L40S", "reported_metric": "up to 3.5x serving throughput vs TensorRT-LLM (Qwen1.5-72B, L40S)", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "QServe co-designs 4-bit weight, 8-bit activation, 4-bit KV quantization with GPU kernels to speed LLM serving.", "relevance": 3, "relevance_reason": "generic efficiency technique connected to 5 core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2509.09090, 2605.28803"}
{"id": "2412.15576", "title": "QUART-Online: Latency-Free Large Multimodal Language Model for Quadruped Robot Learning", "authors": ["Xinyang Tong", "Pengxiang Ding", "Yiguo Fan", "Donglin Wang", "Wenjie Zhang", "Can Cui", "Mingyang Sun", "Han Zhao", "Hongyin Zhang", "Yonghao Dang", "Siteng Huang", "Shangke Lyu"], "year": 2024, "venue": "ICRA 2025", "type": "paper", "url": "https://arxiv.org/abs/2412.15576", "pdf_url": "https://arxiv.org/pdf/2412.15576", "code_url": "https://quart-online.github.io", "clusters": ["C1", "C7"], "keywords": ["quadruped VLA", "action chunk discretization", "latency-free", "QUART-Online"], "model_class": "VLA", "techniques": ["action tokenization", "action chunking", "action chunk discretization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Quadruped MLLM-based VLA that compresses action chunks into discrete codes so inference keeps up with the control frequency without shrinking the language model.", "relevance": 3, "relevance_reason": "Latency reduction via action compression.", "foundational_technique": false, "verified": true, "found_via": "search query: action chunk latency VLA; arXiv API all:\"action chunking\" AND latency"}
{"id": "2402.04396", "title": "QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks", "authors": ["Albert Tseng", "Jerry Chee", "Qing-Yao Sun", "V. Kuleshov", "C. De Sa"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2402.04396", "pdf_url": "https://arxiv.org/pdf/2402.04396", "code_url": "", "clusters": ["C2"], "keywords": ["LLM quantization", "2-4 bit"], "model_class": "other", "techniques": ["post-training quantization", "vector quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Weight-only PTQ using Hadamard incoherence and E8 lattice codebooks for extreme low-bit LLM compression.", "relevance": 3, "relevance_reason": "Foundational quantization technique connected to multiple core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135"}
{"id": "2409.19190", "title": "RAIL: Reachability-Aided Imitation Learning for Safe Policy Execution", "authors": ["Wonsuhk Jung", "D. Anthony", "U. A. Mishra", "N. R. Arachchige", "Matthew Bronars", "Danfei Xu", "Shreyas Kousik"], "year": 2024, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2409.19190", "pdf_url": "https://arxiv.org/pdf/2409.19190", "code_url": "", "clusters": ["C7"], "keywords": ["safe imitation learning"], "model_class": "other", "techniques": ["reachability analysis", "hard constraints"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Studies enforcing hard safety constraints on IL policies with reachability-aided filtering.", "relevance": 3, "relevance_reason": "Safety filtering for learned policies (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2511.06385"}
{"id": "2404.13456", "title": "Real-Time Safe Control of Neural Network Dynamic Models with Sound Approximation", "authors": ["Hanjiang Hu", "Jianglin Lan", "Changliu Liu"], "year": 2024, "venue": "L4DC 2024", "type": "paper", "url": "https://arxiv.org/abs/2404.13456", "pdf_url": "https://arxiv.org/pdf/2404.13456", "code_url": "https://github.com/intelligent-control-lab/BOND", "clusters": ["C7"], "keywords": ["BOND", "neural network dynamic models", "safety index", "real-time safe control"], "model_class": "other", "techniques": ["Bernstein polynomial over-approximation", "worst-case safety index"], "base_model": "", "params": "", "hardware": "", "reported_metric": "10-100x faster than MIP safe control baseline", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Makes safe control with neural-network dynamics models real-time by using a sound Bernstein over-approximation and an offline worst-case safety index, giving a linearized online constraint.", "relevance": 3, "relevance_reason": "Real-time certified safe control with learned models (Liu lab).", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"safety index\" synthesis safe control"}
{"id": "2408.12588", "title": "Real-Time Video Generation with Pyramid Attention Broadcast", "authors": ["Xuan-Lei Zhao", "Xiaolong Jin", "Kai Wang", "Yang You"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2408.12588", "pdf_url": "https://arxiv.org/pdf/2408.12588", "code_url": "", "clusters": ["C8"], "keywords": ["video DiT acceleration", "real-time video generation"], "model_class": "world model", "techniques": ["attention broadcast", "training-free acceleration", "sequence parallel"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 10.5x speedup", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free acceleration of DiT video generation by reusing attention outputs across diffusion steps.", "relevance": 3, "relevance_reason": "Foundational efficiency technique for video world models.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2506.08009, 2603.28489"}
{"id": "2409.15250", "title": "ReVLA: Reverting Visual Domain Limitation of Robotic Foundation Models", "authors": ["Sombit Dey", "Jan-Nico Zaech", "Nikolay Nikolov", "Luc Van Gool", "Danda Pani Paudel"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2409.15250", "pdf_url": "https://arxiv.org/pdf/2409.15250", "code_url": "", "clusters": ["C3"], "keywords": ["ReVLA", "model merging", "visual forgetting"], "model_class": "VLA", "techniques": ["model merging", "encoder reversion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Gradually merges VLA vision encoders back toward their original foundation weights to recover visual OOD generalization.", "relevance": 3, "relevance_reason": "Cheap post-hoc fix of fine-tuning forgetting.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"model merging\""}
{"id": "2403.10672", "title": "Riemannian Flow Matching Policy for Robot Motion Learning", "authors": ["Max Braun", "Noémie Jaquier", "L. Rozo", "T. Asfour"], "year": 2024, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "type": "paper", "url": "https://arxiv.org/abs/2403.10672", "pdf_url": "http://arxiv.org/pdf/2403.10672", "code_url": "", "clusters": ["C1"], "keywords": ["Riemannian flow matching", "RFMP"], "model_class": "diffusion/flow policy", "techniques": ["flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Flow-matching visuomotor policy on Riemannian manifolds that yields smoother trajectories with much lower inference time than diffusion policies.", "relevance": 3, "relevance_reason": "Fast-inference generative policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2503.08558"}
{"id": "2412.09858", "title": "RLDG: Robotic Generalist Policy Distillation via Reinforcement Learning", "authors": ["Charles Xu", "Qiyang Li", "Jianlan Luo", "Sergey Levine"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.09858", "pdf_url": "https://arxiv.org/pdf/2412.09858", "code_url": "", "clusters": ["C3"], "keywords": ["RLDG", "RL-generated data", "distillation into generalist"], "model_class": "VLA", "techniques": ["RL data generation", "fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Trains RL specialists and uses their data to fine-tune generalist policies, beating human-demo fine-tuning on precise tasks.", "relevance": 3, "relevance_reason": "Data-efficient generalist fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2406.02523", "title": "RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots", "authors": ["Nasiriany, Soroush", "Maddukuri, Abhiram", "Zhang, Lance", "Parikh, Adeet", "Lo, Aaron", "Joshi, Abhishek", "Mandlekar, Ajay", "Zhu, Yuke"], "year": 2024, "venue": "RSS 2024", "type": "benchmark", "url": "https://arxiv.org/abs/2406.02523", "pdf_url": "https://arxiv.org/pdf/2406.02523", "code_url": "", "clusters": ["C4"], "keywords": ["RoboCasa", "kitchen simulation", "generative assets"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Kitchen-scale simulation with thousands of assets, 100 evaluation tasks and synthetic demo generation, showing scaling trends with synthetic data.", "relevance": 3, "relevance_reason": "Seed benchmark.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2409.05865", "title": "Robot Utility Models: General Policies for Zero-Shot Deployment in New Environments", "authors": ["Haritheja Etukuru", "N. Naka", "Zi-Jing Hu", "Seungjae Lee", "Julian Mehu", "A. Edsinger", "Chris Paxton", "Soumith Chintala", "Lerrel Pinto", "Nur Muhammad (Mahi) Shafiullah"], "year": 2024, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2409.05865", "pdf_url": "https://arxiv.org/pdf/2409.05865", "code_url": "", "clusters": ["C5"], "keywords": ["Robot Utility Models", "Hello Stretch"], "model_class": "other", "techniques": ["on-device deployment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Zero-shot deployable policies trained from quickly collected data and run on-device on a cheap commodity mobile robot.", "relevance": 3, "relevance_reason": "Onboard deployment on low-cost hardware.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2409.12514"}
{"id": "2406.08545", "title": "RVT-2: Learning Precise Manipulation from Few Demonstrations", "authors": ["Ankit Goyal", "Valts Blukis", "Jie Xu", "Yi-Jie Guo", "Yu-Wei Chao", "Dieter Fox"], "year": 2024, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2406.08545", "pdf_url": "https://arxiv.org/pdf/2406.08545", "code_url": "", "clusters": ["C1"], "keywords": ["RVT-2", "multi-view transformer"], "model_class": "other", "techniques": ["system-level optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2x faster inference, 6x faster training vs RVT", "metric_measures": "relative latency", "benchmarks": ["RLBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Multitask 3D manipulation model improving RVT with architectural and system changes for precision and speed.", "relevance": 3, "relevance_reason": "Reports inference and training speedups", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2406.02436", "title": "Safe, Out-of-Distribution-Adaptive MPC with Conformalized Neural Network Ensembles", "authors": ["J. Contreras", "Olaoluwa Shorinwa", "Mac Schwager"], "year": 2024, "venue": "L4DC 2024", "type": "paper", "url": "https://arxiv.org/abs/2406.02436", "pdf_url": "https://arxiv.org/pdf/2406.02436", "code_url": "", "clusters": ["C7"], "keywords": ["SODA-MPC", "conformalized ensembles", "OOD runtime monitor", "reachability fallback"], "model_class": "other", "techniques": ["ensemble OOD monitor with conformal calibration", "reachability-based fallback"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MPC using learned-ensemble predictions whose runtime OOD monitor, conformally calibrated for false-positive rate, triggers a reachability-based safe fallback controller.", "relevance": 3, "relevance_reason": "Monitor-plus-fallback architecture for learned components.", "foundational_technique": false, "verified": true, "found_via": "Semantic Scholar search: runtime monitoring learned robot controller out-of-distribution"}
{"id": "2405.18554", "title": "Scalable Surrogate Verification of Image-based Neural Network Control Systems using Composition and Unrolling", "authors": ["Feiyang Cai", "Chuchu Fan", "Stanley Bak"], "year": 2024, "venue": "AAAI 2025", "type": "paper", "url": "https://arxiv.org/abs/2405.18554", "pdf_url": "https://arxiv.org/pdf/2405.18554", "code_url": "", "clusters": ["C7"], "keywords": ["image-based controller"], "model_class": "other", "techniques": ["surrogate verification", "cGAN", "reachability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Surrogate verification of image-based NN control using a cGAN image generator, composing dynamics, generator and controller and unrolling to reduce over-approximation.", "relevance": 3, "relevance_reason": "Scales formal analysis toward vision-in-the-loop controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2409.20537", "title": "Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers", "authors": ["Lirui Wang", "Xinlei Chen", "Jialiang Zhao", "Kaiming He"], "year": 2024, "venue": "NeurIPS 2024", "type": "preprint", "url": "https://arxiv.org/abs/2409.20537", "pdf_url": "https://arxiv.org/pdf/2409.20537", "code_url": "", "clusters": ["C3"], "keywords": ["HPT", "embodiment stems", "shared trunk"], "model_class": "other", "techniques": ["modular stems", "frozen trunk transfer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Pretrains a shared trunk across embodiments with per-embodiment stems/heads so adapting to a new robot trains small modules.", "relevance": 3, "relevance_reason": "Cheap cross-embodiment adaptation architecture.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2409.12616", "title": "Semi-Supervised Safe Visuomotor Policy Synthesis using Barrier Certificates", "authors": ["Manan Tayal", "Aditya Singh", "Pushpak Jagtap", "Shishir Kolathaya"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2409.12616", "pdf_url": "https://arxiv.org/pdf/2409.12616", "code_url": "", "clusters": ["C7"], "keywords": ["safe visuomotor policy"], "model_class": "other", "techniques": ["barrier certificates", "semi-supervised learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Semi-supervised synthesis of visuomotor policies with barrier certificates that do not need full safety labels or exact dynamics.", "relevance": 3, "relevance_reason": "Barrier-certificate approach for vision-based policies.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:reachability AND (\"visuomotor policy\" OR \"learned policy\" OR \"robot foundation model\" OR \"vision-language-action\")"}
{"id": "gdm-shaping-future-advanced-robotics-blog-2024", "title": "Shaping the future of advanced robotics (AutoRT, SARA-RT, RT-Trajectory)", "authors": ["Google DeepMind Robotics Team"], "year": 2024, "venue": "blog (Google DeepMind)", "type": "blog", "url": "https://deepmind.google/discover/blog/shaping-the-future-of-advanced-robotics/", "pdf_url": "", "code_url": "", "clusters": ["C1", "C8", "C9"], "keywords": ["SARA-RT", "AutoRT", "robot constitution", "RT-Trajectory"], "model_class": "VLA", "techniques": ["linear attention up-training", "safety rules for data collection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "SARA-RT-2 10.6% more accurate, 14% faster than RT-2", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Jan 2024 post introducing SARA-RT (linear-attention up-training for faster robot transformers), AutoRT fleet data collection with a safety 'robot constitution', and RT-Trajectory.", "relevance": 3, "relevance_reason": "Blog companion for SARA-RT with speed numbers; numbers taken via search summary.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Google DeepMind robotics"}
{"id": "2403.03853", "title": "ShortGPT: Layers in Large Language Models are More Redundant Than You Expect", "authors": ["Xin Men", "Mingyu Xu", "Qingyu Zhang", "Bingning Wang", "Hongyu Lin", "Yaojie Lu", "Xianpei Han", "Weipeng Chen"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2403.03853", "pdf_url": "https://arxiv.org/pdf/2403.03853", "code_url": "", "clusters": ["C2"], "keywords": ["layer redundancy", "block influence", "layer removal"], "model_class": "other", "techniques": ["layer pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows many LLM layers are redundant by a block-influence metric and removes them directly.", "relevance": 3, "relevance_reason": "Basis for depth pruning of VLA backbones.", "foundational_technique": true, "verified": true, "found_via": "search: VLA pruning"}
{"id": "2407.01531", "title": "Sparse Diffusion Policy: A Sparse, Reusable, and Flexible Policy for Robot Learning", "authors": ["Yixiao Wang", "Yifei Zhang", "Mingxiao Huo", "Ran Tian", "Xiang Zhang", "Yichen Xie", "Chenfeng Xu", "Pengliang Ji", "Wei Zhan", "Mingyu Ding", "Masayoshi Tomizuka"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2407.01531", "pdf_url": "https://arxiv.org/pdf/2407.01531", "code_url": "", "clusters": ["C3"], "keywords": ["Sparse Diffusion Policy", "MoE", "continual learning"], "model_class": "diffusion/flow policy", "techniques": ["mixture of experts", "expert reuse"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "MoE diffusion policy that activates task-specific experts so new tasks can be learned without retraining the whole model.", "relevance": 3, "relevance_reason": "Parameter-efficient multitask/continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2405.16406", "title": "SpinQuant: LLM quantization with learned rotations", "authors": ["Zechun Liu", "Changsheng Zhao", "Igor Fedorov", "Bilge Soran", "Dhruv Choudhary", "Raghuraman Krishnamoorthi", "Vikas Chandra", "Yuandong Tian", "Tijmen Blankevoort"], "year": 2024, "venue": "ICLR 2025", "type": "paper", "url": "https://arxiv.org/abs/2405.16406", "pdf_url": "https://arxiv.org/pdf/2405.16406", "code_url": "", "clusters": ["C2"], "keywords": ["learned rotations", "W4A4KV4"], "model_class": "other", "techniques": ["rotation-based quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns the rotation matrices for outlier removal instead of using random Hadamard, improving 4-bit LLM quantization.", "relevance": 3, "relevance_reason": "Rotation-PTQ technique used as VLA baseline.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2404.00456"}
{"id": "2312.11255", "title": "State-action control barrier functions: Imposing safety on learning-based control with low online computational costs", "authors": ["Kanghui He", "Shengling Shi", "Ton van den Boom", "Bart De Schutter"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2312.11255", "pdf_url": "https://arxiv.org/pdf/2312.11255", "code_url": "", "clusters": ["C7"], "keywords": ["state-action CBF", "low online cost", "learning-based control"], "model_class": "other", "techniques": ["state-action control barrier functions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Introduces state-action CBFs whose safety condition is simple to evaluate online, reducing the real-time computational cost of CBF filtering over learned policies.", "relevance": 3, "relevance_reason": "Computational cost of safety filters in the loop.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"hamilton-jacobi\" reachability learned policy safety filter"}
{"id": "2403.02310", "title": "Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve", "authors": ["Amey Agrawal", "Nitin Kedia", "A. Panwar", "Jayashree Mohan", "Nipun Kwatra", "Bhargav S. Gulavani", "Alexey Tumanov", "R. Ramjee"], "year": 2024, "venue": "USENIX Symposium on Operating Systems Design and Implementation", "type": "paper", "url": "https://arxiv.org/abs/2403.02310", "pdf_url": "https://arxiv.org/pdf/2403.02310", "code_url": "", "clusters": ["C6"], "keywords": ["LLM serving"], "model_class": "other", "techniques": ["chunked prefill", "stall-free scheduling"], "base_model": "", "params": "", "hardware": "A100", "reported_metric": "2.6x higher serving capacity vs vLLM (Mistral-7B, single A100)", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LLM serving scheduler that balances throughput and latency via chunked prefills.", "relevance": 3, "relevance_reason": "Foundational serving technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2210.17323"}
{"id": "github-rail-berkeley-tensorrt-openvla", "title": "TensorRT-OpenVLA", "authors": ["RAIL Berkeley"], "year": 2024, "venue": "GitHub", "type": "repo", "url": "https://github.com/rail-berkeley/tensorrt-openvla", "pdf_url": "", "code_url": "https://github.com/rail-berkeley/tensorrt-openvla", "clusters": ["C2", "C5"], "keywords": ["TensorRT-LLM", "OpenVLA", "FP8", "ECoT"], "model_class": "VLA", "techniques": ["TensorRT-LLM compilation", "FP8 quantization (Hopper)"], "base_model": "OpenVLA-7B / Embodied-CoT", "params": "7B", "hardware": "Hopper GPUs for FP8; other NVIDIA GPUs", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Recipe to compile OpenVLA-architecture models (including ECoT) into TensorRT-LLM engines, passing 256 image embeddings as a prompt table, with optional FP8.", "relevance": 3, "relevance_reason": "Reference TensorRT-LLM port of OpenVLA; README gives no speed numbers.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub repo search openvla+tensorrt"}
{"id": "2402.08191", "title": "THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation", "authors": ["Wilbert Pumacay", "Ishika Singh", "Jiafei Duan", "Ranjay Krishna", "Jesse Thomason", "Dieter Fox"], "year": 2024, "venue": "RSS 2024", "type": "benchmark", "url": "https://arxiv.org/abs/2402.08191", "pdf_url": "https://arxiv.org/pdf/2402.08191", "code_url": "", "clusters": ["C4"], "keywords": ["THE COLOSSEUM", "perturbation axes", "RLBench-based", "generalization"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["THE COLOSSEUM", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "20 RLBench-derived tasks under 14 perturbation axes; models lose 30-50% success per factor and >=75% under combined perturbations, with real-world validation.", "relevance": 3, "relevance_reason": "Precursor robustness benchmark; Colosseum V2 extends it for VLAs.", "foundational_technique": false, "verified": true, "found_via": "search query: THE COLOSSEUM"}
{"id": "2410.10088", "title": "The Ingredients for Robotic Diffusion Transformers", "authors": ["Sudeep Dasari", "Oier Mees", "Sebastian Zhao", "Mohan Kumar Srirama", "Sergey Levine"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.10088", "pdf_url": "https://arxiv.org/pdf/2410.10088", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion transformer policy", "design choices", "ALOHA"], "model_class": "diffusion/flow policy", "techniques": ["diffusion transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Identifies architectural design decisions that make diffusion-transformer policies train stably and scale, solving long-horizon ALOHA tasks.", "relevance": 3, "relevance_reason": "Diffusion policy architecture.", "foundational_technique": false, "verified": true, "found_via": "search query: RDT"}
{"id": "2412.03568", "title": "The Matrix: Infinite-Horizon World Generation with Real-Time Moving Control", "authors": ["Rui-Li Feng", "Han Zhang", "Zhilei Shu", "Zhan-Tao Yang", "Long-Xiang Tang", "Zhi-Cai Wang", "Zian Andy Zheng", "Jie Xiao", "Zhiheng Liu", "Rui-Hang Chu", "Yukun Huang", "Yu Liu", "Hong-Yang Zhang"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2412.03568", "pdf_url": "https://arxiv.org/pdf/2412.03568", "code_url": "", "clusters": ["C8"], "keywords": ["real-time world simulator", "infinite horizon"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "16 FPS", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Real-time controllable world simulator generating continuous 720p video streams.", "relevance": 3, "relevance_reason": "Real-time interactive world model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2310.16828"}
{"id": "2407.04988", "title": "The Reachability Problem for Neural-Network Control Systems", "authors": ["Christian Schilling", "Martin Zimmermann"], "year": 2024, "venue": "AISoLA 2023", "type": "paper", "url": "https://arxiv.org/abs/2407.04988", "pdf_url": "https://arxiv.org/pdf/2407.04988", "code_url": "", "clusters": ["C7"], "keywords": ["reachability problem", "NNCS"], "model_class": "other", "techniques": ["decidability analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proves reachability for ReLU NN control systems is undecidable even for trivial plants and small fixed-depth networks, and semi-decidable under automata restrictions.", "relevance": 3, "relevance_reason": "Fundamental limit on exact verification of learned controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2407.20179", "title": "Theia: Distilling Diverse Vision Foundation Models for Robot Learning", "authors": ["Jinghuan Shang", "Karl Schmeckpeper", "Brandon B. May", "M. Minniti", "Tarik Kelestemur", "David Watkins", "Laura Herlant"], "year": 2024, "venue": "CoRL 2024", "type": "paper", "url": "https://arxiv.org/abs/2407.20179", "pdf_url": "https://arxiv.org/pdf/2407.20179", "code_url": "", "clusters": ["C2"], "keywords": ["vision foundation model distillation", "compact encoder"], "model_class": "other", "techniques": ["multi-teacher distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills several vision foundation models into one compact robot-learning encoder that beats its teachers with smaller size and less data.", "relevance": 3, "relevance_reason": "Compact perception backbone via distillation for policies.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246"}
{"id": "github-openvla-issue-66-rtx4090-hz", "title": "Unable to Reproduce 6 Actions/s Inference on RTX4090 (openvla/openvla issue #66)", "authors": ["depetrol", "jonathonsather-bc", "kpertsch"], "year": 2024, "venue": "GitHub issue", "type": "repo", "url": "https://github.com/openvla/openvla/issues/66", "pdf_url": "", "code_url": "", "clusters": ["C4"], "keywords": ["OpenVLA", "reproducibility", "control rate"], "model_class": "VLA", "techniques": [], "base_model": "OpenVLA-7B", "params": "7B", "hardware": "RTX 4090; RTX A6000", "reported_metric": "4.4 actions/s vs 6.02 actions/s on RTX 4090 by different users; 2.06 actions/s on A6000 (bf16, flash-attn)", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Users report 4.4 to 6.0 actions/s for the same model and GPU class, showing that single-number VLA speed claims vary with software stack and setup.", "relevance": 3, "relevance_reason": "Evidence that reported VLA speeds are not reproducible across setups without stack details (Q3).", "foundational_technique": false, "verified": true, "found_via": "gap check: Q4: GitHub issues repo:openvla/openvla 4-bit"}
{"id": "2409.14580", "title": "Updating Robot Safety Representations Online From Natural Language Feedback", "authors": ["L. Santos", "Zi-Rui Li", "Lasse Peters", "Somil Bansal", "A. Bajcsy"], "year": 2024, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2409.14580", "pdf_url": "https://arxiv.org/pdf/2409.14580", "code_url": "", "clusters": ["C7"], "keywords": ["language feedback", "safety constraints"], "model_class": "other", "techniques": ["HJ reachability", "warm-starting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses a VLM to turn language feedback into safety constraints and updates an HJ-reachability safety controller online with warm-starting.", "relevance": 3, "relevance_reason": "Online safety-filter updates at deployment.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2309.05837; cited by 2410.15185"}
{"id": "10.1109/access.2024.3379018", "title": "Utilizing Machine Learning Techniques for Worst-Case Execution Time Estimation on GPU Architectures", "authors": ["Vikash Kumar", "Behnaz Ranjbar", "Akash S. Kumar"], "year": 2024, "venue": "IEEE Access (2024)", "type": "paper", "url": "https://doi.org/10.1109/access.2024.3379018", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["GPU WCET"], "model_class": "other", "techniques": ["WCET estimation", "machine learning", "MPS"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses ML models to estimate WCET of GPU applications run concurrently under NVIDIA MPS partitioning for real-time scheduling.", "relevance": 3, "relevance_reason": "Directly targets WCET estimation for GPU workloads.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2412.18194", "title": "VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks", "authors": ["Zhang, Shiduo", "Xu, Zhe", "Liu, Peiju", "Yu, Xiaopeng", "Li, Yuan", "Gao, Qinghui", "Fei, Zhaoye", "Yin, Zhangyue", "Wu, Zuxuan", "Jiang, Yu-Gang", "Qiu, Xipeng"], "year": 2024, "venue": "ICCV 2025", "type": "benchmark", "url": "https://arxiv.org/abs/2412.18194", "pdf_url": "https://arxiv.org/pdf/2412.18194", "code_url": "", "clusters": ["C4"], "keywords": ["VLABench", "language-conditioned", "long-horizon reasoning"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["VLABench"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "100 task categories with strong randomization and 2000+ objects testing world knowledge, implicit intent and multi-step reasoning for VLAs and LLM-based agents.", "relevance": 3, "relevance_reason": "Seed benchmark; later audited for bugs affecting acceleration evaluation.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2409.12894", "title": "VLATest: Testing and Evaluating Vision-Language-Action Models for Robotic Manipulation", "authors": ["Zhijie Wang", "Zhehua Zhou", "Jia-Yang Song", "Yu-Heng Huang", "Zhan Shu", "Lei Ma"], "year": 2024, "venue": "Proc. ACM Softw. Eng.", "type": "paper", "url": "https://arxiv.org/abs/2409.12894", "pdf_url": "https://dl.acm.org/doi/pdf/10.1145/3729343", "code_url": "", "clusters": ["C4"], "keywords": ["VLATest", "robustness testing"], "model_class": "VLA", "techniques": ["fuzzing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fuzzing framework that generates manipulation scenes to test VLAs, revealing robustness gaps across seven models.", "relevance": 3, "relevance_reason": "Evaluation methodology for VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941; cited by 2410.08001"}
{"id": "2306.11695", "title": "A Simple and Effective Pruning Approach for Large Language Models", "authors": ["Mingjie Sun", "Zhuang Liu", "Anna Bair", "J. Zico Kolter"], "year": 2023, "venue": "ICLR 2024", "type": "paper", "url": "https://arxiv.org/abs/2306.11695", "pdf_url": "https://arxiv.org/pdf/2306.11695", "code_url": "", "clusters": ["C2"], "keywords": ["Wanda", "weight-activation pruning"], "model_class": "other", "techniques": ["unstructured pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Prunes weights by magnitude times input activation norm, no retraining or weight update needed.", "relevance": 3, "relevance_reason": "Pruning baseline used in VLA pruning papers.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2510.08464"}
{"id": "2304.10020", "title": "A Survey on Deep Neural Network Partition over Cloud, Edge and End Devices", "authors": ["Di Xu", "Xiang He", "Tonghua Su", "Zhongjie Wang"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2304.10020", "pdf_url": "https://arxiv.org/pdf/2304.10020", "code_url": "", "clusters": ["C6"], "keywords": ["DNN partition", "survey", "cloud-edge-end"], "model_class": "other", "techniques": ["split computing", "DNN partitioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey with a unified model and five-axis taxonomy (locations, granularity, constraints, objectives, algorithms) of DNN partitioning across cloud, edge and device.", "relevance": 3, "relevance_reason": "General split-computing background that VLA partitioning papers build on.", "foundational_technique": true, "verified": true, "found_via": "search query"}
{"id": "2310.19102", "title": "Atom: Low-bit Quantization for Efficient and Accurate LLM Serving", "authors": ["Yi-Long Zhao", "Chien-Yu Lin", "Kan Zhu", "Zihao Ye", "Lequn Chen", "Size Zheng", "Luis Ceze", "Arvind Krishnamurthy", "Tian-Qi Chen", "Baris Kasikci"], "year": 2023, "venue": "Conference on Machine Learning and Systems", "type": "paper", "url": "https://arxiv.org/abs/2310.19102", "pdf_url": "https://arxiv.org/pdf/2310.19102", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["Atom", "W4A4", "LLM serving"], "model_class": "other", "techniques": ["low-bit quantization", "mixed precision"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 7.73x throughput vs FP16", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "4-bit weight-activation quantization with mixed precision to raise LLM serving throughput.", "relevance": 3, "relevance_reason": "Generic quantization technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323; cited by 2404.00456"}
{"id": "2310.11453", "title": "BitNet: Scaling 1-bit Transformers for Large Language Models", "authors": ["Hongyu Wang", "Shu-Ming Ma", "Li Dong", "Shao-Han Huang", "Huai-Jie Wang", "Ling-Xiao Ma", "Fan Yang", "Rui-Ping Wang", "Yi Wu", "Furu Wei"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2310.11453", "pdf_url": "https://arxiv.org/pdf/2310.11453", "code_url": "", "clusters": ["C2"], "keywords": ["BitLinear", "1-bit transformer"], "model_class": "other", "techniques": ["1-bit quantization-aware training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Introduces BitLinear to train 1-bit-weight transformers from scratch with competitive LM quality and lower energy.", "relevance": 3, "relevance_reason": "Root of BitNet line used by BitVLA.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2506.07530 (references)"}
{"id": "2310.06343", "title": "Boosting Continuous Control with Consistency Policy", "authors": ["Yuhui Chen", "Hao-Ran Li", "Dongbin Zhao"], "year": 2023, "venue": "Adaptive Agents and Multi-Agent Systems", "type": "paper", "url": "https://arxiv.org/abs/2310.06343", "pdf_url": "https://arxiv.org/pdf/2310.06343", "code_url": "https://github.com/cccedric/cpql", "clusters": ["C1"], "keywords": ["consistency policy", "one-step"], "model_class": "diffusion/flow policy", "techniques": ["consistency model", "Q-learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~45x faster inference than Diffusion-QL", "metric_measures": "relative latency", "benchmarks": ["D4RL"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "CPQL uses a consistency-model policy that produces actions in a single step with Q-guided improvement for offline and online RL.", "relevance": 3, "relevance_reason": "relevant: one-step diffusion-style policy for real-time control", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137; cited by 2405.07503, 2406.01586"}
{"id": "2309.08603", "title": "Closing the Loop on Runtime Monitors with Fallback-Safe MPC", "authors": ["Rohan Sinha", "E. Schmerling", "M. Pavone"], "year": 2023, "venue": "IEEE Conference on Decision and Control", "type": "paper", "url": "https://arxiv.org/abs/2309.08603", "pdf_url": "http://arxiv.org/pdf/2309.08603", "code_url": "", "clusters": ["C7"], "keywords": ["runtime monitor", "safe MPC"], "model_class": "other", "techniques": ["runtime monitoring", "fallback-safe MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Couples perception runtime monitors with a fallback-safe MPC so a robot can take safety-preserving actions when learned perception becomes unreliable.", "relevance": 3, "relevance_reason": "relevant: runtime monitoring with safety fallback", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2310.17552, 2410.22689"}
{"id": "2309.12786", "title": "CloudGripper: An Open Source Cloud Robotics Testbed for Robotic Manipulation Research, Benchmarking and Data Collection at Scale", "authors": ["Muhammad Zahid", "Florian T. Pokorny"], "year": 2023, "venue": "ICRA 2024 (under review at time)", "type": "preprint", "url": "https://arxiv.org/abs/2309.12786", "pdf_url": "https://arxiv.org/pdf/2309.12786", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["cloud robotics testbed", "remote experimentation", "32 work cells"], "model_class": "other", "techniques": ["cloud robotics testbed"], "base_model": "", "params": "", "hardware": "rack of 32 low-cost arms, 10 Gbit/s network", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Rack of 32 enclosed low-cost robot arm cells with high-bandwidth networking, built as a community testbed for remote, cloud-connected manipulation experiments and data collection.", "relevance": 3, "relevance_reason": "Infrastructure for studying cloud/edge robot learning at scale.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2310.01382", "title": "Compressing LLMs: The Truth is Rarely Pure and Never Simple", "authors": ["Ajay Jaiswal", "Zhe Gan", "Xianzhi Du", "Bowen Zhang", "Zhangyang Wang", "Yinfei Yang"], "year": 2023, "venue": "ICLR 2024", "type": "paper", "url": "https://arxiv.org/abs/2310.01382", "pdf_url": "https://arxiv.org/pdf/2310.01382", "code_url": "", "clusters": ["C2", "C4"], "keywords": ["LLM-KICK", "perplexity", "pruning", "quantization", "evaluation"], "model_class": "other", "techniques": ["pruning", "quantization"], "base_model": "Vicuna/LLaMA family", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows perplexity hides compression damage: pruning degrades knowledge-intensive tasks even at 25-30% sparsity while quantization holds up better, arguing for task-level evaluation of compressed models.", "relevance": 3, "relevance_reason": "LLM analogue of proxy-metric vs closed-loop gap in VLA compression.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q1: arXiv ti:\"compressing LLMs\" AND ti:truth"}
{"id": "2303.01469", "title": "Consistency Models", "authors": ["Yang Song", "Prafulla Dhariwal", "Mark Chen", "Ilya Sutskever"], "year": 2023, "venue": "ICML 2023", "type": "paper", "url": "https://arxiv.org/abs/2303.01469", "pdf_url": "https://arxiv.org/pdf/2303.01469", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["consistency models", "one-step generation"], "model_class": "other", "techniques": ["consistency distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Generative models mapping any noisy point directly to data, trainable by distillation or standalone, for one/few-step sampling.", "relevance": 3, "relevance_reason": "Foundation of Consistency Policy / ManiCM.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2405.07503"}
{"id": "2309.16984", "title": "Consistency Models as a Rich and Efficient Policy Class for Reinforcement Learning", "authors": ["Zihan Ding", "Chi Jin"], "year": 2023, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2309.16984", "pdf_url": "https://arxiv.org/pdf/2309.16984", "code_url": "", "clusters": ["C1"], "keywords": ["consistency policy", "RL"], "model_class": "diffusion/flow policy", "techniques": ["consistency model", "actor-critic"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses consistency models as a fast yet expressive policy class for offline, offline-to-online and online RL, cutting the sampling cost of diffusion policies.", "relevance": 3, "relevance_reason": "relevant: few-step generative policy", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2405.07503, 2406.01586"}
{"id": "10.1109/mcs.2023.3291885", "title": "Data-Driven Safety Filters: Hamilton-Jacobi Reachability, Control Barrier Functions, and Predictive Methods for Uncertain Systems", "authors": ["K. P. Wabersich", "Andrew J. Taylor", "J. Choi", "K. Sreenath", "C. Tomlin", "A. Ames", "M. Zeilinger"], "year": 2023, "venue": "IEEE Control Systems", "type": "paper", "url": "https://doi.org/10.1109/MCS.2023.3291885", "pdf_url": "", "code_url": "", "clusters": ["C7"], "keywords": ["survey", "uncertain systems"], "model_class": "other", "techniques": ["safety filters", "CBF", "HJ reachability", "predictive safety"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Review of data-driven safety filters (HJ reachability, control barrier functions, predictive methods) for uncertain systems.", "relevance": 3, "relevance_reason": "Core reference for safety filters in C7.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2410.15185, 2502.00935"}
{"id": "2305.18455", "title": "Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models", "authors": ["Weijian Luo", "Tianyang Hu", "Shifeng Zhang", "Jia-Cheng Sun", "Zhen-Guo Li", "Zhihua Zhang"], "year": 2023, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2305.18455", "pdf_url": "http://arxiv.org/pdf/2305.18455", "code_url": "", "clusters": ["C1"], "keywords": ["one-step", "IKL divergence"], "model_class": "other", "techniques": ["diffusion distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Data-free distillation of pretrained diffusion models into one-step generators.", "relevance": 3, "relevance_reason": "Foundational technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2406.12538, 2410.21257"}
{"id": "2311.08983", "title": "Edge Accelerated Robot Navigation With Collaborative Motion Planning", "authors": ["Guoliang Li", "Ruihua Han", "Shuai Wang", "Fei Gao", "Yonina C. Eldar", "Chengzhong Xu"], "year": 2023, "venue": "IEEE/ASME Transactions on Mechatronics", "type": "paper", "url": "https://arxiv.org/abs/2311.08983", "pdf_url": "https://arxiv.org/pdf/2311.08983", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["edge-accelerated navigation", "collaborative motion planning", "model predictive switching"], "model_class": "other", "techniques": ["local/edge planner switching", "resource-aware MPS"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Robots switch between a safe local planner and an aggressive edge-offloaded planner, choosing via model predictive switching under compute and communication constraints.", "relevance": 3, "relevance_reason": "Principled local-vs-edge switching for robot control.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2305.20081", "title": "Efficient Diffusion Policies for Offline Reinforcement Learning", "authors": ["Bingyi Kang", "Xiao Ma", "Chao Du", "Tianyu Pang", "Shuicheng Yan"], "year": 2023, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2305.20081", "pdf_url": "http://arxiv.org/pdf/2305.20081", "code_url": "", "clusters": ["C1"], "keywords": ["offline RL", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": ["efficient training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "training time 5 days -> 5 hours", "metric_measures": "other", "benchmarks": ["D4RL"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Avoids running the full diffusion chain during training to make diffusion policies efficient for offline RL.", "relevance": 3, "relevance_reason": "Diffusion policy efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2406.12538, 2409.00588"}
{"id": "2309.06180", "title": "Efficient Memory Management for Large Language Model Serving with PagedAttention", "authors": ["Woosuk Kwon", "Zhuo-Han Li", "Si-Yuan Zhuang", "Ying Sheng", "Lian-Min Zheng", "C. Yu", "Joseph E. Gonzalez", "Haotong Zhang", "Ion Stoica"], "year": 2023, "venue": "SOSP 2023", "type": "paper", "url": "https://arxiv.org/abs/2309.06180", "pdf_url": "https://arxiv.org/pdf/2309.06180", "code_url": "https://github.com/vllm-project/vllm", "clusters": ["C5"], "keywords": ["KV cache", "paging", "vLLM", "serving"], "model_class": "other", "techniques": ["paged KV cache memory management"], "base_model": "", "params": "", "hardware": "", "reported_metric": "2-4x throughput at same latency", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Manages KV cache in fixed-size pages to eliminate fragmentation and enable sharing, the basis of the vLLM serving engine.", "relevance": 3, "relevance_reason": "Serving substrate used/compared by VLA serving systems (e.g., Robion vs vLLM-Omni).", "foundational_technique": true, "verified": true, "found_via": "seed"}
{"id": "2309.15505", "title": "Finite Scalar Quantization: VQ-VAE Made Simple", "authors": ["Fabian Mentzer", "David C. Minnen", "E. Agustsson", "Michael Tschannen"], "year": 2023, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2309.15505", "pdf_url": "https://arxiv.org/pdf/2309.15505", "code_url": "", "clusters": ["C1"], "keywords": ["VQ-VAE", "tokenization"], "model_class": "other", "techniques": ["finite scalar quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces VQ codebooks with simple per-dimension scalar quantization, matching VQ performance without codebook collapse.", "relevance": 3, "relevance_reason": "Foundational tokenization technique used by action tokenizers.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2506.18123, 2511.07820"}
{"id": "2312.11983", "title": "Fluctuation-based Adaptive Structured Pruning for Large Language Models", "authors": ["Yongqi An", "Xu Zhao", "Tao Yu", "Ming Tang", "Jin-Qiao Wang"], "year": 2023, "venue": "AAAI Conference on Artificial Intelligence", "type": "paper", "url": "https://arxiv.org/abs/2312.11983", "pdf_url": "https://arxiv.org/pdf/2312.11983", "code_url": "", "clusters": ["C2"], "keywords": ["retraining-free", "LLM"], "model_class": "other", "techniques": ["structured pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "FLAP is a retraining-free structured pruning method for LLMs with fluctuation-based importance and bias compensation.", "relevance": 3, "relevance_reason": "generic efficiency technique connected to 4 core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2506.17639, 2602.00780"}
{"id": "2306.17157", "title": "FogROS2-SGC: A ROS2 Cloud Robotics Platform for Secure Global Connectivity", "authors": ["Kaiyuan Chen", "Ryan Hoque", "Karthik Dharmarajan", "Edith LLontop", "Simeon Adebola", "Jeffrey Ichnowski", "John Kubiatowicz", "Ken Goldberg"], "year": 2023, "venue": "ICRA 2024", "type": "paper", "url": "https://arxiv.org/abs/2306.17157", "pdf_url": "https://arxiv.org/pdf/2306.17157", "code_url": "https://github.com/BerkeleyAutomation/FogROS2-SGC", "clusters": ["C6"], "keywords": ["secure connectivity", "global routing", "ROS 2"], "model_class": "other", "techniques": ["secure global routing", "location-independent identifiers"], "base_model": "", "params": "", "hardware": "4 robots and compute nodes up to 3600 km apart", "reported_metric": "19x faster than rosbridge", "metric_measures": "relative latency", "benchmarks": ["grasp planning"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Extends FogROS2 with secure, globally addressable routing so robots and cloud compute across different networks can exchange ROS 2 topics without code changes.", "relevance": 3, "relevance_reason": "Networking infrastructure for robot-cloud offloading; less about model efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2305.12821", "title": "FurnitureBench: Reproducible Real-World Benchmark for Long-Horizon Complex Manipulation", "authors": ["Minho Heo", "Youngwoon Lee", "Doohyun Lee", "Joseph J. Lim"], "year": 2023, "venue": "RSS 2023", "type": "benchmark", "url": "https://arxiv.org/abs/2305.12821", "pdf_url": "https://arxiv.org/pdf/2305.12821", "code_url": "", "clusters": ["C4"], "keywords": ["FurnitureBench", "reproducible real-world", "3D-printable", "long-horizon"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["FurnitureBench", "FurnitureSim", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Reproducible real-world furniture-assembly benchmark with 3D-printable parts, setup guides, systematic initialization and 5000+ demos plus a simulator.", "relevance": 3, "relevance_reason": "Template for reproducible real-robot benchmarks.", "foundational_technique": false, "verified": true, "found_via": "search query: FurnitureBench"}
{"id": "2304.02532", "title": "Goal-Conditioned Imitation Learning using Score-based Diffusion Policies", "authors": ["Moritz Reuss", "M. Li", "Xiaogang Jia", "Rudolf Lioutikov"], "year": 2023, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2304.02532", "pdf_url": "http://arxiv.org/pdf/2304.02532", "code_url": "", "clusters": ["C1"], "keywords": ["BESO", "score-based diffusion", "goal-conditioned IL"], "model_class": "diffusion/flow policy", "techniques": ["few-step diffusion sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "3 denoising steps vs 30+ for other diffusion policies", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Score-based diffusion policy that decouples score learning from sampling, enabling goal-conditioned behaviour generation in only 3 denoising steps.", "relevance": 3, "relevance_reason": "Early few-step diffusion policy", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137"}
{"id": "2310.15478", "title": "How to Train Your Neural Control Barrier Function: Learning Safety Filters for Complex Input-Constrained Systems", "authors": ["Oswin So", "Zachary Serlin", "Makai Mann", "Jake Gonzales", "Kwesi J. Rutledge", "Nicholas Roy", "Chu-Chu Fan"], "year": 2023, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2310.15478", "pdf_url": "https://arxiv.org/pdf/2310.15478", "code_url": "", "clusters": ["C7"], "keywords": ["input constraints"], "model_class": "other", "techniques": ["neural CBF", "safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Method for learning neural control barrier functions as safety filters for input-constrained nonlinear systems.", "relevance": 3, "relevance_reason": "Safety-filter technique (C7).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2505.10547, 2511.18606"}
{"id": "2301.12012", "title": "In-Distribution Barrier Functions: Self-Supervised Policy Filters that Avoid Out-of-Distribution States", "authors": ["Fernando Castañeda", "Haruki Nishimura", "R. McAllister", "K. Sreenath", "Adrien Gaidon"], "year": 2023, "venue": "Conference on Learning for Dynamics & Control", "type": "paper", "url": "https://arxiv.org/abs/2301.12012", "pdf_url": "http://arxiv.org/pdf/2301.12012", "code_url": "", "clusters": ["C7"], "keywords": ["safety filter", "OOD"], "model_class": "other", "techniques": ["barrier functions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learned latent barrier-function filter keeping visuomotor policies in-distribution.", "relevance": 3, "relevance_reason": "Safety filters.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2502.00935, 2503.08558"}
{"id": "2302.00111", "title": "Learning Universal Policies via Text-Guided Video Generation", "authors": ["Yilun Du", "Mengjiao Yang", "Bo Dai", "Hanjun Dai", "Ofir Nachum", "Joshua B. Tenenbaum", "Dale Schuurmans", "Pieter Abbeel"], "year": 2023, "venue": "NeurIPS 2023", "type": "paper", "url": "https://arxiv.org/abs/2302.00111", "pdf_url": "https://arxiv.org/pdf/2302.00111", "code_url": "", "clusters": ["C8"], "keywords": ["UniPi", "video planning", "inverse dynamics"], "model_class": "world model", "techniques": ["text-conditioned video diffusion planner", "inverse dynamics action extraction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["combinatorial manipulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Frames decision making as text-conditioned video generation followed by inverse dynamics; the first video-planning policy and the ancestor of today's slow imagine-then-act WAMs.", "relevance": 3, "relevance_reason": "Seminal video-as-plan work.", "foundational_technique": false, "verified": true, "found_via": "snowball from NVIDIA WAM blog"}
{"id": "2305.17888", "title": "LLM-QAT: Data-Free Quantization Aware Training for Large Language Models", "authors": ["Ze-Chun Liu", "Barlas Oğuz", "Changsheng Zhao", "Ernie Chang", "Pierre Stock", "Yashar Mehdad", "Yangyang Shi", "Raghuraman Krishnamoorthi", "Vikas Chandra"], "year": 2023, "venue": "Annual Meeting of the Association for Computational Linguistics", "type": "paper", "url": "https://arxiv.org/abs/2305.17888", "pdf_url": "http://arxiv.org/pdf/2305.17888", "code_url": "", "clusters": ["C2"], "keywords": ["low-bit LLM"], "model_class": "other", "techniques": ["quantization-aware training", "data-free distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Data-free QAT for LLMs using self-generated data for distillation, enabling lower bit-widths.", "relevance": 3, "relevance_reason": "Foundational quantization technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2506.17639, 2602.15836"}
{"id": "2311.02058", "title": "LOTUS: Continual Imitation Learning for Robot Manipulation Through Unsupervised Skill Discovery", "authors": ["Weikang Wan", "Yi-Feng Zhu", "Rutav Shah", "Yu-Ke Zhu"], "year": 2023, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2311.02058", "pdf_url": "https://arxiv.org/pdf/2311.02058", "code_url": "", "clusters": ["C3"], "keywords": ["continual imitation learning"], "model_class": "other", "techniques": ["continual skill discovery"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Continual IL that grows a skill library from few demos while avoiding forgetting.", "relevance": 3, "relevance_reason": "Continual adaptation (C3).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2506.05985, 2601.09512"}
{"id": "2302.04659", "title": "ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills", "authors": ["Gu, Jiayuan", "Xiang, Fanbo", "Li, Xuanlin", "Ling, Zhan", "Liu, Xiqiang", "Mu, Tongzhou", "Tang, Yihe", "Tao, Stone", "Wei, Xinyue", "Yao, Yunchao", "Yuan, Xiaodi", "Xie, Pengwei", "Huang, Zhiao", "Chen, Rui", "Su, Hao"], "year": 2023, "venue": "ICLR 2023", "type": "benchmark", "url": "https://arxiv.org/abs/2302.04659", "pdf_url": "https://arxiv.org/pdf/2302.04659", "code_url": "", "clusters": ["C4"], "keywords": ["ManiSkill2", "generalizable manipulation", "unified evaluation protocol"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill2"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "20 task families with 2000+ objects and 4M+ demo frames under a unified interface and evaluation protocol across controllers and observation types.", "relevance": 3, "relevance_reason": "Seed benchmark.", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "2310.07896", "title": "NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration", "authors": ["Ajay Sridhar", "Dhruv Shah", "Catherine Glossop", "Sergey Levine"], "year": 2023, "venue": "ICRA 2024", "type": "paper", "url": "https://arxiv.org/abs/2310.07896", "pdf_url": "https://arxiv.org/pdf/2310.07896", "code_url": "", "clusters": ["C3", "C9"], "keywords": ["diffusion policy", "navigation", "goal masking"], "model_class": "diffusion/flow policy", "techniques": ["diffusion policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Single diffusion-policy navigation model handling both goal-reaching and exploration via goal masking, deployed on small mobile robots.", "relevance": 3, "relevance_reason": "Compact diffusion navigation model deployed onboard.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2308.13137", "title": "OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models", "authors": ["Wenqi Shao", "Mengzhao Chen", "Zhaoyang Zhang", "Peng Xu", "Lirui Zhao", "Zhiqiang Li", "Kaipeng Zhang", "Peng Gao", "Yu Qiao", "Ping Luo"], "year": 2023, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2308.13137", "pdf_url": "https://arxiv.org/pdf/2308.13137", "code_url": "", "clusters": ["C2"], "keywords": ["LLM quantization", "W4A4"], "model_class": "other", "techniques": ["post-training quantization", "learnable clipping"], "base_model": "", "params": "", "hardware": "A100-40G", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Post-training LLM quantization with learnable weight clipping and equivalent transformations.", "relevance": 3, "relevance_reason": "Foundational quantization technique connected to 9 core items.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323"}
{"id": "2311.18828", "title": "One-Step Diffusion with Distribution Matching Distillation", "authors": ["Tianwei Yin", "Michael Gharbi", "Richard Zhang", "Eli Shechtman", "Frédo Durand", "William T. Freeman", "Taesung Park"], "year": 2023, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2311.18828", "pdf_url": "https://arxiv.org/pdf/2311.18828", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["distribution matching distillation", "few-step sampling"], "model_class": "other", "techniques": ["distillation", "one-step diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "20 FPS image generation (FP16)", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills a multi-step image diffusion model into a one-step generator via distribution-level matching; widely reused for few-step diffusion policies and world models.", "relevance": 3, "relevance_reason": "generic efficiency technique connected to 19 core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2405.07503, 2410.21257"}
{"id": "2304.01218", "title": "POLAR-Express: Efficient and Precise Formal Reachability Analysis of Neural-Network Controlled Systems", "authors": ["Yixuan Wang", "Weichao Zhou", "Jiameng Fan", "Zhilu Wang", "Jiajun Li", "Xin Chen", "Chao Huang", "Wenchao Li", "Qi Zhu"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2304.01218", "pdf_url": "https://arxiv.org/pdf/2304.01218", "code_url": "", "clusters": ["C7"], "keywords": ["POLAR-Express"], "model_class": "other", "techniques": ["Taylor models", "reachability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Efficient layer-by-layer Taylor-model propagation for NN controllers with improved ReLU handling and parallelism.", "relevance": 3, "relevance_reason": "Faster NNCS reachability tool; used for runtime verification.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2302.04304", "title": "Q-Diffusion: Quantizing Diffusion Models", "authors": ["Xiuyu Li", "Yi-Jia Liu", "Long Lian", "Hua Yang", "Zhen Dong", "Daniel Kang", "Shang-Hang Zhang", "Kurt Keutzer"], "year": 2023, "venue": "IEEE International Conference on Computer Vision", "type": "paper", "url": "https://arxiv.org/abs/2302.04304", "pdf_url": "https://arxiv.org/pdf/2302.04304", "code_url": "", "clusters": ["C2"], "keywords": ["diffusion quantization", "4-bit"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Timestep-aware calibration and split shortcut quantization to run diffusion models at 4-bit post-training.", "relevance": 3, "relevance_reason": "Foundational diffusion quantization technique", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135"}
{"id": "2308.15368", "title": "RED: A Systematic Real-Time Scheduling Approach for Robotic Environmental Dynamics", "authors": ["Zexin Li", "Tao Ren", "Xiaoxi He", "Cong Liu"], "year": 2023, "venue": "RTSS 2023", "type": "paper", "url": "https://arxiv.org/abs/2308.15368", "pdf_url": "https://arxiv.org/pdf/2308.15368", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["RED", "real-time scheduling", "robotic environmental dynamics"], "model_class": "other", "techniques": ["real-time scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Original RTSS version of RED: real-time scheduling of multi-task DNN workloads on resource-limited robots under environment-induced workload changes.", "relevance": 3, "relevance_reason": "Timing guarantees for robot DNN inference.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"deadline\" robot inference neural network real-time scheduling"}
{"id": "2306.15724", "title": "REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction", "authors": ["Zeyi Liu", "Arpit Bahety", "Shuran Song"], "year": 2023, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2306.15724", "pdf_url": "https://arxiv.org/pdf/2306.15724", "code_url": "", "clusters": ["C7"], "keywords": ["failure explanation", "LLM", "RoboFail"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Summarizes multisensory robot experience hierarchically so an LLM can explain failures and plan corrections.", "relevance": 3, "relevance_reason": "Failure detection/explanation (C7 runtime monitoring).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2410.04640, 2410.22689"}
{"id": "10.1145/3579371.3589104", "title": "RoboShape: Using Topology Patterns to Scalably and Flexibly Deploy Accelerators Across Robots", "authors": ["Sabrina M. Neuman", "Radhika Ghosal", "Thomas Bourgeat", "Brian Plancher", "V. Reddi"], "year": 2023, "venue": "International Symposium on Computer Architecture", "type": "paper", "url": "https://doi.org/10.1145/3579371.3589104", "pdf_url": "https://dl.acm.org/doi/pdf/10.1145/3579371.3589104", "code_url": "", "clusters": ["C5"], "keywords": ["robot dynamics", "hardware acceleration"], "model_class": "other", "techniques": ["FPGA accelerators", "topology patterns"], "base_model": "", "params": "", "hardware": "FPGA", "reported_metric": "4.0-4.4x lower compute latency vs CPU; 8.0-15.1x vs GPU (dynamics gradients)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses robot-topology patterns to generate FPGA accelerators for dynamics gradients across diverse robots.", "relevance": 3, "relevance_reason": "Robotics hardware accelerator design", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2407.04292, 2607.24148"}
{"id": "2310.15145", "title": "Robot Fine-Tuning Made Easy: Pre-Training Rewards and Policies for Autonomous Real-World Reinforcement Learning", "authors": ["Jingyun Yang", "Max Sobol Mark", "Brandon Vu", "Archit Sharma", "Jeannette Bohg", "Chelsea Finn"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2310.15145", "pdf_url": "https://arxiv.org/pdf/2310.15145", "code_url": "", "clusters": ["C3"], "keywords": ["RoboFuME", "reset-free fine-tuning", "VLM reward"], "model_class": "other", "techniques": ["offline-to-online RL", "VLM reward model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Reset-free real-world fine-tuning of a pretrained multitask policy with calibrated offline RL and a VLM reward.", "relevance": 3, "relevance_reason": "Low-human-effort adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2304.01089", "title": "RPTQ: Reorder-based Post-training Quantization for Large Language Models", "authors": ["Zhihang Yuan", "Lin Niu", "Jia-Wen Liu", "Wenyu Liu", "Xing-Gang Wang", "Yuzhang Shang", "Guangyu Sun", "Qiang Wu", "Jia-Xiang Wu", "Bingzhe Wu"], "year": 2023, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2304.01089", "pdf_url": "http://arxiv.org/pdf/2304.01089", "code_url": "", "clusters": ["C2"], "keywords": ["RPTQ", "activation quantization"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 80% memory reduction (OPT-175B)", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reorder-based post-training quantization clustering channels for low-bit LLM activations.", "relevance": 3, "relevance_reason": "Generic quantization technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135; cited by 2601.14140"}
{"id": "2306.14896", "title": "RVT: Robotic View Transformer for 3D Object Manipulation", "authors": ["Ankit Goyal", "Jie Xu", "Yi-Jie Guo", "Valts Blukis", "Yu-Wei Chao", "D. Fox"], "year": 2023, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2306.14896", "pdf_url": "http://arxiv.org/pdf/2306.14896", "code_url": "", "clusters": ["C1"], "keywords": ["3D manipulation", "efficient"], "model_class": "other", "techniques": ["multi-view transformer", "virtual view re-rendering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "36x faster training, 2.3x inference speed vs PerAct", "metric_measures": "relative latency", "benchmarks": ["RLBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RVT replaces costly voxel representations with a multi-view transformer over re-rendered virtual views, making 3D manipulation both more accurate and much cheaper to train and run.", "relevance": 3, "relevance_reason": "relevant: compute-efficient 3D manipulation policy", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2410.04640, 2410.08001"}
{"id": "2310.04288", "title": "Searching for Optimal Runtime Assurance via Reachability and Reinforcement Learning", "authors": ["Kristina Miller", "Christopher K. Zeitler", "William Shen", "Kerianne Hobbs", "Sayan Mitra", "John Schierman", "Mahesh Viswanathan"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2310.04288", "pdf_url": "https://arxiv.org/pdf/2310.04288", "code_url": "", "clusters": ["C7"], "keywords": ["RTA design"], "model_class": "other", "techniques": ["runtime assurance", "reachability", "RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Formulates optimal runtime-assurance switching design and solves it with reward shaping and RL while keeping safety guarantees via reachability.", "relevance": 3, "relevance_reason": "Less conservative switching between untrusted and safe controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2301.00774", "title": "SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot", "authors": ["Elias Frantar", "Dan Alistarh"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2301.00774", "pdf_url": "https://arxiv.org/pdf/2301.00774", "code_url": "", "clusters": ["C2"], "keywords": ["one-shot pruning", "50-60% sparsity"], "model_class": "other", "techniques": ["unstructured pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "One-shot pruning of GPT-scale models to 50%+ sparsity without retraining via approximate sparse regression.", "relevance": 3, "relevance_reason": "LLM pruning baseline applied to VLAs (e.g., GLUESTICK study).", "foundational_technique": true, "verified": true, "found_via": "snowball from 2510.08464"}
{"id": "2310.09615", "title": "STORM: Efficient Stochastic Transformer based World Models for Reinforcement Learning", "authors": ["Weipu Zhang", "Gang Wang", "Jian Sun", "Yetian Yuan", "Gao Huang"], "year": 2023, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2310.09615", "pdf_url": "https://arxiv.org/pdf/2310.09615", "code_url": "", "clusters": ["C8"], "keywords": ["STORM", "Atari 100k"], "model_class": "world model", "techniques": ["stochastic transformer world model"], "base_model": "", "params": "", "hardware": "RTX 3090", "reported_metric": "4.3 h training for 1.85 h of interaction", "metric_measures": "other", "benchmarks": ["Atari 100k"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Efficient world model combining a Transformer with VAE-style stochasticity for model-based RL.", "relevance": 3, "relevance_reason": "Efficient world model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104; cited by 2602.00440"}
{"id": "2306.00942", "title": "Train Offline, Test Online: A Real Robot Learning Benchmark", "authors": ["Gaoyue Zhou", "Victoria Dean", "M. K. Srirama", "A. Rajeswaran", "Jyothish Pari", "K. Hatch", "Aryan Jain", "Tian-He Yu", "P. Abbeel", "Lerrel Pinto", "Chelsea Finn", "Abhi Gupta"], "year": 2023, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2306.00942", "pdf_url": "https://arxiv.org/pdf/2306.00942", "code_url": "", "clusters": ["C4"], "keywords": ["real-robot benchmark", "remote evaluation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["TOTO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "TOTO: a shared real-robot manipulation benchmark with offline training data and remote online evaluation of submitted policies.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2405.05941, 2405.15223"}
{"id": "2306.14846", "title": "ViNT: A Foundation Model for Visual Navigation", "authors": ["Dhruv Shah", "Ajay Sridhar", "Nitish Dashora", "Kyle Stachowicz", "Kevin Black", "Noriaki Hirose", "Sergey Levine"], "year": 2023, "venue": "CoRL 2023", "type": "paper", "url": "https://arxiv.org/abs/2306.14846", "pdf_url": "https://arxiv.org/pdf/2306.14846", "code_url": "", "clusters": ["C4", "C9"], "keywords": ["navigation foundation model", "adaptation"], "model_class": "other", "techniques": ["efficient adaptation", "prompt-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Transformer navigation foundation model trained across many robot datasets that adapts efficiently to new tasks and runs on small robots.", "relevance": 3, "relevance_reason": "Seminal compact robot foundation model deployed on edge platforms.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2311.01378", "title": "Vision-Language Foundation Models as Effective Robot Imitators", "authors": ["Xing-Hang Li", "Ming-Huan Liu", "Hanbo Zhang", "Cunjun Yu", "Jie Xu", "Hongtao Wu", "Chi-Hou Cheang", "Ya Jing", "Weinan Zhang", "Huaping Liu", "Hang Li", "Tao Kong"], "year": 2023, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2311.01378", "pdf_url": "https://arxiv.org/pdf/2311.01378", "code_url": "", "clusters": ["C1"], "keywords": ["RoboFlamingo", "OpenFlamingo"], "model_class": "VLA", "techniques": ["VLM adaptation", "policy head decoupling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts an open VLM (OpenFlamingo) to manipulation with light fine-tuning and a separate history-aware policy head, which the authors argue suits open-loop control and deployment on weaker hardware.", "relevance": 3, "relevance_reason": "Cheap VLM-to-policy adaptation with an explicit claim of low-performance platform deployment.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2405.07503, 2406.04339"}
{"id": "2202.09710", "title": "A Barrier Certificate-based Simplex Architecture for Systems with Approximate and Hybrid Dynamics", "authors": ["Amol Damare", "Shouvik Roy", "Roshan Sharma", "Keith DSouza", "Scott A. Smolka", "Scott D. Stoller"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2202.09710", "pdf_url": "https://arxiv.org/pdf/2202.09710", "code_url": "", "clusters": ["C7"], "keywords": ["runtime assurance", "switching condition"], "model_class": "other", "techniques": ["simplex", "barrier certificates"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Simplex design where barrier certificates prove the baseline controller safe and yield cheap Taylor-expansion switching conditions, extended to hybrid systems.", "relevance": 3, "relevance_reason": "Barrier-based switching logic for learned advanced controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "10.4230/oasics.ng-res.2022.1", "title": "Can We Trust AI-Powered Real-Time Embedded Systems? (Invited Paper)", "authors": ["Buttazzo, Giorgio"], "year": 2022, "venue": "NG-RES 2022 (OASIcs)", "type": "paper", "url": "https://drops.dagstuhl.de/entities/document/10.4230/OASIcs.NG-RES.2022.1", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["embedded AI", "safety-critical"], "model_class": "other", "techniques": ["real-time systems", "timing predictability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Invited paper surveying why current AI methods are not built for safety-critical embedded use and outlining research directions for security, safety and time predictability of AI-powered real-time systems.", "relevance": 3, "relevance_reason": "Frames the timing-predictability gap for neural inference on embedded platforms.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2202.03976", "title": "Communication-Control Co-design in Wireless Edge Industrial Systems", "authors": ["Mark Eisen", "Santosh Shukla", "Dave Cavalcanti", "Amit S. Baxi"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2202.03976", "pdf_url": "https://arxiv.org/pdf/2202.03976", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["communication-control co-design", "wireless edge", "industrial robots", "QoS adaptation"], "model_class": "other", "techniques": ["adaptive QoS allocation", "learned co-design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Varies per-robot wireless QoS and control actions jointly so a shared factory channel can support edge-offloaded perception for many industrial plants without instability.", "relevance": 3, "relevance_reason": "Fleet-level network/control co-design for edge-offloaded robots (Intel).", "foundational_technique": false, "verified": true, "found_via": "arXiv API: edge server robot policy"}
{"id": "2206.14176", "title": "DayDreamer: World Models for Physical Robot Learning", "authors": ["Philipp Wu", "Alejandro Escontrela", "Danijar Hafner", "Kenneth Y. Goldberg", "P. Abbeel"], "year": 2022, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2206.14176", "pdf_url": "https://arxiv.org/pdf/2206.14176", "code_url": "", "clusters": ["C8"], "keywords": ["real-robot RL", "world model"], "model_class": "world model", "techniques": ["Dreamer", "MBRL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Applies Dreamer-style latent world models to learn directly on several physical robots within hours, without simulators.", "relevance": 3, "relevance_reason": "Seminal efficient (sample- and compute-light) world-model learning on real robots.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2310.17552, 2410.22689"}
{"id": "2209.01774", "title": "ElasticROS: An Elastically Collaborative Robot Operation System for Fog and Cloud Robotics", "authors": ["Boyi Liu"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2209.01774", "pdf_url": "https://arxiv.org/pdf/2209.01774", "code_url": "", "clusters": ["C6"], "keywords": ["elastic computing", "node-level offloading", "fog/cloud robotics"], "model_class": "other", "techniques": ["dynamic node-level offloading"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "ROS extension that elastically reassigns computation between robot and fog/cloud as conditions change, targeting large-model compute pressure.", "relevance": 3, "relevance_reason": "Adaptive robot/cloud computation split.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "2202.05894", "title": "Failure Prediction with Statistical Guarantees for Vision-Based Robot Control", "authors": ["Alec Farid", "David Snyder", "Allen Z. Ren", "Anirudha Majumdar"], "year": 2022, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2202.05894", "pdf_url": "https://doi.org/10.15607/rss.2022.xviii.042", "code_url": "", "clusters": ["C7"], "keywords": ["failure prediction", "statistical guarantees"], "model_class": "other", "techniques": ["failure prediction", "PAC-Bayes guarantees"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Synthesizes failure predictors for vision-based robot control policies with guaranteed bounds on false-positive and false-negative rates.", "relevance": 3, "relevance_reason": "relevant: runtime failure prediction with guarantees", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2410.04640, 2506.09937"}
{"id": "2208.07339", "title": "LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale", "authors": ["Tim Dettmers", "M. Lewis", "Younes Belkada", "Luke Zettlemoyer"], "year": 2022, "venue": "NeurIPS 2022", "type": "paper", "url": "https://arxiv.org/abs/2208.07339", "pdf_url": "https://arxiv.org/pdf/2208.07339", "code_url": "", "clusters": ["C2"], "keywords": ["INT8 matmul", "outlier decomposition"], "model_class": "other", "techniques": ["8-bit quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Mixed-precision decomposition that keeps outlier features in 16-bit and the rest in INT8, enabling 8-bit inference for large transformers (bitsandbytes).", "relevance": 3, "relevance_reason": "8-bit path used by OpenVLA inference.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2406.09246"}
{"id": "2206.09557", "title": "nuQmm: Quantized MatMul for Efficient Inference of Large-Scale Generative Language Models", "authors": ["Gunho Park", "Baeseong Park", "S. Kwon", "Byeongwook Kim", "Youngjoo Lee", "Dongsoo Lee"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2206.09557", "pdf_url": "http://arxiv.org/pdf/2206.09557", "code_url": "", "clusters": ["C2"], "keywords": ["LLM quantization"], "model_class": "other", "techniques": ["quantized matmul"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lookup-table-based GEMM kernel that speeds up weight-only sub-4-bit quantized LLM inference by avoiding dequantization overhead.", "relevance": 3, "relevance_reason": "verified 2026-09-30 by orchestrator: arXiv abstract fetched (paper later retitled LUT-GEMM). Foundational quantized-kernel technique.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2210.17323, 2211.10438"}
{"id": "2210.03142", "title": "On Distillation of Guided Diffusion Models", "authors": ["Chenlin Meng", "Robin Rombach", "Rui-Qi Gao", "Diederik P. Kingma", "S. Ermon", "Jonathan Ho", "Tim Salimans"], "year": 2022, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2210.03142", "pdf_url": "http://arxiv.org/pdf/2210.03142", "code_url": "", "clusters": ["C2"], "keywords": ["classifier-free guidance", "progressive distillation"], "model_class": "other", "techniques": ["distillation", "few-step sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 256x faster sampling (pixel space); >=10x (latent space)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills classifier-free-guided diffusion models into few-step samplers.", "relevance": 3, "relevance_reason": "Generic few-step diffusion distillation that diffusion-policy acceleration builds on.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2405.07503, 2406.04806"}
{"id": "2202.00512", "title": "Progressive Distillation for Fast Sampling of Diffusion Models", "authors": ["Tim Salimans", "Jonathan Ho"], "year": 2022, "venue": "ICLR 2022", "type": "paper", "url": "https://arxiv.org/abs/2202.00512", "pdf_url": "https://arxiv.org/pdf/2202.00512", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["progressive distillation", "step halving"], "model_class": "other", "techniques": ["step distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Repeatedly distills a diffusion sampler into one with half the steps, reaching few-step sampling.", "relevance": 3, "relevance_reason": "Foundation of diffusion-policy step distillation.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2405.07503"}
{"id": "2203.08098", "title": "RB2: Robotic Manipulation Benchmarking with a Twist", "authors": ["S. Dasari", "Jian-Ren Wang", "Joyce Hong", "Shikhar Bahl", "Yi-Xin Lin", "Austin S. Wang", "Abitha Thankaraj", "K. Chahal", "B. Çalli", "Saurabh Gupta", "David Held", "Lerrel Pinto", "Deepak Pathak", "Vikash Kumar", "Abhi Gupta"], "year": 2022, "venue": "NeurIPS Datasets and Benchmarks", "type": "paper", "url": "https://arxiv.org/abs/2203.08098", "pdf_url": "https://arxiv.org/pdf/2203.08098", "code_url": "", "clusters": ["C4"], "keywords": ["benchmarking", "reproducibility"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Ranking-based benchmarking protocol that trades off reproducibility and accessibility across labs.", "relevance": 3, "relevance_reason": "Real-robot benchmarking methodology (C4).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2405.05941, 2506.18123"}
{"id": "2205.07149", "title": "Robotic Computing on FPGAs: Current Progress, Research Challenges, and Opportunities", "authors": ["Zi-Shen Wan", "A. Lele", "Bo Yu", "Shaoshan Liu", "Yu Wang", "V. Reddi", "Cong Hao", "A. Raychowdhury"], "year": 2022, "venue": "IEEE AICAS 2022", "type": "paper", "url": "https://arxiv.org/abs/2205.07149", "pdf_url": "https://arxiv.org/pdf/2205.07149", "code_url": "", "clusters": ["C5"], "keywords": ["FPGA", "robotic computing", "SWaP"], "model_class": "other", "techniques": ["survey", "FPGA acceleration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Position/survey paper on FPGAs as a low-power reconfigurable substrate for robotic perception, localization, planning and control.", "relevance": 3, "relevance_reason": "Foundational hardware survey for robotic computing (pre-VLA).", "foundational_technique": false, "verified": true, "found_via": "web: VLA accelerator FPGA ASIC design"}
{"id": "2202.11762", "title": "Safe Control with Learned Certificates: A Survey of Neural Lyapunov, Barrier, and Contraction methods", "authors": ["Charles Dawson", "Sicun Gao", "Chuchu Fan"], "year": 2022, "venue": "IEEE T-RO", "type": "paper", "url": "https://arxiv.org/abs/2202.11762", "pdf_url": "https://arxiv.org/pdf/2202.11762", "code_url": "https://github.com/MIT-REALM/neural_clbf", "clusters": ["C7"], "keywords": ["learned certificates", "neural CBF", "neural Lyapunov", "survey"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of learning neural Lyapunov, barrier and contraction certificates alongside controllers to provide safety/stability guarantees for learned control.", "relevance": 3, "relevance_reason": "Foundational survey on certifying learned controllers.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"Safe Control with Learned Certificates\""}
{"id": "2205.01026", "title": "Safety-Critical Manipulation for Collision-Free Food Preparation", "authors": ["Andrew W. Singletary", "W. Guffey", "Tamás G. Molnár", "Ryan W. Sinnet", "A. Ames"], "year": 2022, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2205.01026", "pdf_url": "https://arxiv.org/pdf/2205.01026", "code_url": "", "clusters": ["C7"], "keywords": ["CBF", "safety filter"], "model_class": "other", "techniques": ["control barrier functions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses control barrier functions to efficiently modify pre-validated manipulator trajectories for collision-free operation in dynamic cooking environments with safety guarantees.", "relevance": 3, "relevance_reason": "relevant: computationally efficient CBF safety filtering", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2410.15185, 2511.06385"}
{"id": "2210.09461", "title": "Token Merging: Your ViT But Faster", "authors": ["Daniel Bolya", "Cheng-Yang Fu", "Xiao-Liang Dai", "Pei-Zhao Zhang", "Christoph Feichtenhofer", "Judy Hoffman"], "year": 2022, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2210.09461", "pdf_url": "http://arxiv.org/pdf/2210.09461", "code_url": "", "clusters": ["C1"], "keywords": ["ToMe", "ViT"], "model_class": "other", "techniques": ["token merging"], "base_model": "", "params": "", "hardware": "", "reported_metric": "about 2x throughput (off-the-shelf ViT)", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free Token Merging that progressively fuses similar tokens in ViTs to increase throughput with small accuracy loss.", "relevance": 3, "relevance_reason": "Generic token-reduction technique that VLA token-pruning work builds on.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135; cited by 2502.02175"}
{"id": "2209.00588", "title": "Transformers are Sample Efficient World Models", "authors": ["Vincent Micheli", "Eloi Alonso", "François Fleuret"], "year": 2022, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2209.00588", "pdf_url": "http://arxiv.org/pdf/2209.00588", "code_url": "", "clusters": ["C8"], "keywords": ["IRIS", "Atari 100k"], "model_class": "world model", "techniques": ["transformer world model", "MBRL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Atari 100k"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "IRIS: a discrete-autoencoder plus Transformer world model in which agents learn from imagined rollouts with very little real interaction.", "relevance": 3, "relevance_reason": "Seminal sample-efficient transformer world model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2301.04104, 2310.06114"}
{"id": "2211.11152", "title": "You Need Multiple Exiting: Dynamic Early Exiting for Accelerating Unified Vision Language Model", "authors": ["Sheng Tang", "Yaqing Wang", "Zhenglun Kong", "Tianchi Zhang", "Yao Li", "Caiwen Ding", "Yanzhi Wang", "Yi Liang", "Dongkuan Xu"], "year": 2022, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2211.11152", "pdf_url": "https://arxiv.org/pdf/2211.11152", "code_url": "", "clusters": ["C1"], "keywords": ["vision-language model", "layer skipping"], "model_class": "other", "techniques": ["early exit"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 50% and 40% lower expected inference time at 99%/96% performance", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Dynamic multi-exit early exiting for unified vision-language transformers, skipping encoder and decoder layers per input.", "relevance": 3, "relevance_reason": "Foundational early-exit technique", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2409.11403, 2411.02359"}
{"id": "2206.01861", "title": "ZeroQuant: Efficient and Affordable Post-Training Quantization for Large-Scale Transformers", "authors": ["Z. Yao", "Reza Yazdani Aminabadi", "Minjia Zhang", "Xiaoxia Wu", "Cong-Long Li", "Yuxiong He"], "year": 2022, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2206.01861", "pdf_url": "https://arxiv.org/pdf/2206.01861", "code_url": "", "clusters": ["C2"], "keywords": ["INT8", "transformer quantization"], "model_class": "other", "techniques": ["post-training quantization", "layer-wise distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "End-to-end post-training quantization pipeline for large transformers with fine-grained hardware-friendly schemes and layer-wise KD.", "relevance": 3, "relevance_reason": "Foundational quantization technique cited by core items.", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2210.17323, 2211.10438"}
{"id": "1812.05506", "title": "A predictive safety filter for learning-based control of constrained nonlinear dynamical systems", "authors": ["Kim P. Wabersich", "Melanie N. Zeilinger"], "year": 2021, "venue": "Automatica 2021", "type": "paper", "url": "https://arxiv.org/abs/1812.05506", "pdf_url": "https://arxiv.org/pdf/1812.05506", "code_url": "", "clusters": ["C7"], "keywords": ["predictive safety filter", "learning-based control", "MPC"], "model_class": "other", "techniques": ["MPC-based predictive safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Introduces the predictive safety filter: an MPC-style layer that checks whether a proposed learning-based input keeps a constrained nonlinear system recoverable and minimally modifies it otherwise.", "relevance": 3, "relevance_reason": "Seminal predictive safety filter for learned controllers.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"predictive safety filter\" learning-based"}
{"id": "2103.06624", "title": "Beta-CROWN: Efficient Bound Propagation with Per-neuron Split Constraints for Complete and Incomplete Neural Network Robustness Verification", "authors": ["Shiqi Wang", "Huan Zhang", "Kaidi Xu", "Xue Lin", "Suman Jana", "Cho-Jui Hsieh", "J. Zico Kolter"], "year": 2021, "venue": "NeurIPS 2021", "type": "paper", "url": "https://arxiv.org/abs/2103.06624", "pdf_url": "https://arxiv.org/pdf/2103.06624", "code_url": "", "clusters": ["C7"], "keywords": ["beta-CROWN"], "model_class": "other", "techniques": ["bound propagation", "branch and bound"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Bound propagation that encodes branch-and-bound neuron splits via optimizable beta parameters, giving GPU-parallel complete verification far faster than LP-based BaB.", "relevance": 3, "relevance_reason": "Core of alpha-beta-CROWN verifier.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:Reluplex OR Marabou OR NNV OR beta-CROWN OR VNN-COMP"}
{"id": "10.1109/iros51168.2021.9635877", "title": "Energy-Efficient Mobile Robot Control via Run-time Monitoring of Environmental Complexity and Computing Workload", "authors": ["Sherif A. S. Mohamed", "Mohammad-Hashem Haghbayan", "Antonio Miele", "Onur Mutlu", "Juha Plosila"], "year": 2021, "venue": "IROS 2021", "type": "paper", "url": "https://doi.org/10.1109/iros51168.2021.9635877", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["DVFS", "motor voltage", "battery energy", "Jetson TX2"], "model_class": "other", "techniques": ["joint CPU DVFS and motor control"], "base_model": "", "params": "", "hardware": "Jetson TX2", "reported_metric": "battery energy savings 50.5% / 41% / 30% in low/medium/high-complexity environments vs baselines", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Jointly tunes CPU frequency and motor speed at run time, showing compute and locomotion energy must be optimized together on a battery robot.", "relevance": 3, "relevance_reason": "Quantifies compute share of robot energy budget.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q2: OpenAlex title_and_abstract.search: energy consumption computing mobile robots"}
{"id": "2108.01235", "title": "Interpretable Trade-offs Between Robot Task Accuracy and Compute Efficiency", "authors": ["Bineet Ghosh", "Sandeep Chinchali", "Parasara Sridhar Duggirala"], "year": 2021, "venue": "IROS 2021", "type": "paper", "url": "https://arxiv.org/abs/2108.01235", "pdf_url": "https://arxiv.org/pdf/2108.01235", "code_url": "", "clusters": ["C6"], "keywords": ["model selection", "cloud GPU", "compute budget", "cost-benefit"], "model_class": "other", "techniques": ["fast/slow model selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Decides when a robot should invoke an expensive accurate model (e.g. cloud GPU) over a cheap one by exploiting correlation between their accuracies, within compute and energy budgets.", "relevance": 3, "relevance_reason": "Formal cost-accuracy framing for onboard vs cloud compute invocation.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q2: arXiv abs:robot AND abs:\"cloud GPU\" AND abs:cost"}
{"id": "10.1109/rtss52674.2021.00038", "title": "LaLaRAND: Flexible Layer-by-Layer CPU/GPU Scheduling for Real-Time DNN Tasks", "authors": ["Woosung Kang", "Kilho Lee", "Jinkyu Lee", "Insik Shin", "Hoon Sung Chwa"], "year": 2021, "venue": "RTSS 2021", "type": "paper", "url": "https://doi.org/10.1109/rtss52674.2021.00038", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["LaLaRAND", "real-time DNN"], "model_class": "other", "techniques": ["layer-level scheduling", "schedulability analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Layer-level CPU/GPU real-time scheduling framework for DNN tasks that improves schedulability with worst-case timing guarantees.", "relevance": 3, "relevance_reason": "Schedulability-aware DNN execution on heterogeneous embedded hardware.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2106.07597", "title": "MLPerf Tiny Benchmark", "authors": ["Colby Banbury", "Vijay Janapa Reddi", "Peter Torelli", "Jeremy Holleman", "Nat Jeffries", "Csaba Kiraly", "Pietro Montino", "David Kanter", "Sebastian Ahmed", "Danilo Pau", "Urmish Thakker", "Antonio Torrini", "Peter Warden", "Jay Cordaro", "Giuseppe Di Guglielmo", "Javier Duarte", "Stephen Gibellini", "Videet Parekh", "Honson Tran", "Nhan Tran", "Niu Wenxu", "Xu Xuesong"], "year": 2021, "venue": "arXiv (TinyML benchmark)", "type": "benchmark", "url": "https://arxiv.org/abs/2106.07597", "pdf_url": "https://arxiv.org/pdf/2106.07597", "code_url": "", "clusters": ["C4"], "keywords": ["TinyML", "energy measurement", "reproducibility"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Industry-standard benchmark for ultra-low-power ML measuring accuracy, latency and energy with modular rules for fair cross-stack comparison.", "relevance": 3, "relevance_reason": "Energy-inclusive benchmark methodology template.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q3: arXiv id lookup MLPerf Tiny"}
{"id": "2106.13867", "title": "POLAR: A Polynomial Arithmetic Framework for Verifying Neural-Network Controlled Systems", "authors": ["Chao Huang", "Jiameng Fan", "Zhilu Wang", "Yixuan Wang", "Weichao Zhou", "Jiajun Li", "Xin Chen", "Wenchao Li", "Qi Zhu"], "year": 2021, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2106.13867", "pdf_url": "https://arxiv.org/pdf/2106.13867", "code_url": "https://github.com/ChaoHuang2018/POLAR_Tool", "clusters": ["C7"], "keywords": ["POLAR"], "model_class": "other", "techniques": ["Taylor models", "reachability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Polynomial-arithmetic reachability for NN-controlled systems combining Taylor models with Bernstein forms and symbolic remainders for tighter, faster bounds.", "relevance": 3, "relevance_reason": "Standard NNCS verification tool.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2010.02966", "title": "Reinforcement Learning with Random Delays", "authors": ["Simon Ramstedt", "Yann Bouteiller", "Giovanni Beltrame", "Christopher Pal", "Jonathan Binas"], "year": 2021, "venue": "ICLR 2021", "type": "paper", "url": "https://arxiv.org/abs/2010.02966", "pdf_url": "https://arxiv.org/pdf/2010.02966", "code_url": "", "clusters": ["C7"], "keywords": ["random delays", "DCAC", "delayed RL"], "model_class": "other", "techniques": ["delay-correcting actor-critic", "trajectory resampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["delayed MuJoCo"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Analyzes randomly delayed MDPs and derives Delay-Correcting Actor-Critic, which resamples trajectory fragments in hindsight for off-policy value estimation under action/observation delays.", "relevance": 3, "relevance_reason": "Foundational learning under random delays.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"random delays\" reinforcement learning"}
{"id": "2104.14270", "title": "Resource Allocation and Service Provisioning in Multi-Agent Cloud Robotics: A Comprehensive Survey", "authors": ["Mahbuba Afrin", "Jiong Jin", "Akhlaqur Rahman", "Ashfaqur Rahman", "Jiafu Wan", "Ekram Hossain"], "year": 2021, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2104.14270", "pdf_url": "https://arxiv.org/pdf/2104.14270", "code_url": "", "clusters": ["C6"], "keywords": ["multi-agent cloud robotics", "resource allocation", "survey"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey of resource allocation and service provisioning for multi-robot cloud robotics under latency, data and compute constraints.", "relevance": 3, "relevance_reason": "Fleet/cloud resource allocation background.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2110.00843", "title": "SHARP: Shielding-Aware Robust Planning for Safe and Efficient Human-Robot Interaction", "authors": ["Haimin Hu", "Kensuke Nakamura", "Jaime F. Fisac"], "year": 2021, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2110.00843", "pdf_url": "https://arxiv.org/pdf/2110.00843", "code_url": "", "clusters": ["C7"], "keywords": ["human-robot interaction"], "model_class": "other", "techniques": ["shielding", "robust planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Planner that anticipates future shielding (safety override) events using Bayesian human prediction so the nominal plan triggers fallback less often.", "relevance": 3, "relevance_reason": "Shows how to plan around a runtime shield instead of fighting it.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:fallback AND (\"learned policy\" OR \"robot policy\") AND safe*"}
{"id": "2103.04505", "title": "Split Computing and Early Exiting for Deep Learning Applications: Survey and Research Challenges", "authors": ["Yoshitomo Matsubara", "Marco Levorato", "Francesco Restuccia"], "year": 2021, "venue": "ACM Computing Surveys 2022", "type": "paper", "url": "https://arxiv.org/abs/2103.04505", "pdf_url": "https://arxiv.org/pdf/2103.04505", "code_url": "", "clusters": ["C6"], "keywords": ["split computing", "early exiting", "survey"], "model_class": "other", "techniques": ["split computing", "early exit", "bottleneck injection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Canonical survey of splitting DNNs between mobile device and edge server and of early-exit models, covering bottleneck design and open challenges.", "relevance": 3, "relevance_reason": "Foundational split-computing reference for robot edge-cloud inference.", "foundational_technique": true, "verified": true, "found_via": "search query"}
{"id": "2105.07091", "title": "Verification of Image-based Neural Network Controllers Using Generative Models", "authors": ["Sydney M. Katz", "Anthony L. Corso", "Christopher A. Strong", "Mykel J. Kochenderfer"], "year": 2021, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2105.07091", "pdf_url": "https://arxiv.org/pdf/2105.07091", "code_url": "", "clusters": ["C7"], "keywords": ["image-based control verification"], "model_class": "other", "techniques": ["GAN", "closed-loop reachability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains a GAN mapping states to plausible images and verifies the concatenated generator-controller network with closed-loop reachability tools.", "relevance": 3, "relevance_reason": "Early approach to verifying image-based controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2011.02082", "title": "DeepReach: A Deep Learning Approach to High-Dimensional Reachability", "authors": ["Somil Bansal", "Claire Tomlin"], "year": 2020, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2011.02082", "pdf_url": "https://arxiv.org/pdf/2011.02082", "code_url": "", "clusters": ["C7"], "keywords": ["DeepReach"], "model_class": "other", "techniques": ["HJ reachability", "neural PDE"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Neural PDE solver (sinusoidal networks) for high-dimensional HJ reachability whose cost scales with tube complexity rather than state dimension.", "relevance": 3, "relevance_reason": "Makes reachability-based certification feasible in higher dimensions.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: seminal fetch by arXiv id (HJ reachability, KnowNo, DeepReach)"}
{"id": "2008.11830", "title": "Designing Neural Networks for Real-Time Systems", "authors": ["Hammond Pearce", "Xin Yang", "Partha S. Roop", "Marc Katzef", "Tórur Biskopstø Strøm"], "year": 2020, "venue": "IEEE Embedded Systems Letters 2020", "type": "paper", "url": "https://arxiv.org/abs/2008.11830", "pdf_url": "https://arxiv.org/pdf/2008.11830", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["WCET", "static timing analysis", "neural network to C", "time-predictable"], "model_class": "other", "techniques": ["static WCET analysis", "code generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Compiles Keras networks to analyzable C for a time-predictable processor so worst-case execution times of a neural controller can be statically bounded.", "relevance": 3, "relevance_reason": "Timing certification of neural controllers.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"worst-case execution time\" neural network"}
{"id": "2008.08243", "title": "Enabling Remote Whole-Body Control with 5G Edge Computing", "authors": ["Huaijiang Zhu", "Manali Sharma", "Kai Pfeiffer", "Marco Mezzavilla", "Jia Shen", "Sundeep Rangan", "Ludovic Righetti"], "year": 2020, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2008.08243", "pdf_url": "https://arxiv.org/pdf/2008.08243", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["5G edge", "remote whole-body control", "humanoid/legged"], "model_class": "other", "techniques": ["offloading low-level control to 5G edge"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Studies offloading even low-level whole-body control to a 5G edge server, examining how latency and bandwidth affect feasibility of remote real-time control.", "relevance": 3, "relevance_reason": "Pushes offloading down to the control loop; latency limits.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: cloud humanoid latency"}
{"id": "2012.02328", "title": "MLPerf Mobile Inference Benchmark", "authors": ["Vijay Janapa Reddi", "David Kanter", "Peter Mattson", "Jared Duke", "Thai Nguyen", "Ramesh Chukka"], "year": 2020, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2012.02328", "pdf_url": "https://arxiv.org/pdf/2012.02328", "code_url": "", "clusters": ["C4"], "keywords": ["mobile SoC", "run rules", "cross-vendor"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Open benchmark app with fixed models, datasets, quality targets and run rules for comparing mobile AI chips and stacks; early rounds showed large software-driven gains.", "relevance": 3, "relevance_reason": "Run-rule methodology for heterogeneous accelerators.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q3: arXiv id lookup MLPerf Mobile"}
{"id": "1908.00528", "title": "Neural Simplex Architecture", "authors": ["Dung T. Phan", "Radu Grosu", "Nils Jansen", "Nicola Paoletti", "Scott A. Smolka", "Scott D. Stoller"], "year": 2020, "venue": "NASA Formal Methods Symposium (NFM) 2020", "type": "paper", "url": "https://arxiv.org/abs/1908.00528", "pdf_url": "https://arxiv.org/pdf/1908.00528", "code_url": "", "clusters": ["C7"], "keywords": ["neural simplex architecture", "runtime assurance", "baseline controller switching"], "model_class": "other", "techniques": ["simplex switching", "online retraining of advanced controller"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Runtime assurance for neural controllers that switches to a verified baseline controller when unsafe and supports switching back and correcting the neural controller.", "relevance": 3, "relevance_reason": "Foundational fallback architecture for learned controllers.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"simplex architecture\" learned controller"}
{"id": "2004.05519", "title": "NNV: The Neural Network Verification Tool for Deep Neural Networks and Learning-Enabled Cyber-Physical Systems", "authors": ["Hoang-Dung Tran", "Xiaodong Yang", "Diego Manzanas Lopez", "Patrick Musau", "Luan Viet Nguyen", "Weiming Xiang", "Stanley Bak", "Taylor T. Johnson"], "year": 2020, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2004.05519", "pdf_url": "https://arxiv.org/pdf/2004.05519", "code_url": "", "clusters": ["C7"], "keywords": ["NNV"], "model_class": "other", "techniques": ["reachability", "star sets"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Set-based verification tool (polyhedra, star sets, zonotopes) for DNNs and learning-enabled CPS including closed-loop NN control.", "relevance": 3, "relevance_reason": "Widely used NN/NNCS verification tool.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:Reluplex OR Marabou OR NNV OR beta-CROWN OR VNN-COMP"}
{"id": "2004.07876", "title": "Reach-SDP: Reachability Analysis of Closed-Loop Systems with Neural Network Controllers via Semidefinite Programming", "authors": ["Haimin Hu", "Mahyar Fazlyab", "Manfred Morari", "George J. Pappas"], "year": 2020, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2004.07876", "pdf_url": "https://arxiv.org/pdf/2004.07876", "code_url": "", "clusters": ["C7"], "keywords": ["Reach-SDP"], "model_class": "other", "techniques": ["semidefinite programming", "quadratic constraints"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Forward reachability for linear systems with NN feedback by abstracting activations with quadratic constraints solved as SDPs.", "relevance": 3, "relevance_reason": "Seminal SDP-based closed-loop NN reachability.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "2011.03216", "title": "Task-relevant Representation Learning for Networked Robotic Perception", "authors": ["Manabu Nakanoya", "Sandeep Chinchali", "Alexandros Anemogiannis", "Akul Datta", "Sachin Katti", "Marco Pavone"], "year": 2020, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2011.03216", "pdf_url": "https://arxiv.org/pdf/2011.03216", "code_url": "", "clusters": ["C6"], "keywords": ["task-relevant compression", "networked perception", "bandwidth"], "model_class": "other", "techniques": ["task-aware learned compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 11x more compression than competing methods", "metric_measures": "other", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Learns sensory representations co-designed with a downstream robot perception model so robots transmit only task-relevant bits to remote servers.", "relevance": 3, "relevance_reason": "Bandwidth-aware robot-to-cloud transmission.", "foundational_technique": false, "verified": true, "found_via": "snowball from 1902.05703"}
{"id": "1911.02549", "title": "MLPerf Inference Benchmark", "authors": ["Reddi, Vijay Janapa", "Cheng, Christine", "Kanter, David", "Mattson, Peter", "Schmuelling, Guenther", "Wu, Carole-Jean", "Anderson, Brian", "Breughe, Maximilien", "Charlebois, Mark", "Chou, William", "Chukka, Ramesh", "Coleman, Cody", "Davis, Sam", "Deng, Pan", "Diamos, Greg", "Duke, Jared", "Fick, Dave", "Gardner, J. Scott", "Hubara, Itay", "Idgunji, Sachin", "Jablin, Thomas B.", "Jiao, Jeff", "John, Tom St.", "Kanwar, Pankaj", "Lee, David", "Liao, Jeffery", "Lokhmotov, Anton", "Massa, Francisco", "Meng, Peng", "Micikevicius, Paulius", "Osborne, Colin", "Pekhimenko, Gennady", "Rajan, Arun Tejusve Raghunath", "Sequeira, Dilip", "Sirasao, Ashish", "Sun, Fei", "Tang, Hanlin", "Thomson, Michael", "Wei, Frank", "Wu, Ephrem", "Xu, Lingjie", "Yamada, Koichi", "Yu, Bing", "Yuan, George", "Zhong, Aaron", "Zhang, Peizhao", "Zhou, Yuchen"], "year": 2019, "venue": "ISCA 2020", "type": "paper", "url": "https://arxiv.org/abs/1911.02549", "pdf_url": "https://arxiv.org/pdf/1911.02549", "code_url": "", "clusters": ["C4"], "keywords": ["MLPerf Inference", "single-stream", "multistream", "tail latency", "benchmark rules"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Industry benchmark methodology for ML inference systems with scenario definitions (single-stream, multistream, server, offline), quality targets and tail-latency constraints to ensure comparability.", "relevance": 3, "relevance_reason": "Methodological foundation for latency percentile and scenario-based reporting that VLA benchmarks can borrow.", "foundational_technique": true, "verified": true, "found_via": "WebSearch: MLPerf robotics inference benchmark"}
{"id": "10.1109/rtss46320.2019.00042", "title": "Pipelined Data-Parallel CPU/GPU Scheduling for Multi-DNN Real-Time Inference", "authors": ["Yecheng Xiang", "Hyoseung Kim"], "year": 2019, "venue": "RTSS 2019", "type": "paper", "url": "https://doi.org/10.1109/rtss46320.2019.00042", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["multi-DNN inference", "response time"], "model_class": "other", "techniques": ["real-time scheduling", "pipelining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DART: pipelined data-parallel CPU/GPU scheduling of multiple DNNs that gives deterministic response time to real-time requests while serving best-effort ones.", "relevance": 3, "relevance_reason": "Real-time scheduling of concurrent neural inference.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "1910.01055", "title": "QuaRL: Quantization for Fast and Environmentally Sustainable Reinforcement Learning", "authors": ["Srivatsan Krishnan", "Maximilian Lam", "Sharad Chitlangia", "Zi-Shen Wan", "Gabriel Barth-Maron", "Aleksandra Faust", "V. Reddi"], "year": 2019, "venue": "TMLR", "type": "paper", "url": "https://arxiv.org/abs/1910.01055", "pdf_url": "https://arxiv.org/pdf/1910.01055", "code_url": "", "clusters": ["C2"], "keywords": ["quantized RL", "8-bit actors", "ActorQ", "carbon"], "model_class": "other", "techniques": ["8-bit quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.5-5.41x end-to-end training speedup", "metric_measures": "relative latency", "benchmarks": ["DeepMind Control Suite"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows RL policies tolerate quantization and uses 8-bit actors to speed distributed RL data collection without hurting convergence.", "relevance": 3, "relevance_reason": "Early policy quantization evidence cited by QAIL.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2412.01034 (references)"}
{"id": "1906.10654", "title": "ReachNN: Reachability Analysis of Neural-Network Controlled Systems", "authors": ["Chao Huang", "Jiameng Fan", "Wenchao Li", "Xin Chen", "Qi Zhu"], "year": 2019, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1906.10654", "pdf_url": "https://arxiv.org/pdf/1906.10654", "code_url": "", "clusters": ["C7"], "keywords": ["ReachNN"], "model_class": "other", "techniques": ["reachability", "Bernstein polynomials"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reachability analysis of NN-controlled systems via Bernstein polynomial abstraction with bounded approximation error, supporting general Lipschitz activations.", "relevance": 3, "relevance_reason": "Seminal NNCS reachability tool.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:\"neural network control*\" AND (verification OR reachability OR verified OR verifying)"}
{"id": "1909.12271", "title": "RLBench: The Robot Learning Benchmark & Learning Environment", "authors": ["James, Stephen", "Ma, Zicong", "Arrojo, David Rovick", "Davison, Andrew J."], "year": 2019, "venue": "IEEE RA-L 2020", "type": "benchmark", "url": "https://arxiv.org/abs/1909.12271", "pdf_url": "https://arxiv.org/pdf/1909.12271", "code_url": "https://github.com/stepjam/RLBench", "clusters": ["C4"], "keywords": ["RLBench", "multi-task", "CoppeliaSim", "motion-planned demos"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RLBench"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "100 hand-designed manipulation tasks with unlimited motion-planned demonstrations and multi-camera observations; an early standard for multi-task imitation evaluation.", "relevance": 3, "relevance_reason": "Seed benchmark (pre-2022 foundation).", "foundational_technique": false, "verified": true, "found_via": "seed work"}
{"id": "1903.00999", "title": "RT-Gang: Real-Time Gang Scheduling Framework for Safety-Critical Systems", "authors": ["Waqar Ali", "Heechul Yun"], "year": 2019, "venue": "RTAS 2019", "type": "paper", "url": "https://arxiv.org/abs/1903.00999", "pdf_url": "https://arxiv.org/pdf/1903.00999", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["RT-Gang", "real-time"], "model_class": "other", "techniques": ["gang scheduling", "memory bandwidth throttling", "WCET"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Linux gang scheduler that runs one parallel real-time task at a time and throttles best-effort memory bandwidth to make WCET tight on multicore platforms.", "relevance": 3, "relevance_reason": "Timing-predictability foundation for co-running perception/DNN workloads.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"worst-case execution time\" AND (\"neural network\" OR DNN OR GPU)"}
{"id": "1808.07921", "title": "SOTER: A Runtime Assurance Framework for Programming Safe Robotics Systems", "authors": ["Ankush Desai", "Shromona Ghosh", "Sanjit A. Seshia", "Natarajan Shankar", "Ashish Tiwari"], "year": 2019, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1808.07921", "pdf_url": "https://arxiv.org/pdf/1808.07921", "code_url": "", "clusters": ["C7"], "keywords": ["SOTER", "runtime assurance", "simplex", "robot programming"], "model_class": "other", "techniques": ["RTA module: advanced + certified safe controller"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Robot programming framework with built-in runtime assurance modules that switch between an uncertified high-performance controller and a certified safe controller.", "relevance": 3, "relevance_reason": "Foundational runtime assurance for uncertified (learned) components.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"runtime assurance\" learning-enabled robot"}
{"id": "1809.06716", "title": "A Fog Robotic System for Dynamic Visual Servoing", "authors": ["Nan Tian", "Jinfa Chen", "Mas Ma", "Robert Zhang", "Bill Huang", "Ken Goldberg", "Somayeh Sojoudi"], "year": 2018, "venue": "ICRA 2019 (submitted)", "type": "preprint", "url": "https://arxiv.org/abs/1809.06716", "pdf_url": "https://arxiv.org/pdf/1809.06716", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["fog robotics", "visual servoing", "heartbeat", "asynchronous protocol"], "model_class": "other", "techniques": ["cloud + smart edge device", "asynchronous protocol with heartbeat"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Combines cloud compute with a smart edge device via an asynchronous heartbeat protocol so a service robot can do dynamic visual servoing despite network variability.", "relevance": 3, "relevance_reason": "Early fog-robotics latency-tolerant control.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "10.1145/3173162.3173191", "title": "The Architectural Implications of Autonomous Driving", "authors": ["Shih-Chieh Lin", "Yunqi Zhang", "Chang-Hong Hsu", "Matt Skach", "Md Enamul Haque", "Lingjia Tang", "Jason Mars"], "year": 2018, "venue": "ACM ASPLOS 2018 (DOI 10.1145/3173162)", "type": "paper", "url": "https://doi.org/10.1145/3173162.3173191", "pdf_url": "", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["tail latency", "power", "thermal", "GPU/FPGA/ASIC"], "model_class": "other", "techniques": ["hardware acceleration"], "base_model": "", "params": "", "hardware": "multicore CPU, GPU, FPGA, ASIC", "reported_metric": "tail latency reduced 169x (GPU), 10x (FPGA), 93x (ASIC)", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Formalizes autonomous-driving compute constraints (latency, predictability, storage, thermal, power) and shows accelerators are needed to meet tail-latency targets within power limits.", "relevance": 3, "relevance_reason": "Seminal onboard-compute constraint analysis with tail-latency and power framing.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q2: OpenAlex title_and_abstract.search: architectural implications of autonomous driving"}
{"id": "1803.10122", "title": "World Models", "authors": ["David R Ha", "J. Schmidhuber"], "year": 2018, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/1803.10122", "pdf_url": "http://arxiv.org/pdf/1803.10122", "code_url": "", "clusters": ["C8"], "keywords": ["VAE", "MDN-RNN", "dreaming"], "model_class": "world model", "techniques": ["latent world model", "compact representation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Seminal work training a compact VAE+RNN generative world model and a tiny controller that can even be trained inside the model's imagined rollouts.", "relevance": 3, "relevance_reason": "Foundational compact latent world model underlying later efficient world-model work.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2301.04104, 2411.04983"}
{"id": "1709.07523", "title": "Hamilton-Jacobi Reachability: A Brief Overview and Recent Advances", "authors": ["Somil Bansal", "Mo Chen", "Sylvia Herbert", "Claire J. Tomlin", "S. Herbert", "C. Tomlin"], "year": 2017, "venue": "IEEE Conference on Decision and Control", "type": "preprint", "url": "https://arxiv.org/abs/1709.07523", "pdf_url": "https://arxiv.org/pdf/1709.07523", "code_url": "", "clusters": ["C7"], "keywords": ["tutorial", "formal verification", "safety"], "model_class": "other", "techniques": ["HJ reachability"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Tutorial on Hamilton-Jacobi reachability for safety verification, including GPU-parallel level-set tools and high-dimensional approximations.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: seminal fetch by arXiv id (HJ reachability, KnowNo, DeepReach); snowball round 2: cited by 2302.09182, 2309.05837"}
{"id": "1702.01135", "title": "Reluplex: An Efficient SMT Solver for Verifying Deep Neural Networks", "authors": ["Guy Katz", "Clark Barrett", "David Dill", "Kyle Julian", "Mykel Kochenderfer"], "year": 2017, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1702.01135", "pdf_url": "https://arxiv.org/pdf/1702.01135", "code_url": "", "clusters": ["C7"], "keywords": ["Reluplex", "ACAS Xu"], "model_class": "other", "techniques": ["SMT", "NN verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SMT-based verifier extending simplex to ReLU networks, proving properties of ACAS Xu networks an order of magnitude larger than before.", "relevance": 3, "relevance_reason": "Seminal DNN verification method.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:Reluplex OR Marabou OR NNV OR beta-CROWN OR VNN-COMP"}
{"id": "1609.06408", "title": "Control Barrier Function Based Quadratic Programs for Safety Critical Systems", "authors": ["A. Ames", "Xiangru Xu", "J. Grizzle", "P. Tabuada"], "year": 2016, "venue": "IEEE Transactions on Automatic Control", "type": "paper", "url": "https://arxiv.org/abs/1609.06408", "pdf_url": "https://arxiv.org/pdf/1609.06408", "code_url": "", "clusters": ["C7"], "keywords": ["safety filter", "CLF-CBF QP"], "model_class": "other", "techniques": ["control barrier functions", "quadratic programming"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Seminal formulation unifying control barrier functions and control Lyapunov functions in a real-time QP controller, the basis of most CBF safety filters for learned policies.", "relevance": 3, "relevance_reason": "directly addresses efficiency/real-time/deployment/benchmarking for embodied models", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 1903.11199, 2309.05837"}
{"id": "1510.00149", "title": "Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding", "authors": ["Song Han", "Huizi Mao", "William J. Dally"], "year": 2015, "venue": "ICLR 2016", "type": "paper", "url": "https://arxiv.org/abs/1510.00149", "pdf_url": "https://arxiv.org/pdf/1510.00149", "code_url": "", "clusters": ["C2"], "keywords": ["pruning", "weight sharing", "Huffman coding", "pipeline"], "model_class": "other", "techniques": ["pruning", "trained quantization", "entropy coding"], "base_model": "AlexNet, VGG-16", "params": "", "hardware": "CPU, GPU, mobile GPU", "reported_metric": "35x-49x storage reduction without accuracy loss; 3-4x layerwise speedup, 3-7x energy efficiency", "metric_measures": "relative memory", "benchmarks": ["ImageNet"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Seminal three-stage pipeline showing pruning, quantization and coding can be stacked with retraining between stages to multiply compression.", "relevance": 3, "relevance_reason": "Canonical composition-of-compression baseline cited by VLA compression pipelines.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q1: seed known foundational composition work (arXiv id lookup)"}
{"id": "1503.02531", "title": "Distilling the Knowledge in a Neural Network", "authors": ["Geoffrey Hinton", "Oriol Vinyals", "Jeff Dean"], "year": 2015, "venue": "NIPS 2014 Deep Learning Workshop", "type": "paper", "url": "https://arxiv.org/abs/1503.02531", "pdf_url": "https://arxiv.org/pdf/1503.02531", "code_url": "", "clusters": ["C2"], "keywords": ["knowledge distillation", "soft targets"], "model_class": "other", "techniques": ["knowledge distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Introduces distillation of an ensemble/large model into a small one via temperature-softened output targets.", "relevance": 3, "relevance_reason": "Origin of teacher-student compression.", "foundational_technique": true, "verified": true, "found_via": "snowball"}
{"id": "s2-simulation-evaluating-robot-policies-asynchronous-real-time", "title": "Simulation for Evaluating Robot Policies Should Be Asynchronous and Real-Time", "authors": ["Python Song", "Hyeonho Oh", "Ishika Singh", "Jesse Thomason", "A. Mandlekar"], "year": null, "venue": "unknown (Semantic Scholar record)", "type": "paper", "url": "https://www.semanticscholar.org/paper/3dc6d3bd6abc687a9087096b671bd53e9c31d70b", "pdf_url": "", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["asynchronous simulation", "real-time evaluation", "benchmark"], "model_class": "other", "techniques": ["evaluation methodology"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Position that policy evaluation in simulation should model asynchronous, real-time execution so inference latency affects outcomes (title-only record; abstract not retrieved).", "relevance": 3, "relevance_reason": "Evaluation under latency.", "foundational_technique": false, "verified": false, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "s2-time-optimal-execution-action-chunk-policies", "title": "Time Optimal Execution of Action Chunk Policies Beyond Demonstration Speed", "authors": ["Sunwoo Kim", "Jeongjun Kim", "Joseph J. Lim"], "year": null, "venue": "ICLR 2026 (poster)", "type": "paper", "url": "https://openreview.net/forum?id=INsLvSCJ4z", "pdf_url": "", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["time-optimal execution", "action chunk", "execution speed"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Accelerates any action-chunk imitation policy so the robot executes faster than the demonstrations, addressing speed limits from slow teleoperated data and inference latency.", "relevance": 3, "relevance_reason": "verified 2026-09-30 by orchestrator: OpenReview abstract fetched (ICLR 2026 poster). Execution-speed/latency of chunked policies.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2608.06375", "title": "$\\omega$-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation", "authors": ["Zhe Li", "Zhen-Zhen Zhang", "Yang-Yang Wei", "Wenjie Zhang", "Xi-Chen Yuan", "Peiyuan Zhi", "Gen Li", "Xin-Ying Guo", "Feng Gao", "Jian-Fei Yang", "Shang-Hang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.06375", "pdf_url": "https://arxiv.org/pdf/2608.06375", "code_url": "", "clusters": ["C8"], "keywords": ["humanoid", "loco-manipulation", "latent WAM"], "model_class": "world-action model", "techniques": ["latent prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Latent predictive whole-body world-action model for humanoids that predicts compact future embeddings instead of reconstructing videos.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links; lightweight latent WAM", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2604.24182", "title": "$M^2$-VLA: Boosting Vision-Language Models for Generalizable Manipulation via Layer Mixture and Meta-Skills", "authors": ["Siyao Xiao", "Yuhong Zhang", "Zhifang Liu", "Zihan Gao", "Jingye Zhang", "Sinwai Choo", "Dake Zhong", "Mengzhe Wang", "Xiao Lin", "Xianfeng Zhou", "Jia Jia", "Haoqian Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.24182", "pdf_url": "https://arxiv.org/pdf/2604.24182", "code_url": "", "clusters": ["C3"], "keywords": ["M2-VLA", "mixture of layers", "frozen VLM"], "model_class": "VLA", "techniques": ["frozen backbone", "layer mixture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses a frozen generalist VLM with layer mixture and meta-skills instead of end-to-end fine-tuning.", "relevance": 2, "relevance_reason": "Avoids full fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"catastrophic forgetting\""}
{"id": "2607.25516", "title": "A Causality-aware Infer-diagnose-refine Framework for Test-time Modality Adaptation in VLA Models", "authors": ["Haoyu Zhang", "Yuwei Wu", "Jin Chen", "Gao Zhi", "Zhenxin Diao", "Mingyang Gao", "Kun Wu", "Yongchun Liu", "Fan Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.25516", "pdf_url": "https://arxiv.org/pdf/2607.25516", "code_url": "", "clusters": ["C3"], "keywords": ["test-time adaptation", "counterfactual", "modality importance"], "model_class": "VLA", "techniques": ["training-free test-time refinement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Training-free infer-diagnose-refine loop that estimates visual modality importance counterfactually and refines VLA actions at test time.", "relevance": 2, "relevance_reason": "Training-free adaptation; adds inference cost.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.08307", "title": "A Measurement Study of LLM Inference Trade-offs Across Edge Continuum Hardware", "authors": ["Maysam Khatib", "Moysis Symeonides", "Demetris Trihinas", "George Pallis", "Marios D. Dikaiakos"], "year": 2026, "venue": "WIMS 2026", "type": "paper", "url": "https://arxiv.org/abs/2609.08307", "pdf_url": "https://arxiv.org/pdf/2609.08307", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["Jetson AGX Orin", "near-edge server", "energy", "Pareto"], "model_class": "other", "techniques": ["quantization"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin; near-edge CPU/GPU server; GPT-4o cloud reference", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Controlled measurements of open LLMs and quantized variants on Jetson AGX Orin vs a GPU server vs cloud show GPU server is fastest, Orin lowest energy, and parameter count or file size poorly predicts latency.", "relevance": 2, "relevance_reason": "General (non-robotics) edge-vs-server comparison methodology; applicable to placement decisions.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q3: arXiv abs:\"large language model\" AND abs:Jetson AND abs:benchmark"}
{"id": "2609.21817", "title": "A Sim-to-Real Integration Pipeline for Training and Deployment of Chunk-Based VLA Manipulation Policies", "authors": ["Mathilde Kappel", "Clémence Grislain", "Mohamed Chetouani", "Olivier Sigaud", "Louis Annabi", "Faïz Ben Amar", "Stéphane Doncieux", "Mahdi Khoramshahi"], "year": 2026, "venue": "IROS 2026 Workshop", "type": "preprint", "url": "https://arxiv.org/abs/2609.21817", "pdf_url": "https://arxiv.org/pdf/2609.21817", "code_url": "", "clusters": ["C4"], "keywords": ["sim-to-real protocol", "Franka FR3", "open-loop replay", "closed-loop evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "Franka FR3", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Open protocol that replays sim expert trajectories on a real FR3 to collect real observations and reuses the same stack for closed-loop policy evaluation.", "relevance": 2, "relevance_reason": "Reproducible evaluation stack.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation world model"}
{"id": "2602.01067", "title": "A Systematic Study of Data Modalities and Strategies for Co-training Large Behavior Models for Robot Manipulation", "authors": ["Fanqi Lin", "K. Arora", "Jean-Pierre Mercat", "Haruki Nishimura", "Paarth Shah", "Cheng Xu", "Mengchao Zhang", "Mark Zolotas", "Maya Angeles", "Owen Pfannenstiehl", "Andrew Beaulieu", "Jose A. Barreiros"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2602.01067", "pdf_url": "https://arxiv.org/pdf/2602.01067", "code_url": "", "clusters": ["C3"], "keywords": ["co-training", "large behavior models"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Systematic study of co-training data modalities and strategies for large behavior models.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2501.09747"}
{"id": "2607.00678", "title": "ABot-M0.5: Unified Mobility-and-Manipulation World Action Model", "authors": ["Ronghan Chen", "Yan-Dan Yang", "Zuojin Tang", "Dongjie Huo", "Tong Lin", "Haoning Wu", "Haoyun Liu", "Yuzhi Chen", "Lulu Zheng", "B. Yuan", "Tian-Lun Li", "Mingxin Wang", "Dekang Qi", "Bin Hu", "Wei Mei", "Yuze Xuan", "Haolong Yang", "Yanqing Zhu", "Mu Xu", "Zhiheng Ma", "Xinyuan Chang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2607.00678", "pdf_url": "https://arxiv.org/pdf/2607.00678", "code_url": "", "clusters": ["C8"], "keywords": ["mobile manipulation", "world-action model"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A world-action model unifying mobility and manipulation, jointly predicting future observations and whole-body actions for mobile manipulators.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 24 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2301.04104"}
{"id": "2602.11236", "title": "ABot-M0: VLA Foundation Model for Robotic Manipulation with Action Manifold Learning", "authors": ["Yan-Dan Yang", "Shuang Zeng", "Tong Lin", "Xinyuan Chang", "Dekang Qi", "Junjin Xiao", "Haoyu Liu", "Ronghan Chen", "Yuzhi Chen", "Dongjie Huo", "Feng Xiong", "Xing Wei", "Zhiheng Ma", "Mu Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.11236", "pdf_url": "https://arxiv.org/pdf/2602.11236", "code_url": "", "clusters": ["C1"], "keywords": ["foundation model", "cross-embodiment"], "model_class": "VLA", "techniques": ["action manifold learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "ABot-M0 is a manipulation VLA foundation model built with a large curated cross-embodiment data pipeline and an action-manifold learning objective aimed at 'one brain, many forms'.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310; cited by 2604.24622, 2605.00078"}
{"id": "2609.18243", "title": "Acting in Meters: Learning Metric Interactions for Precise Robotic Manipulation", "authors": ["Li-Jie Wang", "Zheng Lu", "Yi-Ming Wang", "He-Yang Yu", "K. Hoi", "Bo-Wen Hu", "Di Cui", "Tian-Yu Xin", "Hao-Ran Liao", "Wanqi Zhong", "Xing-Jie Fan", "Yi-Zhao Xu", "Zi-Li Wang", "Fei Gao", "Yi-Ming Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18243", "pdf_url": "https://arxiv.org/pdf/2609.18243", "code_url": "", "clusters": ["C1"], "keywords": ["metric interactions", "precise manipulation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Models metric object-level relations explicitly to improve precision of VLAs and world-action models.", "relevance": 2, "relevance_reason": "Context: VLA/WAM (links=13).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2406.09246"}
{"id": "2609.18487", "title": "ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models", "authors": ["Shijie Lian", "Bin Yu", "Zhaolong Shen", "Xiaopeng Lin", "Yichao Du", "Zhirui Zhang", "Laurence T. Yang", "Kai Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18487", "pdf_url": "https://arxiv.org/pdf/2609.18487", "code_url": "", "clusters": ["C1"], "keywords": ["action tokenizer", "rank consistency"], "model_class": "VLA", "techniques": ["action tokenization"], "base_model": "Qwen3-VL-4B", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "LIBERO-Plus", "SIMPLER", "VLA-Arena"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Tokenizer trained to preserve physical distance rankings among actions, evaluated with a new rank-consistency metric.", "relevance": 2, "relevance_reason": "Tokenization fidelity, not efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query: small VLA"}
{"id": "2609.34982", "title": "ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning", "authors": ["Di Zhu", "Ziheng Yan", "Fang Wan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34982", "pdf_url": "https://arxiv.org/pdf/2609.34982", "code_url": "", "clusters": ["C3"], "keywords": ["temporal U-Net", "multi-scale fine-tuning"], "model_class": "VLA", "techniques": ["lightweight head fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Adds a lightweight temporal U-Net over action features to fine-tune VLAs for robustness at minimal compute.", "relevance": 2, "relevance_reason": "Efficient fine-tuning module.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.24124", "title": "ActiveArena: Benchmarking and Understanding Active Perception in Robotic Manipulation", "authors": ["Yi-Bo Li", "En-Shen Zhou", "Rui Chen", "Yan-Jun Ding", "Meng-Zhen Liu", "Yi Han", "Jia-Bo Zhan", "Li-Peng Wang", "Shang-Hang Zhang", "Lu Sheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24124", "pdf_url": "https://arxiv.org/pdf/2609.24124", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "active perception"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Simulator and benchmark for evaluating active perception and memory in manipulation.", "relevance": 2, "relevance_reason": "Context: benchmark (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2605.22896", "title": "Agentic-VLA: Efficient Online Adaptation for Vision-Language-Action Models", "authors": ["Ruofan Jin", "Zaixi Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.22896", "pdf_url": "https://arxiv.org/pdf/2605.22896", "code_url": "", "clusters": ["C3"], "keywords": ["agentic training", "reward synthesis", "curriculum"], "model_class": "VLA", "techniques": ["online RL", "reward synthesis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Agentic framework with adaptive reward synthesis and critic-guided exploration for efficient online VLA adaptation.", "relevance": 2, "relevance_reason": "Online adaptation efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2608.00927", "title": "AgriJetsonBench: External-Power-Referenced TensorRT Benchmarking of Agricultural Vision Models on Jetson Edge Platforms", "authors": ["Hasan Jahanifar", "Hasan Mirzakhaninafchi", "Wesley Porter", "Abolfazl Najar", "Glen C. Rains"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.00927", "pdf_url": "https://arxiv.org/pdf/2608.00927", "code_url": "", "clusters": ["C5", "C9"], "keywords": ["Jetson power measurement", "external power logging", "TensorRT", "Orin Nano Super vs AGX Orin"], "model_class": "other", "techniques": ["energy benchmarking methodology"], "base_model": "", "params": "", "hardware": "Jetson AGX Orin 64GB, Jetson Orin Nano Super", "reported_metric": "At 15 W INT8: YOLO11s 145.82 FPS / 0.0767 J (Nano Super) vs 80.90 FPS / 0.1814 J (AGX Orin)", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Jetson benchmark methodology using external board-input power logging, showing internal telemetry is biased and that under matched power caps a smaller module can be more efficient.", "relevance": 2, "relevance_reason": "Agricultural vision, not foundation-model policy, but gives power-measurement methodology for Jetson.", "foundational_technique": false, "verified": true, "found_via": "web: edge VLA Jetson AGX Orin latency"}
{"id": "2602.12691", "title": "ALOE: Action-Level Off-Policy Evaluation for Vision-Language-Action Model Post-Training", "authors": ["Rushuai Yang", "Hecheng Wang", "Zhichao Wu", "Chiming Liu", "Xiaohan Yan", "Xuan Du", "Shuoyu Yue", "Chuheng Zhang", "Yunlong Wang", "Yongcheng Liu", "Lizhe Qi", "Yi Chen", "Wei Shan", "Maoqing Yao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.12691", "pdf_url": "https://arxiv.org/pdf/2602.12691", "code_url": "", "clusters": ["C3"], "keywords": ["value estimation"], "model_class": "VLA", "techniques": ["off-policy evaluation", "RL post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Off-policy evaluation that estimates current-policy value per action from mixed replay buffers for human-in-the-loop VLA post-training.", "relevance": 2, "relevance_reason": "Better value signals for real-world VLA RL.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (interrupt* OR \"human intervention\" OR \"mid-execution\")"}
{"id": "2609.25961", "title": "An Action Is Worth One Patch: Unified World-Action Modeling with PatchWAM", "authors": ["Tian-Heng Wang", "Zhou-Yao Xie", "Heng-Ji Jia", "Jian-Hua Xu", "Tong Zhang", "Kaicheng Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.25961", "pdf_url": "https://arxiv.org/pdf/2609.25961", "code_url": "", "clusters": ["C8"], "keywords": ["world-action model", "unified generation"], "model_class": "world-action model", "techniques": ["action-as-patch"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus", "RoboTwin 2.0"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PatchWAM encodes continuous actions as extra patches inside a video generative backbone, so one model predicts both future frames and actions without a separate action head.", "relevance": 2, "relevance_reason": "context: world-action model architecture", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2406.09246"}
{"id": "2602.21233", "title": "AngelSlim: A more accessible, comprehensive, and efficient toolkit for large model compression", "authors": ["Rui Cen", "QiangQiang Hu", "Hong Huang", "Hong Liu", "Song Liu", "Xin Luo", "Lin Niu", "Yifan Tan", "Decheng Wu", "Linchuan Xie", "Rubing Yang", "Guanghua Yu", "Jianchen Zhu"], "year": 2026, "venue": "arXiv tech report", "type": "tech-report", "url": "https://arxiv.org/abs/2602.21233", "pdf_url": "https://arxiv.org/pdf/2602.21233", "code_url": "", "clusters": ["C2"], "keywords": ["toolkit", "FP8", "INT2", "speculative decoding", "token pruning"], "model_class": "other", "techniques": ["quantization", "speculative decoding", "sparse attention", "token pruning", "distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "speculative decoding 1.8-2.0x throughput", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Tencent toolkit unifying PTQ (FP8/INT8/2-bit), speculative decoding, sparse attention and multimodal token pruning in one deployment pipeline.", "relevance": 2, "relevance_reason": "Industrial example of combining several compression families in one stack.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q1: arXiv abs:\"speculative decoding\" AND abs:quantization AND abs:compatibility"}
{"id": "2607.18985", "title": "Athena-Brain Technical Report: An Efficient Robot Brain for General Intelligence and Embodied Interaction", "authors": ["Jialian Li", "Junhong Liu", "Yuchen Cao", "Weiran Guo", "Jiaming Song", "Xutao Wang", "Yi Zhao", "Jiangpin Liu", "Jie Chen"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2607.18985", "pdf_url": "https://arxiv.org/pdf/2607.18985", "code_url": "", "clusters": ["C1"], "keywords": ["Athena-Brain"], "model_class": "other", "techniques": ["post-training", "model merge"], "base_model": "", "params": "8B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "8B LLM post-trained as an on-device embodied \"brain\" with concise responses for efficient interaction.", "relevance": 2, "relevance_reason": "Compact on-device high-level planner.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot policy\" OR embodied) AND (Jetson OR \"on-device\" OR onboard)"}
{"id": "2606.16826", "title": "ATOM-Bench: A Real-World Benchmark for Atomic Skills and Compositional Generalization in Manipulation Policies", "authors": ["Zenan Wu", "Bingqing Wei", "Lu Liu", "Zheqi He", "Xi Wang", "Jiakang Liu", "Zehui Li", "Guocai Yao", "Jing-Shu Zheng", "Xi Yang", "Yongtao Wang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.16826", "pdf_url": "https://arxiv.org/pdf/2606.16826", "code_url": "", "clusters": ["C4"], "keywords": ["ATOM-Bench", "compositional generalization", "real-world"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ATOM-Bench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Real-world benchmark of 30 atomic and 24 compositional tasks with 2,700 physical rollouts on five policies; atomic skill strength does not transfer to compositions.", "relevance": 2, "relevance_reason": "Real-robot benchmark design.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2603.08519", "title": "AtomVLA: Scalable Post-Training for Robotic Manipulation via Predictive Latent World Models", "authors": ["Xiaoquan Sun", "Zetian Xu", "Chen Cao", "Zonghe Liu", "Yihan Sun", "Jingrui Pang", "Ruijian Zhang", "Zhen Yang", "Kang Pang", "Dingxin He", "Mingqi Yuan", "Jiayu Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.08519", "pdf_url": "https://arxiv.org/pdf/2603.08519", "code_url": "", "clusters": ["C3"], "keywords": ["AtomVLA", "latent world model", "offline post-training", "subtasks"], "model_class": "VLA", "techniques": ["offline post-training", "latent world model rewards"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Subtask-aware VLA with scalable offline post-training driven by a predictive latent world model.", "relevance": 2, "relevance_reason": "Scalable post-training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"world model\" AND abs:\"post-training\""}
{"id": "2603.14851", "title": "AutoMoT: A Unified Vision-Language-Action Model with Asynchronous Mixture-of-Transformers for End-to-End Autonomous Driving", "authors": ["Wenhui Huang", "Songyan Zhang", "Qihang Huang", "Zhidong Wang", "Zhiqi Mao", "Collister Chua", "Zhan Chen", "Long Chen", "Chen Lv"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.14851", "pdf_url": "https://arxiv.org/pdf/2603.14851", "code_url": "", "clusters": ["C1"], "keywords": ["AutoMoT"], "model_class": "VLA", "techniques": ["mixture of transformers", "asynchronous inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA using an asynchronous mixture of transformers to decouple reasoning and action streams.", "relevance": 2, "relevance_reason": "Asynchronous architecture for latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "github-awesome-vla-deployment", "title": "awesome_vla_deployment: how to make a VLA policy actually run on a real robot", "authors": ["mingqian0850"], "year": 2026, "venue": "GitHub", "type": "repo", "url": "https://github.com/mingqian0850/awesome_vla_deployment", "pdf_url": "", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["deployment timing", "curated list", "symptom-cause-fix"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Curated deployment guide covering data, training recipes and deployment timing/safety organized as symptom to cause to fix.", "relevance": 2, "relevance_reason": "verified 2026-09-30 by orchestrator: README fetched: 171 curated entries on latency, chunking, quantisation, edge hardware and safety layers for real-robot VLA deployment.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: control loop jitter timing variance VLA"}
{"id": "2608.30536", "title": "Behavior-Skill: A Fine-Grained Benchmark for Evaluating Vision-Language-Action Policies in Long-Horizon Tasks", "authors": ["Ma, Chunyun", "Luo, Lun", "Luo, Xingjian", "Feng, Xiexing", "Zhang, Hang", "Liu, Wei", "Qiao, Feng", "Wang, Yaonan", "Lu, Huimin", "Chen, Xieyuanli"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2608.30536", "pdf_url": "https://arxiv.org/pdf/2608.30536", "code_url": "", "clusters": ["C4"], "keywords": ["Behavior-Skill", "skill-level evaluation", "restorable intermediate states", "BEHAVIOR"], "model_class": "VLA", "techniques": [], "base_model": "pi0.5, GR00T", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Behavior-Skill"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Decomposes 50 BEHAVIOR household tasks into 235k skill instances with restorable states and success conditions for skill-level evaluation.", "relevance": 2, "relevance_reason": "Finer-grained evaluation protocol.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: OpenReview CoRL 2025 evaluation protocol VLA benchmark flaws"}
{"id": "2601.12993", "title": "Being-H0.5: Scaling Human-Centric Robot Learning for Cross-Embodiment Generalization", "authors": ["Hao Luo", "Ye Wang", "Wan-Peng Zhang", "Sipeng Zheng", "Ziheng Xi", "Chao-Yi Xu", "Haiweng Xu", "Haoqi Yuan", "Chi Zhang", "Yi-Qin Wang", "Yicheng Feng", "Zongqing Lu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2601.12993", "pdf_url": "https://arxiv.org/pdf/2601.12993", "code_url": "", "clusters": ["C1"], "keywords": ["cross-embodiment", "human-centric"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Human-centric pretraining of a cross-embodiment VLA foundation model using human interaction traces.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2605.00078"}
{"id": "2608.27550", "title": "Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models", "authors": ["Sen-Qiao Yang", "Cheng-Yao Wang", "Yuxin Chen", "Zixuan Wang", "Long-Xiang Tang", "Haokun Gui", "Ji-lu Ye", "Chang-Sheng Lu", "Xiao-Yan Wu", "Ming-Kang Zhu", "Pengguang Chen", "Shu Liu", "Zhuo-Tao Tian", "Heng-Shuang Zhao", "Bei Yu", "Jia-Ya Jia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.27550", "pdf_url": "https://arxiv.org/pdf/2608.27550", "code_url": "", "clusters": ["C3"], "keywords": ["continued pretraining", "representation learning"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Continued-pretraining recipe for VLAs that focuses on representation quality under a fixed robot-data budget.", "relevance": 2, "relevance_reason": "Context: VLA training recipe (links=14).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2405.12213"}
{"id": "2604.24273", "title": "BitRL: Reinforcement Learning with 1-bit Quantized Language Models for Resource-Constrained Edge Deployment", "authors": ["Md. Ashiq Ul Islam Sajid", "Mohammad Sakib Mahmood", "Md. Tareq Hasan", "Md Abdur Rahim", "Rafat Ara", "Md. Arafat Hossain"], "year": 2026, "venue": "IEEE conference (per arXiv comments)", "type": "paper", "url": "https://arxiv.org/abs/2604.24273", "pdf_url": "https://arxiv.org/pdf/2604.24273", "code_url": "", "clusters": ["C2"], "keywords": ["BitNet b1.58", "RL agents", "edge"], "model_class": "other", "techniques": ["ternary quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "10-16x memory reduction, 3-5x energy efficiency", "metric_measures": "relative memory", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Builds RL agents on ternary BitNet LLMs for on-device learning/inference under tight memory and energy budgets.", "relevance": 2, "relevance_reason": "1-bit LLM agents; limited embodied evaluation.", "foundational_technique": false, "verified": true, "found_via": "search: quantization reinforcement learning policy deployment"}
{"id": "2608.12854", "title": "BrainWAM: Action-Space Coordination of Semantic Priors and Predictive Dynamics for Autonomous Driving", "authors": ["Bing Zhan", "Shuyao Shang", "Shuo Lu", "Yuan Xu", "Zhao Wang", "Yida Wang", "Xueyang Zhang", "Kun Zhan", "Jiahao Gu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.12854", "pdf_url": "https://arxiv.org/pdf/2608.12854", "code_url": "", "clusters": ["C8"], "keywords": ["driving WAM"], "model_class": "world-action model", "techniques": ["asynchronous denoising", "rectified flow"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM v1", "NAVSIM v2"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving planner coordinating a VLM semantic pathway and a world-action predictive pathway in action space, with asynchronous rectified-flow inference decoupling video and action denoising to cut latency.", "relevance": 2, "relevance_reason": "Latency-aware WAM/VLA fusion.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=0]"}
{"id": "2609.33007", "title": "CAPEX: Efficiently Distilling Foundation Model Behavior into Deployable Robot Policies through Experience-Adaptive Reasoning", "authors": ["Shivam Aarya", "Zhang Xi-Jia", "Chengyue Huang", "Junhyun Kim", "Huishu Xue", "Hrishit Leen", "Roman Yakunin", "Animesh Garg", "Zsolt Kira"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33007", "pdf_url": "https://arxiv.org/pdf/2609.33007", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["foundation model demonstrator", "adaptive replanning", "distillation", "data distillation", "Diffusion Policy", "ACT"], "model_class": "other", "techniques": ["data distillation", "adaptive query frequency", "behavior distillation via demonstrations"], "base_model": "", "params": "", "hardware": "", "reported_metric": "80% lower cost per successful demonstration", "metric_measures": "other", "benchmarks": ["RoboCasa", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Uses a multimodal foundation model as an autonomous demonstrator, adaptively reducing its calls, to distill behavior into small deployable policies.", "relevance": 2, "relevance_reason": "Distillation to deployable policies, data-side.", "foundational_technique": false, "verified": true, "found_via": "search query: diffusion policy distillation; search: distillation diffusion policy"}
{"id": "2609.23184", "title": "CausalWM: Causal Chain-of-Thought Reasoning for Embodied World Model", "authors": ["Zi-Ming Xu", "Shuang Liang", "R. Han", "Zi-Qiao Xi", "Ming-Xing Rao", "Kun Zhou", "Z. Zhang", "Yu-Chen Yan", "Yuxuan Wei", "Jun-Bo Huang", "Yi-Fei Shao", "Fang Nan", "Biwei Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.23184", "pdf_url": "https://arxiv.org/pdf/2609.23184", "code_url": "", "clusters": ["C8"], "keywords": ["embodied world model", "causal CoT"], "model_class": "world model", "techniques": ["few-step generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "CausalWM: 16B embodied world model with causal chain-of-thought before video prediction; also supports few-step generation.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2405.15223"}
{"id": "2601.04061", "title": "CLAP: Contrastive Latent Action Pretraining for Learning Vision-Language-Action Models from Human Videos", "authors": ["Chu-Bin Zhang", "Jia-Nan Wang", "Zi-Feng Gao", "Yue Su", "Tian-Ru Dai", "Cai Zhou", "Jiwen Lu", "Yansong Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.04061", "pdf_url": "https://arxiv.org/pdf/2601.04061", "code_url": "", "clusters": ["C1"], "keywords": ["latent action pretraining", "human video"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Contrastive latent action pretraining that learns executable action tokens from human videos for VLA training.", "relevance": 2, "relevance_reason": "Context: VLA pretraining (links=13).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.27406", "title": "CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators", "authors": ["Ke-Chen Liu", "Olaoluwa Shorinwa"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.27406", "pdf_url": "https://arxiv.org/pdf/2608.27406", "code_url": "", "clusters": ["C8"], "keywords": ["cross-embodiment video world model", "simulator"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cross-embodiment action-conditioned video world model that serves as a zero-shot physical simulator across robots.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 17 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2405.12213"}
{"id": "2604.07457", "title": "CMP: Robust Whole-Body Tracking for Loco-Manipulation via Competence Manifold Projection", "authors": ["Zi-Yang Cheng", "Hao-Yu Wei", "Hang Yin", "Xiu-Wei Xu", "Bingyao Yu", "Jie Zhou", "Jiwen Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.07457", "pdf_url": "https://arxiv.org/pdf/2604.07457", "code_url": "", "clusters": ["C7"], "keywords": ["competence manifold projection", "whole-body control", "OOD commands", "safety estimator"], "model_class": "other", "techniques": ["single-step manifold inclusion safety scheme"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Projects OOD commands onto a learned competence manifold via a frame-wise safety scheme so a whole-body loco-manipulation policy stays robust to noise and infeasible inputs.", "relevance": 2, "relevance_reason": "Efficient runtime safety for learned WBC.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2510.09459 (S2)"}
{"id": "2605.19138", "title": "COBALT: Crowdsourcing Robot Learning via Cloud-Based Teleoperation with Smartphones", "authors": ["Ayush Agarwal", "Ansh Gandhi", "Jeremy A. Collins", "Omar Rayyan", "Aryan Sarswat", "Ranjani Koushik", "Masoud Moghani", "Ajay Mandlekar", "Animesh Garg"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.19138", "pdf_url": "https://arxiv.org/pdf/2605.19138", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["cloud teleoperation", "GPU sharing", "crowdsourcing"], "model_class": "other", "techniques": ["load-balanced cloud simulation", "video streaming"], "base_model": "", "params": "", "hardware": "", "reported_metric": "20 Hz with sub-100 ms latency for up to 8 users per GPU; 256 simulated clients across 8 GPUs", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Cloud teleoperation platform sharing GPUs across many concurrent users for crowdsourced robot data collection.", "relevance": 2, "relevance_reason": "Cloud GPU multiplexing for robot learning infrastructure.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2607.00288", "title": "Communication-Aware and Safety-Aware UAV Control via Predictive Latent Models", "authors": ["Hamid Shiri", "Mehdi Bennis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.00288", "pdf_url": "https://arxiv.org/pdf/2607.00288", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["JEPA latent control", "UAV safety", "communication-aware"], "model_class": "world model", "techniques": ["JEPA multi-step rollouts with probabilistic safety constraints"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Uses a JEPA latent model's multi-step rollouts to anticipate collision risk and link degradation for jointly safe and energy-aware UAV control.", "relevance": 2, "relevance_reason": "Predictive latent safety control.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2506.09985 (citations)"}
{"id": "2606.00145", "title": "Completion at the Boundary (CaB): Deployable Switching with Completion-Aware Control under Limited Calibration", "authors": ["Y. Sano", "Takeshi Itoga"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.00145", "pdf_url": "https://arxiv.org/pdf/2606.00145", "code_url": "", "clusters": ["C7"], "keywords": ["completion detection", "switching", "low-calibration"], "model_class": "VLA", "techniques": ["completion-aware switching rule"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Studies when to declare an instruction complete and switch in composite VLA tasks under a single globally calibrated rule, since mistimed handoffs cascade into failures.", "relevance": 2, "relevance_reason": "Timing of handoffs in deployed VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2609.31792", "title": "ConflictVLA-Bench: Benchmarking Behavioral Responses of Vision-Language-Action Models to Premise Conflicts", "authors": ["Liyu Hou", "Yuan Wu", "Yi Chang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.31792", "pdf_url": "https://arxiv.org/pdf/2609.31792", "code_url": "https://github.com/EmbodiedAISurvey/ConflictVLA-Bench", "clusters": ["C4"], "keywords": ["ConflictVLA-Bench", "premise conflict"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "LIBERO-based tasks with invalid premises; eight VLAs keep pursuing original targets ('failed persistence'), showing binary outcomes are insufficient for evaluation.", "relevance": 2, "relevance_reason": "Process-level evaluation argument; tangential to efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2609.04552", "title": "Continual Field-Adaptive Models (CFAMs) for Post-Deployment Physical AI", "authors": ["Amarjot Singh", "Tanmay R. Pancholi", "Jainam Kothari", "Shrirang Mahajan", "Ketan Bansal", "Zackory Erickson", "Giuseppe Loianno", "Alexandre M. Bayen", "Jeff Schneider", "Vince Nakayama"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.04552", "pdf_url": "https://arxiv.org/pdf/2609.04552", "code_url": "", "clusters": ["C3"], "keywords": ["CFAM"], "model_class": "other", "techniques": ["gradient-free continual learning", "complementary learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Frozen slow component plus gradient-free \"capsule field\" for one-shot on-device continual skill updates across five embodiments.", "relevance": 2, "relevance_reason": "On-device continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot policy\" OR embodied) AND (Jetson OR \"on-device\" OR onboard)"}
{"id": "2608.13819", "title": "Control Barrier--Value Functions under Partial Observability: Safety Guarantees via Conformal Prediction", "authors": ["Niloofar Jahanshahi", "Mo Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.13819", "pdf_url": "https://arxiv.org/pdf/2608.13819", "code_url": "", "clusters": ["C7"], "keywords": ["partial observability"], "model_class": "other", "techniques": ["CBVF", "conformal prediction", "QP safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Extends control barrier-value functions to partially observed systems by conformally bounding state-estimation error, yielding a QP safety filter with finite-horizon probabilistic guarantees.", "relevance": 2, "relevance_reason": "Safety filtering under perception error.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:conformal AND (ti:robot* OR ti:planning OR ti:control)"}
{"id": "2609.36588", "title": "Cooperative Multi-Agent Vision-Language-Action Models via Reinforced Fine Tuning", "authors": ["Rui-Xiao Xu", "W. Kenny", "Zhi-Qian Liu", "Jia-Ni Guo", "Han-Xiao Li", "Ke-Jian Shi", "Shu-Ning Zhang", "Pu Feng", "Yong-Jia Ma", "Yuqing Ma", "Kai Chen", "Qi Dou", "Yao-Dong Yang", "Xiang-Long Liu", "Si-Min Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.36588", "pdf_url": "https://arxiv.org/pdf/2609.36588", "code_url": "", "clusters": ["C3"], "keywords": ["multi-agent", "cooperative"], "model_class": "VLA", "techniques": ["RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RL fine-tuning for cooperative multi-agent VLAs beyond single-agent pretraining.", "relevance": 2, "relevance_reason": "Context: VLA RL (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2606.02800", "title": "Cosmos 3: Omnimodal World Models for Physical AI", "authors": ["Nvidia Aditi", "Niket Agarwal", "Arslan Ali", "J. Allen", "Martin Antolini", "Adeline Aubame", "Alisson G. Azzolini", "Junjie Bai", "M. Bala", "Yogesh Balaji", "Josh Bapst", "Aarti Basant", "Mukesh Beladiya", "Mohammad Bhat", "Zaid Pervaiz Bhat", "Daniel F. Blick", "Vanni Brighella", "Han Cai", "Tiffany Cai", "Eric Cameracci", "Jiaxin Cao", "Yulong Cao", "Mark T. Carlson", "Carlos Casanova", "Ting-Yun Chang", "Yan Chang", "Yu-Wei Chao", "Prithvijit Chattopadhyay", "Roshan Chaudhari", "Chieh-yun Chen", "Jun-Yu Chen", "Kejiang Chen", "Qizhi Chen", "Wen-Kai Chen", "Xiaotong Chen", "Yu Chen", "An-Chieh Cheng", "Click Cheng", "X. Chia", "Jeana Choi", "Chaeyeon Chung", "Wenyan Cong", "Yin Cui", "Magdalena Dadela", "Nalin Dadhich", "Wenliang Dai", "Joyjit Daw", "Alperen Değirmenci", "Rodrigo Vieira Del Monte", "R. Denommé", "Sameer Dharur", "Marco Di Lucca", "Ke Ding", "Wenhao Ding", "Yifan Ding", "Yuzhu Dong", "Nicole Drumheller", "Yi-Lun Du", "A. Dzhumamuratova", "A. Efitorov", "Hamid Eghbalzadeh", "Naomi Eigbe", "Imad El Hanafi", "H. Eslami", "B. Falk", "Jiao-Jiao Fan", "Jim Fan", "Amol Fasale", "Sergiy Fefilatyev", "Liang Feng", "Francesco Ferroni", "Sanja Fidler", "Xiaojin Fu", "Vikram Fugro", "P. Gaikwad", "TJ Galda", "Katelyn Gao", "Yihuai Gao", "Wen-Hang Ge", "Sreyan Ghosh", "Arushi Goel", "Vivek Goel", "Akash Gokul", "R. Govindaraju", "Jinwei Gu", "Miguel Guerrero", "Elfie Guo", "Aryaman Gupta", "Siddharth Gururani", "Hugo Hadfield", "Song Han", "Ankur Handa", "Zekun Hao", "Mohammad Harrim", "A. Hassani", "Nathan Hayes-Roth", "Yu-Fan He", "C. Helvig", "Cyrus Hogg", "Madison Huang", "Michael Huang", "Sophia Huang", "Yu-Fan Huang", "J. Huffman", "DeLesley S. Hutchins", "Suneel Indupuru", "B. Ivanovic", "Arihant Jain", "J. Jang", "Ryan Ji", "Yanan Jian", "Dongfu Jiang", "Jingyi Jin", "A. Joshi", "N. Joshi", "Pranjali Joshi", "Andy Ju", "Jaehun Jung", "Weiwei Kang", "Scott Kassekert", "Jan Kautz", "Ashna Khetan", "J. Kiczka", "Slawomir Kierat", "Gwanghyun Kim", "Kuno Kim", "Sunny Kim", "Kezhi Kong", "Xinke Kong", "Zhifeng Kong", "Tomasz Kornuta", "Egor Krivov", "Hui Kuang", "Saurav Kumar", "Chia-Wen Kuo", "George Kurian", "Wojciech Kutak", "JF Lafleche", "Himangshu Lahkar", "Omar Laymoun", "Jayjun Lee", "Sang-Gyoung Lee", "G. Leone", "Boyi Li", "Freya Li", "Jia-Ju Li", "Jin-Feng Li", "Ling Li", "Pengcheng Li", "Shang-Ru Li", "Ting Li", "Xiao-Long Li", "Xuan Li", "Zhao-Shuo Li", "Zhi-Qi Li", "Haojie Liang", "Maosheng Liao", "Cheng-Hua Lin", "Tsung-Yi Lin", "Mingyang Liu", "Sifei Liu", "Zihan Liu", "Hai-Ying Lu", "Xiangyu Lu", "Alice Luo", "Ruipu Luo", "Wen-Jie Luo", "Jiangran Lyu", "M. Ma", "Nic Ma", "Qianli Ma", "Dawid Majchrowski", "Louis-Alexandre Marcoux", "Miguel Martin", "Qing Miao", "Ashkan Mirzaei", "Shreyas Misra", "Kaichun Mo", "D. Mohsin", "Hye-Jeong Moon", "Pawel M. Morkisz", "Saeid Motiian", "K. Motkov", "Seungjun Nah", "Yashraj Narang", "Deepak Narayanan", "Thabang Ngazimbi", "Julian Ouyang", "Shubham Pachori", "David Page", "Yatian Pang", "Sehwi Park", "M. Patekar", "M. Patwary", "Marco Pavone", "T. Pham", "Wei Ping", "Soha Pouya", "Shrimai Prabhumoye", "Varun Praveen", "Delin Qu", "Hesam Rabeti", "Morteza Ramezanali", "M. Reeb", "Xuanchi Ren", "Kristen Rumley", "W. Rymer", "Jun Saito", "Yeongho Seol", "J. Shao", "Piyush Shekdar", "Tianwei Shen", "Humphrey Shi", "Min Shi", "Stella Shi", "Kevin J. Shih", "M. Shoeybi", "Mateusz Sieniawski", "Shuran Song", "A. Sotelo", "A. Sotoodeh", "Sunil Srinivasa", "Vignesh Srinivasakumar", "Bartosz Stefaniak", "Rahul Steiger", "Shangkun Sun", "Jiaxiang Tang", "Shitao Tang", "Yan Tang", "Yue Tang", "Tolou Tavakkoli", "Kayley Ting", "Krzysztof Tomala", "Wei-Cheng Tseng", "J. Varghese", "S. Vasil'ev", "Thomas Volk", "Raju Wagwani", "R. Waleffe", "Andrew Z. Wang", "Boxiang Wang", "Hao-Xiang Wang", "Qiaolin Wang", "Shi-Hao Wang", "Shijie Wang", "Tingchang Wang", "Yan Wang", "Yu Wang", "Rohit Watve", "David Wehr", "Fang-Yin Wei", "Xinshuo Weng", "J. Wu", "Kedi Wu", "Hongchi Xia", "S. Xiao", "Tianjun Xiao", "Kevin Xie", "Da-Guang Xu", "Jiashu Xu", "Mengya Xu", "Ruqing Xu", "Xingqian Xu", "Yao Xu", "Dinghao Yang", "Dong Yang", "Han Yang", "Xiao-Dong Yang", "Xu-Ning Yang", "Yichu Yang", "Yurong You", "Zhiding Yu", "Haobo Yuan", "S. Yuen", "Xiaohui Zeng", "Pengcuo Zeren", "Cindy Zha", "Hao-Tian Zhang", "Jenny Zhang", "Jing Zhang", "Liang Zhang", "Paris Zhang", "Shun Zhang", "Xuan Zhang", "Zhizheng Zhang", "A. Zhao", "Yilin Zhao", "Yuliya Zhautouskaya", "Charles Zhou", "Fengzhe Zhou", "Shilin Zhu", "Yu-Ke Zhu", "D. Zhylko", "Artur Zolkowski"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.02800", "pdf_url": "https://arxiv.org/pdf/2606.02800", "code_url": "https://github.com/nvidia/cosmos", "clusters": ["C8"], "keywords": ["world foundation model", "omnimodal"], "model_class": "world model", "techniques": ["mixture-of-transformers", "omnimodal"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cosmos 3 is a family of omnimodal world models jointly generating language, image, video, audio and action in a unified mixture-of-transformers architecture for physical AI.", "relevance": 2, "relevance_reason": "context: major world foundation model", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2606.19531, 2607.01088"}
{"id": "2602.06219", "title": "Coupled Local and Global World Models for Efficient First Order RL", "authors": ["Joseph Amigo", "Rooholla Khorrambakht", "Nicolas Mansard", "Ludovic Righetti"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.06219", "pdf_url": "https://arxiv.org/pdf/2602.06219", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["world-model RL"], "model_class": "world model", "techniques": ["first-order gradients", "surrogate model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Trains RL inside a large diffusion world model using a lightweight latent surrogate for first-order gradients.", "relevance": 2, "relevance_reason": "Tractable policy learning in world models.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "2602.18025", "title": "Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets", "authors": ["Haruki Abe", "Takayuki Osa", "Yusuke Mukuta", "Tatsuya Harada"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.18025", "pdf_url": "https://arxiv.org/pdf/2602.18025", "code_url": "", "clusters": ["C3"], "keywords": ["cross-embodiment", "offline RL"], "model_class": "other", "techniques": ["offline RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Systematic analysis of offline RL for cross-embodiment policy pretraining using suboptimal data.", "relevance": 2, "relevance_reason": "Cross-embodiment data efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "cross-embodiment-robot-foundation-world-models-latent-actions-icml2026", "title": "Cross-Embodiment Robot Foundation World Models with Latent Actions", "authors": ["Huang Huang", "Sriram Yenamandra", "Arjun Majumdar", "Elie Aljalbout", "Tushar Nagarajan", "Tsung-Yen Yang", "Akshara Rai", "Michael Rabbat", "Li Fei-Fei", "Jiajun Wu", "Tingfan Wu", "Franziska Meier"], "year": 2026, "venue": "ICML 2026", "type": "paper", "url": "https://openreview.net/forum?id=S30Y1sBz34", "pdf_url": "", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["foundation world model", "latent actions", "cross-embodiment"], "model_class": "world model", "techniques": ["latent action conditioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LAC-WM: a robot world model conditioned on a learned latent action space shared across embodiments, which adapts better to unseen robots than explicit action conditioning.", "relevance": 2, "relevance_reason": "verified 2026-09-30 by orchestrator: OpenReview record (ICML 2026 regular) and abstract fetched. World-model work from speaker Jiajun Wu's group.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu (SAIL blog ICML 2026)"}
{"id": "2606.30293", "title": "CSAR: Containerized System Architecture for Robotics", "authors": ["Gregorio Ambrosio-Cestero", "Cipriano Galindo", "Javier Gonzalez-Jimenez", "Jose-Raul Ruiz-Sarmiento"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.30293", "pdf_url": "https://arxiv.org/pdf/2606.30293", "code_url": "", "clusters": ["C6"], "keywords": ["containerized architecture", "edge-cloud continuum", "ROS 2"], "model_class": "other", "techniques": ["LXC/LXD containers", "three-layer edge infrastructure"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Container-centric architecture for robotics teams spanning embedded devices, edge servers and cloud with ROS 2/DDS networking.", "relevance": 2, "relevance_reason": "Edge-cloud deployment infrastructure.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2605.07325", "title": "CSR: Infinite-Horizon Real-Time Policies with Massive Cached State Representations", "authors": ["Robin Karlsson", "Go Suzui"], "year": 2026, "venue": "IEEE RA-L", "type": "paper", "url": "https://arxiv.org/abs/2605.07325", "pdf_url": "https://arxiv.org/pdf/2605.07325", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["TTFT", "cached state", "LLM robotics", "prefix stability"], "model_class": "other", "techniques": ["KV cache reuse", "asynchronous state reconciliation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Shows how prefix-stable, incrementally extended state with asynchronous reconciliation keeps LLM time-to-first-token low for continuously running robot cognition.", "relevance": 2, "relevance_reason": "Latency for large (typically server-hosted) LLM robot controllers.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: offload robot foundation model"}
{"id": "2603.09051", "title": "Cutting the Cord: System Architecture for Low-Cost, GPU-Accelerated Bimanual Mobile Manipulation", "authors": ["Artemis Shaw", "Chen Liu", "Justin Costa", "Rane Gray", "Alina Skowronek", "Kevin Diaz", "Nam Bui", "Nikolaus Correll"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.09051", "pdf_url": "https://arxiv.org/pdf/2603.09051", "code_url": "", "clusters": ["C5"], "keywords": ["low-cost mobile manipulator", "Jetson Orin Nano", "power topology", "XLeRobot"], "model_class": "other", "techniques": ["isolated compute power bus", "embedded GPU autonomy"], "base_model": "", "params": "", "hardware": "Jetson Orin Nano", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "Sub-$1300 bimanual mobile manipulator with onboard Jetson Orin Nano and a power topology that isolates compute from motor voltage transients.", "relevance": 2, "relevance_reason": "Onboard compute integration and power design for low-cost robot-learning platforms.", "foundational_technique": false, "verified": true, "found_via": "arXiv: Jetson AND (robot policy OR manipulation OR VLA)"}
{"id": "2606.17574", "title": "DeepInsight: A Unified Evaluation Infrastructure Across the Physical AI Stack", "authors": ["Siyi Li", "Chunyu Sun", "Jiahao Zhang", "Yuchen Kang", "Wuliang Wang", "Yu Qiu", "Rui Jiang", "Haitao Cui", "Jie Chen"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.17574", "pdf_url": "https://arxiv.org/pdf/2606.17574", "code_url": "", "clusters": ["C4"], "keywords": ["DeepInsight"], "model_class": "other", "techniques": ["evaluation infrastructure"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Evaluation infrastructure spanning foundation-model decoding to whole-body control ticks on one runtime with shared task/resource/result abstractions.", "relevance": 2, "relevance_reason": "Cross-layer benchmarking for physical AI stacks.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot policy\" OR embodied) AND (Jetson OR \"on-device\" OR onboard)"}
{"id": "2608.11063", "title": "Deployment Is Not Destiny: Robot Recomposition in the Field with Unseen Software, Hardware, and Compute Payloads", "authors": ["Steven Swanbeck", "Jonathan Salfity", "Jeffery Gunawan", "Corrie Van Sice", "Mitch Pryor", "Robert Blake Anderson"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.11063", "pdf_url": "https://arxiv.org/pdf/2608.11063", "code_url": "", "clusters": ["C6"], "keywords": ["runtime recomposition", "compute payloads", "peer sharing"], "model_class": "other", "techniques": ["plug-and-play payloads", "distributed resource sharing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Framework letting fielded robots integrate new software, hardware and compute payloads at runtime and share them with peers, so constrained robots can use new remote compute.", "relevance": 2, "relevance_reason": "Fleet-level compute sharing.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2609.24976", "title": "DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation", "authors": ["Haoran Yuan", "Zekai Wang", "Boning Shao", "Haoran Lu", "Trevor Darrell", "Ismini Lourentzou", "Wei Zhan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.24976", "pdf_url": "https://arxiv.org/pdf/2609.24976", "code_url": "", "clusters": ["C8"], "keywords": ["DexTacWAM"], "model_class": "world-action model", "techniques": ["tactile compression", "video diffusion WAM"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.29x faster inference, 2.26x faster training", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Visuo-tactile WAM with per-fingertip encoding and a tactile compressor injected into a video diffusion world model.", "relevance": 2, "relevance_reason": "Compressor gives modest speedups.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2603.29844", "title": "DIAL: Decoupling Intent and Action via Latent World Modeling for End-to-End VLA", "authors": ["Yi Chen", "Yuying Ge", "Hui Zhou", "Mingyu Ding", "Yixiao Ge", "Xihui Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.29844", "pdf_url": "https://arxiv.org/pdf/2603.29844", "code_url": "", "clusters": ["C8"], "keywords": ["DIAL"], "model_class": "VLA", "techniques": ["latent world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Decouples intent and action in driving VLAs via latent world modeling for end-to-end control.", "relevance": 2, "relevance_reason": "Latent WM inside VLA.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2609.18392", "title": "DistAL: Distance-based Advantage Learning for VLA Fine-Tuning", "authors": ["Reece O'Mahoney", "Ioannis Havoutis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.18392", "pdf_url": "https://arxiv.org/pdf/2609.18392", "code_url": "", "clusters": ["C3"], "keywords": ["advantage conditioning", "embedding distance reward"], "model_class": "VLA", "techniques": ["advantage-conditioned post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses embedding-space distance as reward for richer value functions in advantage-conditioned VLA fine-tuning.", "relevance": 2, "relevance_reason": "Post-training recipe.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2603.10448", "title": "DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control", "authors": ["Teli Ma", "Jiayu Zheng", "Zifan Wang", "Chun-Li Jiang", "Andrew Cui", "Junwei Liang", "Shuo Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.10448", "pdf_url": "https://arxiv.org/pdf/2603.10448", "code_url": "", "clusters": ["C8"], "keywords": ["video-action model", "DiT"], "model_class": "world-action model", "techniques": ["video diffusion features", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Cascades a video diffusion transformer with an action DiT, conditioning actions on intermediate denoising features instead of full reconstructed frames.", "relevance": 2, "relevance_reason": "Context: prominent video-action model cited by several core items.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2604.26694, 2606.09215"}
{"id": "2608.29434", "title": "Does Latent Planning Survive Point Clouds? Action-Conditioned JEPA World Models for Geometric Observations and Goals", "authors": ["Fabio F. Oberweger", "Michael Schwingshackl", "Markus Murschitz"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.29434", "pdf_url": "https://arxiv.org/pdf/2608.29434", "code_url": "", "clusters": ["C8"], "keywords": ["point-cloud JEPA", "geometric goals", "stable-worldmodel"], "model_class": "world model", "techniques": ["JEPA world models on raycast point clouds", "pose-based goal latents"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Tests whether JEPA latent planning transfers from images to point-cloud observations and pose-specified goals, comparing anti-collapse designs.", "relevance": 2, "relevance_reason": "Extends latent planning modality.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2609.34911", "title": "Don't Throw Away the Tail: Action Upcycling for Policy Acceleration", "authors": ["Kwon, Taesung", "Park, Jangho", "Park, Sunwoo", "Kim, Youngmin", "Jin, Seonghyun", "Jun, Youngjun", "Choi, Kyumin", "Ye, Jong Chul"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34911", "pdf_url": "https://arxiv.org/pdf/2609.34911", "code_url": "", "clusters": ["C2", "C4"], "keywords": ["Action Upcycling", "execution horizon", "policy calls"], "model_class": "diffusion/flow policy", "techniques": ["adaptive execution horizon"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.2-1.7x fewer policy calls", "metric_measures": "other", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Extends the executed prefix of a chunk while predicted velocity stays smooth, cutting policy calls without accuracy loss.", "relevance": 2, "relevance_reason": "C2 primary; policy-call count as metric.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: action chunk size inference latency trade-off"}
{"id": "2608.13489", "title": "DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation", "authors": ["DreamX Team", "Rui Chen", "Xiangxiang Chu", "Geng Li", "Jifan Li", "Qingfeng Shi", "Datao Tang", "Jing Tang", "Jun Wang", "Pengfei Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.13489", "pdf_url": "https://arxiv.org/pdf/2608.13489", "code_url": "https://github.com/AMAP-ML/DreamX-Phi", "clusters": ["C2", "C8"], "keywords": ["DreamX-Phi"], "model_class": "world model", "techniques": ["distribution matching distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Action-conditioned video world model for manipulation with per-arm SE(3) encoding, then distribution-matching distilled to a few-step student.", "relevance": 2, "relevance_reason": "Few-step distillation of robot world models.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "2608.10413", "title": "DriveVLA-M0: Failure-Aware Memory Augmentation for Autonomous Driving", "authors": ["Zebin Xing", "Yupeng Zheng", "Qiang Chen", "Linbo Wang", "Yichen Zhang", "Pengxuan Yang", "Junli Wang", "Deheng Qian", "Xiaoqing Ye", "Junyu Han", "Yifeng Pan", "Qichao Zhang", "Dongbin Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.10413", "pdf_url": "https://arxiv.org/pdf/2608.10413", "code_url": "https://github.com/ZebinX/DriveVLA-M0", "clusters": ["C1"], "keywords": ["failure memory", "driving"], "model_class": "VLA", "techniques": ["retrieval", "LoRA test-time training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA with a failure-case latent memory and LoRA test-time training on retrieved cases for scenario-specific correction.", "relevance": 2, "relevance_reason": "Lightweight test-time correction for driving VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]"}
{"id": "2605.28544", "title": "DriveWAM: Video Generative Priors Enable Scalable World-Action Modeling for Autonomous Driving", "authors": ["Chen Shi", "Jinrui Xu", "Shaoshuai Shi", "Kehua Sheng", "Bo Zhang", "Li Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.28544", "pdf_url": "https://arxiv.org/pdf/2605.28544", "code_url": "", "clusters": ["C8"], "keywords": ["DriveWAM"], "model_class": "world-action model", "techniques": ["selective KV cache", "video diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM", "PhysicalAI-AV"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts a pretrained video DiT into an autoregressive driving video-action policy with selective KV memory to bound long rollouts.", "relevance": 2, "relevance_reason": "KV memory control for WAMs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\")"}
{"id": "2607.04927", "title": "DSWAM: A Dual-System World Action Foundation Model for Fine-Grained Robot Manipulation", "authors": ["Jian Zhu", "Jian-Jun Zhang", "Taiyi Su", "Tianbin Liu", "Zhangyuan Wang", "Kai Xie", "Zi-Tai Huang", "Chong Ma", "Youzhang He", "Tianjian Wang", "Han-Yang Wang", "Wei-Hao Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.04927", "pdf_url": "https://arxiv.org/pdf/2607.04927", "code_url": "", "clusters": ["C1", "C7", "C9"], "keywords": ["dual-system", "world action model", "language planning"], "model_class": "world-action model", "techniques": ["dual-system WAM"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DSWAM adds a VLM planning interface on top of a world-action model for fine-grained multi-step household manipulation.", "relevance": 2, "relevance_reason": "World-action model architecture.", "foundational_technique": false, "verified": true, "found_via": "venue sweep: S2 'efficient world model real-time planning robot manipulation'"}
{"id": "2601.16065", "title": "DTP: A Simple yet Effective Distracting Token Pruning Framework for Vision-Language Action Models", "authors": ["Chenyang Li", "Jieyuan Liu", "Bin Li", "Bo Gao", "Yilin Yuan", "Yangfan He", "Yuchen Li", "Jingqun Tang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.16065", "pdf_url": "https://arxiv.org/pdf/2601.16065", "code_url": "", "clusters": ["C1"], "keywords": ["distracting tokens", "token pruning"], "model_class": "VLA", "techniques": ["token pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SIMPLER"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Detects and prunes image tokens in task-irrelevant regions that attract attention, aiming mainly at success-rate gains rather than speed.", "relevance": 2, "relevance_reason": "Pruning mostly for accuracy, weak efficiency angle.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA token pruning"}
{"id": "2606.11901", "title": "DuoBench: A Reproducible Benchmark for Bimanual Manipulation in Simulation and the Real World", "authors": ["Tobias Jülg", "Seongjin Bien", "Simon Hilber", "Yannik Blei", "Pierre Krack", "Maximilian Li", "Sven Parusel", "Rudolf Lioutikov", "Florian Walter", "Wolfram Burgard"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.11901", "pdf_url": "https://arxiv.org/pdf/2606.11901", "code_url": "", "clusters": ["C4"], "keywords": ["DuoBench", "bimanual", "reproducible", "stage-based evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "FR3 Duo", "reported_metric": "", "metric_measures": "", "benchmarks": ["DuoBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Bimanual benchmark with 11 tasks in sim and partially in real via 3D-printable recipes, with stage-based failure analysis.", "relevance": 2, "relevance_reason": "Reproducible sim+real benchmark.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: CALVIN RLBench ManiSkill3"}
{"id": "2604.00813", "title": "DVGT-2: Vision-Geometry-Action Model for Autonomous Driving at Scale", "authors": ["Sicheng Zuo", "Zixun Xie", "Wenzhao Zheng", "Shaoqing Xu", "Fang Li", "Hanbing Li", "Long Chen", "Zhi-Xin Yang", "Jiwen Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.00813", "pdf_url": "https://arxiv.org/pdf/2604.00813", "code_url": "https://github.com/wzzheng/DVGT", "clusters": ["C1"], "keywords": ["DVGT-2"], "model_class": "VLA", "techniques": ["streaming inference", "caching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM", "nuScenes"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving vision-geometry-action model with sliding-window streaming and historical caches to avoid recomputation.", "relevance": 2, "relevance_reason": "Streaming/caching for efficient driving models.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2606.23623", "title": "dVLA-RL: Reinforcement Learning over Denoising Trajectories for Discrete Diffusion Vision-Language-Action Models", "authors": ["Yuhao Wu", "Yitian Liu", "Weijie Shen", "Mishuo Han", "Wenjie Xu", "Haotian Liang", "Zhongshan Liu", "Yinan Mao", "Lei Xu", "Xinping Guan", "Ru Ying", "Ran Zheng", "Wei Sui", "Xiaokang Yang", "Wenbo Ding", "Yao Mu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.23623", "pdf_url": "https://arxiv.org/pdf/2606.23623", "code_url": "", "clusters": ["C3"], "keywords": ["discrete diffusion VLA", "RL"], "model_class": "VLA", "techniques": ["RL over denoising trajectories"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "First RL fine-tuning method for discrete-diffusion VLAs over their denoising trajectories.", "relevance": 2, "relevance_reason": "RL post-training for new VLA class.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2605.17486", "title": "DyGRO-VLA: Cross-Task Scaling of Vision-Language-Action Models via Dynamic Grouped Residual Optimization", "authors": ["Sixu Lin", "Yunpeng Qing", "Litao Liu", "Ming Zhou", "Ruixing Jin", "Xiaoyi Fan", "Guiliang Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.17486", "pdf_url": "https://arxiv.org/pdf/2605.17486", "code_url": "", "clusters": ["C3"], "keywords": ["cross-task RL", "grouped residual optimization"], "model_class": "VLA", "techniques": ["RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Two-stage RL optimizer preserving cross-task representations to avoid task overfitting during VLA RL.", "relevance": 2, "relevance_reason": "RL post-training generalization.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.33172", "title": "Dynamic Manipulation with World-Action Models via Counterfactual Planning", "authors": ["Sunwoo Park", "Wonbin Lee", "Seonghyun Jin", "Youngmin Kim", "Jangho Park", "Jong Chul Ye"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33172", "pdf_url": "https://arxiv.org/pdf/2609.33172", "code_url": "", "clusters": ["C8"], "keywords": ["dynamic manipulation"], "model_class": "world-action model", "techniques": ["counterfactual observations"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Counterfactual planning for WAMs on moving targets: predicts interaction timing and position and builds a counterfactual observation; runs in real time on one consumer GPU.", "relevance": 2, "relevance_reason": "Real-time dynamic manipulation with frozen WAMs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2606.29699", "title": "Early Warning Signals for OpenVLA Failure under Visual Distribution Shift", "authors": ["Dipesh Mahato", "Rachel Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.29699", "pdf_url": "https://arxiv.org/pdf/2606.29699", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["early warning", "OpenVLA", "linear probe", "visual shift"], "model_class": "VLA", "techniques": ["activation linear probes"], "base_model": "OpenVLA", "params": "", "hardware": "", "reported_metric": "layer-16 probe AUROC 0.972 within failed trajectories", "metric_measures": "other", "benchmarks": ["LIBERO-10"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Finds linear probes on OpenVLA MLP activations give early warning of failure under occlusion but reports calibration issues (frequent false warnings).", "relevance": 2, "relevance_reason": "Cheap internal monitors; calibration caveat.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2606.18239", "title": "EBench: Elemental Diagnosis of Generalist Mobile Manipulation Policies", "authors": ["Ning Gao", "Jinliang Zheng", "Xing Gao", "Hao-Xiang Ma", "Han-Qing Wang", "Yu-Kai Wang", "Jiantong Chen", "Zan-Xin Chen", "Shujie Zhang", "Mingda Jia", "Xuekun Jiang", "Zihou Zhu", "Xinyu Li", "Shuai Wang", "Hao Li", "Wenzhe Cai", "Yu-Qiang Yang", "Xudong Xu", "Zhaoyang Lyu", "Yao Mu", "Tai Wang", "Jiangmiao Pang", "Jia-Qi Zeng", "Weinan Zhang", "Chunhua Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18239", "pdf_url": "https://arxiv.org/pdf/2606.18239", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "mobile manipulation", "diagnosis"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Simulation benchmark that diagnoses generalist mobile-manipulation policies along capability and generalization dimensions.", "relevance": 2, "relevance_reason": "Context: benchmark (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2603.11085", "title": "Edge-Assisted Multi-Robot Visual-Inertial SLAM with Efficient Communication", "authors": ["Xin Liu", "Shuhuan Wen", "Jing Zhao", "Tony Z. Qiu", "Hong Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.11085", "pdf_url": "https://arxiv.org/pdf/2603.11085", "code_url": "", "clusters": ["C6"], "keywords": ["multi-robot VI-SLAM", "robot-edge-cloud", "bandwidth"], "model_class": "other", "techniques": ["feature-only transmission", "lossless compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Layered robot-edge-cloud multi-robot SLAM sending only features and keyframe descriptors to cope with limited bandwidth.", "relevance": 2, "relevance_reason": "Bandwidth-aware robot-edge-cloud pipeline.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:edge-cloud AND abs:robot"}
{"id": "2602.20577", "title": "Efficient and Explainable End-to-End Autonomous Driving via Masked Vision-Language-Action Diffusion", "authors": ["Jiaru Zhang", "Manav Gagvani", "Can Cui", "Juntong Peng", "Ruqi Zhang", "Ziran Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.20577", "pdf_url": "https://arxiv.org/pdf/2602.20577", "code_url": "", "clusters": ["C1"], "keywords": ["driving VLA"], "model_class": "VLA", "techniques": ["masked diffusion", "action codebook"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["nuScenes"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Masked discrete-diffusion driving VLA with a compact codebook of feasible waypoints and action-priority decoding for faster planning than AR decoding.", "relevance": 2, "relevance_reason": "Parallel decoding for driving VLA latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]"}
{"id": "2605.16692", "title": "EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control", "authors": ["Thomas Evers", "Cristian Meo", "Wendelin Bohmer", "Justin Dauwels", "Yaniv Oren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.16692", "pdf_url": "https://arxiv.org/pdf/2605.16692", "code_url": "", "clusters": ["C8"], "keywords": ["TD-MPC", "multi-horizon objective", "dynamics ensemble"], "model_class": "world model", "techniques": ["multi-horizon planning objective", "value-equivalent dynamics ensemble", "pessimistic reanalyze"], "base_model": "", "params": "", "hardware": "", "reported_metric": "reaches BMPC final performance with 57% fewer environment steps", "metric_measures": "other", "benchmarks": ["HumanoidBench", "DMControl"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Reduces model and value error in TD-MPC planning with multi-horizon returns and a dynamics ensemble, setting sample-efficiency SOTA on HumanoidBench/DMC.", "relevance": 2, "relevance_reason": "Sample efficiency rather than compute efficiency of latent MPC.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828"}
{"id": "2606.02246", "title": "Ego-METAS: Egocentric online Multimodal Energy-efficient Temporal Action Segmentation benchmark", "authors": ["Maria Santos-Villafranca", "Jesus Bermudez-cameo", "Alejandro Perez-Yus", "Giovanni Maria Farinella", "Antonino Furnari"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.02246", "pdf_url": "https://arxiv.org/pdf/2606.02246", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["energy budget", "sensor routing", "egocentric", "online segmentation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Ego-METAS"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Online egocentric action-segmentation benchmark where models choose which of 5 sensors to activate under hardware-representative energy budgets.", "relevance": 2, "relevance_reason": "Energy-budgeted embodied perception benchmark.", "foundational_technique": false, "verified": true, "found_via": "search query: embodied AI benchmark efficiency"}
{"id": "2607.03283", "title": "Embodied Operators and Benchmarking: Toward Reusable and Deployable Embodied Intelligence Systems", "authors": ["Junwu Xiong", "Jiaxuan Gao", "Wei Chai", "Renxing Chen", "Yuzhen Li", "Yu Guo", "Yucheng Guo", "Mingxi Luo", "Wenyang Ma", "Yiyun Mou", "Yifei Zhang", "Chen Zhou", "Yongjian Guo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.03283", "pdf_url": "https://arxiv.org/pdf/2607.03283", "code_url": "", "clusters": ["C4", "C9"], "keywords": ["embodied operators", "composable modules", "benchmark framework", "deployability"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Defines reusable 'embodied operators' with standardized I/O contracts, builds a taxonomy, and proposes a multi-dimensional benchmark covering deployability.", "relevance": 2, "relevance_reason": "Composability/deployability benchmark framing.", "foundational_technique": false, "verified": true, "found_via": "search query: policy evaluation real robot scalable"}
{"id": "2606.11324", "title": "Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models", "authors": ["Yifu Yuan", "Yao-Bang Huang", "X. Yao", "Yutong Li", "Shuo Zhang", "Linqi Han", "Peng-Yi Li", "Jiangeng Sun", "Wenting Jia", "Zhao-Xiang Zhang", "Yuhao Liu", "Ruihao Liao", "Yucheng Hu", "Qiyu Wu", "Yuxiao Li", "Zibin Dong", "Fei Ni", "Yang-Heng Zheng", "Shuyang Gu", "Yi Ma", "Hongyao Tang", "Han Hu", "Jian-Ye Hao"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.11324", "pdf_url": "https://arxiv.org/pdf/2606.11324", "code_url": "", "clusters": ["C1"], "keywords": ["embodied foundation model", "reasoning"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Embodied-R1.5 unifies embodied cognition, planning, correction and pointing in one embodied foundation model.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 12 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.01221", "title": "EndoWAM: A Grounded World-Action Model for Generalizable Endoscopic Navigation", "authors": ["Jinsong Lin", "Zikang Pan", "Wanhao Liu", "Chi Kit Ng", "Liangjing Shao", "Zihang Yu", "Ziyu Wang", "Yin Wang", "Jiaxi Wang", "Jeremy Yuen-Chun Teoh", "Zhiyong Xiong", "Huxin Gao", "Hongliang Ren"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.01221", "pdf_url": "https://arxiv.org/pdf/2608.01221", "code_url": "", "clusters": ["C8"], "keywords": ["EndoWAM", "medical robotics"], "model_class": "world-action model", "techniques": ["single denoising pass", "future grounding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Endoscopic navigation WAM coupling a lightweight DiT for future target-region prediction with a discrete action expert, running in a single denoising pass.", "relevance": 2, "relevance_reason": "Single-pass WAM for real-time control.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2609.18164", "title": "Energy-Regularized Imitation Learning for Force- and Work-Aware Robotic Manipulation", "authors": ["Toshiki Otani", "Hiromu Taketsugu", "Norimichi Ukita"], "year": 2026, "venue": "ECCV 2026 Workshop", "type": "preprint", "url": "https://arxiv.org/abs/2609.18164", "pdf_url": "https://arxiv.org/pdf/2609.18164", "code_url": "", "clusters": ["C4"], "keywords": ["mechanical work", "energy-aware imitation", "RLBench"], "model_class": "other", "techniques": [], "base_model": "RVT-2", "params": "", "hardware": "", "reported_metric": "mean mechanical work 208.8 J to 204.4 J (-2.1%), success 86.2% to 86.9%", "metric_measures": "energy", "benchmarks": ["RLBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Defines a joint-space mechanical-work proxy and uses a learned differentiable predictor as a fine-tuning regularizer to reduce motion energy.", "relevance": 2, "relevance_reason": "Motion-energy metric definition.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy energy"}
{"id": "2603.15956", "title": "ExpertGen: Scalable Sim-to-Real Expert Policy Learning from Imperfect Behavior Priors", "authors": ["Zi-Fan Xu", "Ran Gong", "M. Minniti", "K. Sivakumar", "A. Gundogdu", "Eric Rosen", "Riedana Yan", "Tushar Kusnur", "Zixing Wang", "Di Deng", "Peter Stone", "Xiaohan Zhang", "Karl Schmeckpeper"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.15956", "pdf_url": "https://arxiv.org/pdf/2603.15956", "code_url": "", "clusters": ["C4"], "keywords": ["sim-to-real", "expert policies"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Automates expert policy learning in simulation from imperfect priors for sim-to-real data generation.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.01596", "title": "Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation", "authors": ["Haoyuan Deng", "Hai-Chao Liu", "Wenkai Guo", "Yu-Rong Ling", "Zaijia Yang", "Yuanjiang Xue", "Hao-Sheng Sun", "Liang-Zi Wang", "Zi-Wei Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.01596", "pdf_url": "https://arxiv.org/pdf/2609.01596", "code_url": "", "clusters": ["C3"], "keywords": ["contact-rich", "foundation model", "RL post-training"], "model_class": "VLA", "techniques": ["lightweight actor"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Facet-0: robot foundation model for sub-millimetre contact-rich assembly with RL post-training and lightweight on-robot adaptation.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2605.08434", "title": "Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models", "authors": ["Meng Zheng", "Samhita Marri", "Anwesa Choudhuri", "Benjamin Planche", "Zhongpai Gao", "Van Nguyen Nguyen", "Terrence Chen", "Girish Chowdhary", "Ziyan Wu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.08434", "pdf_url": "https://arxiv.org/pdf/2605.08434", "code_url": "", "clusters": ["C3"], "keywords": ["AFIL", "failure-informed learning", "negative guidance"], "model_class": "VLA", "techniques": ["failure trajectories", "dual action heads"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses self-generated failure rollouts as negative guidance when fine-tuning diffusion/flow VLAs.", "relevance": 2, "relevance_reason": "Post-training without human recovery data.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:adapter AND abs:\"parameter\""}
{"id": "2609.21358", "title": "FAN: Foresight Action Normalization for Continual Adaptation of Vision-Language-Action Models", "authors": ["Yijun Hong", "Jiarun Zhu", "Xiaoquan Sun", "Le Xu", "Qijun He", "Xin Jin", "Mingqi Yuan", "Wenjun Zeng", "Jiayu Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21358", "pdf_url": "https://arxiv.org/pdf/2609.21358", "code_url": "", "clusters": ["C3"], "keywords": ["action normalization", "continual adaptation", "coordinate drift"], "model_class": "VLA", "techniques": ["normalization strategy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Shows action-normalization choice drives failures in continual VLA adaptation and proposes foresight action normalization.", "relevance": 2, "relevance_reason": "Practical recipe detail for continual adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"continual learning\""}
{"id": "2603.25661", "title": "Fast-dVLA: Accelerating Discrete Diffusion VLA to Real-Time Performance", "authors": ["Wenxuan Song", "Jiayi Chen", "Shuai Chen", "Jingbo Wang", "Pengxiang Ding", "Han Zhao", "Yikai Qin", "Xinhu Zheng", "Donglin Wang", "Yan Wang", "Haoang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.25661", "pdf_url": "https://arxiv.org/pdf/2603.25661", "code_url": "", "clusters": ["C1"], "keywords": ["discrete diffusion VLA", "capability vectors", "fine-tuning efficiency"], "model_class": "VLA", "techniques": ["parameter merging", "regularized SFT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Despite the title, the abstract describes reducing fine-tuning overhead by merging auxiliary-task 'capability vectors' into pretrained weights plus an orthogonal regularizer.", "relevance": 2, "relevance_reason": "Training efficiency; abstract does not describe inference acceleration.", "foundational_technique": false, "verified": true, "found_via": "search query: discrete diffusion VLA"}
{"id": "2605.30795", "title": "Feat2Go: Visual Feature-Grounded Value Estimation for Embodied Reinforcement Learning", "authors": ["Junyang Shu", "Zhiwei Lin", "Bingqing Wei", "Yongtao Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30795", "pdf_url": "https://arxiv.org/pdf/2605.30795", "code_url": "", "clusters": ["C3"], "keywords": ["Feat2Go", "value estimation", "progress"], "model_class": "VLA", "techniques": ["value model", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Derives progress targets from visual world-model features to train value models for VLA RL.", "relevance": 2, "relevance_reason": "Dense rewards for RL post-training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:GRPO"}
{"id": "2606.08653", "title": "FiberTune: Preserving Action-Fiber Visual Residuals in Vision-Language-Action Fine-Tuning", "authors": ["Haihao Lin", "Xiangsheng Huang", "Xiao Yang", "Weibang Zhou", "Yiqi Zhang", "Bo Yang", "Simin Zeng", "Jiawei Yang", "Zhengyang Wang", "Jiahui Du"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08653", "pdf_url": "https://arxiv.org/pdf/2606.08653", "code_url": "", "clusters": ["C3"], "keywords": ["visual residual collapse", "teacher alignment"], "model_class": "VLA", "techniques": ["representation regularization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Training-time objective preserving teacher-structured visual residuals during VLA fine-tuning, zero inference overhead.", "relevance": 2, "relevance_reason": "Fine-tuning quality recipe.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2609.32069", "title": "Find Something You Can't Do: Agentic Real-World Reinforcement Learning for Self-Improving VLA Models", "authors": ["Yuan Fang", "Zechu Li", "Haolei Tong", "Puze Liu", "Georgia Chalvatzaki"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.32069", "pdf_url": "https://arxiv.org/pdf/2609.32069", "code_url": "", "clusters": ["C3"], "keywords": ["self-improving VLA"], "model_class": "VLA", "techniques": ["real-world RL", "task selection", "self-evaluation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Agentic real-world RL that chooses what to practice from the current scene and self-evaluates outcomes, avoiding manual resets.", "relevance": 2, "relevance_reason": "Autonomous continual improvement of deployed VLAs.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (interrupt* OR \"human intervention\" OR \"mid-execution\")"}
{"id": "2608.26645", "title": "FLARE: A Failure-Aware Framework for Autonomous Correction and Recovery in Visual-Language Robotic Manipulation", "authors": ["Ganlong Zhao", "Zijia Tang", "Xingping Chen", "Zhanghui Kuang", "Ye Tian", "Guanbin Li"], "year": 2026, "venue": "CVPR 2026", "type": "paper", "url": "https://arxiv.org/abs/2608.26645", "pdf_url": "https://arxiv.org/pdf/2608.26645", "code_url": "", "clusters": ["C7"], "keywords": ["retry", "reset"], "model_class": "VLA", "techniques": ["data augmentation", "failure analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds Retry (perturbation and bridging data) and Reset (MLLM-identified OOD failures) mechanisms so VLAs recover from execution errors.", "relevance": 2, "relevance_reason": "Recovery behaviors for VLA failures.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (\"failure recovery\" OR \"error recovery\" OR \"self-correct*\")"}
{"id": "2606.16286", "title": "FlowMPC: Improving Flow Matching policies with World Models", "authors": ["Chandon Hamel"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.16286", "pdf_url": "https://arxiv.org/pdf/2606.16286", "code_url": "", "clusters": ["C8"], "keywords": ["flow matching policy", "MPPI", "TD-MPC2"], "model_class": "world model", "techniques": ["world-model MPPI over policy proposals"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Adds TD-MPC2-style world-model planning over flow-matching policy proposals, improving ManiSkill success without retraining the policy.", "relevance": 2, "relevance_reason": "Small study of test-time world-model planning.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2607.16506", "title": "Foresight Residual RL for Long-Horizon Robot Manipulation with Vision-Language-Action Models", "authors": ["Yuhan Liu", "Xinyu Zhang", "Litao Liu", "Abdeslam Boularias"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.16506", "pdf_url": "https://arxiv.org/pdf/2607.16506", "code_url": "", "clusters": ["C3"], "keywords": ["residual RL", "frozen VLA", "long-horizon assembly"], "model_class": "VLA", "techniques": ["residual RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Residual RL over a frozen VLA with foresight-value rewards for chained contact-rich subtasks.", "relevance": 2, "relevance_reason": "Residual adaptation of frozen VLAs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.06403", "title": "From Foundation to Application: Improving VLA Models in Practice", "authors": ["Wei Wu", "Fangjing Wang", "Fan Lu", "He Sun", "Shi Liu", "Yun-Nan Wang", "Yibin Yan", "Yong Wang", "Shuailei Ma", "Xinyang Wang", "Yibin Liu", "Shuai Yang", "Tianxiang Zhou", "Kejia Zhang", "Lei Zhou", "Che-Shih Su", "Nan Xue", "Bin Tan", "Han Zhang", "Youchao Zhang", "Fei Liao", "Xing Zhu", "Yujun Shen", "Kecheng Zheng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.06403", "pdf_url": "https://arxiv.org/pdf/2607.06403", "code_url": "", "clusters": ["C1"], "keywords": ["LingBot-VLA", "cross-embodiment", "mobile manipulation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["GM-100"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LingBot-VLA 2.0 scales pretraining to ~60k hours, expands the action space to whole-body DoFs and adds predictive dynamics as a proxy task.", "relevance": 2, "relevance_reason": "Context: industrial VLA foundation model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2608.03682, 2608.10232"}
{"id": "2608.23070", "title": "From Generation to Simulation: How Far Are World Models from Being True Simulators?", "authors": ["Tong Wang", "Huan Deng", "Mucheng Yang", "Yang He", "Xiaohui Kuang", "Gang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.23070", "pdf_url": "https://arxiv.org/pdf/2608.23070", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["world models as simulators", "evaluation"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Capability-based assessment of how far generative world models are from replacing simulators, noting missing state feedback and physics guarantees.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links; world-model evaluation", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2405.15223"}
{"id": "2602.01811", "title": "From Knowing to Doing Precisely: A General Self-Correction and Termination Framework for VLA models", "authors": ["Wentao Zhang", "Aolan Sun", "Wentao Mo", "Xiaoyang Qu", "Yuxin Zheng", "Jianzong Wang"], "year": 2026, "venue": "ICASSP 2026", "type": "paper", "url": "https://arxiv.org/abs/2602.01811", "pdf_url": "https://arxiv.org/pdf/2602.01811", "code_url": "", "clusters": ["C7"], "keywords": ["VLA"], "model_class": "VLA", "techniques": ["self-correction", "termination detection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free loop that refines grasp action tokens and adds conditional termination logic so VLAs stop when the task is done.", "relevance": 2, "relevance_reason": "Adds explicit termination/correction to VLA execution.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (\"failure recovery\" OR \"error recovery\" OR \"self-correct*\")"}
{"id": "2602.10719", "title": "From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving", "authors": ["Sining Ang", "Yuguang Yang", "Chenxu Dang", "Canyu Chen", "Cheng Chi", "Haiyan Liu", "Xuanyao Mao", "Jason Bao", "Xuliang", "Bingchuan Sun", "Yan Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.10719", "pdf_url": "https://arxiv.org/pdf/2602.10719", "code_url": "", "clusters": ["C1"], "keywords": ["driving"], "model_class": "VLA", "techniques": ["dual system", "trajectory scoring"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Analyzes VLM vs vision-only encoders in driving policies and runs both branches with a learned trajectory scorer to improve accuracy-cost trade-offs.", "relevance": 2, "relevance_reason": "Hybrid dual-system accuracy/cost trade-off.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]"}
{"id": "2607.00836", "title": "From World Models to World Action Models: A Concise Tutorial for Robotics", "authors": ["Xiaoxiong Zhang", "Xiong Zeng", "Wei Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.00836", "pdf_url": "https://arxiv.org/pdf/2607.00836", "code_url": "", "clusters": ["C8"], "keywords": ["tutorial", "JEPA", "Cosmos", "world models"], "model_class": "world-action model", "techniques": ["tutorial"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Short tutorial defining world models vs world action models and comparing spatial-intelligence models, JEPA and Cosmos.", "relevance": 2, "relevance_reason": "Background tutorial.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.11739", "title": "G0.5: One Autoregressive Stream for Robot Reasoning and Action", "authors": ["Yicheng Liu", "Zibin Dong", "Baijun Ye", "Tian-Yuan Yuan", "Tao Jiang", "An-Qi Yang", "Shicheng Cao", "Hao-Nan Liu", "Yue Sun", "Zi-Han Guo", "Xiao Liu", "Dong Ke", "Changxun Pan", "Chenru Wu", "Tailai Cheng", "Xiaoshu Ren", "Xinlei Zhang", "Ji Cui", "Zi-Jie Zhao", "Haoyu Zhang", "Kai-Ming Xu", "Hao Yang", "Bowen Zhang", "Jiahui Niu", "Shao-Ting Zhu", "Shiduo Zhang", "Hang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.11739", "pdf_url": "https://arxiv.org/pdf/2608.11739", "code_url": "", "clusters": ["C1"], "keywords": ["autoregressive", "reasoning"], "model_class": "VLA", "techniques": ["autoregressive VLA", "unified reasoning-action stream"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "G0.5 is a pretrained autoregressive VLA where one transformer decoder emits both reasoning and action tokens under a single objective, instead of pairing a VLM with a flow action expert.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2608.24714", "title": "GaussianWAM: Distilling Geometry and Semantics from 3D Gaussian Fields into World-Action Models", "authors": ["Zijian Zhang", "Yuqing Jiang", "Weitao Zhou", "Minglei Li", "Jinhao Zhang", "Yao Mu", "Xiaofan Li", "Hao Zhao", "Haibao Yu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.24714", "pdf_url": "https://arxiv.org/pdf/2608.24714", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["GaussianWAM"], "model_class": "world-action model", "techniques": ["representation distillation", "3D Gaussian fields"], "base_model": "FastWAM, Cosmos Policy", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills 3D Gaussian-field geometry and semantic targets into WAM observation representations during training only, adding no inference cost.", "relevance": 2, "relevance_reason": "Training-time distillation with zero inference overhead.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "10.1145/3829079", "title": "Generative AI at the Edge: A Comprehensive Survey of Architectures, Hardware and Applications", "authors": ["Mozhgan Navardi", "Yu-Zhe Fu", "Yueqian Lin", "Hai Li", "Yi-Ran Chen", "Tinoosh Mohsenin"], "year": 2026, "venue": "ACM Computing Surveys", "type": "paper", "url": "https://doi.org/10.1145/3829079", "pdf_url": "https://doi.org/10.1145/3829079", "code_url": "", "clusters": ["C5"], "keywords": ["edge AI", "survey", "hardware"], "model_class": "other", "techniques": ["survey", "edge deployment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of architectures, hardware and applications for deploying generative AI models on resource-constrained edge devices.", "relevance": 2, "relevance_reason": "foundational: edge-deployment survey connected to several core items", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2205.14135"}
{"id": "2608.06332", "title": "GeniWorld: A Generalizable Interactive World Model for Robotic Manipulation via Visual Actions", "authors": ["Chenghao Gu", "Hanyang Yu", "Jingbo Zhang", "Haitao Lin", "Wenyao Zhang", "Jinghe Wang", "Hang-Lei Jin", "Shu-Zhao Xie", "Jing-Yan Jiang", "Zhi Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.06332", "pdf_url": "https://arxiv.org/pdf/2608.06332", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["action-conditioned world model", "policy evaluation"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Interactive action-conditioned world model for manipulation that uses visual actions to generalize, intended for scalable policy learning and evaluation.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 18 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2405.15223"}
{"id": "2602.12099", "title": "GigaBrain-0.5M*: a VLA That Learns From World Model-Based Reinforcement Learning", "authors": ["Boyuan Wang", "Bo-Han Li", "Chaojun Ni", "Guan Huang", "Guosheng Zhao", "Hao Li", "Jie Li", "Jin Lv", "Jingyue Liu", "Lv Feng", "Mingming Yu", "Peng Li", "Qiuping Deng", "Tianze Liu", "Xinyu Zhou", "Xinze Chen", "Xiao-Feng Wang", "Yang Wang", "Yi-Fan Li", "Yifei Nie", "Yi-Long Li", "Yu-Kun Zhou", "Yun Ye", "Zhi-Chao Liu", "Zhengjia Zhu"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2602.12099", "pdf_url": "https://arxiv.org/pdf/2602.12099", "code_url": "", "clusters": ["C8"], "keywords": ["GigaBrain", "world model RL"], "model_class": "VLA", "techniques": ["world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA trained with reinforcement learning inside a video world model to improve anticipation beyond chunked reactive prediction.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2608.18234", "title": "GigaBrain-WBC-0.5: A Behavior World Model for Robust Humanoid Whole-Body Tracking with Environment Interaction", "authors": ["Ziyang Cheng", "Tianshu Tang", "Jinxin Lan", "Xinze Chen", "Yuhan Gong", "Zhichao Liu", "Changzhong Wu", "Yahao Mao", "Zongyan Deng", "Mingxuan Ma", "Huasen Xi", "Yilong Liu", "Yutong Wu", "Xiaofeng Wang", "Borui Zhang", "Bingyao Yu", "Yang Wang", "Yun Ye", "Guan Huang", "Xiaojie Jin", "Zheng Zhu", "Jiwen Lu"], "year": 2026, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2608.18234", "pdf_url": "https://arxiv.org/pdf/2608.18234", "code_url": "", "clusters": ["C8"], "keywords": ["humanoid", "whole-body tracking"], "model_class": "world model", "techniques": ["behavior world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Behavior world model for humanoid whole-body tracking predicting next action, state and behavior distribution; real-time cross-terrain teleoperation.", "relevance": 2, "relevance_reason": "Real-time humanoid world model control.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "2608.19066", "title": "GS-VLA: Plug-and-Play Viewpoint Canonicalization for Frozen VLA Policies via Gaussian Splatting", "authors": ["Yechan Park", "HyunJin Kim"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.19066", "pdf_url": "https://arxiv.org/pdf/2608.19066", "code_url": "", "clusters": ["C3"], "keywords": ["GS-VLA", "viewpoint canonicalization", "Gaussian splatting"], "model_class": "VLA", "techniques": ["observation-space adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Re-renders observations to the training viewpoint via Gaussian splatting so frozen VLAs handle camera shifts without retraining.", "relevance": 2, "relevance_reason": "Training-free deployment adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"catastrophic forgetting\""}
{"id": "2606.18363", "title": "Guava: Distilling Frontier VLMs into a Compact Agent through a Robotic Manipulation Harness", "authors": ["Haowen Liu", "Xirui Li", "Shaoxiong Yao", "Peng Shi", "Tianyi Zhou", "Jia-Bin Huang", "Furong Huang", "Jiayuan Mao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18363", "pdf_url": "https://arxiv.org/pdf/2606.18363", "code_url": "", "clusters": ["C2"], "keywords": ["Guava"], "model_class": "other", "techniques": ["distillation", "tool-use harness"], "base_model": "", "params": "4B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Harness for embodied tool use and distillation of manipulation ability into a 4B model from fewer than 2K simulated trajectories.", "relevance": 2, "relevance_reason": "Distilling frontier VLM agents into a compact model.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"robot polic*\") AND (offload* OR \"edge-cloud\" OR serving OR \"inference runtime\")"}
{"id": "2609.23888", "title": "HapticWAM: Distilling Imagined Touch into a World-Action Model without Inference-Time Tactile Sensing", "authors": ["Mikhail Sannikov", "Ilya Mikhalchuk", "Konstantin Gubernatorov", "Petr Kovalev", "Ogunwoye Faith Oluwatobi", "Dzmitry Tsetserukou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.23888", "pdf_url": "https://arxiv.org/pdf/2609.23888", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["tactile distillation", "WAM", "no inference-time tactile sensing"], "model_class": "world-action model", "techniques": ["teacher-student distillation of tactile contact prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Distills tactile-aware contact prediction from a teacher WAM into a student that needs no tactile sensor at deployment.", "relevance": 2, "relevance_reason": "Distillation for cheaper deployment sensing.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2602.15922 (citations)"}
{"id": "2603.05982", "title": "HarvestFlex: Strawberry Harvesting via Vision-Language-Action Policy Adaptation in the Wild", "authors": ["Ziyang Zhao", "Shuheng Wang", "Zhonghua Miao", "Ya Xiong"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.05982", "pdf_url": "https://arxiv.org/pdf/2603.05982", "code_url": "", "clusters": ["C3"], "keywords": ["agriculture", "strawberry harvesting", "LoRA vs full FT"], "model_class": "VLA", "techniques": ["LoRA", "full fine-tuning", "asynchronous inference"], "base_model": "pi0, pi0.5, WALL-OSS", "params": "", "hardware": "", "reported_metric": "74.0% success, 32.6 s/pick (pi0.5 full FT)", "metric_measures": "other", "benchmarks": ["real greenhouse"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Compares LoRA and full fine-tuning of several VLAs for in-the-wild strawberry harvesting with <4 h of demos.", "relevance": 2, "relevance_reason": "Applied fine-tuning comparison.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:LoRA"}
{"id": "2605.16056", "title": "Health-Conditioned Vision-Language-Action Models for Malfunction-Aware Robot Control", "authors": ["Huseyin Arslan", "Özgür Erkent"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.16056", "pdf_url": "https://arxiv.org/pdf/2605.16056", "code_url": "", "clusters": ["C7"], "keywords": ["health-conditioned VLA", "malfunction-aware", "actuator degradation"], "model_class": "VLA", "techniques": ["health vector conditioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Conditions a VLA on a joint health vector (angle/torque capability) so it adapts actions to degraded actuators.", "relevance": 2, "relevance_reason": "Robustness to hardware faults.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2607.12839", "title": "HeteroMosaic: Exposing and Exploiting Heterogeneous Execution Opportunities for Energy-Efficient Edge LLM Inference", "authors": ["Gregory Jun", "Wesley Pang", "E. Richter", "Mehdi Saeedi", "Aporva Amarnath", "Pallavi Ferrao", "Deming Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.12839", "pdf_url": "https://arxiv.org/pdf/2607.12839", "code_url": "", "clusters": ["C5"], "keywords": ["edge inference", "energy efficiency", "NPU"], "model_class": "other", "techniques": ["heterogeneous scheduling", "NPU/iGPU/CPU"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "HeteroMosaic schedules LLM inference across CPU, integrated GPU and NPU on unified-memory edge SoCs to improve energy efficiency.", "relevance": 2, "relevance_reason": "foundational technique for edge SoC inference", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2503.14734"}
{"id": "2602.06069", "title": "HQP: Sensitivity-Aware Hybrid Quantization and Pruning for Ultra-Low-Latency Edge AI Inference", "authors": ["Dinesh Gopalan", "Ratul Ali"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.06069", "pdf_url": "https://arxiv.org/pdf/2602.06069", "code_url": "", "clusters": ["C2", "C5"], "keywords": ["pruning then INT8", "Fisher sensitivity", "Jetson"], "model_class": "other", "techniques": ["structured pruning", "8-bit PTQ"], "base_model": "MobileNetV3, ResNet-18", "params": "", "hardware": "NVIDIA Jetson edge platforms", "reported_metric": "up to 3.12x inference speedup, 55% model size reduction", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Prunes filters under an accuracy-drop budget before INT8 quantization so the sparse model stays quantization-robust, evaluated on Jetson boards.", "relevance": 2, "relevance_reason": "Small-CNN evidence for ordered prune-then-quantize on Jetson.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q3: arXiv abs:\"hardware-agnostic\" AND abs:efficiency AND abs:metric AND abs:inference"}
{"id": "2607.01067", "title": "Human-Centric Transferable Tactile Pre-Training for Dexterous Robotic Manipulation", "authors": ["Chi Zhang", "Penglin Cai", "Ziheng Xi", "Haoqi Yuan", "Hao Luo", "Wan-Peng Zhang", "Sipeng Zheng", "Chao-Yi Xu", "Zongqing Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.01067", "pdf_url": "https://arxiv.org/pdf/2607.01067", "code_url": "", "clusters": ["C3"], "keywords": ["tactile pretraining"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Human-centric tactile pretraining transferable to dexterous robot VLAs.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2606.17833", "title": "HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning", "authors": ["Wang, Taowen", "Xie, Zikang", "Yang, Bin", "Wang, Yunheng", "Yuan, Zizhao", "Fang, Yuetong", "Feng, Yixiao", "Wang, Yichi", "Chen, Xingyu", "Chen, Haodong", "Wu, Qiwei", "Xu, Weisheng", "Chen, Lihan", "Li, Lusong", "Zeng, Zecui", "Xu, Renjing"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.17833", "pdf_url": "https://arxiv.org/pdf/2606.17833", "code_url": "", "clusters": ["C4"], "keywords": ["HumanoidArena", "humanoid", "whole-body", "Isaac Lab", "perturbations"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["HumanoidArena"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Simulation-first humanoid whole-body benchmark with seven tasks and perturbation plus cross-motion-tracker transfer evaluation.", "relevance": 2, "relevance_reason": "Humanoid benchmark.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: humanoid VLA onboard inference latency benchmark"}
{"id": "2604.07430", "title": "HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents", "authors": ["Tencent Robotics X", "HY Vision Team", ":", "Xumin Yu", "Zuyan Liu", "Ziyi Wang", "He Zhang", "Yongming Rao", "Fangfu Liu", "Yani Zhang", "Ruowen Zhao", "Oran Wang", "Yves Liang", "Haitao Lin", "Minghui Wang", "Yubo Dong", "Kevin Cheng", "Bolin Ni", "Rui Huang", "Han Hu", "Zhengyou Zhang", "Linus", "Shunyu Yao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.07430", "pdf_url": "https://arxiv.org/pdf/2604.07430", "code_url": "https://github.com/Tencent-Hunyuan/HY-Embodied", "clusters": ["C1"], "keywords": ["HY-Embodied"], "model_class": "VLA", "techniques": ["on-policy distillation", "mixture of transformers"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Embodied foundation VLM family (MoT-2B to 32B) using on-policy distillation to transfer capability to the small variant, used to train a VLA.", "relevance": 2, "relevance_reason": "Compact embodied backbone via distillation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2605.01826", "title": "Hybrid Visual Telemetry for Bandwidth-Constrained Robotic Vision: A Pilot Study with HEVC Base Video and JPEG ROI Stills", "authors": ["Natalia Trukhina", "Vadim Vashkelis"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.01826", "pdf_url": "https://arxiv.org/pdf/2605.01826", "code_url": "", "clusters": ["C6"], "keywords": ["bandwidth-constrained vision", "video + ROI telemetry"], "model_class": "other", "techniques": ["hybrid HEVC + JPEG ROI transmission"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Pilot study splitting a fixed uplink budget between low-bitrate video and selectively sent high-detail ROI stills for robotic/UAV perception.", "relevance": 2, "relevance_reason": "Bandwidth-aware transport for remote robot perception.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.09813", "title": "iMaC: Translating Actions into Motion and Contact Images for Embodied World Models", "authors": ["Zhen-Yu Wu", "Xiu-Wei Xu", "Yu-Kun Zhou", "Yifan Li", "Qiuping Deng", "Xiaofeng Wang", "Zheng Zhu", "Bingyao Yu", "Ziwei Wang", "Jiwen Lu", "Haibin Yan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.09813", "pdf_url": "https://arxiv.org/pdf/2606.09813", "code_url": "", "clusters": ["C8"], "keywords": ["action images", "embodied world model"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Represents actions as motion and contact images to condition embodied world models more expressively across embodiments.", "relevance": 2, "relevance_reason": "Context: world model (links=13).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2306.03310"}
{"id": "2606.01205", "title": "ImagineUAV: Aerial Vision-Language Navigation via World-Action Modeling and Kinodynamic Planning", "authors": ["Xuchen Liu", "Jiawei Huang", "Shihao Xia", "Bingxi Liu", "Jinqiang Cui", "Jiankun Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.01205", "pdf_url": "https://arxiv.org/pdf/2606.01205", "code_url": "", "clusters": ["C8"], "keywords": ["UAV", "VLN"], "model_class": "world-action model", "techniques": ["step distillation", "kinodynamic planning"], "base_model": "", "params": "1.3B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Aerial VLN with a latent video diffusion world model, action extractor and kinodynamic planner; step-distilled inference for real time; 1.3B parameters.", "relevance": 2, "relevance_reason": "Step distillation for real-time WM control.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=100]"}
{"id": "2606.13675", "title": "Improving Robotic Generalist Policies via Flow Reversal Steering", "authors": ["Andy Tang", "William Chen", "Andrew Wagenmaker", "Chelsea Finn", "Sergey Levine"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.13675", "pdf_url": "https://arxiv.org/pdf/2606.13675", "code_url": "", "clusters": ["C3"], "keywords": ["policy steering", "flow matching", "latent noise"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Steers flow-matching generalist policies by inverting reasonable actions to latent noise and reusing it to invoke desired behaviors.", "relevance": 2, "relevance_reason": "Context: flow policy steering (links=13).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.26588", "title": "Inference-Time Robot Behavior Steering through Physically-Aware Reconfiguration of Task-Structure", "authors": ["Yiyuan Pan", "Hanjiang Hu", "Shangtao Li", "Xusheng Luo", "Changliu Liu"], "year": 2026, "venue": "CoRL 2026", "type": "paper", "url": "https://arxiv.org/abs/2606.26588", "pdf_url": "https://arxiv.org/pdf/2606.26588", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["inference-time steering", "neural automaton policy", "ReStruct"], "model_class": "other", "techniques": ["task-structure reconfiguration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "ReStruct edits the state-machine structure of a neural-automaton visuomotor policy at test time to satisfy new user preferences while keeping plans physically feasible.", "relevance": 2, "relevance_reason": "Test-time policy steering without retraining.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "1x-neo-onboard-compute-gtc-2026", "title": "Inside 1X's Humanoid Robot Stack: Simulation, AI Training, and Onboard Compute with NVIDIA", "authors": ["1X Technologies"], "year": 2026, "venue": "1X blog", "type": "blog", "url": "https://www.1x.tech/discover/nvidia-gtc-2026", "pdf_url": "", "code_url": "", "clusters": ["C5"], "keywords": ["humanoid", "NEO", "Jetson Thor", "onboard inference"], "model_class": "VLA", "techniques": ["onboard foundation-model inference"], "base_model": "", "params": "", "hardware": "Jetson Thor (onboard); HGX B200 (training)", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": true, "tldr": "1X describes running its NEO humanoid's large neural networks onboard on Jetson Thor with cloud B200 training; no latency or power numbers given.", "relevance": 2, "relevance_reason": "Industry confirmation of onboard humanoid compute choice; little technical detail.", "foundational_technique": false, "verified": true, "found_via": "web: NVIDIA developer blog GR00T N1.6 Jetson Thor"}
{"id": "2601.02456", "title": "InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation", "authors": ["Jun-Hao Cai", "Ze-Tao Cai", "Jia-Fei Cao", "Yi-Lun Chen", "Zeyu He", "Lei Jiang", "Hang Li", "Heng-Jie Li", "Yang Li", "Yufei Liu", "Yanan Lu", "Qi Lv", "Hao-Xiang Ma", "Jiang-Miao Pang", "Yu Qiao", "Zherui Qiu", "Yan-Qing Shen", "Xu Shi", "Yang Tian", "Bo-Lun Wang", "Han-Qing Wang", "Jia-Heng Wang", "Tai Wang", "Xue-Yuan Wei", "Chao Wu", "Yiman Xie", "B. Xing", "Yu-Qiang Yang", "Yu-Yin Yang", "Qiao-Jun Yu", "Feng Yuan", "Jia Zeng", "Jing-Jing Zhang", "Sheng-Han Zhang", "Shi Zhang", "Zhuoma Zhaxi", "Bo Zhou", "Yuan-Zhen Zhou", "Yun-Song Zhou", "Hong Zhu", "Yang-Kun Zhu", "Yu-Chen Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.02456", "pdf_url": "https://arxiv.org/pdf/2601.02456", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["unified understanding generation action"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA that unifies understanding, visual generation and action for manipulation.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2501.03575"}
{"id": "2602.21736", "title": "Joint-Aligned Latent Action: Towards Scalable VLA Pretraining in the Wild", "authors": ["Hao Luo", "Ye Wang", "Wan-Peng Zhang", "Haoqi Yuan", "Yicheng Feng", "Haiweng Xu", "Sipeng Zheng", "Zongqing Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.21736", "pdf_url": "https://arxiv.org/pdf/2602.21736", "code_url": "", "clusters": ["C1"], "keywords": ["latent action", "human video"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Pretraining with jointly aligned latent actions to use in-the-wild human video for VLAs.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2603.29213", "title": "Kilohertz-Safe: A Scalable Framework for Constrained Dexterous Retargeting", "authors": ["Yinxiao Tian", "Ziyi Yang", "Zinan Zhao", "Zhen Kan"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.29213", "pdf_url": "https://arxiv.org/pdf/2603.29213", "code_url": "", "clusters": ["C7"], "keywords": ["retargeting", "teleoperation"], "model_class": "other", "techniques": ["QP", "CBF"], "base_model": "", "params": "", "hardware": "Wuji Hand", "reported_metric": "9.05 ms average latency", "metric_measures": "closed-loop control", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Convex QP retargeting with linearized constraints and CBFs for dexterous teleoperation, reaching about 9 ms average latency with over 95% frames meeting safety criteria.", "relevance": 2, "relevance_reason": "Shows CBF filtering can run inside a tight control loop.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier\" AND (delay* OR latency) AND (learn* OR neural)"}
{"id": "2607.19719", "title": "Koopman Dreamer: Spectrally Constrained Latent Dynamics for Stable World-Model Imagination", "authors": ["Jiaqi Li", "Xinglong Zhang", "Haibin Xie", "Yixing Lan", "Wei Pan", "Xin Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.19719", "pdf_url": "https://arxiv.org/pdf/2607.19719", "code_url": "", "clusters": ["C8"], "keywords": ["Koopman dynamics", "stable imagination", "spectral constraints"], "model_class": "world model", "techniques": ["rotation-scaling Koopman latent core", "low-rank bilinear action terms"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Dreamer variant with spectrally constrained Koopman-style latent dynamics that limit error growth in long imagined rollouts.", "relevance": 2, "relevance_reason": "Structured latent dynamics for stable imagination.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.17982", "title": "LAGO Policy: Latency-Aware Asynchronous Diffusion Policies with Goal-Directed Collision-Free Planning for Smooth Manipulation", "authors": ["Guowei Shi", "Xupeng Xie", "Yiming Luo", "Jian Guo", "Jun Ma", "Boyu Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.17982", "pdf_url": "https://arxiv.org/pdf/2606.17982", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["latency-aware CFG", "asynchronous diffusion policy", "collision-free"], "model_class": "diffusion/flow policy", "techniques": ["latency-aware guidance", "trajectory optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Asynchronous diffusion policy using latency-aware guidance on future actions plus trajectory optimization for smooth, collision-free execution.", "relevance": 2, "relevance_reason": "Async execution smoothness/safety.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2512.01031"}
{"id": "2604.28192", "title": "LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning", "authors": ["Hao Chen", "Jiaming Liu", "Zhonghao Yan", "Nuowei Han", "Renrui Zhang", "Chenyang Gu", "Jialin Gao", "Ziyu Guo", "Siyuan Qian", "Yinxi Wang", "Peng Jia", "Shanghang Zhang", "Pheng-Ann Heng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.28192", "pdf_url": "https://arxiv.org/pdf/2604.28192", "code_url": "", "clusters": ["C3"], "keywords": ["LaST-R1"], "model_class": "VLA", "techniques": ["latent CoT", "policy optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "RL post-training that jointly optimizes latent chain-of-thought and actions with an adaptive reasoning horizon.", "relevance": 2, "relevance_reason": "Adaptive latent reasoning under RL.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=0]"}
{"id": "2609.06079", "title": "LayerRoute: Action-Conditioned Mixture-of-Layers Routing for Vision-Language-Action Policies", "authors": ["Zheng Lu", "Haoran Liao", "Wanqi Zhong", "Yunhe Ni", "Lijie Wang", "Xingjie Fan", "Zhisheng Chen", "Yantang Qu", "Meijia Chen", "Tianyu Xin", "Zirui Song", "Yiming Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.06079", "pdf_url": "https://arxiv.org/pdf/2609.06079", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["LayerRoute"], "model_class": "VLA", "techniques": ["layer routing", "feature reuse"], "base_model": "StarVLA-pi, pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Action-conditioned router mixing cached VLM layer representations and rereading earlier action states, adding under 4% parameters.", "relevance": 2, "relevance_reason": "Better use of cached features at small cost.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\")"}
{"id": "2602.12215", "title": "LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion", "authors": ["Jiangran Lyu", "Kai Liu", "Xuheng Zhang", "Hao-Ran Liao", "Yusen Feng", "Wenxuan Zhu", "Tingrui Shen", "Jiayi Chen", "Jiazhao Zhang", "Yifei Dong", "Wen-Bo Cui", "Senmao Qi", "Shuo Wang", "Yi-Xin Zheng", "Mingqi Yan", "Xuesong Shi", "Haoran Li", "Dong-Bin Zhao", "Mingyu Liu", "Zhizheng Zhang", "Li Yi", "Yizhou Wang", "He Wang"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2602.12215", "pdf_url": "https://arxiv.org/pdf/2602.12215", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["latent dynamics", "unified world model", "robot foundation model"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Scales a latent dynamics/action robot foundation model (~1B) by ingesting heterogeneous embodied data including action-free and low-quality trajectories.", "relevance": 2, "relevance_reason": "Context: robot foundation world-action model (links=20).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2405.12213"}
{"id": "2609.23974", "title": "LEAP-NBV: Lightweight Edge Active-Perception for Foundation-Model Next-Best-View Planning", "authors": ["Boxun Hu", "Jiawei Ge", "Axel Krieger", "Peng Wang", "Tinoosh Mohsenin"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.23974", "pdf_url": "https://arxiv.org/pdf/2609.23974", "code_url": "", "clusters": ["C6"], "keywords": ["edge active perception", "foundation model", "next-best-view"], "model_class": "other", "techniques": ["lightweight edge model guiding foundation-model queries"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Lightweight edge next-best-view planner that decides viewpoints so heavy human-mesh-recovery foundation models are used efficiently on tactical edge platforms.", "relevance": 2, "relevance_reason": "Edge-side efficiency for foundation-model perception.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: offload robot foundation model"}
{"id": "2606.26095", "title": "Learning Action Priors for Cross-embodiment Robot Manipulation", "authors": ["Dong Jing", "Tianqi Zhang", "Jiaqi Liu", "Jinman Zhao", "Zelong Sun", "Li Erran Li", "Zhiwu Lu", "Mingyu Ding", "Jia-Qi Liu", "Jin-Man Zhao", "Ze-Long Sun", "Er-Ran L. Li", "Ming-Yu Ding"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.26095", "pdf_url": "https://arxiv.org/pdf/2606.26095", "code_url": "", "clusters": ["C1"], "keywords": ["cross-embodiment", "action prior"], "model_class": "VLA", "techniques": ["action pretraining", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Pretrains the VLA action module on unconditioned cross-embodiment trajectories with a lightweight flow-matching encoder-decoder before VLM alignment.", "relevance": 2, "relevance_reason": "Action-prior pretraining speeding VLA training.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=100]; snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2609.15883", "title": "Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport", "authors": ["Jaehun Shon", "Jinha Choi", "Jongwook Jeon", "Jongmin Lee"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.15883", "pdf_url": "https://arxiv.org/pdf/2609.15883", "code_url": "https://github.com/Yonsei-DILLab/OptiFlow", "clusters": ["C1"], "keywords": ["OptiFlow"], "model_class": "diffusion/flow policy", "techniques": ["one-step flow", "optimal transport distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Offline RL one-step flow policy distilled from a value-aware reference flow via state-wise entropic optimal transport.", "relevance": 2, "relevance_reason": "One-step flow policies.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=0]"}
{"id": "2606.29501", "title": "Learning Transferable Dynamics Priors from Action to World Modeling", "authors": ["Ze Huang", "Jiahui Zhang", "Hairuo Liu", "Chen-Xi Zhang", "Ran Cheng", "Li Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.29501", "pdf_url": "https://arxiv.org/pdf/2606.29501", "code_url": "", "clusters": ["C8"], "keywords": ["world model as simulator", "policy evaluation"], "model_class": "world model", "techniques": ["action-conditioned diffusion world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Pretrains a multi-view action-conditioned diffusion world model (A2World) and adapts it both into a policy-evaluation simulator and a video-action policy.", "relevance": 2, "relevance_reason": "Context: action-conditioned robot world model foundation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2609.33575"}
{"id": "2607.08182", "title": "LEEVLA: Seeing What Matters in Latent Environment Evolution for Vision-Language-Action", "authors": ["Qi Lyu", "Baicheng Liu", "Xu-Dong Wang", "Jia-Hua Dong", "Lian-Qing Liu", "Zhi Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.08182", "pdf_url": "https://arxiv.org/pdf/2607.08182", "code_url": "", "clusters": ["C1"], "keywords": ["latent environment evolution"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA emphasizing task-critical visual evidence via latent environment-evolution reasoning.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2402.15391"}
{"id": "2606.10862", "title": "LIBERO-Occ: Evaluating and Improving Vision-Language-Action Models under Scene-Induced Occlusion via Viewpoint Imagination", "authors": ["Taishan Li", "Jiwen Zhang", "Siyuan Wang", "Xuanjing Huang", "Zhongyu Wei"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.10862", "pdf_url": "https://arxiv.org/pdf/2606.10862", "code_url": "https://github.com/litsh/Libero-Occ", "clusters": ["C4"], "keywords": ["LIBERO-Occ", "occlusion", "viewpoint imagination"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Occ"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Occlusion-focused extension of LIBERO plus a method that imagines a second viewpoint; SOTA VLAs degrade substantially under scene-induced occlusion.", "relevance": 2, "relevance_reason": "LIBERO robustness variant.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2608.17600", "title": "LIBERO-VIFO: Benchmarking the Capability and Safety of Visual Cue Following in Vision-Language-Action Models", "authors": ["Zhengyan Qian", "Rui Yan", "Alex Jinpeng Wang", "Jinhui Tang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2608.17600", "pdf_url": "https://arxiv.org/pdf/2608.17600", "code_url": "", "clusters": ["C4", "C8"], "keywords": ["LIBERO-VIFO", "visual cue following", "safety"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-VIFO", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Benchmark of eight visual-cue families testing whether VLAs follow authorized cues and ignore unauthorized ones; finds VLAs often act on cues even with no language instruction.", "relevance": 2, "relevance_reason": "LIBERO-derived evaluation variant; tangential to efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query: all:LIBERO AND benchmark"}
{"id": "2604.16788", "title": "LongBench: Evaluating Robotic Manipulation Policies on Real-World Long-Horizon Tasks", "authors": ["Chen, Xueyao", "Jia, Jingkai", "Yang, Tong", "Fu, Yibo", "Li, Wei", "Zhang, Wenqiang"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2604.16788", "pdf_url": "https://arxiv.org/pdf/2604.16788", "code_url": "", "clusters": ["C4"], "keywords": ["LongBench", "real-world long-horizon", "context-dependent"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LongBench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "1000+ real-world long-horizon episodes split into fully observable and ambiguity-driven regimes to diagnose why six policies fail over long horizons.", "relevance": 2, "relevance_reason": "Real-robot benchmark design.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: CALVIN RLBench ManiSkill3"}
{"id": "2609.38154", "title": "LongLive-Plug: Once-for-All Distillation for Video Generation", "authors": ["Shuai Yang", "Luo-Zhou Wang", "Wei Huang", "Zhi-Fei Chen", "Bo-Han Zhang", "Xiao Fu", "Qian-Li Ma", "Chen-Hsuan Lin", "Wei-An Mao", "Bryan Chu", "Song Han", "Yu-Kang Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.38154", "pdf_url": "https://arxiv.org/pdf/2609.38154", "code_url": "", "clusters": ["C2"], "keywords": ["video diffusion distillation", "LoRA"], "model_class": "other", "techniques": ["distillation", "LoRA"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LongLive-Plug learns reusable distillation capabilities (e.g., fast sampling, long-video) as LoRAs on a base video diffusion model that plug into specialized models without re-distilling.", "relevance": 2, "relevance_reason": "foundational technique for accelerating video diffusion", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2501.03575, 2506.08009"}
{"id": "2608.28437", "title": "LUCID: An Agentic AI Framework on Digital-Twin in the Loop for QoS-Guaranteeing Robotic Control", "authors": ["Hyeonsu Lyu", "Minwoo Kim", "Sehyun Ryu", "Hyun Jong Yang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.28437", "pdf_url": "https://arxiv.org/pdf/2608.28437", "code_url": "", "clusters": ["C6"], "keywords": ["cloud robotics QoS", "radio resource management", "digital twin", "agentic"], "model_class": "other", "techniques": ["joint trajectory planning and RRM", "digital twin in the loop"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Agentic framework reconfiguring coupled robot trajectory planning and radio resource allocation with a digital twin to maintain uplink QoS for cloud-controlled fleets.", "relevance": 2, "relevance_reason": "Network QoS co-design for cloud robot fleets.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "2608.17386", "title": "MANIGUARD: A Benchmark and Data Suite for Specification-Grounded Safety Evaluation and Improvement of Robotic Manipulation", "authors": ["Yiyan Peng", "Philip Wang", "Simon Sinong Zhan", "Yiqi Lyu", "Zhenyang Ni", "Jixin Yan", "Fiorelli Wong", "Ruochen Jiao", "Hang Yin", "Xinyu Cao", "Huajie Shao", "Manling Li", "Ruohan Zhang", "Qi Zhu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2608.17386", "pdf_url": "https://arxiv.org/pdf/2608.17386", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["ManiGuard", "specification-grounded safety", "contact-rich"], "model_class": "VLA", "techniques": ["benchmark + safety-annotated data generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmark and data suite of 200 contact-rich household tasks with safety specified independently of task success for foundation-model manipulation.", "relevance": 2, "relevance_reason": "Safety evaluation.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: runtime monitoring robot policy"}
{"id": "2607.19343", "title": "Masked Visual Actions for Unified World Modeling", "authors": ["Hadi Alzayer", "Wenlong Huang", "Haonan Chen", "Christopher Luey", "Lvmin Zhang", "Maneesh Agrawala", "Gordon Wetzstein", "Li Fei-Fei", "Yilun Du", "Jiajun Wu", "Jia-Bin Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.19343", "pdf_url": "https://arxiv.org/pdf/2607.19343", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["masked visual actions", "unified world model", "forward/inverse dynamics"], "model_class": "world-action model", "techniques": ["pixel-space action interface"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Expresses actions as partially revealed trajectories in video so one video model serves as forward dynamics or inverse planner, finetuned on only 15 h of data.", "relevance": 2, "relevance_reason": "World-action model interface from speaker Jiajun Wu's group.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "10.3390/drones10080572", "title": "MiniUAV-VLA: A Compact Vision–Language–Action Model for Cooperative Multi-UAV Search and Elimination via MARL Expert Distillation", "authors": ["Hongwei Han", "Guanghong Gong", "Ni Li"], "year": 2026, "venue": "Drones (MDPI)", "type": "paper", "url": "https://doi.org/10.3390/drones10080572", "pdf_url": "", "code_url": "", "clusters": ["C2"], "keywords": ["compact VLA", "UAV", "MARL expert distillation"], "model_class": "VLA", "techniques": ["policy distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulated multi-UAV"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Distills a QMIX multi-agent expert into a compact centralized VLA controller for simulated multi-UAV missions.", "relevance": 2, "relevance_reason": "Compact distilled VLA, simulation only.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246"}
{"id": "2608.01127", "title": "MiniWorld: Democratizing the Training of Video World Models from Scratch", "authors": ["Yian Zhao", "Ruochong Zheng", "Hongcan Guo", "Yu Yan", "Jian Zhang", "Jie Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.01127", "pdf_url": "https://arxiv.org/pdf/2608.01127", "code_url": "", "clusters": ["C8"], "keywords": ["training from scratch", "causal video world model", "low compute"], "model_class": "world model", "techniques": ["native causal training without bidirectional pretraining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Recipe for training causal video world models from scratch cheaply, avoiding costly post-training/distillation of bidirectional video models.", "relevance": 2, "relevance_reason": "Lowers training cost of world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.25406", "title": "MMaDA-VLA: Large Diffusion Vision-Language-Action Model with Unified Multi-Modal Instruction and Generation", "authors": ["Yang Liu", "Pengxiang Ding", "Teng-Long Jiang", "Xu-Dong Wang", "Wen-Xuan Song", "Minghui Lin", "Hang Zhao", "Hong-Yin Zhang", "Zifeng Zhuang", "Wei Zhao", "Siteng Huang", "Jinkui Shi", "Dong-Lin Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.25406", "pdf_url": "https://arxiv.org/pdf/2603.25406", "code_url": "", "clusters": ["C1"], "keywords": ["discrete diffusion VLA", "unified generation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Native discrete-diffusion VLA that unifies multimodal understanding, generation and action prediction without separate autoregressive/hierarchical modules.", "relevance": 2, "relevance_reason": "Context: diffusion-language VLA (links=17).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2609.17524", "title": "Modality-Autoregressive World-Action Models", "authors": ["Adam Hung", "Bardienus P. Duisterhof", "Deva Ramanan", "Jeffrey Ichnowski"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.17524", "pdf_url": "https://arxiv.org/pdf/2609.17524", "code_url": "", "clusters": ["C8"], "keywords": ["ModAR"], "model_class": "world-action model", "techniques": ["multi-modality prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "~20x fewer training FLOPs than Flex-pi", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "WAM that autoregressively denoises several future modalities (point tracks, DINO, depth) before actions; matches a video-initialized WAM with ~20x fewer training FLOPs.", "relevance": 2, "relevance_reason": "Cheaper WAM training via better predicted modalities.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world action model*\" AND (efficien* OR real-time OR latency OR fast* OR distill*) [start=0]"}
{"id": "2609.09591", "title": "Modality-Decoupled Federated Learning for Privacy-Preserving Embodied Intelligence in 6G", "authors": ["Zhuodong Liu", "Xiangyu Li", "Chunhong Yuan", "Hongyang Du", "Bodong Shang", "Qingqing Wu", "Tony Q. S. Quek", "Mohsen Guizani"], "year": 2026, "venue": "IEEE Wireless Communications Magazine", "type": "paper", "url": "https://arxiv.org/abs/2609.09591", "pdf_url": "https://arxiv.org/pdf/2609.09591", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["6G", "privacy"], "model_class": "VLA", "techniques": ["federated learning", "modality-aware compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Modality-decoupled federated learning for VLAs over 6G that treats vision, language and action pathways differently for privacy, compression and heterogeneity.", "relevance": 2, "relevance_reason": "Communication-efficient distributed VLA adaptation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2607.14455", "title": "Motion Planning with Model-Based Diffusion via Constraint Optimization and Adaptive Scheduling", "authors": ["Zhilin He", "Bowei Li", "Jianlin Dou", "Yuner Zhang", "Changliu Liu"], "year": 2026, "venue": "IROS 2026", "type": "paper", "url": "https://arxiv.org/abs/2607.14455", "pdf_url": "https://arxiv.org/pdf/2607.14455", "code_url": "", "clusters": ["C3", "C9"], "keywords": ["model-based diffusion", "trajectory optimization", "adaptive scheduling"], "model_class": "diffusion/flow policy", "techniques": ["adaptive noise scheduling", "constraint optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Improves model-based diffusion motion planning under hard constraints and compute limits with constraint optimization and adaptive denoising schedules.", "relevance": 2, "relevance_reason": "Diffusion sampling schedule efficiency for planning.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2609.25689", "title": "MotionForge: A Data Generation Pipeline and Large-Scale Benchmark for Long-Horizon Manipulation of Dynamic Objects with Domain Shifts", "authors": ["Mo-Han Liu", "Deng-Chen Mei", "Hao-Tian Xian", "Ru-Yang Han", "Jia-Yi Sun", "Xuan-Yu Chen", "Hai-Tian Zhang", "Lu-Xi Li", "Kai-Min Mao", "Lin Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.25689", "pdf_url": "https://arxiv.org/pdf/2609.25689", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "dynamic objects", "long-horizon"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Data generation pipeline and benchmark for long-horizon manipulation of moving objects under domain shifts.", "relevance": 2, "relevance_reason": "Context: benchmark (links=14).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2608.30237", "title": "Motus2: A Self-Evolving General World Model for Dexterous Manipulation", "authors": ["Hong-Zhe Bi", "Zikun Zhou", "Yihao Tang", "Jin-Li Pang", "Shu-He Huang", "Hai-Tian Liu", "Run-Qing Wang", "Shu-He Huang", "Yicheng Wang", "Yi-Ming Cheng", "Ruowen Zhao", "Zheng-Hua Li", "Heng-Kai Tan", "Xiaolong Liu", "Jingtian Wan", "Jia-Bao Liu", "Min Zhao", "Fan Bao", "Jun Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.30237", "pdf_url": "https://arxiv.org/pdf/2608.30237", "code_url": "", "clusters": ["C8"], "keywords": ["general world model", "dexterous manipulation"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Motus2 couples prediction, action, evaluation and self-improvement in one general world model for dexterous manipulation.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 11 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2605.00475", "title": "MSACT: Multistage Spatial Alignment for Stable Low-Latency Fine Manipulation", "authors": ["Xianbo Cai", "Hideyuki Ichiwara", "Masaki Yoshikawa", "Tetsuya Ogata"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.00475", "pdf_url": "https://arxiv.org/pdf/2605.00475", "code_url": "", "clusters": ["C1"], "keywords": ["ACT", "spatial attention points", "low-latency"], "model_class": "other", "techniques": ["action chunking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adds multistage 2D attention-point extraction to ACT for stable low-latency bimanual fine manipulation.", "relevance": 2, "relevance_reason": "Low-latency ACT variant.", "foundational_technique": false, "verified": true, "found_via": "search query: efficient diffusion policy action chunk"}
{"id": "2605.31460", "title": "On-Device Robotic Planning: Eliminating Inference Redundancy for Efficient Decision-Making", "authors": ["Joonhee Lee", "Hyunseung Shin", "Hyunmi Kim", "Pei Zhang", "JeongGil Ko"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.31460", "pdf_url": "https://arxiv.org/pdf/2605.31460", "code_url": "", "clusters": ["C1", "C2", "C5", "C6"], "keywords": ["on-device planning", "temporal redundancy", "KV-steered routing", "scene gating"], "model_class": "other", "techniques": ["scene gating", "reasoning skipping", "scene-change gating", "KV-steered routing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ALFRED", "real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Skips redundant LLM/VLM reasoning across consecutive robot observations via scene gating and KV-steered routing to make on-device planning practical.", "relevance": 2, "relevance_reason": "On-device angle; mainly an inference-skipping method.", "foundational_technique": false, "verified": true, "found_via": "web: Jetson diffusion policy real-time onboard; snowball from 2602.18397"}
{"id": "2607.01651", "title": "One Demonstration Is Enough for Real-World Robotic Reinforcement Learning", "authors": ["Yuwan Liu", "Hongze Yu", "Song Liu", "Yuhan Wang", "Junge Zhang", "Yaodong Yang", "Yuanpei Chen", "Ceyao Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.01651", "pdf_url": "https://arxiv.org/pdf/2607.01651", "code_url": "", "clusters": ["C3"], "keywords": ["AutoSERL", "single demonstration", "automated intervention"], "model_class": "other", "techniques": ["real-world RL", "automated intervention"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Automates human intervention in real-world RL from a single demonstration via sliding-window guidance and safety recovery.", "relevance": 2, "relevance_reason": "Reduces human cost of real-world RL.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: HIL-SERL SERL sample-efficient real-world robot reinforcement learning fine-tuning pretrained"}
{"id": "2601.20701", "title": "One Step Is Enough: Dispersive MeanFlow Policy Optimization", "authors": ["Guowei Zou", "Haitao Wang", "Hejun Wu", "Yukun Qian", "Yuhang Wang", "Weibing Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.20701", "pdf_url": "https://arxiv.org/pdf/2601.20701", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["MeanFlow", "one-step", "RL fine-tuning"], "model_class": "diffusion/flow policy", "techniques": ["one-step generation", "RL fine-tuning", "dispersive regularization"], "base_model": "", "params": "", "hardware": "high-performance GPU", "reported_metric": ">120 Hz; 5-20x inference speedup", "metric_measures": "relative latency", "benchmarks": ["RoboMimic", "Gym", "Franka real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "One-step MeanFlow policy with RL fine-tuning, giving real-time generation without distillation.", "relevance": 2, "relevance_reason": "RL fine-tuning combined with inference efficiency; small policy.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: ti:\"policy optimization\" AND abs:diffusion AND abs:\"fine-tuning\""}
{"id": "2603.05291", "title": "Online Self-Training for Co-Adaptation in Hierarchical Diffusion Policies", "authors": ["Clemence Grislain", "Mathilde Kappel", "Olivier Sigaud", "Mohamed Chetouani"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.05291", "pdf_url": "https://arxiv.org/pdf/2603.05291", "code_url": "", "clusters": ["C3"], "keywords": ["hierarchical policy"], "model_class": "diffusion/flow policy", "techniques": ["self-training", "filtered distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Self-training for hierarchical diffusion policies that distills jointly successful planner-controller samples back into both levels.", "relevance": 2, "relevance_reason": "Stable online co-adaptation letting a small model beat a larger VLA.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]"}
{"id": "2604.02965", "title": "Open-Loop Planning, Closed-Loop Verification: Speculative Verification for VLA", "authors": ["Wang, Zihua", "Lin, Zhitao", "Li, Ruibo", "Zhang, Yu", "Yang, Xu", "Mi, Siya", "Wei, Xiu-Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.02965", "pdf_url": "https://arxiv.org/pdf/2604.02965", "code_url": "", "clusters": ["C2", "C4"], "keywords": ["speculative verification", "open-loop planning", "closed-loop verifier"], "model_class": "VLA", "techniques": ["lightweight verifier", "adaptive replanning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Heavy VLA plans chunks at low rate while a light verifier checks each step against a closed-loop reference and triggers replanning only when needed.", "relevance": 2, "relevance_reason": "C2 primary.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: closed-loop vs open-loop control frequency VLA"}
{"id": "2609.07398", "title": "OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining", "authors": ["Yu-Ran Wang", "Siqiao Huang", "Mingleyang Li", "Chen-Hao Zhang", "Jia-Qi Liang", "Wei-Yang Jin", "Yue Chen", "Xue-Min Chi", "Dong-Hao Zhou", "Qi-Ze Yu", "Yu-Kai Wang", "Yu-Han Rui", "Shen-Zhe Yao", "Zhen Yuan", "Zheng-Hao Shen", "Ke-Fei Zhu", "Zi-Jie Zhu", "Ning Gao", "Xiaowei Chi", "Guanqi He", "Shanghang Zhang", "Hao Dong", "Lin Shao", "Hang Zhao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.07398", "pdf_url": "https://arxiv.org/pdf/2609.07398", "code_url": "", "clusters": ["C8"], "keywords": ["WAM pretraining", "open stack"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open, modular research stack that factorizes world-action-model design choices and studies them in controlled experiments, releasing a pretrained WAM.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.27144", "title": "PAMAE: Phase-Aware-MoE Action Experts Towards Reliable Flow-Matching Vision-Language-Action Policies", "authors": ["Jiayu Yang", "Tao Yang", "Xiang Chang", "Fei Chao", "Changjing Shang", "Qiang Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.27144", "pdf_url": "https://arxiv.org/pdf/2606.27144", "code_url": "", "clusters": ["C1"], "keywords": ["phase-aware MoE", "flow action expert"], "model_class": "VLA", "techniques": ["mixture of experts"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Replaces a flow VLA's single action expert with a sparse phase-routed expert mixture for more reliable multi-stage manipulation.", "relevance": 2, "relevance_reason": "MoE for reliability rather than efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA mixture-of-experts"}
{"id": "2609.32634", "title": "PF-RL: Progress Field Reinforcement Learning via Goal-Conditioned Value Geometry for Vision-Language-Action Models", "authors": ["Yunpeng Qing", "Yilun Kong", "Sixu Lin", "Ming Zhou", "Yiming Fei", "Shuang Luo", "Yixiao Chi", "Haoming Gu", "Jingyuan Liu", "Changxu Wei", "Zhi Hou", "Changqing Zou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.32634", "pdf_url": "https://arxiv.org/pdf/2609.32634", "code_url": "", "clusters": ["C3"], "keywords": ["progress field", "dense credit", "RFT"], "model_class": "VLA", "techniques": ["RL fine-tuning", "value learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Learns a goal-conditioned progress representation on VLA features to give dense credit for reinforcement fine-tuning.", "relevance": 2, "relevance_reason": "RL post-training credit assignment.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.03598", "title": "PHASER: Phase-Aware and Semantic Experience Replay for Vision-Language-Action Models", "authors": ["Ziyang Chen", "Shaoguang Wang", "Weiyu Guo", "Qianyi Cai", "He Zhang", "Pengteng Li", "Yiren Zhao", "Yandong Guo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.03598", "pdf_url": "https://arxiv.org/pdf/2606.03598", "code_url": "", "clusters": ["C3"], "keywords": ["PHASER", "phase-aware replay"], "model_class": "VLA", "techniques": ["experience replay"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Phase-centric and forgetting-aware replay sampling for continual VLA learning.", "relevance": 2, "relevance_reason": "Continual adaptation efficiency of replay memory.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"continual learning\""}
{"id": "2603.09030", "title": "PlayWorld: Learning Robot World Models from Autonomous Play", "authors": ["Tenny Yin", "Zhi-Ting Mei", "Zhong Zheng", "Miyu Yamane", "David Wang", "Jade Sceats", "Samuel M. Bateman", "Li-Han Zha", "Apurva Badithela", "Olaoluwa Shorinwa", "Anirudha Majumdar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.09030", "pdf_url": "https://arxiv.org/pdf/2603.09030", "code_url": "", "clusters": ["C8"], "keywords": ["action-conditioned video model", "autonomous play"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains action-conditioned robot video world models from autonomous play data for more physically consistent simulation.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2608.23138", "title": "Pointing-VLA: Typed Spatial Grounding Interfaces for Vision-Language-Action Manipulation", "authors": ["Xiwen Chen", "Zelin Li", "Zhiruo Zhou", "Huiming Chen", "Chenwei Wang", "Xiaojun Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.23138", "pdf_url": "https://arxiv.org/pdf/2608.23138", "code_url": "", "clusters": ["C1"], "keywords": ["pointing", "VLA interface"], "model_class": "VLA", "techniques": ["typed readout heads", "spatial grounding"], "base_model": "Embodied-R1", "params": "", "hardware": "", "reported_metric": ">20x lower controller time (NORA-1.5); 6.68-6.90x faster than text decoding", "metric_measures": "relative latency", "benchmarks": ["Bridge/WidowX", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Typed hidden-state readout heads (points, affordance heatmaps, trajectories) on Embodied-R1 replace text-serialized coordinates; the readout transferred to NORA-1.5 cut controller time by over 20x.", "relevance": 2, "relevance_reason": "Removing text decoding of geometry yields large latency gains.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2605.28634", "title": "PrimitiveVLA: Learning Reusable Motion Primitives for Efficient and Generalizable Robotic Manipulation", "authors": ["Yutai Li", "Shaohui Peng", "Jiaming Guo", "Di Huang", "Zihao Zhang", "Yuxuan Guo", "Yunkai Gao", "Siming Lan", "Ling Li", "Xing Hu", "Yunji Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.28634", "pdf_url": "https://arxiv.org/pdf/2605.28634", "code_url": "", "clusters": ["C1"], "keywords": ["primitive-centric"], "model_class": "VLA", "techniques": ["motion primitives", "hierarchical planning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Decomposes demonstrations into reusable motion primitives and reassembles them with a VLM planner at inference for data-efficient VLA adaptation.", "relevance": 2, "relevance_reason": "Data efficiency of VLA adaptation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]"}
{"id": "2609.34497", "title": "QAMM: Adjoint MeanFlow Matching for Few-Step Offline Reinforcement Learning", "authors": ["Yuehu Gong", "Shutong Ding", "Mokai Pan", "Yimiao Zhou", "Jiashu Hou", "Ye Shi", "Yanwei Fu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.34497", "pdf_url": "https://arxiv.org/pdf/2609.34497", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["QAMM"], "model_class": "diffusion/flow policy", "techniques": ["MeanFlow", "adjoint matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["HumanoidMaze"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Turns critic adjoint signals into MeanFlow average-velocity supervision for few-step offline RL flow policies.", "relevance": 2, "relevance_reason": "Few-step flow policy training.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=0]"}
{"id": "2602.13052", "title": "Quantization-Aware Collaborative Inference for Large Embodied AI Models", "authors": ["Zhonghao Lyu", "Ming Xiao", "Mikael Skoglund", "Merouane Debbah", "H. Vincent Poor"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.13052", "pdf_url": "https://arxiv.org/pdf/2602.13052", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["collaborative inference", "quantization bit-width", "rate-distortion", "delay and energy constraints"], "model_class": "other", "techniques": ["quantization", "co-inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Derives rate-distortion bounds for quantized embodied model co-inference and jointly optimizes bit width and compute frequency under delay and energy constraints.", "relevance": 2, "relevance_reason": "Joint latency-energy optimization.", "foundational_technique": false, "verified": true, "found_via": "search query: embodied inference energy"}
{"id": "2605.24890", "title": "QuoVLA: Quotient Space for Vision-Language-Action Models", "authors": ["Xuan Wang", "Yi-Nan Wu", "Hao-Ran Duan", "Jun-Gong Han"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24890", "pdf_url": "https://arxiv.org/pdf/2605.24890", "code_url": "", "clusters": ["C1"], "keywords": ["quotient space"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Theory-driven VLA design that argues pretrained VLM latents already carry usable action information.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2606.17846", "title": "Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models", "authors": ["Haoqi Yuan", "Zhi-Xuan Liang", "An-Jen Chen", "Ye Wang", "Haoyang Li", "Peibin Lin", "Yi-Yang Huang", "Zi-Xing Lei", "Tong Zhang", "Jiazhao Zhang", "Jie Zhang", "Jingyang Fan", "Geng-Ze Zhou", "Qihang Peng", "Chen-Xu Lv", "Xiao-Yue Chen", "An Yang", "Fei Huang", "Junyang Lin", "Dayiheng Liu", "Jing-Ren Zhou", "Chen-Fei Wu", "Xiong-Hui Chen"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.17846", "pdf_url": "https://arxiv.org/pdf/2606.17846", "code_url": "", "clusters": ["C1"], "keywords": ["technical report", "scaling"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Technical report on scaling a manipulation foundation model by aligning heterogeneous robot data under one formulation.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2607.04434"}
{"id": "2605.30280", "title": "Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments", "authors": ["Qiuyue Wang", "Ming-Sheng Li", "Jian Guan", "Ji-lu Ye", "Sicheng Xie", "Yitao Liu", "Junhao Chen", "Zhi-Xuan Liang", "Jie Zhang", "Xintong Hu", "Xuhong Huang", "Pei Lin", "Junyang Lin", "Dayiheng Liu", "Shuai Bai", "Jing Zhou", "Jiazhao Zhang", "Haoqi Yuan", "Geng-Ze Zhou", "Hang Yin", "Ye-Bin Wang", "Yi-Yang Huang", "Zi-Xing Lei", "Wujian Peng", "Delin Chen", "Yingming Zheng", "Jingyang Fan", "Xianwei Zhuang", "Xin Zhou", "Haoyang Li", "An-Jen Chen", "Tong Zhang", "Xue-Jing Liu", "Yuchong Sun", "Rui Chen", "Zhao Li", "Chenxu Lu", "Zhi-Bo Yang", "Tao Yu", "Xiong-Hui Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.30280", "pdf_url": "https://arxiv.org/pdf/2605.30280", "code_url": "", "clusters": ["C1"], "keywords": ["unified VLA", "cross-embodiment"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Qwen-VLA: unified VLA across manipulation and navigation tasks and robot embodiments.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2605.11151", "title": "RankQ: Offline-to-Online Reinforcement Learning via Self-Supervised Action Ranking", "authors": ["Andrew Choi", "Wei Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.11151", "pdf_url": "https://arxiv.org/pdf/2605.11151", "code_url": "", "clusters": ["C3"], "keywords": ["RankQ"], "model_class": "other", "techniques": ["offline-to-online RL", "action ranking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Offline-to-online RL using self-supervised action ranking to stabilize value learning.", "relevance": 2, "relevance_reason": "RL fine-tuning efficiency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=300]"}
{"id": "2606.25206", "title": "RAVEN: Long-Horizon Reasoning & Navigation with a Visuo-Spatio-Temporal Memory", "authors": ["Yi-Xuan Hu", "Zhi-Cheng Zheng", "Li-Han Zha", "Chunwei Xing", "Rajdeep Singh", "Omar Hossain", "Antonio Loquercio", "Dhruv Shah"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.25206", "pdf_url": "https://arxiv.org/pdf/2606.25206", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["agentic memory", "vector database", "long-horizon navigation"], "model_class": "other", "techniques": ["embedding retrieval memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "matches frontier VLMs at 10x lower retrieval cost", "metric_measures": "other", "benchmarks": ["video QA benchmarks", "Unitree Go1"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Stores posed, time-stamped visual embeddings in a vector DB to answer long-horizon queries and navigate, avoiding captioning and cutting retrieval cost.", "relevance": 2, "relevance_reason": "Efficient memory for long-term robot deployment.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2608.15269", "title": "Remember Smarter: Visual History Compressor and Hyperbolic Experience Space for Robotic Memory", "authors": ["Dai Zhou", "Jiexi Yan", "Tong Li", "Yuxuan Wang", "Cheng Deng"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15269", "pdf_url": "https://arxiv.org/pdf/2608.15269", "code_url": "", "clusters": ["C2"], "keywords": ["robot memory"], "model_class": "VLA", "techniques": ["Mamba compression", "asynchronous retrieval"], "base_model": "pi0", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Plug-in memory module compressing multi-view histories with Mamba and storing experience in a hyperbolic latent space, retrieved asynchronously.", "relevance": 2, "relevance_reason": "History compression without expanding VLA context.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=0]"}
{"id": "2605.15480", "title": "Residual Reinforcement Learning for Robot Teleoperation under Stochastic Delays", "authors": ["Kaize Deng", "Zewen Yang"], "year": 2026, "venue": "IFAC World Congress 2026", "type": "paper", "url": "https://arxiv.org/abs/2605.15480", "pdf_url": "https://arxiv.org/pdf/2605.15480", "code_url": "", "clusters": ["C7"], "keywords": ["teleoperation", "stochastic delays", "residual RL", "LSTM state estimator"], "model_class": "other", "techniques": ["LSTM state reconstruction", "residual RL"], "base_model": "", "params": "", "hardware": "Franka Panda", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Combines an LSTM that reconstructs smooth states from stochastically delayed observations with residual RL torque compensation for teleoperation.", "relevance": 2, "relevance_reason": "Delay compensation with learned control.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: delay compensation reinforcement learning robot"}
{"id": "2609.33412", "title": "Resolving State-Representation Mismatch: State-Space Visual Reasoning for Open-Loop VLA Planning", "authors": ["Junhao Xiao", "Haoxiang Zhao", "Menghao Fang", "Jinkui Zhang", "Jinghan Yu", "Xinyu Huang", "Zhiyu Wu", "Kaiming Xu", "Yi Chen", "Youjun Bao", "Zhiyuan Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.33412", "pdf_url": "https://arxiv.org/pdf/2609.33412", "code_url": "", "clusters": ["C1"], "keywords": ["open-loop planning"], "model_class": "VLA", "techniques": ["recurrent state", "prefix caching"], "base_model": "Qwen2.5-VL", "params": "", "hardware": "", "reported_metric": "up to 98.58x faster Maze decoding rollouts", "metric_measures": "relative latency", "benchmarks": ["FrozenLake", "Maze", "MiniBehavior"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Encodes image and instruction once and updates a recurrent latent state per action for open-loop VLA planning, reusing a prefix cache for fast multi-step decoding on grid-world tasks.", "relevance": 2, "relevance_reason": "Prefix reuse for multi-step decoding.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2609.35469", "title": "Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching", "authors": ["Chenyu Zhang", "Yuhang Cao", "Daru Du", "Yingxi Lu", "Jing Shao", "Ruoqu Chen", "Jiajun Liu", "Liu Cao", "Yicheng Liu", "Hang Zhao", "Mengdi Xu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.35469", "pdf_url": "https://arxiv.org/pdf/2609.35469", "code_url": "", "clusters": ["C1"], "keywords": ["CATok"], "model_class": "VLA", "techniques": ["action tokenization", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Causal action tokenizer built by progressively annealing a flow-matching process, giving coarse-to-fine tokens aligned with autoregressive VLA backbones; claims better fidelity-compression and inference efficiency.", "relevance": 2, "relevance_reason": "Action tokenization affecting AR-VLA efficiency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2605.19282", "title": "Rethinking Muon Beyond Pretraining: Spectral Failures and High-Pass Remedies for VLA and RLVR", "authors": ["Chongyu Fan", "Gaowen Liu", "Mingyi Hong", "Ramana Rao Kompella", "Sijia Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.19282", "pdf_url": "https://arxiv.org/pdf/2605.19282", "code_url": "", "clusters": ["C3"], "keywords": ["Pion", "Muon optimizer", "VLA training"], "model_class": "VLA", "techniques": ["optimizer design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Shows Muon's spectral whitening hurts VLA and RLVR training and proposes a high-pass variant at similar cost.", "relevance": 2, "relevance_reason": "Optimizer efficiency for VLA training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:GRPO"}
{"id": "2609.38163", "title": "Rethinking Representations for World-Action Modeling", "authors": ["Hao-Yi Jiang", "Liu Liu", "Xin-Jiang Wang", "Zhi-Hao Sun", "Ze-Qun Chen", "Sen Wang", "Xin-Jie Wang", "Xia Chen", "Jing-Feng Yao", "Wei-Heng Zhao", "Shang-Lin Yuan", "Zhi-Zhong Su", "Wei Sui", "Wen-Yu Liu", "Xing-Gang Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.38163", "pdf_url": "https://arxiv.org/pdf/2609.38163", "code_url": "", "clusters": ["C8"], "keywords": ["DINO features", "compact world state"], "model_class": "world-action model", "techniques": ["representation bottleneck"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Controlled study of representation spaces for world-action models leading to ReWAM, which uses calibrated DINO features and a compact temporal bottleneck instead of generative video pretraining.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 21 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2503.00200"}
{"id": "2608.27033", "title": "Riemann-1.0: An Embodied World Action Model for Physical AI", "authors": ["Hao Sun", "Jiangbo Pei", "Fei Kang", "Ze-Xiang Liu", "Yao-Kun Li", "Bo-Yi Jiang", "Hua Xue", "Cindy Zhou", "Wei Li", "Yichen Wei", "Mengyin An", "Fan-Liang Zhao", "Biao Jiang", "Zi-Le Wang", "Yang Liu", "Yangguang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.27033", "pdf_url": "https://arxiv.org/pdf/2608.27033", "code_url": "", "clusters": ["C8"], "keywords": ["causal autoregressive WAM"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "LIBERO", "RoboCasa-365"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Riemann-1.0: causal autoregressive world-action model jointly modeling multi-view observations, states and actions.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2402.15391"}
{"id": "2605.11564", "title": "RIO: Flexible Real-Time Robot I/O for Cross-Embodiment Robot Learning", "authors": ["Pablo Ortega-Kral", "Eliot Xing", "A. Bucker", "Vernon Luk", "Junseo Kim", "O. Kwon", "Ang-Chen Xie", "Nikhil Sobanbabu", "Yifu Yuan", "Megan Lee", "Deepam Ameria", "Bhaswanth Ayapilla", "Jaycie Bussell", "Guan-Ya Shi", "Jonathan Francis", "Jean Oh", "Arthur Bucker", "Owen Kwon", "Angchen Xie"], "year": 2026, "venue": "arXiv", "type": "repo", "url": "https://arxiv.org/abs/2605.11564", "pdf_url": "https://arxiv.org/pdf/2605.11564", "code_url": "", "clusters": ["C5", "C7", "C9"], "keywords": ["RIO", "robot I/O", "real-time control framework", "policy deployment", "real-time robot I/O", "cross-embodiment"], "model_class": "other", "techniques": ["open-source deployment framework", "real-time I/O middleware"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open-source Python framework with lightweight real-time robot control, teleoperation, sensor and policy deployment components across embodiments.", "relevance": 2, "relevance_reason": "Deployment infrastructure for timing-sensitive policy execution.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2); venue sweep: arXiv API ti:real-time AND VLA AND humanoid"}
{"id": "2602.11075", "title": "RISE: Self-Improving Robot Policy with Compositional World Model", "authors": ["Jia-Zhi Yang", "Kun-Li Channing Lin", "Jinwei Li", "Wencong Zhang", "Tian-Wei Lin", "Long-Yan Wu", "Zhi-Zhong Su", "Hao Zhao", "Ya-Qin Zhang", "Li Chen", "Ping Luo", "Xiang-Yu Yue", "Hongyang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.11075", "pdf_url": "https://arxiv.org/pdf/2602.11075", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["compositional world model", "self-improvement", "imagined RL"], "model_class": "world model", "techniques": ["world-model-based RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses a compositional world model as an imagination environment for RL-based self-improvement of VLA policies, avoiding costly physical on-policy RL.", "relevance": 2, "relevance_reason": "Context: world-model RL post-training for VLAs (links=22).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2609.25889", "title": "Risk-Aware Online Conformal State Probing", "authors": ["Pietro Talli", "P. Popovski", "Osvaldo Simeone"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.25889", "pdf_url": "https://arxiv.org/pdf/2609.25889", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["online conformal prediction", "state probing", "remote control", "worst-case reliability"], "model_class": "other", "techniques": ["online conformal state probing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Remote decision maker jointly decides actions and when to probe the robot state, with online conformal guarantees on missed-query error without distributional assumptions.", "relevance": 2, "relevance_reason": "Reliability guarantees for remotely hosted controllers with stale state.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2605.03821", "title": "RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models", "authors": ["Hao Wu", "Yuqi Li", "Yuan Gao", "Fan Xu", "Fan Zhang", "Kun Wang", "Penghao Zhao", "Qiufeng Wang", "Yizhou Zhao", "Weiyan Wang", "Yingli Tian", "Xian Wu", "Xiaomeng Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.03821", "pdf_url": "https://arxiv.org/pdf/2605.03821", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["RoboAlign-R1"], "model_class": "world model", "techniques": ["reward distillation", "RL post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reward-aligned post-training of robot video world models with a distilled lightweight judge and sliding-window re-encoding for long rollouts.", "relevance": 2, "relevance_reason": "Distilled reward model for efficient WM post-training.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "2603.01517", "title": "RoboGPU: Accelerating GPU Collision Detection for Robotics", "authors": ["Lufei Liu", "Li Xue", "Yuan-Hsi Chou", "Jocelyn Zhao", "Lara Kawasme", "Y. Mohammed", "T. Aamodt"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.01517", "pdf_url": "https://arxiv.org/pdf/2603.01517", "code_url": "", "clusters": ["C5"], "keywords": ["robotics GPU", "collision detection accelerator", "ray tracing accelerator", "Tesla AI5"], "model_class": "other", "techniques": ["embedded accelerator block in GPU"], "base_model": "", "params": "", "hardware": "", "reported_metric": "RoboCore 2.8x faster collision queries than RTA with 48% less energy and 2% more area", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proposes a robotics-focused GPU with an embedded collision-detection block, arguing neural motion planners need dedicated support for trajectory checking.", "relevance": 2, "relevance_reason": "Robotics GPU architecture; not foundation-model inference directly.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2407.04292 (citations)"}
{"id": "2607.18060", "title": "RoboHarness: Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning", "authors": ["Jinbang Huang", "Yuan Hu", "Zhi-Yuan Li", "Ran Qi", "Yi-Xin Xiao", "Zhan-Guang Zhang", "Mark Coates", "Tong-Tong Cao", "Ying-Xue Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.18060", "pdf_url": "https://arxiv.org/pdf/2607.18060", "code_url": "", "clusters": ["C1"], "keywords": ["policy orchestration", "memory"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Orchestrates heterogeneous robot policies with memory-driven planning for long-horizon tasks.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2607.18709", "title": "RoboInter1.5: A Holistic Intermediate Representation Suite for Embodied World Modeling and Robotic Manipulation", "authors": ["Ziqin Wang", "Hao Li", "Weijun Wang", "Jun-Hao Cai", "Jia Zeng", "Yi-Lun Chen", "Jiang-Miao Pang", "Si Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.18709", "pdf_url": "https://arxiv.org/pdf/2607.18709", "code_url": "", "clusters": ["C8"], "keywords": ["dataset", "annotations"], "model_class": "other", "techniques": ["dataset", "intermediate representations"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RoboInter1.5 provides a large annotated suite of intermediate representations for manipulation and embodied world modeling.", "relevance": 2, "relevance_reason": "context: dataset used by VLA/world-model work", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2411.19650"}
{"id": "2605.10921", "title": "RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark", "authors": ["Hua-Shuo Lei", "Wen-Xuan Song", "Hua-Rui Zhang", "Jie-Yuan Pei", "Jiayi Chen", "Hao-Dong Yan", "Han Zhao", "Pengxiang Ding", "Zhi-Peng Zhang", "Lida Huang", "Dong-Lin Wang", "Yan Wang", "Hao-Ang Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.10921", "pdf_url": "https://arxiv.org/pdf/2605.10921", "code_url": "", "clusters": ["C4"], "keywords": ["memory benchmark"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Robotic memory benchmark with multimodal annotations for long-horizon partially observable tasks.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2603.02115", "title": "Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons", "authors": ["Anthony Liang", "Yigit Korkmaz", "Jiahui Zhang", "Minyoung Hwang", "Abrar Anwar", "Sid Kaushik", "Adit Shah", "Alex Huang", "Luke S. Zettlemoyer", "Dieter Fox", "Yu Xiang", "Anqi Li", "Andreea Bobu", "Abhishek Gupta", "Stephen Tu", "Erdem Biyik", "Jesse Zhang"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2603.02115", "pdf_url": "https://arxiv.org/pdf/2603.02115", "code_url": "", "clusters": ["C3"], "keywords": ["reward model", "trajectory comparison", "progress estimation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "General-purpose robot reward model trained with trajectory-comparison supervision so it can learn from failed and suboptimal data at scale.", "relevance": 2, "relevance_reason": "Context: robot foundation reward model used for policy improvement (links=23).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2306.03310"}
{"id": "2603.04639", "title": "RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies", "authors": ["Yin-Pei Dai", "H. Fu", "Jayjun Lee", "Yue-Jiang Liu", "Haoran Zhang", "Jianing Yang", "Chelsea Finn", "Nima Fazeli", "Joyce Chai"], "year": 2026, "venue": "ICML 2026", "type": "benchmark", "url": "https://arxiv.org/abs/2603.04639", "pdf_url": "https://arxiv.org/pdf/2603.04639", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["memory benchmark", "history-dependent tasks", "VLA memory"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboMME"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Standardized benchmark for memory mechanisms in generalist robot policies (counting, occlusion, history-dependent tasks).", "relevance": 2, "relevance_reason": "Benchmark used to study memory/compute trade-offs (e.g., by DIRECT).", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Stanford AI Lab blog (ICML 2026)"}
{"id": "2604.05226", "title": "RoboPlayground: Democratizing Robotic Evaluation through Structured Physical Domains", "authors": ["Yi Ru Wang", "Carter Ung", "Evan Gubarev", "Christopher Tan", "Siddhartha Srinivasa", "Dieter Fox"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2604.05226", "pdf_url": "https://arxiv.org/pdf/2604.05226", "code_url": "", "clusters": ["C4"], "keywords": ["RoboPlayground", "language-authored tasks", "task families"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboPlayground"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Lets users author reproducible manipulation task families in natural language compiled to explicit specs and success predicates, exposing generalization failures hidden by fixed benchmarks.", "relevance": 2, "relevance_reason": "Evaluation methodology.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2603.05504", "title": "RoboPocket: Improve Robot Policies Instantly with Your Phone", "authors": ["Junjie Fang", "Wendi Chen", "Han Xue", "Fangyuan Zhou", "Tian Le", "Yi Wang", "Yuting Zhang", "Jun Lv", "Chuan Wen", "Cewu Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.05504", "pdf_url": "https://arxiv.org/pdf/2603.05504", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["remote inference", "smartphone", "AR foresight", "online finetuning"], "model_class": "diffusion/flow policy", "techniques": ["remote policy inference", "asynchronous online finetuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Phone-based data collection where a remote server runs the policy and shows its predicted trajectory in AR, with asynchronous online finetuning, enabling robot-free policy iteration.", "relevance": 2, "relevance_reason": "Remote inference infrastructure for policy iteration.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: remote inference robot"}
{"id": "2607.06699", "title": "RoboSnap: One-Shot Real-to-Sim Scene Generation for Generalizable Robot Learning and Evaluation", "authors": ["Shujie Zhang", "Jingkun Yi", "Weipeng Zhong", "Zirui Zhou", "Yangkun Zhu", "Hanqing Wang", "Xudong Xu", "Weinan Zhang", "Chunhua Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.06699", "pdf_url": "https://arxiv.org/pdf/2607.06699", "code_url": "", "clusters": ["C4"], "keywords": ["RoboSnap", "real-to-sim", "Gaussian splatting", "DROID-Sim"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DROID-Sim", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Turns a single image into a simulation-ready scene with physics-stable foreground and splatted background, giving sim-real correlated evaluation; releases 564 DROID scenes.", "relevance": 2, "relevance_reason": "Real-to-sim evaluation infrastructure.", "foundational_technique": false, "verified": true, "found_via": "search query: policy evaluation real robot scalable"}
{"id": "2609.26292", "title": "RoboTwin-Phys: Do WAMs and VLAs Understand the Physical World?", "authors": ["Jiaqi Zhang", "Feng Ye", "Mingjia Yang", "Zhihong Chen", "Mingkang Xiang", "Xinglin Yao", "Yanbin Li", "Siwei Ma", "Chuanmin Jia"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2609.26292", "pdf_url": "https://arxiv.org/pdf/2609.26292", "code_url": "", "clusters": ["C4"], "keywords": ["RoboTwin-Phys", "physical parameters", "mass", "friction"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin-Phys"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Varies 13 physical attributes continuously in RoboTwin to test WAM and VLA robustness to physics shifts; large robustness gaps appear.", "relevance": 2, "relevance_reason": "Benchmark variant.", "foundational_technique": false, "verified": true, "found_via": "search query: robot policy evaluation world model"}
{"id": "2604.19092", "title": "RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation", "authors": ["Feng Jiang", "Yang Chen", "Kyle Xu", "Yuchen Liu", "Haifeng Wang", "Zhenhao Shen", "Jasper Lu", "Shengze Huang", "Yuanfei Wang", "Chen Xie", "Ruihai Wu"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2604.19092", "pdf_url": "https://arxiv.org/pdf/2604.19092", "code_url": "", "clusters": ["C3", "C4"], "keywords": ["RoboWM-Bench", "world model evaluation", "executability"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboWM-Bench"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Converts world-model-generated manipulation videos into actions and executes them in real-to-sim scenes to test physical executability; visual plausibility and executability diverge.", "relevance": 2, "relevance_reason": "Benchmark for world models used in embodied pipelines.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2608.23320", "title": "ROS2SmolVLA: Enabling Small Vision-Language-Action Models for Integration into Industrial-Grade Lightweight Robots", "authors": ["Nils Mandischer", "Noah Böckmann", "Ludwig Holl", "Lars Mikelsons"], "year": 2026, "venue": "IFSM 2026", "type": "paper", "url": "https://arxiv.org/abs/2608.23320", "pdf_url": "https://arxiv.org/pdf/2608.23320", "code_url": "", "clusters": ["C1"], "keywords": ["SmolVLA", "ROS 2", "industrial robot"], "model_class": "VLA", "techniques": ["deployment integration"], "base_model": "SmolVLA", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["UR10e real robot"], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Open-source ROS 2 interface running SmolVLA on premise with a UR10e industrial arm.", "relevance": 2, "relevance_reason": "Deployment integration of small VLA.", "foundational_technique": false, "verified": true, "found_via": "search query: lightweight VLA"}
{"id": "2602.00743", "title": "SA-VLA: Spatially-Aware Flow-Matching for Vision-Language-Action Reinforcement Learning", "authors": ["Xu Pan", "Zhenglin Wan", "Xingrui Yu", "Xianwei Zheng", "Youkai Ke", "Ming Sun", "Rui Wang", "Ziwei Wang", "Ivor Tsang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.00743", "pdf_url": "https://arxiv.org/pdf/2602.00743", "code_url": "", "clusters": ["C3"], "keywords": ["SA-VLA"], "model_class": "VLA", "techniques": ["RL fine-tuning", "dense rewards"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Spatially-aware RL adaptation for flow-matching VLAs using spatial representations, geometric dense rewards and annealed exploration.", "relevance": 2, "relevance_reason": "Stable RL fine-tuning of VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=100]"}
{"id": "2608.09303", "title": "SAFE-CHEM: Uncertainty-Aware Policy Switching for Robust Robotic Chemistry", "authors": ["Laura Jones", "S. Shahzad", "Ayesha Sana", "Gabriella Pizzuto"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.09303", "pdf_url": "https://arxiv.org/pdf/2608.09303", "code_url": "", "clusters": ["C7"], "keywords": ["SAFE-CHEM", "uncertainty-aware policy switching", "ensemble", "lab automation"], "model_class": "other", "techniques": ["ensemble uncertainty", "policy switching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Switches away from a learned chemistry-lab policy when ensemble uncertainty signals unsafe extrapolation.", "relevance": 2, "relevance_reason": "Uncertainty-triggered fallback.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2510.09459 (S2)"}
{"id": "2503.07885", "title": "Safety Guardrails for LLM-Enabled Robots", "authors": ["Zachary Ravichandran", "Alexander Robey", "Vijay Kumar", "George J. Pappas", "Hamed Hassani"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.07885", "pdf_url": "https://arxiv.org/pdf/2503.07885", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["RoboGuard", "LLM guardrails", "temporal logic synthesis"], "model_class": "other", "techniques": ["root-of-trust LLM", "temporal logic control synthesis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-stage guardrail that grounds safety rules into temporal-logic specs with a protected LLM and repairs unsafe LLM plans via control synthesis, resisting jailbreaks.", "relevance": 2, "relevance_reason": "Plan-level safety guardrail for foundation-model robots.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: RoboGuard"}
{"id": "2603.13300", "title": "Safety-Guided Flow (SGF): A Unified Framework for Negative Guidance in Safe Generation", "authors": ["Mingyu Kim", "Young-Heon Kim", "Mijung Park"], "year": 2026, "venue": "ICLR 2026", "type": "paper", "url": "https://arxiv.org/abs/2603.13300", "pdf_url": "https://arxiv.org/pdf/2603.13300", "code_url": "https://github.com/MingyuKim87/SGF", "clusters": ["C7"], "keywords": ["safe generation", "flow models"], "model_class": "other", "techniques": ["negative guidance", "CBF analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Unifies negative-guidance safety methods for diffusion/flow models with an MMD potential and uses CBF analysis to identify a critical time window when safety guidance matters.", "relevance": 2, "relevance_reason": "Theory on when guidance-based safety needs to act during sampling.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"control barrier function\" AND (diffusion OR \"action chunk*\" OR \"vision-language-action\")"}
{"id": "2606.10305", "title": "SARM2: Multi-Task Stage Aware Reward Modeling for Self Improving Robotic Manipulation", "authors": ["Qianzhong Chen", "Hau Zheng", "Justin Yu", "Suning Huang", "Jiankai Sun", "Ken Goldberg", "Chuan Wen", "Pieter Abbeel", "Yide Shentu", "Philipp Wu", "Mac Schwager"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.10305", "pdf_url": "https://arxiv.org/pdf/2606.10305", "code_url": "", "clusters": ["C3"], "keywords": ["SARM2"], "model_class": "VLA", "techniques": ["reward modeling", "on-policy improvement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Multi-task stage-aware dense reward model (primitive stage estimator plus MMoE value head) enabling on-policy VLA self-improvement from cheap rollouts.", "relevance": 2, "relevance_reason": "Dense rewards for efficient VLA RL.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=0]"}
{"id": "2606.27375", "title": "Scalable Behavior Cloning with Open Data, Training, and Evaluation", "authors": ["Arthur Allshire", "Himanshu Gaurav Singh", "Ritvik Singh", "Adam Rashid", "Hongsuk Choi", "David McAllister", "Justin Yu", "Yiyuan Chen", "Huang Huang", "Pieter Abbeel", "Xi Chen", "Rocky Duan", "Phillip Isola", "Jitendra Malik", "Fred Shentu", "Guanya Shi", "Philipp Wu", "Angjoo Kanazawa"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.27375", "pdf_url": "https://arxiv.org/pdf/2606.27375", "code_url": "", "clusters": ["C4"], "keywords": ["ABC", "open stack", "sim-real correlated evaluation", "behavior cloning"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ABC sim", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Open BC stack with a 3,500-hour teleop dataset and a co-training recipe giving correlated sim and real evaluation for ablating design choices.", "relevance": 2, "relevance_reason": "Reproducible training/evaluation toolkit.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA evaluation simulation real correlation"}
{"id": "2609.22085", "title": "SeeQ: Training Generalist Value Functions for Long-Horizon Robotic Manipulation", "authors": ["Saksham Singh", "Zheyuan Hu", "Max Sobol Mark", "Jeffrey Yu", "Zackory Erickson", "Aviral Kumar"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.22085", "pdf_url": "https://arxiv.org/pdf/2609.22085", "code_url": "", "clusters": ["C3"], "keywords": ["SeeQ", "generalist value function", "long-horizon"], "model_class": "VLA", "techniques": ["Q-function", "action ranking"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Trains generalist Q-functions to rank candidate actions and guide generalist policies on long-horizon tasks.", "relevance": 2, "relevance_reason": "Value-guided improvement without policy retraining.", "foundational_technique": false, "verified": true, "found_via": "arXiv API title lookup (seed snowball)"}
{"id": "2604.13542", "title": "Self-adaptive Multi-Access Edge Architectures: A Robotics Case", "authors": ["Mahyar T Moghaddam", "Joakim Leed", "Anders Frandsen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.13542", "pdf_url": "https://arxiv.org/pdf/2604.13542", "code_url": "", "clusters": ["C6"], "keywords": ["multi-access edge", "self-adaptive architecture", "offloading"], "model_class": "other", "techniques": ["adaptive scaling", "computation offloading"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Self-adaptive MEC architecture with agents that scale infrastructure and offload computation for a mixed human-robot environment.", "relevance": 2, "relevance_reason": "Edge offloading architecture for robotics.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2602.21633", "title": "Self-Correcting VLA: Online Action Refinement via Sparse World Imagination", "authors": ["Chenyv Liu", "Wentao Tan", "Lei Zhu", "Fengling Li", "Jingjing Li", "Guoli Yang", "Heng Tao Shen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.21633", "pdf_url": "https://arxiv.org/pdf/2602.21633", "code_url": "https://github.com/Kisaragi0/SC-VLA", "clusters": ["C7", "C8"], "keywords": ["world action model", "self-correction"], "model_class": "VLA", "techniques": ["predictive heads", "online refinement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Auxiliary predictive heads forecast task progress and trajectory trends to drive online action refinement in a VLA.", "relevance": 2, "relevance_reason": "Self-correction from sparse imagination.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"vision-language-action\" AND (\"failure recovery\" OR \"error recovery\" OR \"self-correct*\")"}
{"id": "2608.26659", "title": "Simple Actors and Deep Critics for Scalable Reinforcement Learning", "authors": ["Guhyeon Kang", "Jaehwi Lee", "Minhae Kwon"], "year": 2026, "venue": "CIKM 2026", "type": "paper", "url": "https://arxiv.org/abs/2608.26659", "pdf_url": "https://arxiv.org/pdf/2608.26659", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["LAC"], "model_class": "other", "techniques": ["deep critic", "light actor"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 4x lower inference latency vs diffusion/flow baselines", "metric_measures": "relative latency", "benchmarks": ["OGBench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Offline RL recipe putting capacity in a deep critic (residual MLP, n-step targets, categorical loss) with a light deterministic actor, avoiding multi-step generative actors.", "relevance": 2, "relevance_reason": "Inference-cheap actors without distillation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"flow policy\") AND (\"one-step\" OR \"few-step\" OR consistency) [start=0]"}
{"id": "2606.08278", "title": "SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation", "authors": ["Songlin Wei", "Zhe-Yuan Ni", "Jie Liu", "Zhen-Yu Zhao", "Jun-Jie Ye", "Hongyi Jing", "Junkai Xia", "Xia-Wei Liu", "Michael Leong", "Liang Heng", "Dingtao Huang", "Yue Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08278", "pdf_url": "https://arxiv.org/pdf/2606.08278", "code_url": "", "clusters": ["C4"], "keywords": ["humanoid", "benchmark"], "model_class": "other", "techniques": ["simulation benchmark"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SIMPLE is a simulation testbed for training and reproducibly evaluating humanoid whole-body loco-manipulation policies.", "relevance": 2, "relevance_reason": "context: humanoid policy benchmark", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2607.08354", "title": "SkillPlug: Unsupervised Skill Mining for Few-Shot Adaptation in Robotic Manipulation", "authors": ["Zi-han Ding", "Ziwei Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.08354", "pdf_url": "https://arxiv.org/pdf/2607.08354", "code_url": "", "clusters": ["C3"], "keywords": ["skill mining", "few-shot adaptation", "plug-in"], "model_class": "VLA", "techniques": ["skill library", "few-shot fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Mines a reusable skill library and specializes a policy to new tasks from a handful of demos via a skill-conditioning plug-in.", "relevance": 2, "relevance_reason": "Few-shot adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.16788", "title": "SoK: Security and Privacy of Foundation-Model-Powered Robots", "authors": ["Xueluan Gong", "Chen Chen", "Jinxin Liu", "Qian Wang", "Kwok-Yan Lam"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.16788", "pdf_url": "https://arxiv.org/pdf/2606.16788", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["security", "privacy", "foundation-model robots", "SoK"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Systematization of security and privacy risks of foundation-model-powered robots across models, execution pipelines and ecosystems, relevant to cloud-offloaded perception data.", "relevance": 2, "relevance_reason": "Privacy/security considerations of cloud vs onboard deployment.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2601.06474", "title": "SparseOccVLA: Bridging Occupancy and Vision-Language Models via Sparse Queries for Unified 4D Scene Understanding and Planning", "authors": ["Chenxu Dang", "Jie Wang", "Guang Li", "Zhiwen Hou", "Zihan You", "Hangjun Ye", "Jie Ma", "Long Chen", "Yan Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.06474", "pdf_url": "https://arxiv.org/pdf/2601.06474", "code_url": "", "clusters": ["C1"], "keywords": ["occupancy", "driving VLA"], "model_class": "VLA", "techniques": ["sparse queries", "token reduction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["OmniDrive-nuScenes", "Occ3D-nuScenes"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA using sparse occupancy queries as the only vision-language bridge to avoid token explosion, plus an anchor-diffusion planner.", "relevance": 2, "relevance_reason": "Compact scene tokens for driving VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]"}
{"id": "2604.11757", "title": "StarVLA-α: Reducing Complexity in Vision-Language-Action Systems", "authors": ["Ji-lu Ye", "Ning Gao", "Sen-Qiao Yang", "Jinliang Zheng", "Zixuan Wang", "Yuxin Chen", "Pengguang Chen", "Yi-Lun Chen", "Shu Liu", "Jiaya Jia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.11757", "pdf_url": "https://arxiv.org/pdf/2604.11757", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["baseline", "design choices", "controlled study"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A deliberately simple VLA baseline used to study design choices under controlled conditions, reducing architectural and pipeline complexity.", "relevance": 2, "relevance_reason": "Context: widely referenced simplified VLA baseline (links=19).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2604.05014", "title": "StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing", "authors": ["S. Community"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.05014", "pdf_url": "https://arxiv.org/pdf/2604.05014", "code_url": "https://github.com/starVLA/starVLA", "clusters": ["C4"], "keywords": ["codebase", "reproducibility", "VLA toolkit"], "model_class": "VLA", "techniques": ["modular codebase"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "StarVLA is a modular, Lego-style open codebase that puts many VLA architectures and world-model variants behind shared interfaces and evaluation protocols for controlled comparison.", "relevance": 2, "relevance_reason": "context: shared VLA codebase/evaluation infrastructure that enables fair comparisons", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2511.19433, 2602.01166"}
{"id": "2608.15139", "title": "StructRL: Structured Action-Space Exploration for Flow-Based VLAs", "authors": ["Jiarui Yang", "Bin Zhu", "Jingjing Chen", "Na Zou", "Yanwei Fu", "Jianggang Zhu", "Yu-Gang Jiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15139", "pdf_url": "https://arxiv.org/pdf/2608.15139", "code_url": "", "clusters": ["C3"], "keywords": ["structured exploration", "flow VLA RL", "noise dilution"], "model_class": "VLA", "techniques": ["online RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Moves RL stochasticity for flow VLAs to action space with structured noise to avoid dilution by denoising.", "relevance": 2, "relevance_reason": "RL fine-tuning efficiency.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.24448", "title": "Supervise What Survives: Geometry-Guided VLA Adaptation from Synthetic Robot Videos", "authors": ["Danze Chen", "Yanzhe Chen", "Qiming Huang", "Zhijun Cao", "Chen Gao", "Mike Zheng Shou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.24448", "pdf_url": "https://arxiv.org/pdf/2606.24448", "code_url": "", "clusters": ["C3"], "keywords": ["synthetic robot videos", "geometry supervision"], "model_class": "VLA", "techniques": ["synthetic-data adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Supervises only the geometry that survives video generation when adapting VLAs with synthetic human-to-robot videos.", "relevance": 2, "relevance_reason": "Reduces teleop data for adaptation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2605.07288", "title": "Sword: Style-Robust World Models as Simulators via Dynamic Latent Bootstrapping for VLA Policy Post-Training", "authors": ["Jiaxuan Gao", "Yongjian Guo", "Zhong Guan", "Wen Huang", "Wanlun Ma", "Xi Xiao", "Junwu Xiong", "Sheng Wen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.07288", "pdf_url": "https://arxiv.org/pdf/2605.07288", "code_url": "", "clusters": ["C3"], "keywords": ["Sword", "style-robust world model", "latent bootstrapping"], "model_class": "VLA", "techniques": ["world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Style-robust world-model simulator reducing hallucination for VLA RL post-training on LIBERO.", "relevance": 2, "relevance_reason": "World-model-based post-training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"world model\" AND abs:\"post-training\""}
{"id": "2605.08190", "title": "Synergistic Simplex: Cooperative Runtime Assurance for Safety-Critical Autonomous Systems", "authors": ["Ayoosh Bansal", "Mikael Yeghiazaryan", "Artyom Khachatryan", "Tianyi Zhu", "Hunmin Kim", "Naira Hovakimyan", "Lui Sha"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.08190", "pdf_url": "https://arxiv.org/pdf/2605.08190", "code_url": "", "clusters": ["C7"], "keywords": ["synergistic simplex", "runtime assurance", "ML components"], "model_class": "other", "techniques": ["bidirectional ML-monitor integration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Extends simplex runtime assurance so verifiable safety monitors can use ML outputs while keeping formal guarantees, improving performance of the safe fallback.", "relevance": 2, "relevance_reason": "Runtime assurance for ML-enabled autonomy.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"simplex architecture\" learned controller"}
{"id": "2609.21650", "title": "SynthDemo-RL: Breaking the Zero-Reward Barrier in VLA Adaptation with LLM-Guided Synthetic Demonstrations", "authors": ["Hiroaki Kingetsu", "Hiroaki Kurihara", "Kaoru Yokoo", "Kenji Fukumizu", "Manohar Kaul"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.21650", "pdf_url": "https://arxiv.org/pdf/2609.21650", "code_url": "", "clusters": ["C3"], "keywords": ["synthetic demonstrations", "zero-reward", "PPO"], "model_class": "VLA", "techniques": ["SFT", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Generates privileged-state synthetic demos to bootstrap RL where sparse rewards are never observed.", "relevance": 2, "relevance_reason": "Adaptation without human demos.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2607.02840", "title": "TACO: TActile World Model as a Self-COrrector for Scalable Robot Policy Post-Training", "authors": ["Shengbang Liu", "Yueru Jia", "Yuyang Yan", "Jiaming Liu", "Xinran Zhang", "Qiuxuan Feng", "Yandong Guo", "Shiji Zhou", "Boxin Shi", "Shanghang Zhang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.02840", "pdf_url": "https://arxiv.org/pdf/2607.02840", "code_url": "", "clusters": ["C3"], "keywords": ["TACO", "tactile world model", "self-correction"], "model_class": "VLA", "techniques": ["world-model data synthesis", "post-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Tactile world model synthesizes corrective training data for scalable post-training on contact-rich tasks.", "relevance": 2, "relevance_reason": "Reduces human intervention cost.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"world model\" AND abs:\"post-training\""}
{"id": "2605.07354", "title": "Task-Oriented Communication for Human Action Understanding via Edge-Cloud Co-Inference", "authors": ["Jingyi Liu", "Cheng Yuan", "Lijun He", "Jun Zhang", "Jiawei Shao"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.07354", "pdf_url": "https://arxiv.org/pdf/2605.07354", "code_url": "", "clusters": ["C6"], "keywords": ["task-oriented communication", "motion tokens", "edge-cloud co-inference"], "model_class": "other", "techniques": ["VQ-VAE motion tokenization", "edge-cloud co-inference"], "base_model": "", "params": "", "hardware": "", "reported_metric": "payload ~1% and latency ~20% of video-codec baselines", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Edge extracts pose and sends compact VQ-VAE motion tokens (as low as 9 bits/frame) to a cloud VLM for human action understanding.", "relevance": 2, "relevance_reason": "Task-oriented edge-cloud transport pattern relevant to robot perception offload.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2608.13026", "title": "Temporal GRPO: Beyond Trajectory-Level Credit in Vision-Language-Action Reinforcement Learning", "authors": ["Yao Zhou", "Hang Gao", "Fengge Wu", "Changwen Zheng", "Wenwen Qiang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.13026", "pdf_url": "https://arxiv.org/pdf/2608.13026", "code_url": "", "clusters": ["C3"], "keywords": ["Temporal GRPO", "stage credit"], "model_class": "VLA", "techniques": ["GRPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Stage-aligned advantages in GRPO VLA post-training for better sample efficiency.", "relevance": 2, "relevance_reason": "RL post-training sample efficiency.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:GRPO"}
{"id": "2604.18107", "title": "Test-Time Perturbation Learning with Delayed Feedback for Vision-Language-Action Models", "authors": ["Zehua Zang", "Xi Wang", "Fuchun Sun", "Xiao Xu", "Lixiang Lium", "Jiahuan Zhou", "Jiangmeng Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.18107", "pdf_url": "https://arxiv.org/pdf/2604.18107", "code_url": "", "clusters": ["C3"], "keywords": ["test-time adaptation", "perturbation learning", "no fine-tuning"], "model_class": "VLA", "techniques": ["test-time augmentation", "action voting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Verifier-free test-time adaptation using uncertainty-based augmentation and voting with an adaptive budget, leaving the VLA untouched.", "relevance": 2, "relevance_reason": "Adaptation without fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"test-time adaptation\""}
{"id": "2604.03191", "title": "The Compression Gap: Why Discrete Tokenization Limits Vision-Language-Action Model Scaling", "authors": ["Takuya Shiba"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.03191", "pdf_url": "https://arxiv.org/pdf/2604.03191", "code_url": "", "clusters": ["C2"], "keywords": ["action tokenization"], "model_class": "VLA", "techniques": ["information bottleneck analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Argues and shows on LIBERO that discrete action codebooks become the information bottleneck so vision-encoder upgrades stop helping tokenized VLAs.", "relevance": 2, "relevance_reason": "Informs compression/tokenization choices.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=0]"}
{"id": "2608.23224", "title": "Think Only When Needed: Prompt-Authority Control for Selective Slow-Path Intervention in Vision-Language-Action Manipulation", "authors": ["Zhiruo Zhou", "Zelin Li", "Xiwen Chen", "Jiazhuo Li", "Chenwei Wang", "Huiming Chen", "Xiaojun Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.23224", "pdf_url": "https://arxiv.org/pdf/2608.23224", "code_url": "", "clusters": ["C1"], "keywords": ["frozen VLA", "selective intervention"], "model_class": "VLA", "techniques": ["prompt gating", "retrieval augmentation"], "base_model": "pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Plus", "real robot (PiPER)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Interface that only lets retrieval-generated prompts change a frozen VLA's input when a compatibility rule authorizes a canonical instruction, otherwise restoring the original prompt.", "relevance": 2, "relevance_reason": "Guards frozen VLAs against prompt-form collapse from slow-path interventions.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2609.19666", "title": "Towards High-DoF Dexterous Manipulation through VLA Post-Training", "authors": ["Junlei Zhu", "Shenzhe Yao", "Chaogui Huang", "Wenkai Zhu", "Jingwei Peng", "Guanqi He", "Soren Schwertfeger", "Jiahao Chen", "Yide Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.19666", "pdf_url": "https://arxiv.org/pdf/2609.19666", "code_url": "", "clusters": ["C3"], "keywords": ["dexterous VLA"], "model_class": "VLA", "techniques": ["action codec", "DAgger", "residual RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Four-step post-training pipeline for dexterous-hand VLAs: learned temporal hand-action codec, SFT, DAgger with smooth takeover, and latent residual RL.", "relevance": 2, "relevance_reason": "Efficient real-world VLA post-training for high-DoF hands.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=0]"}
{"id": "2607.11270", "title": "Towards Predictive, Aligned, and Scalable Robot Learning", "authors": ["Peijun Tang", "Shang-Ping Xie", "Baifu Huang", "Bin Sun", "Hao Yang", "Kuncheng Luo", "Weiqi Jin", "Shilin Fang", "Jia-Nan Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.11270", "pdf_url": "https://arxiv.org/pdf/2607.11270", "code_url": "", "clusters": ["C8"], "keywords": ["latent world-action model", "modality alignment"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Lumo-2: latent world-action model reasoning over world dynamics in latent space, positioned as lightweight vs video WAMs.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2609.31313", "title": "Towards VLA-Dreamer: Refining VLA Behavior Using World Models", "authors": ["Parsa Mastouri Kashani", "Jan-Gerrit Habekost", "Stefan Wermter"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.31313", "pdf_url": "https://arxiv.org/pdf/2609.31313", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["VLA-Dreamer"], "model_class": "world model", "techniques": ["latent world model", "JEPA-style loss"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Concept paper proposing a world model trained in the VLA vision-encoder embedding space to test whether those embeddings support action-conditioned prediction and improve sample efficiency.", "relevance": 2, "relevance_reason": "Early-stage idea for world-model refinement of VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=0]"}
{"id": "2608.24101", "title": "TrAct: Bridging Robot Control and Visual Prediction with Visual Tracks", "authors": ["Zhihang Cao", "Howard Ji", "Kevin Zhang", "Kuangzhi Ge", "Fei-Fei Li", "Jiajun Wu", "Huang Huang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.24101", "pdf_url": "https://arxiv.org/pdf/2608.24101", "code_url": "", "clusters": ["C7", "C9"], "keywords": ["world model", "visual tracks", "VLAT", "decision-making"], "model_class": "world-action model", "techniques": ["track-conditioned video prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses predicted point tracks as an embodiment-agnostic interface between a VLA-and-track model and a video world model for decision making.", "relevance": 2, "relevance_reason": "World-action model design from speaker's group; efficiency not the focus.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "dyna-2-infrastructure-blog-2026", "title": "Training Dyna-2 at million-hour scale, repeatably", "authors": ["Dyna Robotics"], "year": 2026, "venue": "blog (Dyna Robotics)", "type": "blog", "url": "https://www.dyna.co/research/dyna-2-infrastructure", "pdf_url": "", "code_url": "", "clusters": ["C9"], "keywords": ["training infrastructure", "data pipeline", "MCAP", "optimizer sharding"], "model_class": "world-action model", "techniques": ["storage compression", "caching", "topology-aware optimizer sharding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "sample reads ~2.9x faster; ingestion 31x; 98% GPU utilization; optimizer ~3x faster", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Engineering post on scaling robot-data training: topic-grouped MCAP storage, Airflow ingestion, warehouse-built manifests, Alluxio NVMe caching and intra-node optimizer sharding.", "relevance": 2, "relevance_reason": "Training-side efficiency, not onboard inference.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Dyna"}
{"id": "2601.09881", "title": "Transition Matching Distillation for Fast Video Generation", "authors": ["Weili Nie", "Julius Berner", "Nanye Ma", "Chao Liu", "Saining Xie", "Arash Vahdat"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.09881", "pdf_url": "https://arxiv.org/pdf/2601.09881", "code_url": "", "clusters": ["C2", "C8"], "keywords": ["video diffusion", "few-step"], "model_class": "other", "techniques": ["few-step distillation", "transition matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Transition Matching Distillation compresses multi-step video diffusion/flow models into few-step generators by matching denoising transitions.", "relevance": 2, "relevance_reason": "foundational technique for fast video world models", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2501.03575; cited by 2606.05254"}
{"id": "2605.01950", "title": "TRAP: Tail-aware Ranking Attack for World-Model Planning", "authors": ["Siyuan Duan", "Ke Zhang", "Xizhao Luo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.01950", "pdf_url": "https://arxiv.org/pdf/2605.01950", "code_url": "", "clusters": ["C8"], "keywords": ["ranking attack", "world-model planning", "backdoor"], "model_class": "world model", "techniques": ["tail-aware trajectory ranking attack"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Identifies that world-model planners can be hijacked by perturbing the ranking of a few decision-critical imagined trajectories, and builds an attack exploiting it.", "relevance": 2, "relevance_reason": "Planning-security risk for world models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2606.03127", "title": "TTT-VLA: Test-Time Latent Prompt Optimization for Vision-Language-Action Models", "authors": ["Wenbo Zhang", "Jianxiong Li", "Shuai Yang", "Sijin Chen", "Jiajun Liu", "Lingqiao Liu", "Xiao Ma"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.03127", "pdf_url": "https://arxiv.org/pdf/2606.03127", "code_url": "", "clusters": ["C3"], "keywords": ["TTT"], "model_class": "VLA", "techniques": ["test-time training", "latent prompt optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Test-time training that optimizes only a learned latent prompt with a self-supervised proxy task on current-environment data.", "relevance": 2, "relevance_reason": "Cheap deployment-time adaptation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]"}
{"id": "2606.10382", "title": "UMI-Bench 1.0: An Open and Reproducible Real-World Benchmark for Tabletop Robotic Manipulation with UMI Data", "authors": ["Shi Jin", "Yuntian Wang", "Yuhui Duan", "Di Wu", "Gaoqi Dong", "Xiaohang Liu", "Xiaotong Li", "Hongfei Jia", "Zehao Zhang", "Tianyu Wang", "Zhongjie Jia", "Yuanqi Yao", "Chenjia Bai", "Zhaxizhuoma", "Siao Liu", "Nieqing Cao", "Jin Wang", "Chao Yu", "Yan Ding"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2606.10382", "pdf_url": "https://arxiv.org/pdf/2606.10382", "code_url": "", "clusters": ["C4"], "keywords": ["UMI-Bench", "real-world benchmark", "reproducible"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["UMI-Bench", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Local-first real-robot benchmark for UMI-style policies aligning data collection, reset, execution and logging under one auditable protocol.", "relevance": 2, "relevance_reason": "Real-robot evaluation protocol.", "foundational_technique": false, "verified": true, "found_via": "search query: robot manipulation benchmark reproducibility evaluation"}
{"id": "2609.15840", "title": "Uncertainty-Guided Sparse Refinement for Action Chunking Transformer Policies", "authors": ["Chen-Yang Wang", "Yun-Tian Wang", "Xiao-Xiong Yang", "Ding Jiang", "Siao Liu", "Yang Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.15840", "pdf_url": "https://arxiv.org/pdf/2609.15840", "code_url": "", "clusters": ["C7"], "keywords": ["uncertainty-guided refinement", "ACT", "sparse refinement"], "model_class": "other", "techniques": ["coarse-to-fine refinement at uncertain timesteps"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Refines only the few uncertain timesteps within an ACT chunk rather than the whole chunk, targeting critical prediction errors cheaply.", "relevance": 2, "relevance_reason": "Uncertainty-targeted compute for chunked policies.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2503.08558 (S2)"}
{"id": "2608.03563", "title": "Unified Visuomotor Targets: Supervising VLAs Beyond Physical Actions", "authors": ["Zhen-Yang Feng", "Unnat Jain"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.03563", "pdf_url": "https://arxiv.org/pdf/2608.03563", "code_url": "", "clusters": ["C3"], "keywords": ["supervision targets"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Changes VLA training targets to unified visuomotor targets beyond physical actions for better, more efficiently trained policies.", "relevance": 2, "relevance_reason": "Context: VLA training (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2606.08775", "title": "Unifying Object-Centric World Models and Diffusion Policy: A Hierarchical Framework for Multi-Stage Robotic Tasks", "authors": ["Raktim Gautam Goswami", "P. Krishnamurthy", "Yann LeCun", "F. Khorrami"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.08775", "pdf_url": "https://arxiv.org/pdf/2606.08775", "code_url": "", "clusters": ["C8"], "keywords": ["object-centric world model", "MPC"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Hierarchical framework coupling an object-centric world model with diffusion policy for multi-stage tasks.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2310.16828"}
{"id": "2608.22869", "title": "UniMem: Unifying Multimodal Memory and Control for Vision-Language-Action Models", "authors": ["Lars Osterberg", "Maggie Wang", "Mac Schwager"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.22869", "pdf_url": "https://arxiv.org/pdf/2608.22869", "code_url": "", "clusters": ["C1"], "keywords": ["UniMem"], "model_class": "VLA", "techniques": ["keyframe caching", "event-triggered memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Single-backbone VLA memory with an event classifier, keyframe encoder and keyframe caching to reduce rollout overhead.", "relevance": 2, "relevance_reason": "Memory with caching for faster inference.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv (VLA OR \"diffusion policy\" OR \"world action model\") AND (cach* OR \"KV cache\")"}
{"id": "2604.09330", "title": "VAG: Dual-Stream Video-Action Generation for Embodied Data Synthesis", "authors": ["Xiaolei Lang", "Yangyue Wang", "Yu-Kun Zhou", "Chaojun Ni", "Ke-Rui Li", "Jiagang Zhu", "T. Liu", "Jiajun Lv", "Xing-Xing Zuo", "Yun Ye", "Guan Huang", "Xiaofeng Wang", "Zheng Zhu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.09330", "pdf_url": "https://arxiv.org/pdf/2604.09330", "code_url": "", "clusters": ["C8"], "keywords": ["video-action generation", "synthetic data", "dual-stream"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Dual-stream model that jointly generates videos and aligned actions to synthesize embodied training data for policy learning.", "relevance": 2, "relevance_reason": "Context: world-action generative model (links=21).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2604.04502", "title": "Veo-Act: Enhancing VLA Policies with Frontier Video Models", "authors": ["Zhongrui Zhang", "Cheng‐Chuan Yang", "Qin Lu", "Yanjiang Guo", "Jianke Zhang", "Yucheng Hu", "Jianyu Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.04502", "pdf_url": "https://arxiv.org/pdf/2604.04502", "code_url": "", "clusters": ["C8"], "keywords": ["video models", "VLA"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses frontier video generation models to complement VLA policies for better generalization.", "relevance": 2, "relevance_reason": "Context: (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2501.16664"}
{"id": "2605.05989", "title": "Verifiable Model-Free Safety Filters via Reinforcement Learning", "authors": ["Bihui Yin", "Yiwen Lu", "Yuchen Jiang", "Yilin Mo"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.05989", "pdf_url": "https://arxiv.org/pdf/2605.05989", "code_url": "", "clusters": ["C7"], "keywords": ["model-free safety filter", "unrolled QP", "certificate"], "model_class": "other", "techniques": ["RL-learned QP safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns the parameters of an unrolled QP safety filter via deep RL without system identification while retaining a formal persistent-safety certificate.", "relevance": 2, "relevance_reason": "Learned but certifiable filter.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"predictive safety filter\" learning-based"}
{"id": "2602.07399", "title": "VGAS: Value-Guided Action-Chunk Selection for Few-Shot Vision-Language-Action Adaptation", "authors": ["Changhua Xu", "En Yu", "Junyu Xuan", "Jie Lu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.07399", "pdf_url": "https://arxiv.org/pdf/2602.07399", "code_url": "", "clusters": ["C3"], "keywords": ["best-of-N", "value-guided selection", "few-shot"], "model_class": "VLA", "techniques": ["inference-time selection", "value model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Few-shot VLA adaptation via value-guided best-of-N action-chunk selection.", "relevance": 2, "relevance_reason": "Few-shot adaptation; adds inference compute.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2604.13733", "title": "Vision-Language-Action Jump-Starting for Reinforcement Learning Robotic Agents", "authors": ["Angelo Moroncelli", "Roberto Zanetti", "Marco Maccarini", "Loris Roveda"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2604.13733", "pdf_url": "https://arxiv.org/pdf/2604.13733", "code_url": "", "clusters": ["C3"], "keywords": ["jump-start RL", "VLA guidance"], "model_class": "VLA", "techniques": ["RL with VLA guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Uses a VLA as a transient source of action suggestions to jump-start a fast on-policy RL controller.", "relevance": 2, "relevance_reason": "Distills VLA knowledge into efficient RL policy.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2606.18247", "title": "Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement", "authors": ["Ming-Tong Zhang", "Dhruv Shah"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.18247", "pdf_url": "https://arxiv.org/pdf/2606.18247", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["inference-time steering", "verifier", "self-improvement"], "model_class": "VLA", "techniques": ["generator-verifier", "test-time compute"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "VERITAS pairs a generalist policy with a gradient-free visual verifier to choose actions at inference time and to filter self-generated rollouts for fine-tuning.", "relevance": 2, "relevance_reason": "Inference-time compute vs quality trade-off for policies.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2608.15816", "title": "ViTaR: Visuo-Tactile Residual Adaptation for Foundation VLA Manipulation", "authors": ["Yi Wang", "Renjun Wu", "Jinyan Liu", "Xuesong Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2608.15816", "pdf_url": "https://arxiv.org/pdf/2608.15816", "code_url": "", "clusters": ["C3"], "keywords": ["ViTaR", "tactile residual", "execution modulation"], "model_class": "VLA", "techniques": ["residual adaptation", "frozen VLA"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Adds tactile feedback as a bounded residual modulator of a frozen VLA's execution for contact-rich tasks.", "relevance": 2, "relevance_reason": "Lightweight modality adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"catastrophic forgetting\""}
{"id": "2602.10098", "title": "VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model", "authors": ["Jing-Wen Sun", "Wenyao Zhang", "Ze-Kun Qi", "Shaojie Ren", "Zezhi Liu", "Han-Xin Zhu", "Guang-Zhong Sun", "Xin Jin", "Zhi-Bo Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.10098", "pdf_url": "https://arxiv.org/pdf/2602.10098", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["JEPA", "latent actions", "video pretraining"], "model_class": "VLA", "techniques": ["JEPA", "latent world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a JEPA-style latent world-model objective to VLA pretraining so latent actions track state transitions rather than pixel changes.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links; latent world model for VLA", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2402.15391"}
{"id": "2607.01586", "title": "VLAFlow: A Unified Training Framework for Vision-Language-Action Models via Co-training and Future Latent Alignment", "authors": ["Guo-Yang Xia", "Feng-Fa Li", "H. Ji", "Lei Ren", "Fangxiang Feng", "Kun Zhan", "Yan Xie"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.01586", "pdf_url": "https://arxiv.org/pdf/2607.01586", "code_url": "", "clusters": ["C4"], "keywords": ["controlled comparison", "training framework"], "model_class": "VLA", "techniques": ["flow matching", "co-training", "controlled comparison"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLAFlow is a unified flow-matching framework for controlled comparisons of VLA pretraining objectives on a shared heterogeneous corpus, adding future-latent alignment.", "relevance": 2, "relevance_reason": "context: controlled benchmarking of VLA training choices", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2602.18532", "title": "VLANeXt Family: A Systematic Study of VLA Models from Core Recipes to Emerging Paradigms", "authors": ["Xiao-Ming Wu", "Kang Liao", "Yihang Luo", "Bin Fan", "Jian-Jian Jiang", "Runze Yang", "Zhonghua Wu", "Wei-Shi Zheng", "Chen Change Loy"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2602.18532", "pdf_url": "https://arxiv.org/pdf/2602.18532", "code_url": "", "clusters": ["C3"], "keywords": ["VLANeXt", "design space", "recipes"], "model_class": "VLA", "techniques": ["recipe study"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Unified re-examination of VLA design choices from RT-2-style baseline to modern recipes.", "relevance": 2, "relevance_reason": "Recipe study relevant to training choices.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2602.12063", "title": "VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model", "authors": ["Yanjiang Guo", "Tony Lee", "L. Shi", "Jianyu Chen", "Percy Liang", "Chelsea Finn"], "year": 2026, "venue": "ICML 2026", "type": "preprint", "url": "https://arxiv.org/abs/2602.12063", "pdf_url": "https://arxiv.org/pdf/2602.12063", "code_url": "", "clusters": ["C4", "C7", "C9"], "keywords": ["world model data generation", "co-improvement", "action-conditioned video"], "model_class": "world model", "techniques": ["world-model rollouts for policy improvement"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Iteratively improves a VLA and an action-conditioned video world model together so the world model can supply reliable synthetic rollouts.", "relevance": 2, "relevance_reason": "World models to reduce real-robot data cost.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Stanford AI Lab blog (ICML 2026)"}
{"id": "2601.03309", "title": "VLM4VLA: Revisiting Vision-Language-Models in Vision-Language-Action Models", "authors": ["Jianke Zhang", "Xiaoyu Chen", "Qiuyue Wang", "Mingsheng Li", "Yanjiang Guo", "Yucheng Hu", "Jiajun Zhang", "Shuai Bai", "Junyang Lin", "Jianyu Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2601.03309", "pdf_url": "https://arxiv.org/pdf/2601.03309", "code_url": "", "clusters": ["C1"], "keywords": ["VLM choice", "minimal adaptation"], "model_class": "VLA", "techniques": ["analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Finds general VLM capability poorly predicts VLA performance and the vision module is the bottleneck, via a minimal VLM-to-VLA pipeline.", "relevance": 2, "relevance_reason": "Backbone selection insight.", "foundational_technique": false, "verified": true, "found_via": "search query: small VLA"}
{"id": "2603.22003", "title": "VP-VLA: Visual Prompting as an Interface for Vision-Language-Action Models", "authors": ["Zixuan Wang", "Yuxin Chen", "Yuqi Liu", "Ji-lu Ye", "Pengguang Chen", "Chang-Sheng Lu", "Shu Liu", "Jiaya Jia"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2603.22003", "pdf_url": "https://arxiv.org/pdf/2603.22003", "code_url": "", "clusters": ["C1"], "keywords": ["visual prompting"], "model_class": "VLA", "techniques": ["dual-system"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Dual-system VLA using visual prompts as an interface between planning and control.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941, 2405.12213"}
{"id": "2607.29613", "title": "WCM: A World Critic Model for Vision-Language-Action Reinforcement Learning", "authors": ["Senyu Fei", "Xiaopeng Yu", "Siyin Wang", "Xianzhong Zhao", "Jingjing Gong", "Xipeng Qiu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.29613", "pdf_url": "https://arxiv.org/pdf/2607.29613", "code_url": "", "clusters": ["C3"], "keywords": ["world critic model", "history-aware critic"], "model_class": "VLA", "techniques": ["critic learning", "RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "World-modeling objective to train history-aware critics for VLA RL post-training.", "relevance": 2, "relevance_reason": "RL post-training component.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"world model\" AND abs:\"post-training\""}
{"id": "2606.10267", "title": "What Matters in Orchestrating Robot Policies: A Systematic Study of Hierarchical VLA Agents", "authors": ["Jiaheng Hu", "Mohit Shridhar", "Caden Lu", "Dhruv Shah", "Hao-Tien Lewis Chiang", "Jie Tan", "Annie Xie"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.10267", "pdf_url": "https://arxiv.org/pdf/2606.10267", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["hierarchical VLA", "planner-controller", "benchmark"], "model_class": "VLA", "techniques": ["hierarchical control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmarks design choices for VLM-planner/VLA-controller hierarchies across short, long and reasoning-heavy manipulation tasks.", "relevance": 2, "relevance_reason": "Slow-planner/fast-controller hierarchy is a key efficiency architecture pattern.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2605.13105", "title": "What to Ignore, What to React: Visually Robust RL Fine-Tuning of VLA Models", "authors": ["Yuanfang Peng", "Jingjing Fu", "Chuheng Zhang", "Li Zhao", "Jiang Bian", "Mingyu Liu", "Ling Zhang", "Jun Zhang", "Rui Wang"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.13105", "pdf_url": "https://arxiv.org/pdf/2605.13105", "code_url": "", "clusters": ["C3"], "keywords": ["PAIR-VLA"], "model_class": "VLA", "techniques": ["PPO", "auxiliary invariance losses"], "base_model": "OpenVLA, pi0.5", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill3"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PPO fine-tuning of VLAs with paired invariance/sensitivity objectives over visual variants to improve robustness under visual shift.", "relevance": 2, "relevance_reason": "RL fine-tuning robustness.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=0]"}
{"id": "2607.28256", "title": "When Robots Exchange Meaning: A Demo of Goal-Oriented Semantic Communications for Collaborative Robotics", "authors": ["Peizheng Li", "Xinyi Lin", "Sajida Gufran", "Adnan Aijaz"], "year": 2026, "venue": "IEEE CSCN 2026 (demo)", "type": "paper", "url": "https://arxiv.org/abs/2607.28256", "pdf_url": "https://arxiv.org/pdf/2607.28256", "code_url": "", "clusters": ["C6"], "keywords": ["semantic communication", "robot-edge testbed", "VQ-VAE"], "model_class": "other", "techniques": ["learned visual compression", "edge semantic mapping"], "base_model": "", "params": "", "hardware": "mobile robot + Jetson Orin edge node", "reported_metric": "42.67x payload reduction vs model-input RGB bytes", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Demo testbed where a robot encodes RGB into VQ-VAE tokens and a Jetson edge node reconstructs and builds a semantic map.", "relevance": 2, "relevance_reason": "Robot-edge semantic transport testbed.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2603.03740", "title": "Whole-Body Safe Control of Robotic Systems with Koopman Neural Dynamics", "authors": ["Sebin Jung", "Abulikemu Abuduweili", "Jiaxing Li", "Changliu Liu"], "year": 2026, "venue": "ICRA 2026", "type": "paper", "url": "https://arxiv.org/abs/2603.03740", "pdf_url": "https://arxiv.org/pdf/2603.03740", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["Koopman", "safe set algorithm", "QP", "real-time safety"], "model_class": "other", "techniques": ["Koopman linearization", "safe set algorithm"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Learns a Koopman embedding so tracking and safety constraints of high-DoF robots reduce to a single real-time QP.", "relevance": 2, "relevance_reason": "Real-time safety filtering for whole-body control.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2609.31428", "title": "Wireless Evidence Acquisition for Multimodal Inference Constrained by Temporal Windows", "authors": ["Alessandro Compagnoni", "Anup Mishra", "Carla Fabiana Chiasserini", "Elad Michael Schiller", "Petar Popovski"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2609.31428", "pdf_url": "https://arxiv.org/pdf/2609.31428", "code_url": "", "clusters": ["C6"], "keywords": ["uplink scheduling", "temporal window", "multimodal fusion", "edge inference"], "model_class": "other", "techniques": ["evidence-gain-driven uplink scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Schedules which sensor observations to send to an edge server within a decision deadline by their expected error reduction rather than throughput.", "relevance": 2, "relevance_reason": "Deadline-aware uplink for edge-inferred robot decisions.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: edge server robot policy"}
{"id": "2607.27599", "title": "World Action Planner: Generalizable Decision-Making with Action-Conditioned World Models", "authors": ["Xiang-Cheng Zhang", "Yilun Du"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.27599", "pdf_url": "https://arxiv.org/pdf/2607.27599", "code_url": "", "clusters": ["C8"], "keywords": ["action-conditioned world model", "planning"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Uses action-conditioned world models for planning to generalize decision making beyond imitation policies.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2605.22164", "title": "World Model Control by Trajectory Reachability Metrics", "authors": ["Liangyu Li", "Shengzhi Wang", "Libin Qiu", "Mingliang Xiong", "Qingwen Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.22164", "pdf_url": "https://arxiv.org/pdf/2605.22164", "code_url": "", "clusters": ["C8"], "keywords": ["trajectory reachability metrics", "latent planning cost"], "model_class": "world model", "techniques": ["learned temporal pairwise cost for ranking predicted endpoints"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Adds a small learned reachability cost to rank latent-world-model rollouts, raising LeWorldModel TwoRoom success from 7% to 97%.", "relevance": 2, "relevance_reason": "Planning cost for latent world models.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828 (citations)"}
{"id": "2605.17912", "title": "WorldArena 2.0: Extending Embodied World Model Benchmarking on Modality, Functionality and Platform", "authors": ["Shang, Yu", "Tang, Yinzhou", "Ma, Yiding", "Li, Zhuohang", "Jin, Lei", "Su, Weikang", "Jin, Xin", "Wang, Zhaolu", "Wang, Ziyou", "Zhang, Xin", "Su, Haisheng", "He, Weizhen", "Wu, Wei", "Duan, Haoyi", "Wetzstein, Gordon", "Liu, Xihui", "Shah, Dhruv", "Zhang, Zhaoxiang", "Chen, Zhibo", "Zhu, Jun", "Tian, Yonghong", "Chua, Tat-Seng", "Zhu, Wenwu", "Gao, Chen", "Li, Yong", "Yu Shang", "Yinzhou Tang", "Yiding Ma", "Zhuohang Li", "Lei Jin", "Weikang Su", "Xin Jin", "Zhaolu Wang", "Ziyou Wang", "Xin Zhang", "Haisheng Su", "Weizhen He"], "year": 2026, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2605.17912", "pdf_url": "https://arxiv.org/pdf/2605.17912", "code_url": "", "clusters": ["C3", "C4", "C7", "C9"], "keywords": ["WorldArena 2.0", "visuotactile", "interactive RL environment", "real-world platforms", "world model benchmark"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["WorldArena 2.0"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Extends WorldArena to visuotactile modalities, interactive RL use and real-platform evaluation of world models.", "relevance": 2, "relevance_reason": "World-model benchmarking; Dhruv Shah co-author.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: WorldArena benchmark; author sweep: Dhruv Shah"}
{"id": "2607.03461", "title": "WorldBagel: Uncovering the Power of Unified Multimodal Models for Vision-Language-Action-World Modeling", "authors": ["Zelin Zhao", "Min Shi", "Bo Yuan", "Haotian Xue", "Jialuo Li", "Lama Moukheiber", "Humphrey Shi", "Yong-Xin Chen"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.03461", "pdf_url": "https://arxiv.org/pdf/2607.03461", "code_url": "", "clusters": ["C8"], "keywords": ["unified multimodal model", "VLAW"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Builds a vision-language-action-world model on a unified multimodal generative backbone.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2605.24892", "title": "X-Foresight: A Joint Vision-Action Causal Forecasting Network via Predictive World Modeling", "authors": ["Baolu Li", "Jingyu Qian", "Rui Guo", "Yilun Chen", "Hanpeng Liu", "Yuan Lin", "Junhong Zhou", "Ruixin Liu", "Liu Yang", "Yutong Zheng", "Zhenli Zhang", "Sean Li", "Chaoda Zheng", "Boyang Wang", "Tenglong", "Gu", "Zhuangzhuang Ding", "Pengkun Zheng", "Yu Zhang", "Xianming Liu"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2605.24892", "pdf_url": "https://arxiv.org/pdf/2605.24892", "code_url": "", "clusters": ["C8"], "keywords": ["X-Foresight"], "model_class": "world-action model", "techniques": ["predictive world model", "chunk-wise AR"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Integrates a chunk-wise autoregressive predictive world model into the VLA to learn world knowledge jointly with real-time action control.", "relevance": 2, "relevance_reason": "World modeling inside VLAs with real-time control claims.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=200]"}
{"id": "2606.14752", "title": "X-Tokenizer: A Multimodal Action Tokenizer for Vision-Language-Action Pretraining", "authors": ["Miracle Kang", "Lights Shi", "Lucy Liang", "Roy Gan", "Dongxiu Liu", "Pushi Zhang", "Sylas Chen", "Shawn Qin", "Yinan Zheng", "Jinliang Zheng", "Hao Wang", "Xianyuan Zhan", "Hang Su"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2606.14752", "pdf_url": "https://arxiv.org/pdf/2606.14752", "code_url": "", "clusters": ["C1"], "keywords": ["semantic residual quantization", "action tokenizer"], "model_class": "VLA", "techniques": ["action tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin 2.0", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Action tokenizer whose first quantization level forms a coarse 'action language' used as a semantic supervision interface in VLA pretraining.", "relevance": 2, "relevance_reason": "Tokenization for pretraining quality.", "foundational_technique": false, "verified": true, "found_via": "search query: action tokenization"}
{"id": "2607.15330", "title": "Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories", "authors": ["Jun Guo", "Piao-Piao Jin", "Jason Li", "Pei-Yan Li", "Yingyan Li", "Fu-Teng Liu", "Wan-Li Peng", "Optimus Qin", "Yifei Su", "Nan Sun", "Qiao Sun", "Run-Ze Suo", "He-Yun Wang", "Yun-Hong Wang", "Ru-Jie Wu", "Cao-Yu Xia", "Li-Na Zhang", "Jack Zhao", "Guo-Liang Chen", "Wen-Long Chen", "Xin He", "Bin Li", "Qing Li", "Zhuo Li", "Heng-Xu Qu", "Wen-Xuan Song", "Di-Yun Xiang", "Yi-Fan Xie", "Pei-Ran Xu", "Hangjun Ye", "Wen Ye", "Han Zhao", "Quan-Yun Zhou"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.15330", "pdf_url": "https://arxiv.org/pdf/2607.15330", "code_url": "", "clusters": ["C3"], "keywords": ["foundation VLA", "UMI data", "scaling"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa365", "RoboDojo", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Foundation VLA pretrained on over 100k hours of UMI-collected trajectories with auto-labelled language, showing scaling and data-efficient fine-tuning.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.12213"}
{"id": "2606.13769", "title": "μ0: A Scalable 3D Interaction-Trace World Model", "authors": ["Seungjae Lee", "Y. Jung", "Jusuk Lee", "Jonghu Shin", "A. Shahidzadeh", "Yao-Chih Lee", "H. J. Kim", "Jia-Bin Huang", "Furong Huang"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.13769", "pdf_url": "https://arxiv.org/pdf/2606.13769", "code_url": "", "clusters": ["C8"], "keywords": ["world model", "embodiment-agnostic"], "model_class": "world model", "techniques": ["3D interaction traces"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "World model over 3D interaction traces instead of pixels, aimed at scalable embodiment-agnostic robot learning.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 10 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2410.24164"}
{"id": "2606.01027", "title": "τ0-WM: A Unified Video-Action World Model for Robotic Manipulation", "authors": ["Peng-Fei Zhou", "Shengcong Chen", "Di Chen", "Jia-Xu Wang", "Rong-Jun Jin", "Bin Zhu", "Yike Pan", "Song-En Gu", "Kuanning Wang", "S. Nan", "Xingyu Qiu", "Chen-Hao Qiu", "Pu Yang", "Yunuo Cai", "Jianxiong Gao", "Yifan Li", "Yanwei Fu", "Xiangyu Yue", "Zhi Chen", "Jian-Lan Luo"], "year": 2026, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2606.01027", "pdf_url": "https://arxiv.org/pdf/2606.01027", "code_url": "", "clusters": ["C8"], "keywords": ["video-action world model"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Unified video-action world model combining policy learning, video prediction and action evaluation.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2501.03575; cited by 2607.27933"}
{"id": "2603.12263", "title": "Ψ0: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation", "authors": ["Songlin Wei", "Hongyi Jing", "Boqian Li", "Zhenyu Zhao", "Jiageng Mao", "Zhe-Yuan Ni", "Sicheng He", "Jie Liu", "Xia-Wei Liu", "Kai Kang", "Sheng Zang", "Wei-Duo Yuan", "Marco Pavone", "Di Huang", "Yue Wang"], "year": 2026, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2603.12263", "pdf_url": "https://arxiv.org/pdf/2603.12263", "code_url": "", "clusters": ["C1"], "keywords": ["humanoid", "loco-manipulation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open humanoid loco-manipulation foundation model trained with a staged human/humanoid data strategy.", "relevance": 2, "relevance_reason": "Context: humanoid foundation model (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2506.22242", "title": "4D-VLA: Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration", "authors": ["Jiahui Zhang", "Yurui Chen", "Yueming Xu", "Ze Huang", "Yanpeng Zhou", "Yuan Yuan", "Xinyue Cai", "Guowei Huang", "Xing-Yue Quan", "Hang Xu", "Li Zhang"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2506.22242", "pdf_url": "https://arxiv.org/pdf/2506.22242", "code_url": "", "clusters": ["C1"], "keywords": ["spatiotemporal pretraining"], "model_class": "VLA", "techniques": ["4D inputs"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds depth and temporal cues to VLA inputs to align robot and scene coordinates, improving pretraining efficiency.", "relevance": 2, "relevance_reason": "Context: VLA pretraining method cited by core items.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2506.07530, 2509.09372"}
{"id": "2507.01925", "title": "A Survey on Vision-Language-Action Models: An Action Tokenization Perspective", "authors": ["Yi-Fan Zhong", "Fengshuo Bai", "Shaofei Cai", "Xuchuan Huang", "Zhan Chen", "Xiaowei Zhang", "Yuan-Fei Wang", "Shao-Yang Guo", "Tianrui Guan", "Ka Nam Lui", "Zhiquan Qi", "Yitao Liang", "Yuan-Pei Chen", "Yao-Dong Yang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2507.01925", "pdf_url": "https://arxiv.org/pdf/2507.01925", "code_url": "", "clusters": ["C1"], "keywords": ["survey", "action tokens"], "model_class": "VLA", "techniques": ["action tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey organizing VLA models by how they tokenize actions (language, code, affordance, latent, raw) and the trade-offs of each.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 13 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2509.09674, 2510.04898"}
{"id": "2507.00917", "title": "A Survey: Learning Embodied Intelligence from Physical Simulators and World Models", "authors": ["Xiao-Xiao Long", "Qing-Rui Zhao", "Kaiwen Zhang", "Zihao Zhang", "Dingrui Wang", "Yumeng Liu", "Zheng-Jie Shu", "Yi Lu", "Shouzheng Wang", "Xinzhe Wei", "Wei Li", "Wei Yin", "Yao Yao", "Jiangtian Pan", "Qiu Shen", "Ruigang Yang", "Xun Cao", "Qionghai Dai"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2507.00917", "pdf_url": "https://arxiv.org/pdf/2507.00917", "code_url": "", "clusters": ["C8"], "keywords": ["survey", "simulators", "world models"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of embodied intelligence learned from physical simulators and world models.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 14 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2510.22201", "title": "ACG: Action Coherence Guidance for Flow-based Vision-Language-Action models", "authors": ["Minho Park", "Kinam Kim", "Junha Hyung", "Hyojin Jang", "Hoiyeong Jin", "Jooyeol Yun", "Hojoon Lee", "Jaegul Choo"], "year": 2025, "venue": "ICRA 2026", "type": "paper", "url": "https://arxiv.org/abs/2510.22201", "pdf_url": "https://arxiv.org/pdf/2510.22201", "code_url": "https://github.com/DAVIAN-Robotics/ACG", "clusters": ["C1"], "keywords": ["action coherence"], "model_class": "VLA", "techniques": ["test-time guidance", "flow matching"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa", "DexMimicGen", "real robot (SO-101)"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Training-free test-time guidance that increases temporal coherence of flow-based VLA action chunks, reducing jitter and drift.", "relevance": 2, "relevance_reason": "Improves execution quality of open-loop chunks without retraining.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:timing AND (\"vision-language-action\" OR \"robot policy\") AND (deadline* OR jitter)"}
{"id": "2512.03538", "title": "AdaPower: Specializing World Foundation Models for Predictive Manipulation", "authors": ["Yuhang Huang", "Shilong Zou", "Jiazhao Zhang", "Xinwang Liu", "Ruizhen Hu", "Kai Xu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.03538", "pdf_url": "https://arxiv.org/pdf/2512.03538", "code_url": "", "clusters": ["C8"], "keywords": ["AdaPower"], "model_class": "world model", "techniques": ["test-time training", "MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts world foundation models to specialist predictive models via temporal-spatial test-time training and memory persistence inside MPC, boosting pretrained VLAs.", "relevance": 2, "relevance_reason": "Lightweight WM adaptation.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv \"world model*\" AND robot* AND (distill* OR \"real-time\" OR lightweight)"}
{"id": "2503.06669", "title": "AgiBot World Colosseo: A Large-Scale Manipulation Platform for Scalable and Intelligent Embodied Systems", "authors": ["AgiBot-World-Contributors", "Qingwen Bu", "Ji-Song Cai", "Li Chen", "X. Cui", "Yan Ding", "Si-Yuan Feng", "Shenyuan Gao", "Xindong He", "Xuan Hu", "Xu Huang", "Shu Jiang", "Yuxin Jiang", "Cheng Jing", "Hongyang Li", "Jialun Li", "Chiming Liu", "Yi Liu", "Yuxiang Lu", "Jianlan Luo", "Ping Luo", "Yao Mu", "Yue Niu", "Yi-Xuan Pan", "Jiang-Miao Pang", "Yu Qiao", "Guanghui Ren", "Cheng-Xing Ruan", "Jiaqi Shan", "Yongjian Shen", "Cheng Shi", "Ming Shi", "Mo-Di Shi", "Chong-Hao Sima", "Jia-Yi Song", "Huijie Wang", "Wenhao Wang", "Dafeng Wei", "Chengen Xie", "Guofeng Xu", "Jun-Chi Yan", "Cunbiao Yang", "Lei Yang", "Shu-Kai Yang", "Maoqing Yao", "Jiansheng Zeng", "Chi Zhang", "Qingli Zhang", "Bin Zhao", "Chengyu Zhao", "Jia-Qi Zhao", "Jianchao Zhu"], "year": 2025, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "type": "paper", "url": "https://arxiv.org/abs/2503.06669", "pdf_url": "https://arxiv.org/pdf/2503.06669", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "large-scale robot data", "GO-1"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["AgiBot World"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Large-scale manipulation platform and dataset with over a million trajectories across 217 tasks, plus a generalist policy trained on it.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2512.11362", "title": "An Anatomy of Vision-Language-Action Models: From Modules to Milestones and Challenges", "authors": ["Chao Xu", "Suyu Zhang", "Yang Liu", "Baigui Sun", "Weihong Chen", "Bo Xu", "Qi Liu", "Juncheng Wang", "Shujun Wang", "Shan Luo", "Jan Peters", "Athanasios V. Vasilakos", "Stefanos Zafeiriou", "Jiankang Deng"], "year": 2025, "venue": "arXiv", "type": "paper", "url": "https://arxiv.org/abs/2512.11362", "pdf_url": "https://arxiv.org/pdf/2512.11362", "code_url": "", "clusters": ["C4", "C9"], "keywords": ["survey", "VLA anatomy", "dataset and evaluation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Structured VLA survey organized by modules, milestones and five challenges including dataset and evaluation.", "relevance": 2, "relevance_reason": "Survey with an evaluation chapter.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA acceleration evaluation"}
{"id": "2502.01070", "title": "An Inquiry into Datacenter TCO for LLM Inference with FP8", "authors": ["Jiwoo Kim", "Joonhyung Lee", "Gunho Park", "Byeongwook Kim", "Se Jung Kwon", "Dongsoo Lee", "Youngjoo Lee"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.01070", "pdf_url": "https://arxiv.org/pdf/2502.01070", "code_url": "", "clusters": ["C6"], "keywords": ["TCO", "FP8", "thin GEMM", "Gaudi", "H100"], "model_class": "other", "techniques": ["FP8 quantization"], "base_model": "", "params": "", "hardware": "Intel Gaudi 2/3, NVIDIA H100/H200", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "TCO framework for LLM inference accelerators showing thin-GEMM (decode) utilization, not peak FLOPs, drives cost, with FP8 changing the comparison.", "relevance": 2, "relevance_reason": "General cloud-inference TCO method (applicable to batch-1 robot decode on cloud GPUs); not robotics.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q2: arXiv abs:\"total cost of ownership\" AND abs:inference"}
{"id": "2510.02457", "title": "Assessing the Potential for Catastrophic Failure in Dynamic Post-Training Quantization", "authors": ["Logan Frank", "Paul Ardis"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.02457", "pdf_url": "https://arxiv.org/pdf/2510.02457", "code_url": "", "clusters": ["C2", "C7"], "keywords": ["dynamic PTQ", "catastrophic failure", "safety-critical"], "model_class": "other", "techniques": ["post-training quantization robustness analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns adversarial network/bit-width pairs to reveal input distributions under which dynamic PTQ fails catastrophically, relevant to safety-critical deployment.", "relevance": 2, "relevance_reason": "Generic PTQ failure analysis, not robot-evaluated.", "foundational_technique": false, "verified": true, "found_via": "search: quantization reinforcement learning policy deployment"}
{"id": "2509.23931", "title": "AutoPrune: Each Complexity Deserves a Pruning Policy", "authors": ["Hanshi Wang", "Yuhao Xu", "Zekun Xu", "Jin Gao", "Yufan Liu", "Weiming Hu", "Ke Wang", "Zhipeng Zhang"], "year": 2025, "venue": "NeurIPS 2025", "type": "paper", "url": "https://arxiv.org/abs/2509.23931", "pdf_url": "https://arxiv.org/pdf/2509.23931", "code_url": "https://github.com/AutoLab-SAI-SJTU/AutoPrune", "clusters": ["C2"], "keywords": ["AutoPrune"], "model_class": "other", "techniques": ["token pruning", "adaptive budget"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free complexity-adaptive visual token pruning that sets a logistic retention schedule from visual-text mutual information under a compute budget.", "relevance": 2, "relevance_reason": "Token pruning method evaluated on VLMs and driving VLA settings.", "foundational_technique": true, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND token* AND (prun* OR reduc* OR compress*) [start=100]"}
{"id": "2511.18960", "title": "AVA-VLA: Improving Vision-Language-Action models with Active Visual Attention", "authors": ["Lei Xiao", "Ji-Feng Li", "Juntao Gao", "Feiyang Ye", "Yan Jin", "Jingjing Qian", "Jing Zhang", "Yonghui Wu", "Xiao-Yuan Yu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.18960", "pdf_url": "https://arxiv.org/pdf/2511.18960", "code_url": "", "clusters": ["C1"], "keywords": ["active visual attention", "POMDP"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Reformulates VLA learning as partially observable and adds history-aware active visual attention.", "relevance": 2, "relevance_reason": "Context: VLA (links=14).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2507.15597", "title": "Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos", "authors": ["Hao Luo", "Yicheng Feng", "Wan-Peng Zhang", "Sipeng Zheng", "Ye Wang", "Haoqi Yuan", "Jia-Zhen Liu", "Chao-Yi Xu", "Qin Jin", "Zongqing Lu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.15597", "pdf_url": "https://arxiv.org/pdf/2507.15597", "code_url": "", "clusters": ["C1"], "keywords": ["human video pretraining", "dexterous VLA"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Being-H0: dexterous VLA pretrained on large-scale human videos.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2512.11315", "title": "Benchmarking the Generality of Vision-Language-Action Models", "authors": ["Pranav Guruprasad", "Sudipta Chowdhury", "Harsh Sikka", "Mridul Sharma", "Helen Lu", "Sean Rivera", "Aryan Khurana", "Hangliang Ren", "Yangyue Wang"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2512.11315", "pdf_url": "https://arxiv.org/pdf/2512.11315", "code_url": "", "clusters": ["C4"], "keywords": ["MultiNet", "generality", "cross-domain"], "model_class": "VLA", "techniques": [], "base_model": "GPT-5, Pi0, Magma", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MultiNet v1.0"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Unified benchmark across six capability regimes including continuous control; no evaluated model is consistently general.", "relevance": 2, "relevance_reason": "Benchmark breadth.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA leaderboard"}
{"id": "2511.22555", "title": "Beyond Success: Refining Elegant Robot Manipulation from Mixed-Quality Data via Just-in-Time Intervention", "authors": ["Yanbo Mao", "Jian-Long Fu", "Ruoxuan Zhang", "Hongxia Xie", "Meibao Yao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.22555", "pdf_url": "https://arxiv.org/pdf/2511.22555", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "execution quality"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO-Elegant"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Introduces LIBERO-Elegant and a just-in-time intervention method to refine execution quality from mixed-quality data.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.05941"}
{"id": "2506.07961", "title": "BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models", "authors": ["Pei-Yan Li", "Yixiang Chen", "Hongtao Wu", "Xiao Ma", "Xiangnan Wu", "Yan Huang", "Liang Wang", "Tao Kong", "Tie-Niu Tan"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2506.07961", "pdf_url": "https://arxiv.org/pdf/2506.07961", "code_url": "", "clusters": ["C1"], "keywords": ["3D VLA", "sample efficiency"], "model_class": "VLA", "techniques": ["2D heatmap actions", "3D-to-2D projection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "3D VLA that projects point clouds to multi-view images and predicts 2D heatmaps, aligning inputs and outputs with the VLM for sample-efficient learning.", "relevance": 2, "relevance_reason": "Context: sample-efficient 3D VLA cited by several core items.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2512.00903, 2604.04161"}
{"id": "2510.14959", "title": "CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions", "authors": ["Lizhi Yang", "Blake Werner", "Massimiliano de Sa", "Aaron D. Ames"], "year": 2025, "venue": "ICRA 2026", "type": "paper", "url": "https://arxiv.org/abs/2510.14959", "pdf_url": "https://arxiv.org/pdf/2510.14959", "code_url": "https://github.com/lzyang2000/cbf-rl-navigation-demo", "clusters": ["C3", "C7"], "keywords": ["safety filter", "training-time safety"], "model_class": "other", "techniques": ["CBF", "safe RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Enforces CBFs during RL training (policy modification plus filtered rollouts) so the learned policy internalizes safety instead of relying only on a deployment filter.", "relevance": 2, "relevance_reason": "Reduces reliance on runtime filters; relevant contrast to chunk-level filtering.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"safety filter\" AND (\"learned policy\" OR \"generative policy\" OR \"diffusion policy\" OR VLA)"}
{"id": "2509.04095", "title": "Cloud-Assisted Remote Control for Aerial Robots: From Theory to Proof-of-Concept Implementation", "authors": ["Achilleas Santi Seisa", "Viswa Narayanan Sankaranarayanan", "Gerasimos Damigos", "Sumeet Gajanan Satpute", "George Nikolakopoulos"], "year": 2025, "venue": "CCGridW 2025", "type": "paper", "url": "https://arxiv.org/abs/2509.04095", "pdf_url": "https://arxiv.org/pdf/2509.04095", "code_url": "", "clusters": ["C6"], "keywords": ["aerial robots", "cloud-assisted control", "testing framework"], "model_class": "other", "techniques": ["cloud/edge offloaded control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Framework and proof-of-concept for remotely controlling aerial robots from cloud and edge resources, examining latency and resource management.", "relevance": 2, "relevance_reason": "Cloud-offloaded control testbed.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "2512.22939", "title": "ColaVLA: Leveraging Cognitive Latent Reasoning for Hierarchical Parallel Trajectory Planning in Autonomous Driving", "authors": ["Qihang Peng", "Xuesong Chen", "Chenye Yang", "Shaoshuai Shi", "Hongsheng Li"], "year": 2025, "venue": "CVPR 2026", "type": "paper", "url": "https://arxiv.org/abs/2512.22939", "pdf_url": "https://arxiv.org/pdf/2512.22939", "code_url": "", "clusters": ["C1"], "keywords": ["driving VLA"], "model_class": "VLA", "techniques": ["latent reasoning", "parallel decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["nuScenes"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Moves driving reasoning into latent meta-action embeddings with two VLM passes and a single-pass hierarchical parallel trajectory decoder.", "relevance": 2, "relevance_reason": "Removes AR CoT latency in driving VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (efficient OR lightweight) [start=400]"}
{"id": "2507.16731", "title": "Collaborative Inference and Learning between Edge SLMs and Cloud LLMs: A Survey of Algorithms, Execution, and Open Challenges", "authors": ["Senyao Li", "Haozhao Wang", "Wenchao Xu", "Rui Zhang", "Song Guo", "Jingling Yuan", "Xian Zhong", "Tianwei Zhang", "Ruixuan Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.16731", "pdf_url": "https://arxiv.org/pdf/2507.16731", "code_url": "", "clusters": ["C6"], "keywords": ["edge SLM + cloud LLM", "collaborative inference", "survey"], "model_class": "other", "techniques": ["task assignment", "token-level collaboration", "speculative decoding across edge-cloud"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Survey and taxonomy of edge-small-model / cloud-large-model collaboration for both inference and training.", "relevance": 2, "relevance_reason": "Generic LLM edge-cloud collaboration that robot systems borrow from.", "foundational_technique": true, "verified": true, "found_via": "search query"}
{"id": "2501.10513", "title": "ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments", "authors": ["Rohit Dwivedula", "Sadanand Modak", "Aditya Akella", "Joydeep Biswas", "Daehyeok Kim", "Christopher J. Rossbach"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2501.10513", "pdf_url": "https://arxiv.org/pdf/2501.10513", "code_url": "", "clusters": ["C6"], "keywords": ["onboard resource allocation", "configuration tuning"], "model_class": "other", "techniques": ["runtime profiling", "automatic configuration tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Adaptively reconfigures onboard robot software stacks via runtime profiling to meet performance specifications instead of static over-provisioning.", "relevance": 2, "relevance_reason": "Onboard compute management; tangential.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2510.01711", "title": "Contrastive Representation Regularization for Vision-Language-Action Models", "authors": ["Taeyoung Kim", "Ji-Min Lee", "Myungkyu Koo", "Dongyoung Kim", "Kyungmin Lee", "Changyeon Kim", "Younggyo Seo", "Jinwoo Shin"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.01711", "pdf_url": "https://arxiv.org/pdf/2510.01711", "code_url": "", "clusters": ["C3"], "keywords": ["contrastive regularization"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Robot-state-aware contrastive loss that regularizes VLA representations.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164"}
{"id": "2508.08574", "title": "DeepFleet: Multi-Agent Foundation Models for Mobile Robots", "authors": ["Ameya Agaskar", "Sriram Siva", "William Pickering", "Kyle O'Brien", "Charles Kekeh", "Alexandre Ormiga Galvao Barbosa", "Ang Li", "Brianna Gallo Sarker", "Alicia Chua", "Mayur Nemade", "Charun Thattai", "Jiaming Di", "Isaac Iyengar", "Ramya Dharoor", "Dino Kirouani", "Jimmy Erskine", "Tamir Hegazy", "Scott Niekum", "Usman A. Khan", "Federico Pecora", "Joseph W. Durham"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.08574", "pdf_url": "https://arxiv.org/pdf/2508.08574", "code_url": "", "clusters": ["C6"], "keywords": ["fleet foundation model", "warehouse", "multi-agent"], "model_class": "other", "techniques": ["multi-agent foundation models"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Amazon warehouse fleet data"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Amazon foundation models trained on movement logs of hundreds of thousands of warehouse robots to predict and coordinate fleet behavior.", "relevance": 2, "relevance_reason": "Fleet-level foundation model, not about inference efficiency.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2510.13375", "title": "DepthVLA: Enhancing Vision-Language-Action Models with Depth-Aware Spatial Reasoning", "authors": ["Tian-Yuan Yuan", "Yicheng Liu", "Chenhao Lu", "Zhuo-Guang Chen", "Tao Jiang", "Hang Zhao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.13375", "pdf_url": "https://arxiv.org/pdf/2510.13375", "code_url": "", "clusters": ["C1"], "keywords": ["depth", "spatial reasoning"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a depth-aware spatial reasoning expert to VLAs for better 3D precision.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2502.02858", "title": "Dexterous Safe Control for Humanoids in Cluttered Environments via Projected Safe Set Algorithm", "authors": ["Rui Chen", "Yifan Sun", "Changliu Liu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.02858", "pdf_url": "https://arxiv.org/pdf/2502.02858", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["safe set algorithm", "limb-level collision avoidance", "humanoid"], "model_class": "other", "techniques": ["projected safe set algorithm"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "p-SSA relaxes conflicting limb-level safety constraints to keep humanoid safe control feasible in clutter.", "relevance": 2, "relevance_reason": "Safety filter for humanoid deployment.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2502.05855", "title": "DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control", "authors": ["Junjie Wen", "Yi-Chen Zhu", "Jinming Li", "Zhibin Tang", "Chaomin Shen", "Fei-Fei Feng"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2502.05855", "pdf_url": "https://arxiv.org/pdf/2502.05855", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["diffusion action expert", "cross-embodiment", "efficient training"], "model_class": "VLA", "techniques": ["plug-in diffusion expert", "curriculum training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds a 1B-parameter diffusion action expert to a VLM and trains it with a three-stage embodiment curriculum, targeting more efficient training and adaptation across embodiments.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 37 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2503.16146", "title": "Distributed Split Computing Using Diffusive Metrics for UAV Swarms", "authors": ["Talip Tolga Sarı", "Gökhan Seçinti", "Angelo Trotta"], "year": 2025, "venue": "Journal of Systems Architecture", "type": "paper", "url": "https://arxiv.org/abs/2503.16146", "pdf_url": "https://arxiv.org/pdf/2503.16146", "code_url": "", "clusters": ["C6"], "keywords": ["UAV swarm", "distributed split computing", "early exit"], "model_class": "other", "techniques": ["diffusive-metric split computing", "early exit"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Fully distributed split computing across a UAV swarm that forwards partial inferences to under-used neighbours, with early exits for load surges.", "relevance": 2, "relevance_reason": "Fleet-level distributed inference.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: split computing robot"}
{"id": "2510.16281", "title": "Do What You Say: Steering Vision-Language-Action Models via Runtime Reasoning-Action Alignment Verification", "authors": ["Yi-Lin Wu", "Anqi Li", "Tucker Hermans", "Fabio Ramos", "A. Bajcsy", "Claudia P'erez-D'Arpino"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.16281", "pdf_url": "https://arxiv.org/pdf/2510.16281", "code_url": "", "clusters": ["C4", "C7"], "keywords": ["reasoning-action alignment", "runtime verification", "policy steering"], "model_class": "VLA", "techniques": ["sample candidates + outcome prediction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free runtime steering that samples multiple action sequences from a reasoning VLA and selects the one whose predicted outcome matches its textual plan.", "relevance": 2, "relevance_reason": "Runtime verification of VLA outputs.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2410.04640 (S2)"}
{"id": "2507.04447", "title": "DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge", "authors": ["Wenyao Zhang", "Hong-Si Liu", "Zekun Qi", "Yun-Nan Wang", "Xinqiang Yu", "Jiazhao Zhang", "Runpei Dong", "Jiawei He", "Fan Lu", "He Wang", "Zhizheng Zhang", "Li Yi", "Wen-Jun Zeng", "Xin Jin"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2507.04447", "pdf_url": "https://arxiv.org/pdf/2507.04447", "code_url": "", "clusters": ["C8"], "keywords": ["world-model augmented VLA", "inverse dynamics"], "model_class": "VLA", "techniques": ["world knowledge forecasting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA that forecasts compact dynamic, spatial and semantic world knowledge before predicting actions via inverse dynamics.", "relevance": 2, "relevance_reason": "Context: frequently used world-model-augmented VLA baseline.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2503.02310, 2509.09372"}
{"id": "2510.12796", "title": "DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving", "authors": ["Yingyan Li", "Shuyao Shang", "Weisong Liu", "Bing Zhan", "Haochen Wang", "Yuqi Wang", "Yuntao Chen", "Xiaoman Wang", "Yasong An", "Chufeng Tang", "Lu Hou", "Lue Fan", "Zhaoxiang Zhang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.12796", "pdf_url": "https://arxiv.org/pdf/2510.12796", "code_url": "", "clusters": ["C8"], "keywords": ["driving VLA"], "model_class": "VLA", "techniques": ["world-model supervision", "lightweight action expert"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds future-image world-model prediction as dense supervision for driving VLAs and a lightweight action expert for real-time inference.", "relevance": 2, "relevance_reason": "World modeling plus a light action head for latency.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv abs:\"vision-language-action\" AND (accelerat* OR \"inference speed\" OR \"inference latency\" OR speedup) [start=100]"}
{"id": "2510.27607", "title": "Dual-Stream Diffusion for World-Model Augmented Vision-Language-Action Model", "authors": ["John Won", "Kyungmin Lee", "Huiwon Jang", "Dongyoung Kim", "Jinwoo Shin"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.27607", "pdf_url": "https://arxiv.org/pdf/2510.27607", "code_url": "", "clusters": ["C8"], "keywords": ["world-model augmented VLA", "inference-time scaling"], "model_class": "world-action model", "techniques": ["dual-stream diffusion", "asynchronous sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboCasa", "GR-1"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Multimodal diffusion transformer with separate state and action streams and decoupled noise, plus asynchronous action/vision sampling for inference-time scaling.", "relevance": 2, "relevance_reason": "Context: world-model-augmented VLA cited by several core items.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2601.16163, 2602.15922"}
{"id": "dyna-1-pretraining-blog-2025", "title": "DYNA-1 Pre-Training: Zero-Shot Dexterity Is Here", "authors": ["Dyna Robotics"], "year": 2025, "venue": "blog (Dyna Robotics)", "type": "blog", "url": "https://www.dyna.co/research/pre-training", "pdf_url": "", "code_url": "", "clusters": ["C4", "C9"], "keywords": ["pre-training", "zero-shot dexterity", "data-efficient fine-tuning"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Announces a DYNA-1 base model doing laundry folding and package sorting without post-training and adapting from about an hour of demonstrations; no technical specs.", "relevance": 2, "relevance_reason": "Data-efficient adaptation claim; low technical content.", "foundational_technique": false, "verified": true, "found_via": "blog sweep: Dyna"}
{"id": "10.1109/iccwamtip68645.2025.11352704", "title": "Efficient Inference for Vision-Language-Action Models: A Comprehensive Review of Acceleration Techniques", "authors": ["Fu-Cheng Zhou"], "year": 2025, "venue": "ICCWAMTIP 2025", "type": "paper", "url": "https://doi.org/10.1109/ICCWAMTIP68645.2025.11352704", "pdf_url": "", "code_url": "", "clusters": ["C1", "C2"], "keywords": ["survey", "VLA acceleration"], "model_class": "VLA", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Review of VLA inference acceleration organized by input optimization and model-side techniques, framed around sub-100 ms control requirements.", "relevance": 2, "relevance_reason": "Minor survey.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2411.02359"}
{"id": "2505.11709", "title": "EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video", "authors": ["Ryan Hoque", "Peide Huang", "David J. Yoon", "Mouli Sivapurapu", "Jian Zhang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2505.11709", "pdf_url": "https://arxiv.org/pdf/2505.11709", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "egocentric video", "dexterous"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large egocentric human video dataset with 3D hand and finger tracking, aimed at learning dexterous manipulation from passive human data.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2402.15391; cited by 2510.24795"}
{"id": "2507.12440", "title": "EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos", "authors": ["Ruihan Yang", "Qin-Xi Yu", "Ye-Cheng Wu", "Rui Yan", "Bo-Rui Li", "An-Chieh Cheng", "Xueyan Zou", "Yun-Hao Fang", "Xu-Xin Cheng", "Ri-Zhao Qiu", "Hongxu Yin", "Sifei Liu", "Song Han", "Yao Lu", "Xiaolong Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.12440", "pdf_url": "https://arxiv.org/pdf/2507.12440", "code_url": "", "clusters": ["C1"], "keywords": ["egocentric human video", "VLA"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Trains VLAs on egocentric human videos and maps human actions to robot actions.", "relevance": 2, "relevance_reason": "Context: VLA (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2509.17750", "title": "EigenSafe: A Spectral Framework for Learning-Based Probabilistic Safety Assessment", "authors": ["Inkyu Jang", "Jonghae Park", "Sihyun Cho", "Chams E. Mballo", "Claire J. Tomlin", "H. Jin Kim"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.17750", "pdf_url": "https://arxiv.org/pdf/2509.17750", "code_url": "", "clusters": ["C7"], "keywords": ["probabilistic safety"], "model_class": "other", "techniques": ["operator theory", "safety critic"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns the dominant eigenpair of a safety-probability operator to assess and constrain stochastic learned policies.", "relevance": 2, "relevance_reason": "Probabilistic safety assessment of learned policies.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:reachability AND (\"visuomotor policy\" OR \"learned policy\" OR \"robot foundation model\" OR \"vision-language-action\")"}
{"id": "2509.15273", "title": "Embodied Arena: A Comprehensive, Unified, and Evolving Evaluation Platform for Embodied AI", "authors": ["Fei Ni", "Min Zhang", "Pengyi Li", "Yifu Yuan", "Lingfeng Zhang", "Yuecheng Liu", "Peilong Han", "Longxin Kou", "Shaojin Ma", "Jinbin Qiao", "David Gamaliel Arcos Bravo", "Yuening Wang", "Xiao Hu", "Zhanguang Zhang", "Xianze Yao", "Yutong Li", "Zhao Zhang", "Ying Wen", "Ying-Cong Chen", "Xiaodan Liang", "Liang Lin", "Bin He", "Haitham Bou-Ammar", "He Wang", "Huazhe Xu"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2509.15273", "pdf_url": "https://arxiv.org/pdf/2509.15273", "code_url": "", "clusters": ["C4"], "keywords": ["Embodied Arena", "unified evaluation platform", "capability taxonomy", "leaderboard"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["22 integrated benchmarks"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Unified evolving platform integrating 22 embodied benchmarks (EQA, navigation, task planning) under a capability taxonomy with leaderboards for 30+ models.", "relevance": 2, "relevance_reason": "Unified evaluation infrastructure (non-manipulation focus).", "foundational_technique": false, "verified": true, "found_via": "search query: Embodied Arena"}
{"id": "2501.01895", "title": "EnerVerse: Envisioning Embodied Future Space for Robotics Manipulation", "authors": ["Siyuan Huang", "Liliang Chen", "Peng-Fei Zhou", "Shengcong Chen", "Yue Liao", "Zhengkai Jiang", "Yue Hu", "Peng Gao", "Hongsheng Li", "Maoqing Yao", "Guanghui Ren"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2501.01895", "pdf_url": "https://arxiv.org/pdf/2501.01895", "code_url": "", "clusters": ["C8"], "keywords": ["video diffusion", "multi-view", "future space"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Chunk-wise autoregressive multi-view video diffusion model that imagines future embodied space for manipulation planning.", "relevance": 2, "relevance_reason": "Context: generative robot world model (links=14).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2508.21112", "title": "EO-1: An Open Unified Embodied Foundation Model for General Robot Control", "authors": ["Delin Qu", "Haoming Song", "Qizhi Chen", "Zhao-Qing Chen", "Xianqiang Gao", "Dong Wang", "Xinyi Ye", "Qi Lv", "Mo-Di Shi", "Guanghui Ren", "Cheng Ruan", "Maoqing Yao", "Haoran Yang", "Jiacheng Bao", "Bin Zhao", "Xue-Long Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.21112", "pdf_url": "https://arxiv.org/pdf/2508.21112", "code_url": "", "clusters": ["C1"], "keywords": ["foundation model", "interleaved pretraining"], "model_class": "VLA", "techniques": ["interleaved reasoning-action pretraining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "EO-1 is an open unified embodied foundation model trained on interleaved vision-text-action data to combine multimodal reasoning with robot control.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310; cited by 2512.20188, 2601.09708"}
{"id": "2511.16166", "title": "EvoVLA: Self-Evolving Vision-Language-Action Model", "authors": ["Zeting Liu", "Zida Yang", "Zeyu Zhang", "Hao Tang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.16166", "pdf_url": "https://arxiv.org/pdf/2511.16166", "code_url": "https://github.com/AIGeeksGroup/EvoVLA", "clusters": ["C3"], "keywords": ["self-evolving", "intrinsic reward", "long-horizon"], "model_class": "VLA", "techniques": ["RL fine-tuning", "intrinsic rewards"], "base_model": "OpenVLA-OFT", "params": "", "hardware": "", "reported_metric": "1.5x sample efficiency", "metric_measures": "other", "benchmarks": ["Discoverse-L", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Self-supervised RL post-training with stage-aligned rewards and pose-based curiosity for long-horizon VLA tasks.", "relevance": 2, "relevance_reason": "Sample-efficiency claim for post-training; efficiency not central.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: parameter-efficient fine-tuning vision-language-action model arXiv"}
{"id": "10.1109/ictc66702.2025.11389105", "title": "Exploring Policy Distillation for Vision-Language-Action Generalist Robotic Policies", "authors": ["Ingook Jang", "Samyeul Noh", "Seonghyun Kim"], "year": 2025, "venue": "ICTC 2025", "type": "paper", "url": "https://doi.org/10.1109/ICTC66702.2025.11389105", "pdf_url": "", "code_url": "", "clusters": ["C2"], "keywords": ["policy distillation overview", "RPD", "RLDG", "HumanVLA"], "model_class": "VLA", "techniques": ["policy distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Short overview contrasting distilling VLA generalists into compact RL experts, using RL experts to train generalists, and distilling privileged teachers into VLAs.", "relevance": 2, "relevance_reason": "Brief overview of VLA policy distillation.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246"}
{"id": "2505.15660", "title": "Exploring the Limits of Vision-Language-Action Manipulations in Cross-task Generalization", "authors": ["Jia-Ming Zhou", "Ke Ye", "Jiayi Liu", "Teli Ma", "Zi-Fan Wang", "Ronghe Qiu", "Kun-Yu Lin", "Zhi-Lin Zhao", "Junwei Liang"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2505.15660", "pdf_url": "https://arxiv.org/pdf/2505.15660", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "cross-task generalization", "AGNOSTOS"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Simulation benchmark probing zero-shot cross-task generalization of VLAs.", "relevance": 2, "relevance_reason": "Context: VLA benchmark (links=14).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2509.06951", "title": "F1: A Vision-Language-Action Model Bridging Understanding and Generation to Actions", "authors": ["Qi Lv", "Weijie Kong", "Hao Li", "Jia Zeng", "Zherui Qiu", "Delin Qu", "Haoming Song", "Qizhi Chen", "Xiang Deng", "Jiang-Miao Pang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.06951", "pdf_url": "https://arxiv.org/pdf/2509.06951", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["visual foresight", "understanding+generation"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "F1 VLA integrates visual foresight generation into the decision pipeline, bridging understanding, future prediction and action.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 18 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2507.17903", "title": "Federated Learning for Large-Scale Cloud Robotic Manipulation: Opportunities and Challenges", "authors": ["Obaidullah Zaland", "Chanh Nguyen", "Florian T. Pokorny", "Monowar Bhuyan"], "year": 2025, "venue": "ICMLC 2025", "type": "paper", "url": "https://arxiv.org/abs/2507.17903", "pdf_url": "https://arxiv.org/pdf/2507.17903", "code_url": "", "clusters": ["C3", "C6"], "keywords": ["federated learning", "cloud robotic manipulation"], "model_class": "other", "techniques": ["federated learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Position paper on federated learning for large-scale cloud robotic manipulation fleets, outlining opportunities and challenges.", "relevance": 2, "relevance_reason": "Fleet-scale distributed learning perspective.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "2501.01005", "title": "FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving", "authors": ["Zihao Ye", "Lequn Chen", "Ruihang Lai", "Wuwei Lin", "Yineng Zhang", "Stephanie Wang", "Tian-Qi Chen", "Baris Kasikci", "Vinod Grover", "Arvind Krishnamurthy", "Luis Ceze"], "year": 2025, "venue": "MLSys 2025", "type": "paper", "url": "https://arxiv.org/abs/2501.01005", "pdf_url": "https://arxiv.org/pdf/2501.01005", "code_url": "https://github.com/flashinfer-ai/flashinfer", "clusters": ["C5"], "keywords": ["attention engine", "JIT", "CUDA graphs", "block-sparse KV"], "model_class": "other", "techniques": ["customizable attention kernels", "load-balanced scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "JIT-compiled, customizable attention kernel library with block-sparse KV formats and CUDA-graph-compatible scheduling for LLM serving.", "relevance": 2, "relevance_reason": "Kernel library underlying LLM/VLA serving stacks.", "foundational_technique": true, "verified": true, "found_via": "snowball from 2407.08608"}
{"id": "2503.08950", "title": "FP3: A 3D Foundation Policy for Robotic Manipulation", "authors": ["Rujia Yang", "Geng Chen", "Chuan Wen", "Yang Gao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.08950", "pdf_url": "https://arxiv.org/pdf/2503.08950", "code_url": "", "clusters": ["C3"], "keywords": ["3D foundation policy", "point cloud", "data-efficient fine-tuning"], "model_class": "diffusion/flow policy", "techniques": ["fine-tuning", "LoRA"], "base_model": "", "params": "1.3B", "hardware": "", "reported_metric": ">90% on new task with 80 demos", "metric_measures": "other", "benchmarks": ["real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Point-cloud diffusion-transformer foundation policy that fine-tunes to new tasks with ~80 demonstrations.", "relevance": 2, "relevance_reason": "Data-efficient fine-tuning of a foundation policy.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: adapter tuning robot foundation model prompt tuning visuomotor policy efficient transfer"}
{"id": "2509.19870", "title": "FreezeVLA: Action-Freezing Attacks against Vision-Language-Action Models", "authors": ["Xin Wang", "Jie Li", "Ze-Jia Weng", "Yi-Xu Wang", "Yi-Feng Gao", "Tian-Yu Pang", "Chao Du", "Yan Teng", "Ying-Chun Wang", "Zu-Xuan Wu", "Xing-Jun Ma", "Yu-Gang Jiang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.19870", "pdf_url": "https://arxiv.org/pdf/2509.19870", "code_url": "", "clusters": ["C7", "C8"], "keywords": ["FreezeVLA", "action-freezing attack", "adversarial images"], "model_class": "VLA", "techniques": ["min-max adversarial optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "76.2% average attack success", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adversarial images that freeze VLAs into inaction regardless of later instructions, exposing a safety risk during critical interventions.", "relevance": 2, "relevance_reason": "Timing-related safety failure (inaction).", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.09937 (S2)"}
{"id": "2506.09930", "title": "From Intention to Execution: Probing the Generalization Boundaries of Vision-Language-Action Models", "authors": ["Irving Fang", "Jue-Xiao Zhang", "Shengbang Tong", "Chen Feng"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.09930", "pdf_url": "https://arxiv.org/pdf/2506.09930", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "generalization boundaries"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmark suite probing where VLA generalization breaks from intention to execution.", "relevance": 2, "relevance_reason": "Context: VLA benchmark (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2502.13406", "title": "Generative Predictive Control: Flow Matching Policies for Dynamic and Difficult-to-Demonstrate Tasks", "authors": ["Vince Kurtz", "Joel W. Burdick"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.13406", "pdf_url": "https://arxiv.org/pdf/2502.13406", "code_url": "", "clusters": ["C7"], "keywords": ["generative predictive control", "flow matching", "sampling-based MPC", "fast dynamics"], "model_class": "diffusion/flow policy", "techniques": ["flow matching + sampling-based predictive control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Connects sampling-based predictive control with generative flow policies to learn fast-dynamics tasks from simulation instead of demonstrations.", "relevance": 2, "relevance_reason": "Generative policies for dynamic, time-critical tasks.", "foundational_technique": false, "verified": true, "found_via": "snowball: citations of 2506.07339 (S2)"}
{"id": "2511.19861", "title": "GigaWorld-0: World Models as Data Engine to Empower Embodied AI", "authors": ["GigaWorld Team", "An-Gen Ye", "Boyuan Wang", "Chaojun Ni", "Guan Huang", "Guosheng Zhao", "Haoyun Li", "Jiagang Zhu", "Ke-Rui Li", "Mengyuan Xu", "Qiuping Deng", "Si-Ting Wang", "Wen-Kang Qin", "Xinze Chen", "Xiao-Feng Wang", "Yan-Kai Wang", "Yu Cao", "Yi-Fan Chang", "Yuan Xu", "Yun Ye", "Yang Wang", "Yu-Kun Zhou", "Zheng-Yu Zhang", "Zhehao Dong", "Zheng Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.19861", "pdf_url": "https://arxiv.org/pdf/2511.19861", "code_url": "", "clusters": ["C8"], "keywords": ["world model data engine"], "model_class": "world model", "techniques": ["video generation", "3DGS"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "World-model framework combining video generation and 3D generative/physics modules to synthesize embodied training data for VLAs.", "relevance": 2, "relevance_reason": "Context: world-model-as-data-engine platform.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2603.28489, 2604.16405"}
{"id": "2507.15493", "title": "GR-3 Technical Report", "authors": ["Chilam Cheang", "Sijin Chen", "Zhongren Cui", "Yingdong Hu", "Liqun Huang", "Tao Kong", "Hang Li", "Yifeng Li", "Yuxiao Liu", "Xiao Ma", "Hao Niu", "Wenxuan Ou", "Wanli Peng", "Zeyu Ren", "Haixin Shi", "Jiawen Tian", "Hongtao Wu", "Xin Xiao", "Yuyang Xiao", "Jiafeng Xu", "Yichu Yang"], "year": 2025, "venue": "arXiv", "type": "tech-report", "url": "https://arxiv.org/abs/2507.15493", "pdf_url": "https://arxiv.org/pdf/2507.15493", "code_url": "", "clusters": ["C1"], "keywords": ["GR-3", "ByteDance", "generalist VLA"], "model_class": "VLA", "techniques": ["co-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "ByteDance Seed's large VLA with web/human-VR/robot co-training and a bimanual mobile robot platform.", "relevance": 2, "relevance_reason": "Industrial VLA; efficiency secondary.", "foundational_technique": false, "verified": true, "found_via": "search query: on-device VLA"}
{"id": "2512.01801", "title": "GR-RL: Going Dexterous and Precise for Long-Horizon Robotic Manipulation", "authors": ["Yunfei Li", "Xiao Ma", "Jia-Feng Xu", "Yu-Chen Cui", "Zhongren Cui", "Zhigang Han", "Liqun Huang", "Tao Kong", "Yuxiao Liu", "H. Niu", "Wanli Peng", "Jingchao Qiao", "Zeyu Ren", "Haixin Shi", "Zhi Su", "Jiawen Tian", "Yuyang Xiao", "Shen-Yu Zhang", "Li-Wei Zheng", "Hang Li", "Yonghui Wu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.01801", "pdf_url": "https://arxiv.org/pdf/2512.01801", "code_url": "", "clusters": ["C3"], "keywords": ["dexterous", "long-horizon", "RL post-training"], "model_class": "VLA", "techniques": ["RL fine-tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Multi-stage pipeline that turns a generalist VLA into a precise long-horizon dexterous specialist by filtering demos and applying online RL.", "relevance": 2, "relevance_reason": "Context: major VLA RL post-training work (links=23).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2508.07648", "title": "Grasp-HGN: Grasping the Unexpected", "authors": ["Mehrshad Zandigohar", "Mallesham Dasari", "Gunar Schirner"], "year": 2025, "venue": "ACM TECS", "type": "paper", "url": "https://arxiv.org/abs/2508.07648", "pdf_url": "https://arxiv.org/pdf/2508.07648", "code_url": "", "clusters": ["C6"], "keywords": ["prosthetic hand", "grasp VLM", "hybrid edge-cloud"], "model_class": "other", "techniques": ["edge fast model + cloud VLM"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Hybrid Grasp Network uses a fast edge grasp estimator with a cloud grasp VLM for accuracy on unseen objects in prosthetic control.", "relevance": 2, "relevance_reason": "Edge/cloud hybrid for embodied grasping.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:edge-cloud AND abs:robot"}
{"id": "2505.03233", "title": "GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data", "authors": ["Shengliang Deng", "Mi Yan", "Songlin Wei", "Hai-Liang Ma", "Yu-Xin Yang", "Jiayi Chen", "Zhi-Qi Zhang", "Tao Yang", "Xuheng Zhang", "He-Ming Cui", "Zhizheng Zhang", "He Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.03233", "pdf_url": "https://arxiv.org/pdf/2505.03233", "code_url": "", "clusters": ["C1"], "keywords": ["grasping foundation model", "synthetic data"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Grasping VLA pretrained on billion-scale synthetic action data, transferring zero-shot to real scenes.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2511.06619", "title": "How Do VLAs Effectively Inherit from VLMs?", "authors": ["Chuheng Zhang", "Rushuai Yang", "Xiaoyu Chen", "Kaixin Wang", "Li Zhao", "Yi Chen", "Jiang Bian"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.06619", "pdf_url": "https://arxiv.org/pdf/2511.06619", "code_url": "", "clusters": ["C3"], "keywords": ["GrinningFace", "VLM prior inheritance", "fine-tuning strategies"], "model_class": "VLA", "techniques": ["fine-tuning strategy study", "co-training", "LoRA"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Diagnostic benchmark studying how fine-tuning choices affect VLM-knowledge transfer into VLAs.", "relevance": 2, "relevance_reason": "Informs adaptation recipes.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"robot\" AND abs:\"parameter-efficient fine-tuning\""}
{"id": "2503.10631", "title": "HybridVLA: Collaborative Diffusion and Autoregression in a Unified Vision-Language-Action Model", "authors": ["Jia-Ming Liu", "Hao Chen", "Pengju An", "Zhuoyang Liu", "Renrui Zhang", "Chenyang Gu", "Xiaoqi Li", "Ziyu Guo", "Si-Xiang Chen", "Meng-Zhen Liu", "Chengkai Hou", "Meng-Di Zhao", "K. Zhou", "Pheng-Ann Heng", "Shanghang Zhang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2503.10631", "pdf_url": "https://arxiv.org/pdf/2503.10631", "code_url": "", "clusters": ["C1"], "keywords": ["action ensemble", "diffusion head"], "model_class": "VLA", "techniques": ["hybrid diffusion+autoregression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Combines autoregressive token prediction and diffusion denoising for actions inside a single LLM and adaptively ensembles the two predictions.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 33 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.07503"}
{"id": "2509.11766", "title": "Igniting VLMs toward the Embodied Space", "authors": ["Andy Zhai", "Brae Liu", "B. Fang", "Chalse Cai", "Ellie Ma", "Ethan Yin", "Hao Wang", "Hugo Zhou", "James Wang", "Lights Shi", "Lucy Liang", "Mark Wang", "Qian Wang", "Roy Gan", "Ryan Yu", "Shalfun Li", "S.-T. Liu", "Sylas Chen", "V. Chen", "Zunnan Xu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.11766", "pdf_url": "https://arxiv.org/pdf/2509.11766", "code_url": "", "clusters": ["C1"], "keywords": ["WALL-OSS", "embodied foundation model"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "End-to-end embodied foundation model that tightly couples VLM understanding with action generation for embodied tasks.", "relevance": 2, "relevance_reason": "Context: open embodied foundation model (links=15).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2507.17520", "title": "InstructVLA: Vision-Language-Action Instruction Tuning from Understanding to Manipulation", "authors": ["Shuai Yang", "Hao Li", "Yi-Lun Chen", "Bin Wang", "Yang Tian", "Tai Wang", "Han-Qing Wang", "Feng Zhao", "Yiyi Liao", "Jiang-Miao Pang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.17520", "pdf_url": "https://arxiv.org/pdf/2507.17520", "code_url": "", "clusters": ["C3"], "keywords": ["instruction tuning", "catastrophic forgetting", "multimodal reasoning"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA trained with instruction tuning to keep the VLM's multimodal reasoning while learning precise manipulation, limiting catastrophic forgetting.", "relevance": 2, "relevance_reason": "Context: frequently compared VLA (links=18).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2306.03310"}
{"id": "2506.09581", "title": "Integrating Quantized LLMs into Robotics Systems as Edge AI to Leverage their Natural Language Processing Capabilities", "authors": ["Miguel Á. González-Santamarta", "Francisco J. Rodríguez-Lera", "David Sobrín-Hidalgo", "Ángel Manuel Guerrero-Higueras", "Vicente MatellÁn-Olivera"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.09581", "pdf_url": "https://arxiv.org/pdf/2506.09581", "code_url": "", "clusters": ["C5"], "keywords": ["llama.cpp", "ROS 2", "quantized LLM", "edge robotics"], "model_class": "other", "techniques": ["quantized LLM runtime integration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "llama_ros integrates llama.cpp-quantized LLMs into ROS 2 for onboard planning and explainability on resource-constrained robots.", "relevance": 2, "relevance_reason": "Tooling for edge LLM deployment on robots.", "foundational_technique": false, "verified": true, "found_via": "arXiv: llama.cpp AND (robot OR VLA)"}
{"id": "2505.02152", "title": "Interleave-VLA: Enhancing Robot Manipulation with Interleaved Image-Text Instructions", "authors": ["Cunxin Fan", "Xiaosong Jia", "Yihang Sun", "Yi-Xiao Wang", "Jiang-Lan Wei", "Ziyang Gong", "Xiang-Yu Zhao", "Masayoshi Tomizuka", "Xue Yang", "Junchi Yan", "Mingyu Ding"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.02152", "pdf_url": "https://arxiv.org/pdf/2505.02152", "code_url": "", "clusters": ["C1"], "keywords": ["interleaved instructions"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA that accepts interleaved image-text instructions for better zero-shot generalization.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2405.05941"}
{"id": "2511.16651", "title": "InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy", "authors": ["Yang Tian", "Yu-Yin Yang", "Yiman Xie", "Ze-Tao Cai", "Xu Shi", "Ning Gao", "Hangxu Liu", "Xuekun Jiang", "Zherui Qiu", "Feng Yuan", "Ya-Ping Li", "Ping Wang", "Jun-Hao Cai", "Jia Zeng", "Hao Dong", "Jiang-Miao Pang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.16651", "pdf_url": "https://arxiv.org/pdf/2511.16651", "code_url": "", "clusters": ["C4"], "keywords": ["synthetic data", "dataset"], "model_class": "other", "techniques": ["synthetic dataset"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "InternData-A1 is a large high-fidelity synthetic dataset shown to match real-robot data for pretraining a VLA.", "relevance": 2, "relevance_reason": "context: major pretraining dataset", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2410.24164; cited by 2510.24795, 2601.21998"}
{"id": "2510.13778", "title": "InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy", "authors": ["Xinyi Chen", "Yi-Lun Chen", "Yanwei Fu", "Ning Gao", "Jia-Ya Jia", "Wei-Yang Jin", "Hao Li", "Yao Mu", "Jiang-Miao Pang", "Yu Qiao", "Yang Tian", "Bin Wang", "Bo-Lun Wang", "Fangjing Wang", "Han-Qing Wang", "Tai Wang", "Ziqin Wang", "Xue-Yuan Wei", "Chao Wu", "Shuai Yang", "Ji-lu Ye", "Jun-Qiu Yu", "Jia Zeng", "Jing-Jing Zhang", "Jin-Yu Zhang", "Shi Zhang", "Feng Zheng", "Bo-Wen Zhou", "Yang-Kun Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.13778", "pdf_url": "https://arxiv.org/pdf/2510.13778", "code_url": "", "clusters": ["C1"], "keywords": ["spatially guided VLA", "SimplerEnv", "LIBERO"], "model_class": "VLA", "techniques": ["spatial grounding pretraining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SimplerEnv", "LIBERO"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Two-stage VLA recipe: large-scale spatial grounding pretraining followed by spatially prompted action post-training.", "relevance": 2, "relevance_reason": "Context: open generalist VLA used as a baseline.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2605.09948, 2606.04233"}
{"id": "2509.02596", "title": "Introducing LCOAI: A Standardized Economic Metric for Evaluating AI Deployment Costs", "authors": ["Eliseo Curcio"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.02596", "pdf_url": "https://arxiv.org/pdf/2509.02596", "code_url": "", "clusters": ["C6"], "keywords": ["levelized cost", "CAPEX/OPEX", "self-hosted vs API"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Proposes a levelized cost per valid inference (CAPEX+OPEX) metric, analogous to LCOE, to compare API and self-hosted deployments.", "relevance": 2, "relevance_reason": "General economic metric template applicable to $/robot-hour comparisons; not robotics.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q2: arXiv abs:\"total cost of ownership\" AND abs:inference"}
{"id": "2508.06571", "title": "IRL-VLA: Training an Vision-Language-Action Policy via Reward World Model", "authors": ["Anqing Jiang", "Yu Gao", "Yiru Wang", "Zhigang Sun", "Shuo Wang", "Yuwen Heng", "Hao Sun", "Shichen Tang", "Lijuan Zhu", "Jinhao Chai", "Jijun Wang", "Zichong Gu", "Hao Jiang", "Li Sun"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.06571", "pdf_url": "https://arxiv.org/pdf/2508.06571", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["IRL-VLA"], "model_class": "VLA", "techniques": ["inverse RL", "reward world model", "PPO"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["NAVSIM v2"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Driving VLA trained with PPO against a lightweight IRL-learned reward world model instead of a sensor simulator.", "relevance": 2, "relevance_reason": "Efficient closed-loop RL via reward world model.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=100]"}
{"id": "2509.16853", "title": "ISCS: Parameter-Guided Feature Pruning for Resource-Constrained Embodied Perception", "authors": ["Jinhao Wang", "Nam Ling", "Wei Wang", "Wei Jiang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.16853", "pdf_url": "https://arxiv.org/pdf/2509.16853", "code_url": "", "clusters": ["C6"], "keywords": ["feature pruning", "edge-robot collaboration", "channel importance"], "model_class": "other", "techniques": ["salient channel selection for transmission"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Selects structure-critical encoder channels from weight statistics so robots transmit only salient features to edge servers.", "relevance": 2, "relevance_reason": "Bandwidth reduction for edge-robot split perception.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: split computing robot"}
{"id": "2505.07417", "title": "LA-IMR: Latency-Aware, Predictive In-Memory Routing and Proactive Autoscaling for Tail-Latency-Sensitive Cloud Robotics", "authors": ["Seo, Eunil", "Nguyen, Chanh", "Elmroth, Erik"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.07417", "pdf_url": "https://arxiv.org/pdf/2505.07417", "code_url": "", "clusters": ["C4", "C5"], "keywords": ["LA-IMR", "P99 latency", "cloud robotics", "autoscaling"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "other", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Latency model decomposing processing, network and queuing to route and autoscale inference for cloud robotics, targeting 99th-percentile delay.", "relevance": 2, "relevance_reason": "Tail-latency framing for offloaded robot inference.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: tail latency p99 VLA robot inference deadline"}
{"id": "2512.15840", "title": "Large Video Planner Enables Generalizable Robot Control", "authors": ["Boyuan Chen", "Tianyuan Zhang", "Hao-Ran Geng", "Kiwhan Song", "Cai-Yi Zhang", "Peihao Li", "William T. Freeman", "Jitendra Malik", "Pieter Abbeel", "Russ Tedrake", "Vincent Sitzmann", "Yilun Du"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2512.15840", "pdf_url": "https://arxiv.org/pdf/2512.15840", "code_url": "", "clusters": ["C8"], "keywords": ["video planner", "generalist control"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large video-generation planner used as the backbone for generalizable robot control instead of an MLLM-based VLA.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 13 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2410.24164, 2501.03575"}
{"id": "2509.18428", "title": "Latent Action Pretraining Through World Modeling", "authors": ["Bahey Tharwat", "Yara Nasser", "Ali Abouzeid", "Ian Reid"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.18428", "pdf_url": "https://arxiv.org/pdf/2509.18428", "code_url": "", "clusters": ["C8"], "keywords": ["latent actions", "world modeling"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns latent action representations through world modeling for VLA pretraining on unlabeled data.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2510.27114", "title": "Learning Generalizable Visuomotor Policy through Dynamics-Alignment", "authors": ["Dohyeok Lee", "Jung-Min Lee", "Munkyung Kim", "Seokhun Ju", "Jin Woo Koo", "Kyungjae Lee", "Dohyeong Kim", "Taehyun Cho", "Jungwoo Lee"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.27114", "pdf_url": "https://arxiv.org/pdf/2510.27114", "code_url": "", "clusters": ["C8"], "keywords": ["dynamics alignment", "video prediction"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Aligns visuomotor policies with action-conditioned dynamics to improve generalization.", "relevance": 2, "relevance_reason": "Context: (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2106.09685, 2301.04104"}
{"id": "2512.23864", "title": "Learning to Feel the Future: DreamTacVLA for Contact-Rich Manipulation", "authors": ["Guo Ye", "Ze-Xi Zhang", "Xu Zhao", "Shang Wu", "Hao-Ran Lu", "Shihan Lu", "Han Liu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.23864", "pdf_url": "https://arxiv.org/pdf/2512.23864", "code_url": "", "clusters": ["C1"], "keywords": ["tactile", "contact-rich"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA that predicts future tactile signals for contact-rich manipulation.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2511.05642", "title": "Lite VLA: Efficient Vision-Language-Action Control on CPU-Bound Edge Robots", "authors": ["Justin Williams", "Kishor Datta Gupta", "Roy George", "Mrinmoy Sarkar"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.05642", "pdf_url": "https://arxiv.org/pdf/2511.05642", "code_url": "", "clusters": ["C5"], "keywords": ["CPU-only", "small VLM", "mobile robot", "GPS-denied"], "model_class": "VLA", "techniques": ["small VLM on embedded CPU"], "base_model": "", "params": "", "hardware": "CPU-bound embedded robot", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Shows a compact VLM running on a mobile robot's CPU for concurrent scene reasoning and movement without cloud connectivity.", "relevance": 2, "relevance_reason": "CPU-only onboard deployment; limited quantitative detail.", "foundational_technique": false, "verified": true, "found_via": "arXiv: vision-language-action AND Jetson/edge/onboard"}
{"id": "2509.06932", "title": "LLaDA-VLA: Vision Language Diffusion Action Models", "authors": ["Yu-Qing Wen", "He-Bei Li", "Kefan Gu", "Yucheng Zhao", "Tian-Cai Wang", "Xiaoyan Sun"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.06932", "pdf_url": "https://arxiv.org/pdf/2509.06932", "code_url": "", "clusters": ["C1"], "keywords": ["diffusion VLA"], "model_class": "VLA", "techniques": ["masked diffusion LM"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Builds a VLA on a masked-diffusion language model backbone rather than an autoregressive VLM.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 11 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.05941, 2405.12213"}
{"id": "2507.01957", "title": "Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation", "authors": ["Zhuoyang Zhang", "Luke J. Huang", "Chengyue Wu", "Shang Yang", "Kelly Peng", "Yao Lu", "Song Han"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.01957", "pdf_url": "https://arxiv.org/pdf/2507.01957", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["parallel decoding", "autoregressive generation", "latency"], "model_class": "other", "techniques": ["parallel decoding", "flexible generation order"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Parallelizes next-patch autoregressive image generation with learnable position queries and locality-aware ordering to cut the number of sequential steps.", "relevance": 2, "relevance_reason": "Parallel-decoding technique transferable to autoregressive action tokens.", "foundational_technique": true, "verified": true, "found_via": "author sweep: Zhuoyang Zhang"}
{"id": "2511.16175", "title": "Mantis: A Versatile Vision-Language-Action Model with Disentangled Visual Foresight", "authors": ["Yi Yang", "Xueqi Li", "Yiyang Chen", "Jin Song", "Yi-Han Wang", "Zipeng Xiao", "Jia-Di Su", "Qiao-Ben You", "Peng-Fei Liu", "Zhi-Jie Deng"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2511.16175", "pdf_url": "https://arxiv.org/pdf/2511.16175", "code_url": "", "clusters": ["C1"], "keywords": ["DiT head", "training cost"], "model_class": "VLA", "techniques": ["disentangled visual foresight"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Mantis offloads visual-foresight prediction to a separate DiT head so the VLA backbone keeps language reasoning while training converges quickly.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 11 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "10.1038/s41586-025-08744-2", "title": "Mastering diverse control tasks through world models", "authors": ["Danijar Hafner", "J. Pašukonis", "Jimmy Ba", "Timothy P. Lillicrap"], "year": 2025, "venue": "Nature", "type": "paper", "url": "https://doi.org/10.1038/s41586-025-08744-2", "pdf_url": "https://doi.org/10.1038/s41586-025-08744-2", "code_url": "", "clusters": ["C8"], "keywords": ["DreamerV3", "world model", "MBRL"], "model_class": "world model", "techniques": ["model-based RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DreamerV3: a single world-model-based RL algorithm with fixed hyperparameters that solves a wide range of control domains.", "relevance": 2, "relevance_reason": "Seminal world-model foundation that efficient world-model work builds on", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2310.16828"}
{"id": "2508.19236", "title": "MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation", "authors": ["Hao Shi", "Bin Xie", "Ying-Fei Liu", "Lin Sun", "Feng-Rong Liu", "Tian-Cai Wang", "Erjin Zhou", "Hao-Qiang Fan", "Xiang-Yu Zhang", "Gao Huang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.19236", "pdf_url": "https://arxiv.org/pdf/2508.19236", "code_url": "", "clusters": ["C1"], "keywords": ["memory", "non-Markovian", "long-horizon"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds working- and episodic-style memory banks to a VLA so it can handle temporally dependent, long-horizon manipulation.", "relevance": 2, "relevance_reason": "Context: frequently compared memory-augmented VLA (links=29).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2512.08580", "title": "Mind to Hand: Purposeful Robotic Control via Embodied Reasoning", "authors": ["Peijun Tang", "Shang-Ping Xie", "Bin Sun", "Baifu Huang", "Kuncheng Luo", "Hao Yang", "Weiqi Jin", "Jia-Nan Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.08580", "pdf_url": "https://arxiv.org/pdf/2512.08580", "code_url": "", "clusters": ["C1"], "keywords": ["reasoning VLA"], "model_class": "VLA", "techniques": ["embodied reasoning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Lumo-1 is a generalist VLA that unifies multimodal reasoning with robot action by building on a pretrained VLM.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2510.27065", "title": "MLPerf Automotive", "authors": ["Shojaei, Radoyeh", "Djurdjevic, Predrag", "El-Khamy, Mostafa", "Goel, James", "Mecklenburg, Kasper", "Muyan-Özçelik, Pınar", "Owens, John", "John, Tom St.", "Suh, Jinho", "Suresh, Arjun"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2510.27065", "pdf_url": "https://arxiv.org/pdf/2510.27065", "code_url": "", "clusters": ["C4"], "keywords": ["MLPerf Automotive", "real-time", "safety", "standardized benchmark"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "First MLCommons standardized benchmark for automotive ML acceleration covering perception, end-to-end driving and infotainment with reproducible real-time-aware metrics.", "relevance": 2, "relevance_reason": "Adjacent-domain precedent for real-time embodied benchmarking.", "foundational_technique": true, "verified": true, "found_via": "WebSearch: MLPerf robotics inference benchmark"}
{"id": "2512.01358", "title": "Modality-Augmented Fine-Tuning of Foundation Robot Policies for Cross-Embodiment Manipulation on GR1 and G1", "authors": ["Junsung Park", "Hogun Kee", "Songhwai Oh"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.01358", "pdf_url": "https://arxiv.org/pdf/2512.01358", "code_url": "", "clusters": ["C3"], "keywords": ["humanoid", "GR1", "Unitree G1", "modality augmentation", "cross-embodiment fine-tuning"], "model_class": "VLA", "techniques": ["fine-tuning", "modality augmentation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["GR1", "Unitree G1"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Fine-tunes foundation policies to humanoids with added contact/depth modalities, improving over standard fine-tuning.", "relevance": 2, "relevance_reason": "Cross-embodiment fine-tuning practice.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: adapter tuning robot foundation model prompt tuning visuomotor policy efficient transfer"}
{"id": "2511.21584", "title": "Model-Based Policy Adaptation for Closed-Loop End-to-End Autonomous Driving", "authors": ["Haohong Lin", "Yunzhi Zhang", "Wenhao Ding", "Jiajun Wu", "Ding Zhao"], "year": 2025, "venue": "NeurIPS 2025", "type": "paper", "url": "https://arxiv.org/abs/2511.21584", "pdf_url": "https://arxiv.org/pdf/2511.21584", "code_url": "", "clusters": ["C7", "C8", "C9"], "keywords": ["policy adapter", "Q-value model", "counterfactual data", "closed-loop driving"], "model_class": "diffusion/flow policy", "techniques": ["diffusion policy adapter", "value-guided selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adapts pretrained end-to-end driving policies at deployment using counterfactual simulated data, a diffusion adapter and a multi-step Q model for safer closed-loop behavior.", "relevance": 2, "relevance_reason": "Deployment-time safety adaptation of foundation-style policies.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "2503.13446", "title": "MoManipVLA: Transferring Vision-language-action Models for General Mobile Manipulation", "authors": ["Zhenyu Wu", "Yuheng Zhou", "Xiuwei Xu", "Ziwei Wang", "Haibin Yan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.13446", "pdf_url": "https://arxiv.org/pdf/2503.13446", "code_url": "", "clusters": ["C3"], "keywords": ["MoManipVLA", "mobile manipulation transfer"], "model_class": "VLA", "techniques": ["policy adaptation", "trajectory optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Transfers fixed-base VLAs to mobile manipulation with an efficient adaptation framework using few demos.", "relevance": 2, "relevance_reason": "Efficient embodiment adaptation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"training cost\""}
{"id": "2511.01266", "title": "MotionStream: Real-Time Video Generation with Interactive Motion Controls", "authors": ["Joonghyuk Shin", "Zhengqi Li", "Richard Zhang", "Jun-Yan Zhu", "Jaesik Park", "E. Schechtman", "Xun Huang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.01266", "pdf_url": "https://arxiv.org/pdf/2511.01266", "code_url": "", "clusters": ["C8"], "keywords": ["real-time video generation", "distillation"], "model_class": "other", "techniques": ["causal distillation", "streaming video generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 29 FPS streaming, sub-second latency on a single GPU", "metric_measures": "throughput", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "MotionStream distills a motion-controlled text-to-video model into a causal streaming generator for real-time interactive video generation.", "relevance": 2, "relevance_reason": "foundational technique for real-time video (world-model) generation", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2402.15391, 2506.08009; cited by 2602.06949, 2603.05449"}
{"id": "2512.13030", "title": "Motus: A Unified Latent Action World Model", "authors": ["Hong-Zhe Bi", "Heng-Kai Tan", "Shenghao Xie", "Zeyu Wang", "Shu-He Huang", "Hai-Tian Liu", "Ruowen Zhao", "Yao Feng", "Chendong Xiang", "Yinze Rong", "Hong-Yang Zhao", "Han-Yu Liu", "Zhi-Zhong Su", "Lei Ma", "Hang Su", "Jun Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.13030", "pdf_url": "https://arxiv.org/pdf/2512.13030", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["unified world-action model", "latent action"], "model_class": "world-action model", "techniques": ["mixture-of-transformers", "latent actions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "A mixture-of-transformers model that switches between world-model, VLA, inverse-dynamics and video-generation modes, learning latent actions from optical flow.", "relevance": 2, "relevance_reason": "Context: widely cited unified world-action model baseline.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2510.06710, 2601.21998"}
{"id": "2509.19480", "title": "OmniVLA: An Omni-Modal Vision-Language-Action Model for Robot Navigation", "authors": ["Noriaki Hirose", "Catherine Glossop", "Dhruv Shah", "Sergey Levine"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.19480", "pdf_url": "https://arxiv.org/pdf/2509.19480", "code_url": "", "clusters": ["C1", "C9"], "keywords": ["navigation VLA", "omni-modal goal conditioning"], "model_class": "VLA", "techniques": ["modality fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Navigation VLA conditioned on any mix of 2D pose, goal image and language via randomized modality fusion; used as the base for AsyncVLA.", "relevance": 2, "relevance_reason": "Base model for edge deployment work.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2502.16495", "title": "Orchestrating Joint Offloading and Scheduling for Low-Latency Edge SLAM", "authors": ["Yao Zhang", "Yuyi Mao", "Hui Wang", "Zhiwen Yu", "Song Guo", "Jun Zhang", "Liang Wang", "Bin Guo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.16495", "pdf_url": "https://arxiv.org/pdf/2502.16495", "code_url": "", "clusters": ["C6"], "keywords": ["edge-assisted SLAM", "offloading and scheduling", "time-varying requirements"], "model_class": "other", "techniques": ["joint offloading and scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Architecture that jointly orchestrates offloading and scheduling for multi-client edge-assisted visual SLAM under stochastic inputs and changing requirements.", "relevance": 2, "relevance_reason": "Edge offloading scheduler for robot workloads.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: edge server robot policy"}
{"id": "2509.19012", "title": "Pure Vision Language Action (VLA) Models: A Comprehensive Survey", "authors": ["Da-Peng Zhang", "Jing Sun", "Chenghui Hu", "Xiaoyang Wu", "Zhenlong Yuan", "Rui Zhou", "Fei Shen", "Qingguo Zhou"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2509.19012", "pdf_url": "https://arxiv.org/pdf/2509.19012", "code_url": "", "clusters": ["C1"], "keywords": ["survey", "VLA taxonomy"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Broad survey of pure vision-language-action models, organizing architectures, training paradigms, datasets and open challenges.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 42 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2501.14823", "title": "Quantifying Energy and Cost Benefits of Hybrid Edge Cloud: Analysis of Traditional and Agentic Workloads", "authors": ["Siavash Alamouti"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2501.14823", "pdf_url": "https://arxiv.org/pdf/2501.14823", "code_url": "", "clusters": ["C6"], "keywords": ["hybrid edge cloud", "energy and cost", "agentic workloads"], "model_class": "other", "techniques": ["cost/energy analysis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 75% energy savings and >80% cost reduction (analysis)", "metric_measures": "energy", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Analytical study claiming hybrid edge-cloud placement saves most energy and cost versus centralized cloud for IoT and agentic/robotic workloads.", "relevance": 2, "relevance_reason": "TCO/energy framing of edge vs cloud (not robot-specific experiments).", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:edge-cloud AND abs:robot"}
{"id": "2506.10826", "title": "RationalVLA: A Rational Vision-Language-Action Model with Dual System", "authors": ["Wen-Xuan Song", "Jiayi Chen", "Wenxue Li", "Xu He", "Han Zhao", "Can Cui", "Pengxiang Ding", "Shiyan Su", "Feilong Tang", "Xuelian Cheng", "Dong-Lin Wang", "Zongyuan Ge", "Xin-Hu Zheng", "Zhe Liu", "He-Sheng Wang", "Hao-Ang Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2506.10826", "pdf_url": "https://arxiv.org/pdf/2506.10826", "code_url": "", "clusters": ["C1"], "keywords": ["rational instruction handling"], "model_class": "VLA", "techniques": ["dual-system"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Dual-system VLA that can reject or reason about ambiguous or infeasible instructions.", "relevance": 2, "relevance_reason": "Context: dual-system VLA (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2312.01990"}
{"id": "2508.10333", "title": "ReconVLA: Reconstructive Vision-Language-Action Model as Effective Robot Perceiver", "authors": ["Wen-Xuan Song", "Ziyang Zhou", "Han Zhao", "Jiayi Chen", "Pengxiang Ding", "Hao-Dong Yan", "Yuxin Huang", "Feilong Tang", "Dong-Lin Wang", "Hao-Ang Li"], "year": 2025, "venue": "AAAI Conference on Artificial Intelligence", "type": "paper", "url": "https://arxiv.org/abs/2508.10333", "pdf_url": "https://arxiv.org/pdf/2508.10333", "code_url": "", "clusters": ["C1"], "keywords": ["reconstruction", "visual grounding"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds reconstructive objectives so VLAs focus visual attention on target regions.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310, 2405.12213"}
{"id": "2505.07395", "title": "ReinboT: Amplifying Robot Visual-Language Manipulation with Reinforcement Learning", "authors": ["Hongyin Zhang", "Zifeng Zhuang", "Han Zhao", "Pengxiang Ding", "Hongchao Lu", "Donglin Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.07395", "pdf_url": "https://arxiv.org/pdf/2505.07395", "code_url": "", "clusters": ["C3"], "keywords": ["ReinboT", "dense return", "offline RL"], "model_class": "VLA", "techniques": ["return-conditioned training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "VLA that predicts dense returns to exploit mixed-quality data RL-style.", "relevance": 2, "relevance_reason": "Data efficiency of training.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"few-shot\""}
{"id": "2510.22909", "title": "Rethinking Inference Placement for Deep Learning across Edge and Cloud Platforms: A Multi-Objective Optimization Perspective and Future Directions", "authors": ["Zongshun Zhang", "Ibrahim Matta"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.22909", "pdf_url": "https://arxiv.org/pdf/2510.22909", "code_url": "", "clusters": ["C6"], "keywords": ["inference placement", "multi-objective", "edge-cloud"], "model_class": "other", "techniques": ["model partitioning", "placement optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Position/survey on placing and partitioning deep-learning inference across devices, edge and cloud as a multi-objective optimization.", "relevance": 2, "relevance_reason": "General inference placement background.", "foundational_technique": true, "verified": true, "found_via": "search query"}
{"id": "2511.17441", "title": "RoboCOIN: An Open-Sourced Bimanual Robotic Data COllection for INtegrated Manipulation", "authors": ["Shihan Wu", "Xue-Cheng Liu", "Shao-Xuan Xie", "Peng-Wei Wang", "Xinghang Li", "Bowen Yang", "Zhe Li", "Kai Zhu", "Hongyu Wu", "Yi-Heng Liu", "Zhaoye Long", "Yue Wang", "Chong Liu", "Dihan Wang", "Ziqiang Ni", "Xiang Yang", "You Liu", "Ruoxuan Feng", "Runtian Xu", "Lei Zhang", "Denghang Huang", "Cheng-Hao Jin", "Anlan Yin", "Xinlong Wang", "Zhen-Zhu Sun", "Jun-Kai Zhao", "Mengfei Du", "Mingyu Cao", "Xian-Sheng Chen", "Hong-Yang Cheng", "Xiao-Jie Zhang", "Yan-Kai Fu", "Ning Chen", "Cheng Chi", "Si-Xiang Chen", "Huaihai Lyu", "Xiaoshuai Hao", "Ye-Quan Wang", "Bo-Nan Lei", "Dong Liu", "Xi Yang", "Yan-Ce Jiao", "Tengfei Pan", "Yun-Yun Zhang", "Song Wang", "Zi-Qian Zhang", "Xu Liu", "Ji Zhang", "Cao-Wei Meng", "Zhizheng Zhang", "Jiyang Gao", "Song Wang", "Xiao-Kun Leng", "Zhiqiang Xie", "Zhen-Yu Zhou", "Peng Huang", "Wu Yang", "Yan-Dong Guo", "Yi-Cheng Zhu", "Su-Fen Zheng", "Hao Cheng", "Xin-Xin Ding", "Yang Yue", "Huanqian Wang", "Chi Chen", "Jin-Li Pang", "Yuxi Qian", "Hao Geng", "Lian-Li Gao", "Haiyuan Li", "Bin Fang", "Gao Huang", "Yaodong Yang", "Hao Dong", "He Wang", "Hang Zhao", "Yadong Mu", "Di Hu", "Hao Zhao", "Tiejun Huang", "Shang-Hang Zhang", "Yong-Wen Lin", "Zhong-Yuan Wang", "Guo-Cai Yao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.17441", "pdf_url": "https://arxiv.org/pdf/2511.17441", "code_url": "", "clusters": ["C4"], "keywords": ["bimanual dataset"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open bimanual manipulation dataset collected across heterogeneous platforms.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2510.26536", "title": "RoboOS-NeXT: A Unified Memory-based Framework for Lifelong, Scalable, and Robust Multi-Robot Collaboration", "authors": ["Huajie Tan", "Cheng Chi", "Xiansheng Chen", "Yuheng Ji", "Zhongxia Zhao", "Xiaoshuai Hao", "Yaoxu Lyu", "Mingyu Cao", "Junkai Zhao", "Huaihai Lyu", "Enshen Zhou", "Ning Chen", "Yankai Fu", "Cheng Peng", "Wei Guo", "Dong Liang", "Zhuo Chen", "Mengsi Lyu", "Chenrui He", "Yulong Ao", "Yonghua Lin", "Pengwei Wang", "Zhongyuan Wang", "Shanghang Zhang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.26536", "pdf_url": "https://arxiv.org/pdf/2510.26536", "code_url": "", "clusters": ["C6"], "keywords": ["multi-robot collaboration", "shared memory", "brain-cerebellum"], "model_class": "other", "techniques": ["centralized brain planning + local controllers", "shared spatio-temporal-embodiment memory"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Successor to RoboOS adding a shared spatio-temporal-embodiment memory used by a central brain model for planning while robots' local controllers execute.", "relevance": 2, "relevance_reason": "Hierarchical central/local multi-robot architecture.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: multi-robot VLA"}
{"id": "2504.13059", "title": "RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins", "authors": ["Yao Mu", "Tian-Xing Chen", "Zan-Xin Chen", "Shi-Jia Peng", "Zhiqian Lan", "Zeyu Gao", "Zhi-Xuan Liang", "Qiaojun Yu", "Yude Zou", "Min Xu", "Lunkai Lin", "Zhiqiang Xie", "Mingyu Ding", "Ping Luo"], "year": 2025, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2504.13059", "pdf_url": "https://arxiv.org/pdf/2504.13059", "code_url": "", "clusters": ["C4"], "keywords": ["dual-arm benchmark", "digital twins"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RoboTwin"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Generative digital-twin framework producing dual-arm manipulation data and a sim/real-aligned benchmark.", "relevance": 2, "relevance_reason": "Context: RoboTwin benchmark widely used for VLA evaluation.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2505.19017, 2507.00435"}
{"id": "2504.18904", "title": "RoboVerse: Towards a Unified Platform, Dataset and Benchmark for Scalable and Generalizable Robot Learning", "authors": ["Haoran Geng", "Fei-Shi Wang", "Songlin Wei", "Yuyang Li", "Bangjun Wang", "Bo-Shi An", "C. Cheng", "Haozhe Lou", "Peihao Li", "Yen-Jen Wang", "Yutong Liang", "Dylan Goetting", "Chaoyi Xu", "Haozhe Chen", "Yuxi Qian", "Yiran Geng", "Jiageng Mao", "Weikang Wan", "Ming-Tong Zhang", "Jiangran Lyu", "Siheng Zhao", "Jiazhao Zhang", "Jialiang Zhang", "Chengyang Zhao", "Haoran Lu", "Yufei Ding", "Ran Gong", "Yu-Ran Wang", "Yuxuan Kuang", "Ruihai Wu", "Baoxiong Jia", "C. Sferrazza", "Hao Dong", "Si-Yuan Huang", "Yue Wang", "Jitendra Malik", "Pieter Abbeel"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2504.18904", "pdf_url": "https://arxiv.org/pdf/2504.18904", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "simulation", "dataset"], "model_class": "other", "techniques": ["simulation platform", "dataset"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RoboVerse unifies simulators, a large synthetic dataset and standardized benchmarks for scalable robot learning and evaluation.", "relevance": 2, "relevance_reason": "context: unified platform and benchmark", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2507.05331, 2509.09674"}
{"id": "2511.01331", "title": "RobustVLA: Robustness-Aware Reinforcement Post-Training for Vision-Language-Action Models", "authors": ["Hongyin Zhang", "Shuo Zhang", "Junxi Jin", "Qixin Zeng", "Runze Li", "Donglin Wang"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.01331", "pdf_url": "https://arxiv.org/pdf/2511.01331", "code_url": "", "clusters": ["C3"], "keywords": ["robustness-aware RL", "lightweight post-training"], "model_class": "VLA", "techniques": ["online RL", "regularization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Lightweight online RL post-training that adds robustness to observation and actuation perturbations.", "relevance": 2, "relevance_reason": "Post-training robustness.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2509.15212", "title": "RynnVLA-001: Using Human Demonstrations to Improve Robot Manipulation", "authors": ["Yuming Jiang", "Siteng Huang", "Shengke Xue", "Yaxi Zhao", "Jun Cen", "Si-Cong Leng", "Kehan Li", "Jiayan Guo", "Kexiang Wang", "Mingxiu Chen", "Fan Wang", "Deli Zhao", "Xin Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.15212", "pdf_url": "https://arxiv.org/pdf/2509.15212", "code_url": "", "clusters": ["C1"], "keywords": ["video generative pretraining", "human demonstrations"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RynnVLA-001: VLA pretrained with ego-centric human video generation before action learning.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705, 2410.24164"}
{"id": "2511.17502", "title": "RynnVLA-002: A Unified Vision-Language-Action and World Model", "authors": ["Jun Cen", "Siteng Huang", "Yuqian Yuan", "Kehan Li", "Hangjie Yuan", "Chaohui Yu", "Bo-Han Hou", "Yuming Jiang", "Jiayan Guo", "Xin Li", "Hao Luo", "Fan Wang", "Deli Zhao", "Hao Chen"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2511.17502", "pdf_url": "https://arxiv.org/pdf/2511.17502", "code_url": "", "clusters": ["C8"], "keywords": ["unified VLA-world model"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Couples a VLA and an action-conditioned world model in one framework so that future-image prediction and action generation improve each other.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2510.06710"}
{"id": "spark-safe-protective-assistive-robot-kit-2025", "title": "SPARK: Safe Protective and Assistive Robot Kit", "authors": ["Yifan Sun", "Rui Chen", "Kai S Yun", "Yikuan Fang", "Sebin Jung", "Feihan Li", "Bowei Li", "Weiye Zhao", "Changliu Liu"], "year": 2025, "venue": "IFAC Symposium on Robotics 2025", "type": "repo", "url": "https://intelligent-control-lab.github.io/spark/", "pdf_url": "", "code_url": "https://github.com/intelligent-control-lab/spark", "clusters": ["C8", "C9"], "keywords": ["safe control toolbox", "humanoid", "benchmark"], "model_class": "other", "techniques": ["safety filters"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Open-source toolbox and benchmark of safe-control algorithms for humanoids and other robots.", "relevance": 2, "relevance_reason": "Tooling for safety layers around learned policies.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2510.12276", "title": "Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model", "authors": ["Fuhao Li", "Wen-Xuan Song", "Han Zhao", "Jingbo Wang", "Pengxiang Ding", "Dong-Lin Wang", "L. Zeng", "Hao-Ang Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.12276", "pdf_url": "https://arxiv.org/pdf/2510.12276", "code_url": "", "clusters": ["C1"], "keywords": ["spatial representation", "alignment"], "model_class": "VLA", "techniques": ["representation alignment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Spatial Forcing aligns intermediate VLA visual features with a 3D foundation model's representations, giving spatial awareness without explicit depth inputs.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310; cited by 2511.19433, 2603.16195"}
{"id": "2501.15830", "title": "SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model", "authors": ["Delin Qu", "Haoming Song", "Qizhi Chen", "Yuan-Qi Yao", "Xinyi Ye", "Yani Ding", "Zhi-Gang Wang", "Jia-Yuan Gu", "Bin Zhao", "Dong Wang", "Xuelong Li"], "year": 2025, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2501.15830", "pdf_url": "https://arxiv.org/pdf/2501.15830", "code_url": "", "clusters": ["C1"], "keywords": ["spatial representations", "action grids", "3D position encoding"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA that injects 3D spatial encodings into observations and discretizes motion with adaptive action grids for cross-embodiment manipulation.", "relevance": 2, "relevance_reason": "Context: widely used VLA baseline (links=58).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2512.05107", "title": "STARE-VLA: Progressive Stage-Aware Reinforcement for Fine-Tuning Vision-Language-Action Models", "authors": ["Feng Xu", "Guangyao Zhai", "Xin Kong", "Tingzhong Fu", "Daniel F. N. Gordon", "Xueli An", "Benjamin Busam"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.05107", "pdf_url": "https://arxiv.org/pdf/2512.05107", "code_url": "", "clusters": ["C3"], "keywords": ["stage-aware RL", "TPO", "PPO"], "model_class": "VLA", "techniques": ["preference optimization", "PPO", "stage decomposition"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["SimplerEnv", "ManiSkill"], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Decomposes long-horizon trajectories into stages for dense stage-aligned rewards in TPO/PPO fine-tuning of VLAs.", "relevance": 2, "relevance_reason": "RL fine-tuning credit assignment.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"preference optimization\""}
{"id": "2508.06553", "title": "Static and Plugged: Make Embodied Evaluation Simple", "authors": ["Jiahao Xiao", "Jianbo Zhang", "BoWen Yan", "Shengyu Guo", "Tongrui Ye", "Kaiwei Zhang", "Zicheng Zhang", "Xiaohong Liu", "Zhengxue Cheng", "Lei Fan", "Chuyi Li", "Guangtao Zhai"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2508.06553", "pdf_url": "https://arxiv.org/pdf/2508.06553", "code_url": "", "clusters": ["C4"], "keywords": ["StaticEmbodiedBench", "static evaluation", "leaderboard"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["StaticEmbodiedBench"], "real_robot_eval": false, "onboard_deployment": null, "tldr": "Plug-and-play benchmark using static scene representations (42 scenarios, 8 dimensions) to evaluate 19 VLMs and 11 VLAs without interactive simulation.", "relevance": 2, "relevance_reason": "Low-cost evaluation methodology, but open-loop.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA acceleration evaluation"}
{"id": "2512.02834", "title": "Steering Vision-Language-Action Models as Anti-Exploration: A Test-Time Scaling Approach", "authors": ["Siyuan Yang", "Yang Zhang", "Haoran He", "Ling Pan", "Xiu Li", "Chenjia Bai", "Xuelong Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.02834", "pdf_url": "https://arxiv.org/pdf/2512.02834", "code_url": "", "clusters": ["C3"], "keywords": ["anti-exploration", "test-time scaling", "noise selection"], "model_class": "VLA", "techniques": ["test-time steering"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Test-time steering of flow/diffusion VLAs toward stable success modes after SFT, without further training.", "relevance": 2, "relevance_reason": "Training-free improvement; extra inference compute.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:\"vision-language-action\" AND abs:\"offline reinforcement learning\""}
{"id": "2508.15201", "title": "Survey of Vision-Language-Action Models for Embodied Manipulation", "authors": ["Hao-Ran Li", "Yuhui Chen", "Wen-Bo Cui", "Weiheng Liu", "Kai Liu", "Ming-Cai Zhou", "Zheng-Tao Zhang", "Dong-Bin Zhao"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2508.15201", "pdf_url": "https://arxiv.org/pdf/2508.15201", "code_url": "", "clusters": ["C1"], "keywords": ["survey", "VLA"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Survey of VLA models for manipulation covering architectures, datasets, pre-/post-training and evaluation, with deployment challenges.", "relevance": 2, "relevance_reason": "Context survey of VLA landscape.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2510.03827"}
{"id": "2502.03550", "title": "TD-M(PC)$^2$: Improving Temporal Difference MPC Through Policy Constraint", "authors": ["Haotian Lin", "Pengcheng Wang", "Jeff Schneider", "Guanya Shi"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.03550", "pdf_url": "https://arxiv.org/pdf/2502.03550", "code_url": "", "clusters": ["C8"], "keywords": ["TD-MPC", "value overestimation", "policy constraint"], "model_class": "world model", "techniques": ["policy regularization reducing OOD value queries"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["DMControl", "HumanoidBench (61-DoF)"], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Fixes value overestimation in TD-MPC2 caused by planner/policy mismatch with a zero-overhead policy constraint, especially helping high-DoF humanoids.", "relevance": 2, "relevance_reason": "Latent MPC improvement without added compute.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2310.16828"}
{"id": "2512.19007", "title": "The 6th International Verification of Neural Networks Competition (VNN-COMP 2025): Summary and Results", "authors": ["Konstantin Kaulen", "Tobias Ladner", "Stanley Bak", "Christopher Brix", "Hai Duong", "Thomas Flinkow", "Taylor T. Johnson", "Lukas Koller", "Edoardo Manino", "ThanhVu H Nguyen", "Haoze Wu"], "year": 2025, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2512.19007", "pdf_url": "https://arxiv.org/pdf/2512.19007", "code_url": "", "clusters": ["C7"], "keywords": ["VNN-COMP"], "model_class": "other", "techniques": ["benchmarking", "NN verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Report of VNN-COMP 2025: 8 tools on 16 regular and 9 extended benchmarks with standardized ONNX/VNN-LIB formats and equal-cost hardware.", "relevance": 2, "relevance_reason": "State of NN verification tooling as of 2025.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:Reluplex OR Marabou OR NNV OR beta-CROWN OR VNN-COMP"}
{"id": "10.1007/s11241-025-09452-w", "title": "The bottlenecks of AI: challenges for embedded and real-time research in a data-centric age", "authors": ["Tarek Abdelzaher", "Y. Charlie Hu", "Denizhan Kara", "Tomoyoshi Kimura", "Ashitabh Misra", "Vishakha Ramani", "Olivier Tardieu", "Tianshi Wang", "Maggie Wigness", "Alaa Youssef"], "year": 2025, "venue": "Real-Time Systems (2025)", "type": "paper", "url": "https://doi.org/10.1007/s11241-025-09452-w", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["real-time AI", "embedded"], "model_class": "other", "techniques": ["position"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Perspective arguing that data-centric AI shifts embedded/real-time bottlenecks to scarce sensor data at training and resource economy and scheduling at inference.", "relevance": 2, "relevance_reason": "Research agenda for real-time neural inference.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2507.16815", "title": "ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning", "authors": ["Chi-Pin Huang", "Yueh-Hua Wu", "Min-Hung Chen", "Y. Wang", "Fu-En Yang"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2507.16815", "pdf_url": "https://arxiv.org/pdf/2507.16815", "code_url": "", "clusters": ["C1"], "keywords": ["dual-system", "reasoning", "latent planning"], "model_class": "VLA", "techniques": ["dual-system", "latent planning", "RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "ThinkAct is a dual-system VLA where a multimodal LLM produces reinforced visual latent plans that condition a downstream action model for long-horizon manipulation.", "relevance": 2, "relevance_reason": "context: dual-system reasoning VLA frequently compared against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310; cited by 2503.02310, 2506.07639"}
{"id": "2503.18278", "title": "TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model", "authors": ["Cheng Yang", "Yang Sui", "Jinqi Xiao", "Lingyi Huang", "Yu Gong", "Chendi Li", "J. Yan", "Yu Bai", "P. Sadayappan", "X. Hu", "Bo Yuan"], "year": 2025, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2503.18278", "pdf_url": "https://arxiv.org/pdf/2503.18278", "code_url": "", "clusters": ["C1"], "keywords": ["token pruning", "VLM"], "model_class": "other", "techniques": ["token pruning", "inference-time optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "TopV prunes visual tokens in VLMs via an inference-time optimization compatible with FlashAttention and KV caching for faster, lower-memory inference.", "relevance": 2, "relevance_reason": "foundational technique for VLA token pruning", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2210.17323; cited by 2511.18950, 2606.29350"}
{"id": "2503.10559", "title": "Towards Safe Path Tracking Using the Simplex Architecture", "authors": ["Georg Jäger", "Nils-Jonathan Friedrich", "Hauke Petersen", "Benjamin Noack"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2503.10559", "pdf_url": "https://arxiv.org/pdf/2503.10559", "code_url": "", "clusters": ["C7"], "keywords": ["path tracking", "runtime assurance"], "model_class": "other", "techniques": ["simplex", "RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Simplex-based path-tracking controller combining an RL tracker with a high-assurance controller as a safe testbed for learned planners.", "relevance": 2, "relevance_reason": "Simplex applied to learned robot controllers.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"runtime assurance\" OR abs:\"simplex architecture\""}
{"id": "2505.08243", "title": "Training Strategies for Efficient Embodied Reasoning", "authors": ["William Chen", "Suneel Belkhale", "Suvir Mirchandani", "Oier Mees", "Danny Driess", "Karl Pertsch", "Sergey Levine"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.08243", "pdf_url": "https://arxiv.org/pdf/2505.08243", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["embodied reasoning", "training strategies", "inference speed"], "model_class": "VLA", "techniques": ["reasoning distillation", "training recipe"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": false, "tldr": "Analyzes why robot CoT helps and derives training strategies keeping gains while speeding inference.", "relevance": 2, "relevance_reason": "Training recipe with deployment speed angle.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2502.19645 (S2 citations)"}
{"id": "2508.19257", "title": "TTF-VLA: Temporal Token Fusion via Pixel-Attention Integration for Vision-Language-Action Models", "authors": ["Chenghao Liu", "Jiachen Zhang", "Chengxuan Li", "Zhimu Zhou", "Shixin Wu", "Songfang Huang", "Huiling Duan"], "year": 2025, "venue": "AAAI 2026", "type": "paper", "url": "https://arxiv.org/abs/2508.19257", "pdf_url": "https://arxiv.org/pdf/2508.19257", "code_url": "", "clusters": ["C1"], "keywords": ["temporal token fusion", "training-free"], "model_class": "VLA", "techniques": ["temporal token fusion"], "base_model": "OpenVLA, VLA-Cache", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "SIMPLER", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Fuses historical and current visual tokens where frames are unchanged, mainly boosting robustness; notes that query reuse could enable acceleration.", "relevance": 2, "relevance_reason": "Temporal reuse, accuracy-oriented.", "foundational_technique": false, "verified": true, "found_via": "search query: VLA caching inference"}
{"id": "2511.01718", "title": "Unified Diffusion VLA: Vision-Language-Action Model via Joint Discrete Denoising Diffusion Process", "authors": ["Jiayi Chen", "Wenxuan Song", "Pengxiang Ding", "Ziyang Zhou", "Han Zhao", "Feilong Tang", "Donglin Wang", "Haoang Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2511.01718", "pdf_url": "https://arxiv.org/pdf/2511.01718", "code_url": "", "clusters": ["C1"], "keywords": ["unified diffusion VLA", "joint discrete denoising"], "model_class": "VLA", "techniques": ["discrete diffusion decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Jointly denoises future-image and action tokens in one discrete diffusion process, with inference-time techniques for efficiency.", "relevance": 2, "relevance_reason": "Diffusion VLA; efficiency secondary.", "foundational_technique": false, "verified": true, "found_via": "search query: hybrid AR diffusion VLA"}
{"id": "2512.24125", "title": "Unified Embodied VLM Reasoning with Robotic Action via Autoregressive Discretized Pre-training", "authors": ["Yi Liu", "Sukai Wang", "Dafeng Wei", "Xiaowei Cai", "Lin-Qing Zhong", "Jiange Yang", "Guanghui Ren", "Jin-Yu Zhang", "Maoqing Yao", "Chuan-Kang Li", "Xindong He", "Liliang Chen", "Jianlan Luo"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2512.24125", "pdf_url": "https://arxiv.org/pdf/2512.24125", "code_url": "", "clusters": ["C1"], "keywords": ["embodied reasoning", "discretized pretraining"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Unifies embodied VLM reasoning with action prediction via autoregressive discretized pretraining.", "relevance": 2, "relevance_reason": "Context: VLA (links=15).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2506.19850", "title": "Unified Vision-Language-Action Model", "authors": ["Yu-Quan Wang", "Xing-Hang Li", "Wen-Xuan Wang", "Junbo Zhang", "Yingyan Li", "Yun-Tao Chen", "Xinlong Wang", "Zhao-Xiang Zhang"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2506.19850", "pdf_url": "https://arxiv.org/pdf/2506.19850", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["UniVLA", "autoregressive"], "model_class": "VLA", "techniques": ["discrete tokenization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN", "LIBERO", "SimplerEnv"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Native multimodal VLA that tokenizes vision, language and actions into one autoregressive sequence and adds world-model post-training on video.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2506.07530"}
{"id": "2504.02792", "title": "Unified World Models: Coupling Video and Action Diffusion for Pretraining on Large Robotic Datasets", "authors": ["Chu-Ning Zhu", "Raymond Yu", "Si-Yuan Feng", "B. Burchfiel", "Paarth Shah", "Abhishek Gupta"], "year": 2025, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2504.02792", "pdf_url": "https://arxiv.org/pdf/2504.02792", "code_url": "", "clusters": ["C8"], "keywords": ["unified world model", "video diffusion", "pretraining"], "model_class": "world-action model", "techniques": ["joint video-action diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Unified World Models couple video diffusion and action diffusion in one transformer so a single model can pretrain on both action-labeled robot data and action-free video, acting as policy, dynamics or inverse-dynamics model.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2505.15659, 2506.09985"}
{"id": "2501.10105", "title": "Universal Actions for Enhanced Embodied Foundation Models", "authors": ["Jinliang Zheng", "Jianxiong Li", "Dong-Xiu Liu", "Yinan Zheng", "Zhihao Wang", "Zhonghong Ou", "Yu Liu", "Jingjing Liu", "Ya-Qin Zhang", "Xianyuan Zhan"], "year": 2025, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2501.10105", "pdf_url": "https://arxiv.org/pdf/2501.10105", "code_url": "", "clusters": ["C1"], "keywords": ["universal actions", "cross-embodiment"], "model_class": "VLA", "techniques": ["universal action space"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "UniAct learns a shared tokenized universal action space across heterogeneous embodiments, enabling a small embodied foundation model that decodes to robot-specific actions.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2502.19417, 2506.17639"}
{"id": "2506.15693", "title": "Verifiable Safety Q-Filters via Hamilton-Jacobi Reachability and Multiplicative Q-Networks", "authors": ["Jiaxing Li", "Hanjiang Hu", "Yujie Yang", "Changliu Liu"], "year": 2025, "venue": "IEEE Control Systems Letters 2025", "type": "paper", "url": "https://arxiv.org/abs/2506.15693", "pdf_url": "https://arxiv.org/pdf/2506.15693", "code_url": "", "clusters": ["C8", "C9"], "keywords": ["safety Q-filter", "HJ reachability", "formal verification"], "model_class": "other", "techniques": ["verifiable model-free safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Synthesizes formally verified model-free safety Q-filters using HJ reachability and multiplicative Q-networks.", "relevance": 2, "relevance_reason": "Verified safety filters.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Changliu Liu"}
{"id": "2507.12898", "title": "Vidar: Embodied Video Diffusion Model for Generalist Manipulation", "authors": ["Yao Feng", "Heng-Kai Tan", "Xin-Yi Mao", "Chendong Xiang", "Guodong Liu", "Shu-He Huang", "Hang Su", "Jun Zhu"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.12898", "pdf_url": "https://arxiv.org/pdf/2507.12898", "code_url": "", "clusters": ["C8"], "keywords": ["video diffusion", "inverse dynamics", "generalist manipulation"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Embodied video diffusion prior plus a masked inverse dynamics adapter for transferring manipulation to new embodiments with few demos.", "relevance": 2, "relevance_reason": "Context: video world-model policy (links=13).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213, 2410.24164"}
{"id": "2512.06963", "title": "VideoVLA: Video Generators Can Be Generalizable Robot Manipulators", "authors": ["Yi-Chao Shen", "Fangyun Wei", "Zhi-Ying Du", "Yaobo Liang", "Yan Lu", "Jiao-Long Yang", "Nan-Ning Zheng", "Bai-Ning Guo"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2512.06963", "pdf_url": "https://arxiv.org/pdf/2512.06963", "code_url": "", "clusters": ["C8"], "keywords": ["video-to-policy", "joint video-action"], "model_class": "world-action model", "techniques": ["video diffusion backbone"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Turns a pretrained video generation DiT into a VLA that jointly predicts future frames and action sequences.", "relevance": 2, "relevance_reason": "Context: video-generator-as-policy baseline cited by core items.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2601.21998, 2602.15882"}
{"id": "2507.23682", "title": "villa-X: Enhancing Latent Action Modeling in Vision-Language-Action Models", "authors": ["Xiaoyu Chen", "Hangxing Wei", "Pushi Zhang", "Chu-Heng Zhang", "Kaixin Wang", "Yanjiang Guo", "Ru-Shuai Yang", "Yucen Wang", "Xinquan Xiao", "Li Zhao", "Jianyu Chen", "Jiang Bian"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.23682", "pdf_url": "https://arxiv.org/pdf/2507.23682", "code_url": "", "clusters": ["C1"], "keywords": ["latent action"], "model_class": "VLA", "techniques": ["latent actions"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "villa-X improves latent-action learning and integrates it into VLA pretraining for better manipulation.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310; cited by 2602.06949, 2607.26657"}
{"id": "2507.10672", "title": "Vision Language Action Models in Robotic Manipulation: A Systematic Review", "authors": ["Muhayy Ud Din", "Waseem Akram", "L. S. Saoud", "Jan Rosell", "Irfan Hussain"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2507.10672", "pdf_url": "https://arxiv.org/pdf/2507.10672", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["systematic review", "VLA", "datasets"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Systematic review of ~100 VLA models, their datasets and simulation platforms with a taxonomy of architectures.", "relevance": 2, "relevance_reason": "Context review of VLA models and benchmarks.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2510.03827, 2510.04898"}
{"id": "2502.07855", "title": "Vision-Language Models for Edge Networks: A Comprehensive Survey", "authors": ["Ahmed Sharshar", "Latif U. Khan", "Waseem Ullah", "Mohsen Guizani"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2502.07855", "pdf_url": "https://arxiv.org/pdf/2502.07855", "code_url": "", "clusters": ["C2", "C6"], "keywords": ["VLM", "edge networks", "survey"], "model_class": "other", "techniques": ["compression", "edge deployment"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Survey of techniques and challenges for deploying vision-language models on resource-limited edge devices.", "relevance": 2, "relevance_reason": "Background survey for edge VLM deployment.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2509.25718", "title": "VLA Model Post-Training via Action-Chunked PPO and Self Behavior Cloning", "authors": ["Si-Cheng Wang", "Tian-Yu Xiang", "Xiao-Hu Zhou", "Mei-Jiang Gui", "Xiao-Liang Xie", "Shi-Qi Liu", "Shuang-Yi Wang", "Ao-Qun Jin", "Zeng-Guang Hou"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2509.25718", "pdf_url": "https://arxiv.org/pdf/2509.25718", "code_url": "", "clusters": ["C3"], "keywords": ["VLA RL"], "model_class": "VLA", "techniques": ["action-chunked PPO", "self behavior cloning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MetaWorld"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Action-chunked PPO with an adaptively weighted self behavior cloning loss from a growing buffer of successful trials for VLA post-training.", "relevance": 2, "relevance_reason": "Chunk-level RL for VLAs.", "foundational_technique": false, "verified": true, "found_via": "recency sweep: arXiv VLA AND (\"reinforcement learning\" OR RL) AND (efficient OR \"fine-tuning\") [start=100]"}
{"id": "2510.13054", "title": "VLA-0: Building State-of-the-Art VLAs with Zero Modification", "authors": ["Ankit Goyal", "Hugo Hadfield", "Xu-Ning Yang", "Valts Blukis", "Fabio Ramos"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.13054", "pdf_url": "https://arxiv.org/pdf/2510.13054", "code_url": "", "clusters": ["C1"], "keywords": ["minimal VLA", "text actions"], "model_class": "VLA", "techniques": ["actions as text"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "real robot"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "VLA-0 shows that simply emitting actions as text from an unmodified VLM, with a careful training recipe, yields a state-of-the-art VLA without new tokens or action heads.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2603.16952, 2606.07895"}
{"id": "2506.17561", "title": "VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models", "authors": ["Chongkai Gao", "Zixuan Liu", "Zhenghao Chi", "Jun-Shan Huang", "Xin Fei", "Yiwen Hou", "Yuxuan Zhang", "Yudi Lin", "Zhirui Fang", "Zeyu Jiang", "Lin Shao"], "year": 2025, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2506.17561", "pdf_url": "https://arxiv.org/pdf/2506.17561", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["planning representations", "hierarchical VLA"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Controlled study of planning paradigms and representations inside VLAs across many benchmarks.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "10.1007/s11241-025-09451-x", "title": "When machine learning and neural networks marry real-time scheduling", "authors": ["Zhishan Guo"], "year": 2025, "venue": "Real-Time Systems (2025)", "type": "paper", "url": "https://doi.org/10.1007/s11241-025-09451-x", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["real-time scheduling", "machine learning"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Short overview of how ML/NN and real-time scheduling intersect, including real-time constraints on ML models and accelerator/safety challenges.", "relevance": 2, "relevance_reason": "Background on real-time guarantees for neural workloads.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2509.19080", "title": "World4RL: Diffusion World Models for Policy Refinement With Reinforcement Learning for Robotic Manipulation", "authors": ["Zhen Jiang", "Kai Liu", "Yu-Xing Qin", "Shuai Tian", "Yu-Peng Zheng", "Ming-Cai Zhou", "Zheng-Tao Zhang", "Chao Yu", "Hao-Ran Li", "Dong-Bin Zhao"], "year": 2025, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2509.19080", "pdf_url": "https://arxiv.org/pdf/2509.19080", "code_url": "", "clusters": ["C3", "C8"], "keywords": ["diffusion world model", "policy refinement"], "model_class": "world model", "techniques": ["world-model RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "World4RL: refines imitation policies with RL inside a diffusion world model instead of on hardware.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2310.16828"}
{"id": "2506.21539", "title": "WorldVLA: Towards Autoregressive Action World Model", "authors": ["Jun Cen", "Chaohui Yu", "Hangjie Yuan", "Yuming Jiang", "Siteng Huang", "Jiayan Guo", "Xin Li", "Yibing Song", "Hao Luo", "Fan Wang", "Deli Zhao", "Hao Chen"], "year": 2025, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2506.21539", "pdf_url": "https://arxiv.org/pdf/2506.21539", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["autoregressive world-action model", "unified generation"], "model_class": "world-action model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Unifies a VLA policy and an image-predicting world model inside one autoregressive model so action and future-frame generation reinforce each other.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 52 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2510.10274", "title": "X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model", "authors": ["Jinliang Zheng", "Jianxiong Li", "Zhihao Wang", "Dong-Xiu Liu", "Xirui Kang", "Yuchun Feng", "Yinan Zheng", "Jia-Yi Zou", "Yi-Lun Chen", "Jia Zeng", "Ya-Qin Zhang", "Jiang-Miao Pang", "Jingjing Liu", "Tai Wang", "Xianyuan Zhan"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2510.10274", "pdf_url": "https://arxiv.org/pdf/2510.10274", "code_url": "", "clusters": ["C1"], "keywords": ["cross-embodiment", "soft prompt", "compact VLA"], "model_class": "VLA", "techniques": ["soft prompting", "flow matching"], "base_model": "", "params": "0.9B", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Cross-embodiment flow-matching VLA that handles heterogeneous robot data with per-source learnable soft-prompt embeddings on a plain Transformer encoder; a 0.9B model reaches strong results across many sim and real benchmarks.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2507.00435, 2510.06710"}
{"id": "2402.10885", "title": "3D Diffuser Actor: Policy Diffusion with 3D Scene Representations", "authors": ["Tsung-Wei Ke", "Nikolaos Gkanatsios", "Katerina Fragkiadaki"], "year": 2024, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2402.10885", "pdf_url": "https://arxiv.org/pdf/2402.10885", "code_url": "", "clusters": ["C1"], "keywords": ["3D diffusion policy", "RLBench", "CALVIN"], "model_class": "diffusion/flow policy", "techniques": ["3D denoising transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RLBench", "CALVIN"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diffusion policy with a 3D relative-position denoising transformer over scene features; SOTA on RLBench and CALVIN.", "relevance": 2, "relevance_reason": "Context: standard 3D diffusion-policy baseline.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2406.01586, 2410.08001"}
{"id": "2403.09631", "title": "3D-VLA: A 3D Vision-Language-Action Generative World Model", "authors": ["Haoyu Zhen", "Xiao-Wen Qiu", "Peihao Chen", "Jincheng Yang", "Xin Yan", "Yilun Du", "Yining Hong", "Chuang Gan"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2403.09631", "pdf_url": "https://arxiv.org/pdf/2403.09631", "code_url": "", "clusters": ["C8"], "keywords": ["3D", "generative world model"], "model_class": "world-action model", "techniques": ["3D generative world model"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "3D-VLA builds a 3D-aware generalist model on a 3D LLM with diffusion models that imagine goal images/point clouds to guide action prediction.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2406.09246, 2410.08001"}
{"id": "2410.18825", "title": "A generic approach for reactive stateful mitigation of application failures in distributed robotics systems deployed with Kubernetes", "authors": ["Florian Mirus", "Frederik Pasch", "Nikhil Singhal", "Kay-Ulrich Scholl"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.18825", "pdf_url": "https://arxiv.org/pdf/2410.18825", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["Kubernetes", "failure mitigation", "stateful robotics apps"], "model_class": "other", "techniques": ["reactive stateful failover"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Generic mechanism for resilient, stateful failure recovery of robot applications offloaded to edge/cloud via Kubernetes.", "relevance": 2, "relevance_reason": "Reliability of offloaded robot workloads.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: offloading manipulation robot"}
{"id": "2405.17418", "title": "A Self-Correcting Vision-Language-Action Model for Fast and Slow System Manipulation", "authors": ["Chenxuan Li", "Jiaming Liu", "Guanqun Wang", "Xiaoqi Li", "Sixiang Chen", "Liang Heng", "Chuyan Xiong", "Jiaxin Ge", "Renrui Zhang", "Kaichen Zhou", "Shanghang Zhang"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2405.17418", "pdf_url": "https://arxiv.org/pdf/2405.17418", "code_url": "", "clusters": ["C1"], "keywords": ["fast and slow system", "self-correction"], "model_class": "VLA", "techniques": ["dual-system architecture"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Single VLA with a fast direct pose-prediction mode and a slow chain-of-thought reflection mode invoked after failures.", "relevance": 2, "relevance_reason": "Fast-slow but correction-focused.", "foundational_technique": false, "verified": true, "found_via": "search query: fast slow VLA"}
{"id": "2410.13126", "title": "ALOHA Unleashed: A Simple Recipe for Robot Dexterity", "authors": ["Tony Z. Zhao", "Jonathan Tompson", "Danny Driess", "Pete Florence", "Kamyar Ghasemipour", "Chelsea Finn", "Ayzaan Wahid"], "year": 2024, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2410.13126", "pdf_url": "https://arxiv.org/pdf/2410.13126", "code_url": "", "clusters": ["C4"], "keywords": ["ALOHA 2", "bimanual dexterity", "diffusion policy"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Shows that large-scale data on ALOHA 2 plus diffusion policies yields strong bimanual dexterous skills.", "relevance": 2, "relevance_reason": "Context: major real-robot bimanual system (links=17).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705"}
{"id": "2406.07539", "title": "BAKU: An Efficient Transformer for Multi-Task Policy Learning", "authors": ["Siddhant Haldar", "Zhuoran Peng", "Lerrel Pinto"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2406.07539", "pdf_url": "https://arxiv.org/pdf/2406.07539", "code_url": "", "clusters": ["C1"], "keywords": ["data-efficient policy", "LIBERO"], "model_class": "other", "techniques": ["action chunking", "multi-task transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "Meta-World"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Simple multi-task transformer combining observation trunks, action chunking and action heads for data-efficient policy learning.", "relevance": 2, "relevance_reason": "Context: lightweight multi-task policy baseline.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2411.04983, 2505.23705"}
{"id": "2403.09227", "title": "BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation", "authors": ["Li, Chengshu", "Zhang, Ruohan", "Wong, Josiah", "Gokmen, Cem", "Srivastava, Sanjana", "Martín-Martín, Roberto", "Wang, Chen", "Levine, Gabrael", "Ai, Wensi", "Martinez, Benjamin", "Yin, Hang", "Lingelbach, Michael", "Hwang, Minjune", "Hiranaka, Ayano", "Garlanka, Sujay", "Aydin, Arman", "Lee, Sharon", "Sun, Jiankai", "Anvari, Mona", "Sharma, Manasi", "Bansal, Dhruva", "Hunter, Samuel", "Kim, Kyu-Young", "Lou, Alan", "Matthews, Caleb R", "Villa-Renteria, Ivan", "Tang, Jerry Huayang", "Tang, Claire", "Xia, Fei", "Li, Yunzhu", "Savarese, Silvio", "Gweon, Hyowon", "Liu, C. Karen", "Wu, Jiajun", "Fei-Fei, Li"], "year": 2024, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2403.09227", "pdf_url": "https://arxiv.org/pdf/2403.09227", "code_url": "", "clusters": ["C4"], "keywords": ["BEHAVIOR-1K", "OmniGibson", "household activities"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["BEHAVIOR-1K"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "1,000 everyday activities in 50 scenes on the OmniGibson simulator with a sim-to-real transfer study; basis of the B1K challenge.", "relevance": 2, "relevance_reason": "Large embodied benchmark.", "foundational_technique": false, "verified": true, "found_via": "search query: BEHAVIOR-1K benchmark"}
{"id": "2410.08852", "title": "Conformalized Interactive Imitation Learning: Handling Expert Shift and Intermittent Feedback", "authors": ["Michelle Zhao", "Reid Simmons", "Henny Admoni", "Aaditya Ramdas", "Andrea Bajcsy"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.08852", "pdf_url": "https://arxiv.org/pdf/2410.08852", "code_url": "", "clusters": ["C3", "C7"], "keywords": ["uncertainty", "expert shift"], "model_class": "other", "techniques": ["online conformal prediction", "interactive imitation learning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Online conformal prediction adapted to intermittent human labels so interactive IL uncertainty adapts to expert shift during deployment.", "relevance": 2, "relevance_reason": "Online-calibrated uncertainty for deciding when to seek help.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:conformal AND (\"vision-language-action\" OR \"diffusion policy\" OR \"action chunk*\" OR \"imitation learning\")"}
{"id": "2409.03166", "title": "Continual Robot Skill and Task Learning via Dialogue", "authors": ["Weiwei Gu", "Suresh Kondepudi", "Anmol Gupta", "Lixiao Huang", "Nakul Gopalan"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2409.03166", "pdf_url": "https://arxiv.org/pdf/2409.03166", "code_url": "", "clusters": ["C3"], "keywords": ["ACT-LoRA", "continual skill learning", "dialogue"], "model_class": "other", "techniques": ["LoRA", "ACT"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["simulation", "real robot"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "ACT policy with LoRA for few-shot continual skill learning, integrated with LLM dialogue to request new skills.", "relevance": 2, "relevance_reason": "LoRA-based continual adaptation of small policies.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: TAIL task-specific adapters imitation learning LoRA robot continual"}
{"id": "2412.01663", "title": "Dadu‐E: Rethinking the Role of Large Language Model in Robotic Computing Pipelines", "authors": ["Wen-Hao Sun", "Sai Hou", "Zixuan Wang", "Bo Yu", "Shao-Shan Liu", "Xu Yang", "Shuai Liang", "Yi-Ming Gan", "Yinhe Han"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2412.01663", "pdf_url": "https://arxiv.org/pdf/2412.01663", "code_url": "", "clusters": ["C5"], "keywords": ["lightweight LLM planner", "robotic computing pipeline"], "model_class": "other", "techniques": ["lightweight LLM", "skill encapsulation", "closed-loop feedback"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Closed-loop planning framework that uses a lightweight LLM with encapsulated skills and feedback to cut compute cost in robotic pipelines.", "relevance": 2, "relevance_reason": "Compute-cost-aware robotic pipeline from Dadu group.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2407.04292 (citations)"}
{"id": "2410.18647", "title": "Data Scaling Laws in Imitation Learning for Robotic Manipulation", "authors": ["Fanqi Lin", "Ying-Dong Hu", "Pingyue Sheng", "Chuan Wen", "Jiacheng You", "Yang Gao"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2410.18647", "pdf_url": "https://arxiv.org/pdf/2410.18647", "code_url": "", "clusters": ["C4"], "keywords": ["data scaling laws", "imitation learning"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Large real-robot study showing policy generalization follows power laws in the number of environments and objects.", "relevance": 2, "relevance_reason": "Context: data-scaling reference for robot policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2503.00200, 2504.16054"}
{"id": "2406.03102", "title": "DEER: A Delay-Resilient Framework for Reinforcement Learning with Variable Delays", "authors": ["Bo Xia", "Yilun Kong", "Yongzhe Chang", "Bo Yuan", "Zhiheng Li", "Xueqian Wang", "Bin Liang"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2406.03102", "pdf_url": "https://arxiv.org/pdf/2406.03102", "code_url": "", "clusters": ["C7"], "keywords": ["DEER", "variable delays", "delayed RL"], "model_class": "other", "techniques": ["pretrained delay encoder"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Encodes delayed states and variable-length past action sequences with a pretrained encoder to make RL resilient to random delays.", "relevance": 2, "relevance_reason": "Delay-robust learned control.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"random delays\" reinforcement learning"}
{"id": "2403.12945", "title": "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset", "authors": ["Alexander Khazatsky", "Karl Pertsch", "S. Nair", "Ashwin Balakrishna", "S. Dasari", "Siddharth Karamcheti", "Soroush Nasiriany", "M. K. Srirama", "Lawrence Yunliang Chen", "Kirsty Ellis", "P. Fagan", "Joey Hejna", "Masha Itkina", "Marion Lepert", "Y. Ma", "Patrick Miller", "Jimmy Wu", "Suneel Belkhale", "S. Dass", "Huy Ha", "Arhan Jain", "Abraham Lee", "Youngwoon Lee", "Marius Memmel", "Sungjae Park", "Ilija Radosavovic", "Kaiyuan Wang", "Albert Zhan", "Kevin Black", "Cheng Chi", "K. Hatch", "Shan Lin", "Jingpei Lu", "Jean-Pierre Mercat", "Abdul Rehman", "Pannag R. Sanketi", "Archit Sharma", "C. Simpson", "Q. Vương", "H. Walke", "Blake Wulfe", "Ted Xiao", "J. Yang", "Arefeh Yavary", "Tony Zhao", "Christopher Agia", "R. Baijal", "Mateo Guaman Castro", "Daphne Chen", "Qiuyu Chen", "T. Chung", "Jaimyn Drake", "E. P. Foster", "Jensen Gao", "D. Herrera", "Minho Heo", "Kyle Hsu", "Jiaheng Hu", "Muhammad Zubair Irshad", "Donovon Jackson", "Charlotte Le", "Yunshuang Li", "K. Lin", "Roy Lin", "Ze-Han Ma", "Abhiram Maddukuri", "Suvir Mirchandani", "D. Morton", "Tony Nguyen", "Abigail O'Neill", "R. Scalise", "Derick Seale", "V. Son", "Stephen Tian", "E. Tran", "Andrew Wang", "Y. Wu", "Annie Xie", "Jingyun Yang", "Patrick Yin", "Yunchu Zhang", "O. Bastani", "Glen Berseth", "Jeannette Bohg", "Ken Goldberg", "Abhinav Gupta", "Abhishek Gupta", "D. Jayaraman", "Joseph J. Lim", "Jitendra Malik", "Roberto Mart'in-Mart'in", "S. Ramamoorthy", "Dorsa Sadigh", "Shuran Song", "Jia-Jun Wu", "Michael C. Yip", "Yu-Ke Zhu", "T. Kollar", "Sergey Levine", "Chelsea Finn"], "year": 2024, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2403.12945", "pdf_url": "https://arxiv.org/pdf/2403.12945", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "in-the-wild manipulation", "DROID"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large multi-site, in-the-wild robot manipulation dataset collected with a shared hardware setup; widely used to pretrain and evaluate generalist policies.", "relevance": 2, "relevance_reason": "Context: major dataset that VLA efficiency work trains/evaluates on (links=82).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2404.03325", "title": "Embodied Neuromorphic Artificial Intelligence for Robotics: Perspectives, Challenges, and Research Development Stack", "authors": ["Rachmad Vidya Wicaksana Putra", "Alberto Marchisio", "Fakhreddine Zayer", "Jorge Dias", "M. Shafique"], "year": 2024, "venue": "ICARCV 2024", "type": "paper", "url": "https://arxiv.org/abs/2404.03325", "pdf_url": "https://arxiv.org/pdf/2404.03325", "code_url": "", "clusters": ["C5"], "keywords": ["neuromorphic", "spiking neural networks", "energy-efficient robotics"], "model_class": "other", "techniques": ["neuromorphic computing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Perspective on using spiking neuromorphic hardware for energy-efficient embodied robotics, outlining cross-layer optimization and benchmark needs.", "relevance": 2, "relevance_reason": "Alternative hardware direction, not foundation-model specific.", "foundational_technique": false, "verified": true, "found_via": "web: processing-in-memory OR NPU accelerator embodied AI"}
{"id": "2401.06118", "title": "Extreme Compression of Large Language Models via Additive Quantization", "authors": ["Vage Egiazarian", "Andrei Panferov", "Denis Kuznedelev", "Elias Frantar", "Artem Babenko", "Dan Alistarh"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2401.06118", "pdf_url": "https://arxiv.org/pdf/2401.06118", "code_url": "", "clusters": ["C2"], "keywords": ["quantization", "2-bit"], "model_class": "other", "techniques": ["additive quantization", "extreme compression"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "AQLM applies multi-codebook additive quantization to LLM weights, reaching strong accuracy at 2-3 bits per parameter.", "relevance": 2, "relevance_reason": "foundational technique for extreme weight quantization", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323; cited by 2404.00456, 2605.24011"}
{"id": "2409.10197", "title": "Fit and Prune: Fast and Training-free Visual Token Pruning for Multi-modal Large Language Models", "authors": ["Wei-Hao Ye", "Qiong Wu", "Wenhao Lin", "Yi-Yi Zhou"], "year": 2024, "venue": "AAAI Conference on Artificial Intelligence", "type": "paper", "url": "https://arxiv.org/abs/2409.10197", "pdf_url": "https://arxiv.org/pdf/2409.10197", "code_url": "", "clusters": ["C1"], "keywords": ["token pruning", "MLLM"], "model_class": "other", "techniques": ["visual token pruning", "training-free"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Fit and Prune is a training-free visual token pruning method for multimodal LLMs that fits a pruning recipe to a compute budget from attention statistics.", "relevance": 2, "relevance_reason": "foundational technique for VLA token pruning", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2503.20384, 2509.09090"}
{"id": "2401.08553", "title": "FMB: A functional manipulation benchmark for generalizable robotic learning", "authors": ["Jian-Lan Luo", "Charles Xu", "Fangchen Liu", "Liam Tan", "Zipeng Lin", "Jeffrey Wu", "Pieter Abbeel", "Sergey Levine"], "year": 2024, "venue": "Int. J. Robotics Res.", "type": "paper", "url": "https://arxiv.org/abs/2401.08553", "pdf_url": "https://doi.org/10.1177/02783649241276017", "code_url": "", "clusters": ["C4"], "keywords": ["real-world benchmark", "manipulation"], "model_class": "other", "techniques": ["benchmark"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["FMB"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "FMB is a real-world functional manipulation benchmark with reproducible hardware/object setups for long-horizon skill composition and generalization studies.", "relevance": 2, "relevance_reason": "context: real-robot benchmark", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705; cited by 2405.12213, 2406.09246"}
{"id": "2410.10803", "title": "Generalizable Humanoid Manipulation with 3D Diffusion Policies", "authors": ["Yanjie Ze", "Zixuan Chen", "Wenhao Wang", "Tianyi Chen", "Xialin He", "Ying Yuan", "Xue Bin Peng", "Jiajun Wu"], "year": 2024, "venue": "IROS 2025", "type": "paper", "url": "https://arxiv.org/abs/2410.10803", "pdf_url": "https://arxiv.org/pdf/2410.10803", "code_url": "", "clusters": ["C3", "C9"], "keywords": ["iDP3", "3D diffusion policy", "humanoid", "egocentric point cloud"], "model_class": "diffusion/flow policy", "techniques": ["3D diffusion policy"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Improved 3D Diffusion Policy using egocentric point clouds lets a full-size humanoid generalize manipulation across scenes from single-scene data.", "relevance": 2, "relevance_reason": "Onboard diffusion policy on humanoid; efficiency is secondary.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Jiajun Wu"}
{"id": "2410.06158", "title": "GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation", "authors": ["Chi-Lam Cheang", "Guangzeng Chen", "Ya Jing", "Tao Kong", "Hang Li", "Yifeng Li", "Yuxiao Liu", "Hongtao Wu", "Jia-Feng Xu", "Yichu Yang", "Hanbo Zhang", "Min-Zhao Zhu"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.06158", "pdf_url": "https://arxiv.org/pdf/2410.06158", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["video-language-action", "generalist policy"], "model_class": "world-action model", "techniques": ["video pretraining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Generalist manipulation agent pretrained on tens of millions of internet video clips, then fine-tuned for joint video generation and action prediction.", "relevance": 2, "relevance_reason": "Context: major video-pretrained robot foundation model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2407.04292, 2411.19650"}
{"id": "2406.14540", "title": "IRASim: A Fine-Grained World Model for Robot Manipulation", "authors": ["Fang-Qi Zhu", "Hongtao Wu", "Song Guo", "Yuxiao Liu", "Chi-Hou Cheang", "Tao Kong"], "year": 2024, "venue": "IEEE International Conference on Computer Vision", "type": "paper", "url": "https://arxiv.org/abs/2406.14540", "pdf_url": "https://arxiv.org/pdf/2406.14540", "code_url": "", "clusters": ["C8"], "keywords": ["video world model", "fine-grained interaction"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "IRASim: action-conditioned video world model focused on fine-grained robot-object interactions.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2409.13319", "title": "Knowledge-Based Ultra-Low-Latency Semantic Communications for Robotic Edge Intelligence", "authors": ["Qunsong Zeng", "Zhanwei Wang", "You Zhou", "Hai Wu", "Lin Yang", "Kaibin Huang"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2409.13319", "pdf_url": "https://arxiv.org/pdf/2409.13319", "code_url": "", "clusters": ["C6"], "keywords": ["semantic communications", "split inference", "6G", "remote brain"], "model_class": "other", "techniques": ["split inference", "ultra-low-latency feature transmission"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Air-interface protocol where an edge-server knowledge graph acts as a remote brain; robots upload features over short uncoded packets, exploiting classifier margin to tolerate bit errors for low latency.", "relevance": 2, "relevance_reason": "Split inference/communication for robotic edge intelligence; not policy models.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2405.03917", "title": "KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization", "authors": ["Tian-Yi Zhang", "Jonah Yi", "Zhaozhuo Xu", "Anshumali Shrivastava"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2405.03917", "pdf_url": "https://arxiv.org/pdf/2405.03917", "code_url": "", "clusters": ["C2"], "keywords": ["KV cache", "quantization"], "model_class": "other", "techniques": ["KV cache quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Coupled Quantization compresses the LLM KV cache to about 1 bit per channel by jointly quantizing correlated channels.", "relevance": 2, "relevance_reason": "foundational technique for KV-cache compression", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2210.17323; cited by 2603.16952, 2609.26425"}
{"id": "2406.11815", "title": "LLARVA: Vision-Action Instruction Tuning Enhances Robot Learning", "authors": ["Dantong Niu", "Yuvan Sharma", "Giscard Biamby", "Jerome Quenum", "Yutong Bai", "Baifeng Shi", "Trevor Darrell", "Roei Herzig"], "year": 2024, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2406.11815", "pdf_url": "https://arxiv.org/pdf/2406.11815", "code_url": "", "clusters": ["C1"], "keywords": ["instruction tuning"], "model_class": "VLA", "techniques": ["instruction tuning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LLARVA instruction-tunes a multimodal model on vision-action data with 2D visual traces as auxiliary targets for robot learning.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213; cited by 2409.12514, 2502.02175"}
{"id": "2403.15388", "title": "LLaVA-Prumerge: Adaptive Token Reduction for Efficient Large Multimodal Models", "authors": ["Yuzhang Shang", "Mu Cai", "Bing-Xin Xu", "Yong Jae Lee", "Yan Yan"], "year": 2024, "venue": "IEEE International Conference on Computer Vision", "type": "paper", "url": "https://arxiv.org/abs/2403.15388", "pdf_url": "https://arxiv.org/pdf/2403.15388", "code_url": "", "clusters": ["C1"], "keywords": ["token pruning", "LMM"], "model_class": "other", "techniques": ["visual token reduction"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PruMerge adaptively selects and merges visual tokens in multimodal LLMs based on attention sparsity, greatly reducing token count with little accuracy loss.", "relevance": 2, "relevance_reason": "foundational technique for VLA token pruning", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2506.10100, 2506.12723"}
{"id": "2412.13211", "title": "ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks", "authors": ["Shukla, Arth", "Tao, Stone", "Su, Hao"], "year": 2024, "venue": "arXiv", "type": "benchmark", "url": "https://arxiv.org/abs/2412.13211", "pdf_url": "https://arxiv.org/pdf/2412.13211", "code_url": "", "clusters": ["C4"], "keywords": ["ManiSkill-HAB", "GPU-accelerated", "home rearrangement", "sim speed"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["MS-HAB"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "GPU-accelerated Home Assistant Benchmark with realistic low-level control running over 3x faster than prior magical-grasp versions, plus RL/IL baselines and filtered demo generation.", "relevance": 2, "relevance_reason": "Fast simulation benchmark.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: ManiSkill3 VLA evaluation GPU parallel rollouts"}
{"id": "2401.14461", "title": "Marabou 2.0: A Versatile Formal Analyzer of Neural Networks", "authors": ["Haoze Wu", "Omri Isac", "Aleksandar Zeljić", "Teruhiro Tagomori", "Matthew Daggitt", "Wen Kokke", "Idan Refaeli", "Guy Amir", "Kyle Julian", "Shahaf Bassan", "Pei Huang", "Ori Lahav", "Min Wu", "Min Zhang", "Ekaterina Komendantskaya", "Guy Katz", "Clark Barrett"], "year": 2024, "venue": "CAV 2024", "type": "paper", "url": "https://arxiv.org/abs/2401.14461", "pdf_url": "https://arxiv.org/pdf/2401.14461", "code_url": "", "clusters": ["C7"], "keywords": ["Marabou"], "model_class": "other", "techniques": ["SMT", "NN verification"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "System description of Marabou 2.0 neural network formal analyzer.", "relevance": 2, "relevance_reason": "Standard NN verification tool.", "foundational_technique": true, "verified": true, "found_via": "gap check: chunk certification: arXiv ti:Reluplex OR Marabou OR NNV OR beta-CROWN OR VNN-COMP"}
{"id": "2401.02117", "title": "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation", "authors": ["Zipeng Fu", "Tony Zhao", "Chelsea Finn"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2401.02117", "pdf_url": "https://arxiv.org/pdf/2401.02117", "code_url": "", "clusters": ["C4"], "keywords": ["ALOHA", "bimanual mobile manipulation", "low-cost hardware"], "model_class": "other", "techniques": ["action chunking", "co-training"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Low-cost whole-body teleoperation rig for bimanual mobile manipulation; shows co-training with static ALOHA data greatly boosts ACT-style policies.", "relevance": 2, "relevance_reason": "Context: standard low-cost platform and dataset for ACT/chunking policies.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2405.05941, 2405.12213"}
{"id": "2412.04445", "title": "Moto: Latent Motion Token as the Bridging Language for Learning Robot Manipulation from Videos", "authors": ["Yi Chen", "Yu-Ying Ge", "Weiliang Tang", "Yi-Zhuo Li", "Yi-Xiao Ge", "Mingyu Ding", "Shan Ying", "Xi-Hui Liu"], "year": 2024, "venue": "IEEE International Conference on Computer Vision", "type": "paper", "url": "https://arxiv.org/abs/2412.04445", "pdf_url": "https://arxiv.org/pdf/2412.04445", "code_url": "", "clusters": ["C1"], "keywords": ["video pretraining", "latent actions"], "model_class": "other", "techniques": ["latent motion tokens", "autoregressive pretraining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns discrete latent motion tokens from videos, pretrains an autoregressive model over them, then maps to robot actions.", "relevance": 2, "relevance_reason": "Context: latent-action tokenization baseline cited by several core items.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2505.15659, 2607.08639"}
{"id": "2412.03572", "title": "Navigation World Models", "authors": ["Amir Bar", "Gaoyue Zhou", "Danny Tran", "Trevor Darrell", "Yann LeCun"], "year": 2024, "venue": "Computer Vision and Pattern Recognition", "type": "paper", "url": "https://arxiv.org/abs/2412.03572", "pdf_url": "https://arxiv.org/pdf/2412.03572", "code_url": "", "clusters": ["C8"], "keywords": ["navigation world model", "planning"], "model_class": "world model", "techniques": ["conditional diffusion transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Controllable video world model (CDiT, up to 1B params) that predicts egocentric observations from navigation actions and is used to plan/rank trajectories.", "relevance": 2, "relevance_reason": "Context: widely used navigation world model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2502.00622, 2504.19322"}
{"id": "2410.19989", "title": "On-Robot Reinforcement Learning with Goal-Contrastive Rewards", "authors": ["Ondrej Biza", "Thomas Weng", "Lingfeng Sun", "Karl Schmeckpeper", "Tarik Kelestemur", "Yecheng Jason Ma", "Robert Platt", "Jan-Willem van de Meent", "Lawson L. S. Wong"], "year": 2024, "venue": "ICRA 2025", "type": "paper", "url": "https://arxiv.org/abs/2410.19989", "pdf_url": "https://arxiv.org/pdf/2410.19989", "code_url": "", "clusters": ["C4", "C9"], "keywords": ["on-robot RL", "goal-contrastive rewards", "passive video"], "model_class": "other", "techniques": ["dense reward learning from video"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "GCR learns dense rewards from action-free videos to make on-robot RL cheaper in robot runtime.", "relevance": 2, "relevance_reason": "On-robot learning efficiency; speaker Jason Ma co-author.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Yecheng Jason Ma"}
{"id": "2310.08864", "title": "Open X-Embodiment: Robotic Learning Datasets and RT-X Models", "authors": ["Abigail O'Neill", "Abdul Rehman", "Abhiram Maddukuri", "Abhishek Gupta", "A. Padalkar", "Abraham Lee", "A. Pooley", "Agrim Gupta", "A. Mandlekar", "Ajinkya Jain", "Albert Tung", "Alex Bewley", "Alex Herzog", "A. Irpan", "Alexander Khazatsky", "Anant Rai", "Anchit Gupta", "Andrew Wang", "Anikait Singh", "Animesh Garg", "Aniruddha Kembhavi", "Annie Xie", "Anthony Brohan", "A. Raffin", "Archit Sharma", "Arefeh Yavary", "Arhan Jain", "A. Balakrishna", "Ayzaan Wahid", "Ben Burgess-Limerick", "Beomjoon Kim", "B. Schölkopf", "Blake Wulfe", "Brian Ichter", "Cewu Lu", "Charles Xu", "Charlotte Le", "Chelsea Finn", "Chen Wang", "Chenfeng Xu", "Cheng Chi", "Chenguang Huang", "Christine Chan", "Christopher Agia", "Chuer Pan", "C. Fu", "Coline Devin", "Danfei Xu", "Daniel Morton", "Danny Driess", "Daphne Chen", "Deepak Pathak", "Dhruv Shah", "Dieter Büchler", "D. Jayaraman", "Dmitry Kalashnikov", "Dorsa Sadigh", "Edward Johns", "E. Foster", "Fangchen Liu", "Federico Ceola", "Fei Xia", "Feiyu Zhao", "F. Stulp", "Gaoyue Zhou", "G. Sukhatme", "Gautam Salhotra", "Ge Yan", "Gilbert Feng", "Giulio Schiavi", "Glen Berseth", "G. Kahn", "Guanzhi Wang", "Hao Su", "Haoshu Fang", "Haochen Shi", "Henghui Bao", "Heni Ben Amor", "Henrik I. Christensen", "Hiroki Furuta", "H. Walke", "Hong-Jie Fang", "Huy Ha", "Igor Mordatch", "Ilija Radosavovic", "Isabel Leal", "Jacky Liang", "Jad Abou-Chakra", "Jaehyung Kim", "Jaimyn Drake", "Jan Peters", "Jan Schneider", "Jasmine Hsu", "Jeannette Bohg", "Jeff Bingham", "Jeffrey Wu", "Jensen Gao", "Jiaheng Hu", "Jia-Jun Wu", "Jialin Wu", "Jiankai Sun", "Jian-Lan Luo", "Jiayuan Gu", "Jie Tan", "Jihoon Oh", "Jimmy Wu", "Jingpei Lu", "Jingyun Yang", "Jitendra Malik", "João Silvério", "Joey Hejna", "Jonathan Booher", "Jonathan Tompson", "Jonathan Yang", "Jordi Salvador", "Joseph J. Lim", "Junhyek Han", "Kaiyuan Wang", "Kanishka Rao", "Karl Pertsch", "Karol Hausman", "Keegan Go", "K. Gopalakrishnan", "Ken Goldberg", "Kendra Byrne", "Kenneth Oslund", "Kento Kawaharazuka", "Kevin Black", "Kevin Lin", "Kevin Zhang", "Kiana Ehsani", "Kiran Lekkala", "Kirsty Ellis", "Krishan Rana", "K. Srinivasan", "Kuan Fang", "K. Singh", "Kuo-Hao Zeng", "K. Hatch", "Kyle Hsu", "L. Itti", "Lawrence Yunliang Chen", "Lerrel Pinto", "Fei-Fei Li", "Liam Tan", "L. Fan", "Lionel Ott", "Lisa Lee", "Luca Weihs", "Magnum Chen", "Marion Lepert", "Marius Memmel", "Masayoshi Tomizuka", "Masha Itkina", "Mateo Guaman Castro", "Max Spero", "Maximilian Du", "Michael Ahn", "Michael C. Yip", "Ming-Tong Zhang", "Mingyu Ding", "Minho Heo", "M. K. Srirama", "Mohit Sharma", "Moo Jin Kim", "Muhammad Zubair Irshad", "Naoaki Kanazawa", "Nicklas Hansen", "N. Heess", "Nikhil J. Joshi", "Niko Suenderhauf", "Ning Liu", "Norman Di Palo", "N. Shafiullah", "Oier Mees", "Oliver Kroemer", "O. Bastani", "Pannag R. Sanketi", "Patrick Miller", "Patrick Yin", "Paul Wohlhart", "Peng Xu", "P. Fagan", "P. Mitrano", "P. Sermanet", "Pieter Abbeel", "Priya Sundaresan", "Qiuyu Chen", "Quan Vuong", "Rafael Rafailov", "Ran Tian", "Ria Doshi", "Roberto Martín-Martín", "R. Baijal", "Rosario Scalise", "Rose Hendrix", "Roy Lin", "Runjia Qian", "Ruo-Han Zhang", "R. Mendonca", "Rutav Shah", "Ryan Hoque", "Ryan C. Julian", "Samuel Bustamante", "Sean Kirmani", "Sergey Levine", "Shan Lin", "Sherry Moore", "Shikhar Bahl", "Shivin Dass", "Shubham D. Sonawani", "Shuran Song", "Si-Chun Xu", "Siddhant Haldar", "Siddharth Karamcheti", "S. Adebola", "Simon Guist", "Soroush Nasiriany", "Stefan Schaal", "Stefan Welker", "Stephen Tian", "S. Ramamoorthy", "S. Dasari", "Suneel Belkhale", "Sungjae Park", "Suraj Nair", "Suvir Mirchandani", "Takayuki Osa", "Tanmay Gupta", "Tanmay Gupta", "Tatsuya Matsushima", "Ted Xiao", "Thomas Kollar", "Tian-He Yu", "Tianli Ding", "Todor Davchev", "Tony Zhao", "Travis Armstrong", "Trevor Darrell", "T. Chung", "Vidhi Jain", "Vincent Vanhoucke", "Wei Zhan", "Wenxuan Zhou", "Wolfram Burgard", "Xi Chen", "Xiaolong Wang", "Xinghao Zhu", "Xin-Yang Geng", "Xi-Yuan Liu", "Liangwei Xu", "Xuanlin Li", "Yao Lu", "Y. Ma", "Yejin Kim", "Yevgen Chebotar", "Yifan Zhou", "Yifeng Zhu", "Y. Wu", "Ying Xu", "Yixuan Wang", "Yonatan Bisk", "Yoonyoung Cho", "Youngwoon Lee", "Yu-Chen Cui", "Yue Cao", "Yueh-Hua Wu", "Yu-Jin Tang", "Yu-Ke Zhu", "Yunchu Zhang", "Yunfan Jiang", "Yunshuang Li", "Yunzhu Li", "Yusuke Iwasawa", "Yutaka Matsuo", "Ze-Han Ma", "Zhuo Xu", "Zichen Jeff Cui", "Zichen Zhang", "Zipeng Lin", "H. Amor", "Henrik I Christensen", "K. Majd", "Q. Vuong", "S. Schaal", "Tatsuya Harada", "T. Matsushima", "T. Darrell"], "year": 2024, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://www.semanticscholar.org/paper/5fc3830322f7eae2934c912c391ee5edded86a15", "pdf_url": "https://arxiv.org/pdf/2310.08864", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "cross-embodiment", "RT-X", "cross-embodiment dataset"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Pools demonstrations from 22 robot types across 21 labs into one standardized dataset and trains RT-1-X/RT-2-X generalist policies that transfer across embodiments.", "relevance": 2, "relevance_reason": "Context: the dominant pretraining corpus most efficient VLAs train or evaluate on.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2409.12514, 2410.05562; snowball round 2: cited by 2310.05905, 2405.05941; snowball round 2: cited by 2405.15223, 2409.12514"}
{"id": "2411.18179", "title": "Prediction with Action: Visual Policy Learning via Joint Denoising Process", "authors": ["Yanjiang Guo", "Yucheng Hu", "Jianke Zhang", "Yen-Jen Wang", "Xiaoyu Chen", "Chaochao Lu", "Jianyu Chen"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2411.18179", "pdf_url": "https://arxiv.org/pdf/2411.18179", "code_url": "", "clusters": ["C8"], "keywords": ["joint denoising", "video prediction"], "model_class": "world-action model", "techniques": ["joint denoising"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PAD jointly denoises future images and robot actions in one diffusion transformer, unifying video prediction and diffusion policy.", "relevance": 2, "relevance_reason": "context: joint prediction-action diffusion model", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2412.14803, 2503.00200"}
{"id": "2412.15109", "title": "Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation", "authors": ["Yang Tian", "Si-Zhe Yang", "Jia Zeng", "Ping Wang", "Da-Hua Lin", "Hao Dong", "Jiang-Miao Pang"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2412.15109", "pdf_url": "https://arxiv.org/pdf/2412.15109", "code_url": "", "clusters": ["C1", "C8"], "keywords": ["Seer", "inverse dynamics", "visual foresight"], "model_class": "VLA", "techniques": ["predictive inverse dynamics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["LIBERO", "CALVIN", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Seer: end-to-end policy that predicts future visual states and infers actions via inverse dynamics, strong on LIBERO and CALVIN.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2405.16005", "title": "PTQ4DiT: Post-training Quantization for Diffusion Transformers", "authors": ["Junyi Wu", "Haoxuan Wang", "Yuzhang Shang", "Mubarak Shah", "Yan Yan"], "year": 2024, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2405.16005", "pdf_url": "https://arxiv.org/pdf/2405.16005", "code_url": "", "clusters": ["C2"], "keywords": ["quantization", "DiT"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "PTQ4DiT is a post-training quantization method for diffusion transformers that handles salient channels and timestep-varying activations.", "relevance": 2, "relevance_reason": "foundational technique for quantizing diffusion transformers", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2601.20130, 2602.20309"}
{"id": "2410.11305", "title": "QSpec: Speculative Decoding with Complementary Quantization Schemes", "authors": ["Juntao Zhao", "Wenhao Lu", "Sheng Wang", "Lingpeng Kong", "Chuan Wu"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.11305", "pdf_url": "https://arxiv.org/pdf/2410.11305", "code_url": "https://github.com/hku-netexplo-lab/QSpec", "clusters": ["C2"], "keywords": ["W4A4 drafting", "weight-only verification", "KV reuse"], "model_class": "other", "techniques": ["speculative decoding", "quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "up to 1.64x speedup without quality loss; up to 1.55x over prior speculative methods in batched settings", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Uses a low-precision activation-weight quantized pass to draft and a weight-only quantized pass of the same model to verify, sharing weights and KV cache.", "relevance": 2, "relevance_reason": "Example of techniques designed to compose rather than conflict.", "foundational_technique": true, "verified": true, "found_via": "gap check: Q1: arXiv id lookup of known speculative-decoding x quantization composition works"}
{"id": "2403.14968", "title": "Real-time Safety Index Adaptation for Parameter-varying Systems via Determinant Gradient Ascend", "authors": ["Rui Chen", "Weiye Zhao", "Ruixuan Liu", "Weiyang Zhang", "Changliu Liu"], "year": 2024, "venue": "American Control Conference 2024", "type": "paper", "url": "https://arxiv.org/abs/2403.14968", "pdf_url": "https://arxiv.org/pdf/2403.14968", "code_url": "", "clusters": ["C7"], "keywords": ["safety index adaptation", "determinant gradient ascent"], "model_class": "other", "techniques": ["closed-form safety index update"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Closed-form determinant-gradient update to keep a safety index valid as dynamics vary, avoiding expensive re-synthesis at runtime.", "relevance": 2, "relevance_reason": "Real-time safe control (Liu lab).", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety index\" synthesis safe control"}
{"id": "2404.12377", "title": "RoboDreamer: Learning Compositional World Models for Robot Imagination", "authors": ["Siyuan Zhou", "Yilun Du", "Jiaben Chen", "Yan-Dong Li", "Dit-Yan Yeung", "Chuang Gan"], "year": 2024, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2404.12377", "pdf_url": "https://arxiv.org/pdf/2404.12377", "code_url": "", "clusters": ["C8"], "keywords": ["video world model", "compositional"], "model_class": "world model", "techniques": ["compositional video diffusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RoboDreamer factorizes language instructions into primitives to learn a compositional text-to-video world model that generalizes to unseen instruction combinations for planning.", "relevance": 2, "relevance_reason": "context: video world model for robot imagination", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2506.18897, 2509.21027"}
{"id": "2403.04440", "title": "RoboKube: Establishing a New Foundation for the Cloud Native Evolution in Robotics", "authors": ["Yu Liu", "Aitor Hernandez Herranz", "Roberto C. Sundin"], "year": 2024, "venue": "ICARA 2024", "type": "paper", "url": "https://arxiv.org/abs/2403.04440", "pdf_url": "https://arxiv.org/pdf/2403.04440", "code_url": "", "clusters": ["C6"], "keywords": ["Kubernetes", "cloud native", "device-cloud continuum", "ROS"], "model_class": "other", "techniques": ["container orchestration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Reviews cloudification practice in robotics and proposes a Kubernetes-based platform for deploying ROS applications across the device-cloud continuum.", "relevance": 2, "relevance_reason": "Fleet deployment infrastructure.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2412.13877", "title": "RoboMIND: Benchmark on Multi-embodiment Intelligence Normative Data for Robot Manipulation", "authors": ["Kun Wu", "Chengkai Hou", "Jiaming Liu", "Zhengping Che", "Xiao-Zhu Ju", "Zhuqin Yang", "Meng Li", "Yi-Nuo Zhao", "Zhiyuan Xu", "Guang Yang", "Shichao Fan", "Xin-Hua Wang", "Fei Liao", "Zhen Zhao", "Guang-Yu Li", "Zhao Jin", "Lecheng Wang", "Jilei Mao", "Ning Liu", "Peifeng Ren", "Qiang Zhang", "Yaoxu Lyu", "Meng-Zhen Liu", "Jingyang He", "Yulin Luo", "Zeyu Gao", "Chen-Xuan Li", "Chenyang Gu", "Yan-Kai Fu", "Disheng Wu", "Xing-Yu Wang", "Si-Xiang Chen", "Zhenyu Wang", "Pengju An", "Siyuan Qian", "Shan-Shan Zhang", "Jianfeng Tang"], "year": 2024, "venue": "Robotics", "type": "paper", "url": "https://arxiv.org/abs/2412.13877", "pdf_url": "https://doi.org/10.15607/rss.2025.xxi.152", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "multi-embodiment"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large multi-embodiment teleoperated manipulation dataset with standardized collection, including failure cases.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2505.19017"}
{"id": "2407.08693", "title": "Robotic Control via Embodied Chain-of-Thought Reasoning", "authors": ["Michał Zawalski", "William Chen", "Karl Pertsch", "Oier Mees", "Chelsea Finn", "Sergey Levine"], "year": 2024, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2407.08693", "pdf_url": "https://arxiv.org/pdf/2407.08693", "code_url": "", "clusters": ["C1"], "keywords": ["embodied CoT", "reasoning VLA"], "model_class": "VLA", "techniques": ["chain-of-thought"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Trains VLAs to produce grounded multi-step reasoning (plans, subtasks, object boxes) before acting, improving generalization.", "relevance": 2, "relevance_reason": "Context: reasoning-VLA baseline whose inference cost motivated later efficient-reasoning work.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2409.12514, 2502.19417"}
{"id": "2409.09882", "title": "Safe Control of Quadruped in Varying Dynamics via Safety Index Adaptation", "authors": ["Kai S. Yun", "Rui Chen", "Chase Dunaway", "John M. Dolan", "Changliu Liu"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2409.09882", "pdf_url": "https://arxiv.org/pdf/2409.09882", "code_url": "", "clusters": ["C7"], "keywords": ["safety index adaptation", "quadruped", "varying dynamics"], "model_class": "other", "techniques": ["real-time safety index adaptation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Adapts a safety index online as quadruped payload changes so the safe control law keeps forward invariance and convergence guarantees in real time.", "relevance": 2, "relevance_reason": "Real-time adaptive safe control (Liu lab).", "foundational_technique": false, "verified": true, "found_via": "arXiv search: \"safety index\" synthesis safe control"}
{"id": "2408.11812", "title": "Scaling Cross-Embodied Learning: One Policy for Manipulation, Navigation, Locomotion and Aviation", "authors": ["Ria Doshi", "H. Walke", "Oier Mees", "S. Dasari", "Sergey Levine"], "year": 2024, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2408.11812", "pdf_url": "https://arxiv.org/pdf/2408.11812", "code_url": "", "clusters": ["C1"], "keywords": ["CrossFormer", "cross-embodiment"], "model_class": "other", "techniques": ["cross-embodiment transformer"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Transformer policy trained on 900K trajectories from 20 embodiments that controls arms, wheeled robots, quadrupeds and drones with one set of weights.", "relevance": 2, "relevance_reason": "Context: cross-embodiment generalist policy.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2503.05833, 2503.24278"}
{"id": "2401.16013", "title": "SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning", "authors": ["Jianlan Luo", "Zheyuan Hu", "Charles Xu", "You Liang Tan", "Jacob Berg", "Archit Sharma", "Stefan Schaal", "Chelsea Finn", "Abhishek Gupta", "Sergey Levine"], "year": 2024, "venue": "ICRA 2024", "type": "preprint", "url": "https://arxiv.org/abs/2401.16013", "pdf_url": "https://arxiv.org/pdf/2401.16013", "code_url": "https://github.com/rail-berkeley/serl", "clusters": ["C3"], "keywords": ["SERL", "software suite", "sample-efficient RL"], "model_class": "other", "techniques": ["off-policy RL", "reward classifiers", "resets"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["real robot assembly"], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Open library for sample-efficient real-world robotic RL with carefully engineered off-policy RL, reward and reset components.", "relevance": 2, "relevance_reason": "Infrastructure underlying real-world RL fine-tuning.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: HIL-SERL SERL sample-efficient real-world robot reinforcement learning fine-tuning pretrained"}
{"id": "10.1109/isca59077.2024.00047", "title": "Tartan: Microarchitecting a Robotic Processor", "authors": ["M. Bakhshalipour", "P. Gibbons"], "year": 2024, "venue": "ISCA 2024", "type": "paper", "url": "https://dl.acm.org/doi/abs/10.1109/ISCA59077.2024.00047", "pdf_url": "", "code_url": "https://github.com/cmu-roboarch/tartan", "clusters": ["C5"], "keywords": ["robotic CPU", "oriented vectorization", "robot-semantic prefetching", "RoWild"], "model_class": "other", "techniques": ["CPU microarchitecture for robotics"], "base_model": "", "params": "", "hardware": "", "reported_metric": "1.2x (legacy) up to 2.11x (approximable, optimized) average speedup on RoWild suite", "metric_measures": "relative latency", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "CPU microarchitecture with vectorization, approximate acceleration, robot-aware prefetching and cache partitioning for end-to-end robot software.", "relevance": 2, "relevance_reason": "Robot processor design; classical workloads.", "foundational_technique": false, "verified": true, "found_via": "web: Tartan microarchitecting a robotic processor"}
{"id": "2410.01515", "title": "Task-Oriented Edge-Assisted Cooperative Data Compression, Communications and Computing for UGV-Enhanced Warehouse Logistics", "authors": ["Jiaming Yang", "Zhen Meng", "Xiangmin Xu", "Kan Chen", "Emma Liying Li", "Philip Guodong G. Zhao"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2410.01515", "pdf_url": "https://arxiv.org/pdf/2410.01515", "code_url": "", "clusters": ["C6"], "keywords": ["UGV", "edge-assisted control", "point cloud compression", "DRL"], "model_class": "other", "techniques": ["task-oriented compression", "edge offloading of control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Offloads navigation/following control of two cooperating warehouse UGVs to an edge server and uses DRL to adapt point-cloud compression ratios to task needs.", "relevance": 2, "relevance_reason": "Bandwidth/compute co-design for offloaded robot control.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2402.13201", "title": "Tiny Reinforcement Learning for Quadruped Locomotion using Decision Transformers", "authors": ["Orhan Eren Akgün", "Néstor Cuevas", "Matheus Farias", "Daniel Garces"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2402.13201", "pdf_url": "https://arxiv.org/pdf/2402.13201", "code_url": "", "clusters": ["C2"], "keywords": ["decision transformer", "low-cost robot", "quadruped"], "model_class": "other", "techniques": ["compact model design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["quadruped locomotion"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Casts imitation learning as sequence modeling with a small decision transformer made deployable on ultra-low-cost robot hardware.", "relevance": 2, "relevance_reason": "Tiny policies for constrained robots.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2406.09246"}
{"id": "2402.14606", "title": "Towards Diverse Behaviors: A Benchmark for Imitation Learning with Human Demonstrations", "authors": ["Xiao-Gang Jia", "Denis Blessing", "Xin-Kai Jiang", "Moritz Reuss", "Atalay Donat", "Rudolf Lioutikov", "Gerhard Neumann"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2402.14606", "pdf_url": "https://arxiv.org/pdf/2402.14606", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "diverse behaviors", "human demonstrations"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Benchmark for imitation learning from diverse human demonstrations that measures behaviour diversity.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2410.01345", "title": "Towards Generalizable Vision-Language Robotic Manipulation: A Benchmark and LLM-guided 3D Policy", "authors": ["Ricardo Garcia", "Shizhe Chen", "Cordelia Schmid"], "year": 2024, "venue": "ICRA 2025", "type": "benchmark", "url": "https://arxiv.org/abs/2410.01345", "pdf_url": "https://arxiv.org/pdf/2410.01345", "code_url": "", "clusters": ["C4"], "keywords": ["GemBench", "generalization levels", "3D-LOTUS"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["GemBench"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Benchmark with seven action primitives and four generalization levels plus a 3D policy and an LLM-guided variant.", "relevance": 2, "relevance_reason": "Generalization benchmark.", "foundational_technique": false, "verified": true, "found_via": "search query: GemBench"}
{"id": "2412.10345", "title": "TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies", "authors": ["Ruijie Zheng", "Yongyuan Liang", "Shuaiyi Huang", "Jianfeng Gao", "Hal Daumé", "A. Kolobov", "Furong Huang", "Jianwei Yang"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2412.10345", "pdf_url": "https://arxiv.org/pdf/2412.10345", "code_url": "", "clusters": ["C1"], "keywords": ["visual trace", "prompting"], "model_class": "VLA", "techniques": ["visual trace prompting"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "TraceVLA overlays visual traces of past motion on the image input so a VLA gains spatio-temporal awareness, improving manipulation over OpenVLA.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2405.12213; cited by 2503.14734, 2505.15659"}
{"id": "2402.10329", "title": "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots", "authors": ["Cheng Chi", "Zhen-Jia Xu", "Chuer Pan", "Eric Cousineau", "B. Burchfiel", "Si-Yuan Feng", "Russ Tedrake", "Shuran Song"], "year": 2024, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2402.10329", "pdf_url": "https://arxiv.org/pdf/2402.10329", "code_url": "", "clusters": ["C4"], "keywords": ["data collection", "handheld gripper", "in-the-wild demos"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Handheld-gripper data collection and policy learning framework that transfers in-the-wild human demonstrations to deployable robot policies, with latency-aware deployment design.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2401.16529", "title": "Unleashing the Power of Preemptive Priority-based Scheduling for Real-Time GPU Tasks", "authors": ["Yidi Wang", "Cong Liu", "Daniel Wong", "Hyoseung Kim"], "year": 2024, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2401.16529", "pdf_url": "https://arxiv.org/pdf/2401.16529", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["real-time GPU"], "model_class": "other", "techniques": ["GPU preemption", "priority scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Kernel-thread and IOCTL-based techniques enabling preemptive priority-based GPU scheduling for real-time tasks.", "relevance": 2, "relevance_reason": "Precursor to GCAPS for analyzable GPU timing.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"real-time\" AND Jetson AND (\"response time\" OR deadline* OR schedulab*)"}
{"id": "10.1016/j.sysarc.2024.103300", "title": "Using MAST for modeling and response-time analysis of real-time applications with GPUs", "authors": ["Iosu Gomez", "Unai Díaz de Cerio", "Jorge Parra", "Juan M. Rivas", "J. Javier Gutiérrez", "Michael González Harbour"], "year": 2024, "venue": "Journal of Systems Architecture (2024)", "type": "paper", "url": "https://doi.org/10.1016/j.sysarc.2024.103300", "pdf_url": "", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["GPU real-time analysis"], "model_class": "other", "techniques": ["response-time analysis", "MAST"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Applies the MAST model and offset-based response-time analysis to real-time applications using GPUs, validated on a railway application.", "relevance": 2, "relevance_reason": "Response-time analysis methodology for GPU-accelerated embedded systems.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: OpenAlex \"real-time GPU scheduling DNN inference embedded autonomous timing predictability\""}
{"id": "2406.02540", "title": "ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation", "authors": ["Tianchen Zhao", "Tongcheng Fang", "Hao-Feng Huang", "Rui Wan", "Widyadewi Soedarmadji", "En-hao Liu", "Shi-Yao Li", "Zinan Lin", "Guohao Dai", "Shengen Yan", "Huazhong Yang", "Xue-Fei Ning", "Yu Wang"], "year": 2024, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2406.02540", "pdf_url": "https://arxiv.org/pdf/2406.02540", "code_url": "", "clusters": ["C2"], "keywords": ["quantization", "diffusion transformer"], "model_class": "other", "techniques": ["post-training quantization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "ViDiT-Q is a post-training quantization method for diffusion transformers in image and video generation, targeting edge deployment.", "relevance": 2, "relevance_reason": "foundational technique for quantizing diffusion transformers (video world models)", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2205.14135; cited by 2602.20309, 2603.28489"}
{"id": "2412.14058", "title": "What matters in building vision–language–action models for generalist robots", "authors": ["Xing-Hang Li", "Pei-Yan Li", "Long Qian", "Minghuan Liu", "Dong Wang", "Ji-Rong Liu", "Bingyi Kang", "Xiao Ma", "Xin-Long Wang", "Dingkun Guo", "Tao Kong", "Hanbo Zhang", "Huaping Liu"], "year": 2024, "venue": "Nature Machine Intelligence", "type": "paper", "url": "https://arxiv.org/abs/2412.14058", "pdf_url": "https://arxiv.org/pdf/2412.14058", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["RoboVLMs", "design study"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large empirical study of which design choices matter when turning VLMs into VLAs (backbone, action formulation, history).", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 11 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2401.00025", "title": "Any-point Trajectory Modeling for Policy Learning", "authors": ["Chuan Wen", "Xing-Yu Lin", "John So", "Kai Chen", "Q. Dou", "Yang Gao", "Pieter Abbeel"], "year": 2023, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2401.00025", "pdf_url": "https://arxiv.org/pdf/2401.00025", "code_url": "", "clusters": ["C1"], "keywords": ["point trajectories", "video"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Predicts future trajectories of arbitrary points from video as an intermediate to learn policies with few action labels.", "relevance": 2, "relevance_reason": "Context: (links=11).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2306.03310"}
{"id": "2308.12952", "title": "BridgeData V2: A Dataset for Robot Learning at Scale", "authors": ["H. Walke", "Kevin Black", "Abraham Lee", "Moo Jin Kim", "Maximilian Du", "Chongyi Zheng", "Tony Zhao", "Philippe Hansen-Estruch", "Q. Vuong", "Andre Wang He", "Vivek Myers", "Kuan Fang", "Chelsea Finn", "S. Levine"], "year": 2023, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2308.12952", "pdf_url": "https://arxiv.org/pdf/2308.12952", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "WidowX"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large multi-environment manipulation dataset on a low-cost WidowX arm, used to train and evaluate many generalist policies.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2310.06114"}
{"id": "2306.02551", "title": "Conformal Predictive Safety Filter for RL Controllers in Dynamic Environments", "authors": ["Kegan J. Strawn", "Nora Ayanian", "Lars Lindemann"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2306.02551", "pdf_url": "https://arxiv.org/pdf/2306.02551", "code_url": "", "clusters": ["C7"], "keywords": ["conformal predictive safety filter", "RL navigation", "pedestrian prediction"], "model_class": "other", "techniques": ["conformal prediction intervals", "learned safety filter"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Predicts other agents with conformal uncertainty intervals and learns a safety filter to make pretrained RL navigation policies uncertainty-aware.", "relevance": 2, "relevance_reason": "Safety filter for learned policies.", "foundational_technique": false, "verified": true, "found_via": "arXiv search: conformal prediction robot policy failure"}
{"id": "10.1109/ijcnn54540.2023.10191108", "title": "FastAct: A Lightweight Actor Compression Framework for Fast Policy Learning", "authors": ["Hongjie Zhang", "Haoming Ma", "Zhenyu Chen"], "year": 2023, "venue": "IJCNN 2023", "type": "paper", "url": "https://doi.org/10.1109/IJCNN54540.2023.10191108", "pdf_url": "", "code_url": "", "clusters": ["C2"], "keywords": ["actor compression", "RL training speed"], "model_class": "other", "techniques": ["dynamic compression (pruning/quantization)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Accelerates RL training by dynamically compressing the actor network with a compressor and scheduler for faster experience generation.", "relevance": 2, "relevance_reason": "Policy compression in RL; not foundation-model scale.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2412.01034 (references)"}
{"id": "2304.09831", "title": "FastRLAP: A System for Learning High-Speed Driving via Deep RL and Autonomous Practicing", "authors": ["Kyle Stachowicz", "Dhruv Shah", "Arjun Bhorkar", "Ilya Kostrikov", "Sergey Levine"], "year": 2023, "venue": "CoRL 2023", "type": "paper", "url": "https://arxiv.org/abs/2304.09831", "pdf_url": "https://arxiv.org/pdf/2304.09831", "code_url": "", "clusters": ["C4", "C9"], "keywords": ["high-speed driving", "online RL"], "model_class": "other", "techniques": ["sample-efficient RL", "pretrained representations"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Learns aggressive RC-car driving from vision in the real world with sample-efficient online RL initialized from prior navigation data.", "relevance": 2, "relevance_reason": "Real-time control learned onboard; tangential.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2309.17080", "title": "GAIA-1: A Generative World Model for Autonomous Driving", "authors": ["Anthony Hu", "Lloyd Russell", "Hudson Yeo", "Zak Murez", "George Fedoseev", "Alex Kendall", "Jamie Shotton", "Gianluca Corrado"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2309.17080", "pdf_url": "https://arxiv.org/pdf/2309.17080", "code_url": "", "clusters": ["C8"], "keywords": ["driving world model", "generative"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Generative world model for autonomous driving that predicts future video from video, text and action inputs.", "relevance": 2, "relevance_reason": "Context: seminal driving world model (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2205.14135, 2301.04104"}
{"id": "2308.10094", "title": "Learning and Communications Co-Design for Remote Inference Systems: Feature Length Selection and Transmission Scheduling", "authors": ["Md Kamran Chowdhury Shisher", "Bo Ji", "I-Hong Hou", "Yin Sun"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2308.10094", "pdf_url": "https://arxiv.org/pdf/2308.10094", "code_url": "", "clusters": ["C6"], "keywords": ["remote inference", "age of information", "feature length", "scheduling"], "model_class": "other", "techniques": ["learning-communication co-design", "AoI-aware scheduling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Jointly optimizes transmitted feature length and transmission scheduling to minimize time-averaged remote inference error measured with Age of Information.", "relevance": 2, "relevance_reason": "Theory for timeliness-aware remote inference (robot motion example).", "foundational_technique": false, "verified": true, "found_via": "arXiv API: remote inference robot"}
{"id": "2305.11627", "title": "LLM-Pruner: On the Structural Pruning of Large Language Models", "authors": ["Xinyin Ma", "Gongfan Fang", "Xinchao Wang"], "year": 2023, "venue": "Neural Information Processing Systems", "type": "paper", "url": "https://arxiv.org/abs/2305.11627", "pdf_url": "https://arxiv.org/pdf/2305.11627", "code_url": "https://github.com/horseee/LLM-Pruner", "clusters": ["C2"], "keywords": ["pruning", "LLM"], "model_class": "other", "techniques": ["structural pruning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "LLM-Pruner removes coupled structures in LLMs using gradient-based importance and recovers quality with brief LoRA tuning.", "relevance": 2, "relevance_reason": "foundational technique for model pruning", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cites 2210.17323; cited by 2411.02359, 2506.17639"}
{"id": "2310.17596", "title": "MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations", "authors": ["A. Mandlekar", "Soroush Nasiriany", "Bowen Wen", "Iretiayo Akinola", "Yashraj Narang", "L. Fan", "Yu-Ke Zhu", "Dieter Fox"], "year": 2023, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2310.17596", "pdf_url": "https://arxiv.org/pdf/2310.17596", "code_url": "", "clusters": ["C4"], "keywords": ["MimicGen", "synthetic demos"], "model_class": "other", "techniques": ["data generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Automatically synthesizes large demonstration datasets from a handful of human demos by adapting object-centric segments to new scenes.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2405.12213"}
{"id": "2309.14320", "title": "MUTEX: Learning Unified Policies from Multimodal Task Specifications", "authors": ["Rutav Shah", "Roberto Mart'in-Mart'in", "Yuke Zhu"], "year": 2023, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2309.14320", "pdf_url": "https://arxiv.org/pdf/2309.14320", "code_url": "", "clusters": ["C1"], "keywords": ["multimodal task specification"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Unified policy that follows tasks specified via speech, text, images or video.", "relevance": 2, "relevance_reason": "Context: multimodal policy (links=13).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2306.03310"}
{"id": "2311.16098", "title": "On Bringing Robots Home", "authors": ["Nur Muhammad (Mahi) Shafiullah", "Anant Rai", "Haritheja Etukuru", "Yiqian Liu", "Is-han Misra", "Soumith Chintala", "Lerrel Pinto"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2311.16098", "pdf_url": "https://arxiv.org/pdf/2311.16098", "code_url": "", "clusters": ["C4"], "keywords": ["home robots", "Dobb-E", "dataset"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Low-cost home robot learning framework with a demonstration tool, dataset and pretrained model.", "relevance": 2, "relevance_reason": "Context: (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137, 2304.13705"}
{"id": "2303.03378", "title": "PaLM-E: An Embodied Multimodal Language Model", "authors": ["Danny Driess", "F. Xia", "Mehdi S. M. Sajjadi", "Corey Lynch", "A. Chowdhery", "Brian Ichter", "Ayzaan Wahid", "Jonathan Tompson", "Q. Vuong", "Tian-He Yu", "Wen-Long Huang", "Yevgen Chebotar", "P. Sermanet", "Daniel Duckworth", "S. Levine", "Vincent Vanhoucke", "Karol Hausman", "Marc Toussaint", "Klaus Greff", "A. Zeng", "Igor Mordatch", "Peter R. Florence"], "year": 2023, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2303.03378", "pdf_url": "http://arxiv.org/pdf/2303.03378", "code_url": "", "clusters": ["C1"], "keywords": ["embodied multimodal LLM", "PaLM"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Google's embodied multimodal language model that injects continuous sensor/state embeddings into a large LLM for robot planning and VQA; a large-scale reference point efficiency work contrasts against.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 70 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2306.00978, 2307.15818"}
{"id": "s2-pseudoinverse-guided-diffusion-iclr-2023", "title": "Pseudoinverse-Guided Diffusion Models for Inverse Problems", "authors": ["Jiaming Song", "Arash Vahdat", "Morteza Mardani", "Jan Kautz"], "year": 2023, "venue": "ICLR 2023", "type": "paper", "url": "https://www.semanticscholar.org/paper/9c81be0c478bfc0a48eedb8769326fe289a11acc", "pdf_url": "", "code_url": "", "clusters": ["C7"], "keywords": ["guided diffusion", "inpainting", "inverse problems"], "model_class": "other", "techniques": ["pseudoinverse guidance"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Training-free guidance for solving inverse problems (e.g., inpainting) with pretrained diffusion models; the guidance mechanism RTC adapts to inpaint action chunks.", "relevance": 2, "relevance_reason": "Foundational technique behind RTC inpainting.", "foundational_technique": true, "verified": true, "found_via": "snowball: references of 2506.07339 (S2)"}
{"id": "2307.00595", "title": "RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot", "authors": ["Haoshu Fang", "Hong-Jie Fang", "Zhenyu Tang", "Jirong Liu", "Chen-Xi Wang", "Junbo Wang", "Haoyi Zhu", "Cewu Lu"], "year": 2023, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2307.00595", "pdf_url": "https://arxiv.org/pdf/2307.00595", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "one-shot"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Diverse real-robot dataset with multimodal sensing for one-shot skill learning.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705; cited by 2405.12213"}
{"id": "2309.01918", "title": "RoboAgent: Generalization and Efficiency in Robot Manipulation via Semantic Augmentations and Action Chunking", "authors": ["Homanga Bharadhwaj", "Jay Vakil", "Mohit Sharma", "Abhi Gupta", "Shubham Tulsiani", "Vikash Kumar"], "year": 2023, "venue": "IEEE International Conference on Robotics and Automation", "type": "paper", "url": "https://arxiv.org/abs/2309.01918", "pdf_url": "https://arxiv.org/pdf/2309.01918", "code_url": "", "clusters": ["C1", "C3"], "keywords": ["data efficiency", "multi-task"], "model_class": "other", "techniques": ["action chunking", "semantic augmentation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "RoboAgent trains a multi-task manipulation agent from a small dataset using generative semantic augmentation and action-chunked transformer policies.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 23 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2304.13705"}
{"id": "2306.11706", "title": "RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation", "authors": ["Konstantinos Bousmalis", "Giulia Vezzani", "Dushyant Rao", "Coline Devin", "Alex X. Lee", "Maria Bauzá", "Todor Davchev", "Yuxiang Zhou", "Agrim Gupta", "A. Raju", "Antoine Laurens", "Claudio Fantacci", "Valentin Dalibard", "Martina Zambelli", "M. F. Martins", "Rugile Pevceviciute", "M. Blokzijl", "Misha Denil", "Nathan Batchelor", "Thomas Lampe", "Emilio Parisotto", "Konrad Zolna", "Scott E. Reed", "Sergio Gómez Colmenarejo", "Jon Scholz", "A. Abdolmaleki", "Oliver Groth", "Jean-Baptiste Regli", "Oleg O. Sushkov", "Thomas Rothörl", "J. Chen", "Yusuf Aytar", "Dave Barker", "Joy Ortiz", "M. Riedmiller", "J. Springenberg", "R. Hadsell", "F. Nori", "N. Heess"], "year": 2023, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2306.11706", "pdf_url": "http://arxiv.org/pdf/2306.11706", "code_url": "", "clusters": ["C3"], "keywords": ["self-improving agent", "generalist"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Multi-embodiment, goal-conditioned decision-transformer agent that improves itself by generating new training data; a context model for cross-embodiment adaptation.", "relevance": 2, "relevance_reason": "verified 2026-09-30 by orchestrator: arXiv abstract fetched. Context: major robot foundation model.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2307.15818, 2310.05905"}
{"id": "2308.07053", "title": "RobotKube: Orchestrating Large-Scale Cooperative Multi-Robot Systems with Kubernetes and ROS", "authors": ["Bastian Lampe", "Lennart Reiher", "Lukas Zanger", "Timo Woopen", "Raphael van Kempen", "Lutz Eckstein"], "year": 2023, "venue": "IEEE ITSC 2023", "type": "paper", "url": "https://arxiv.org/abs/2308.07053", "pdf_url": "https://arxiv.org/pdf/2308.07053", "code_url": "", "clusters": ["C6"], "keywords": ["Kubernetes", "multi-robot orchestration", "microservices"], "model_class": "other", "techniques": ["container orchestration"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Automates orchestration of containerized ROS microservices across robots and infrastructure for large cooperative multi-robot systems.", "relevance": 2, "relevance_reason": "Fleet software orchestration.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2205.09778"}
{"id": "2312.13139", "title": "Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation", "authors": ["Hongtao Wu", "Ya Jing", "Chi-Hou Cheang", "Guangzeng Chen", "Jia-Feng Xu", "Xing-Hang Li", "Minghuan Liu", "Hang Li", "Tao Kong"], "year": 2023, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2312.13139", "pdf_url": "https://arxiv.org/pdf/2312.13139", "code_url": "", "clusters": ["C8"], "keywords": ["video pretraining", "GPT-style policy"], "model_class": "VLA", "techniques": ["video generative pretraining"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN", "real robot"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "GR-1: a GPT-style policy pretrained on large-scale video prediction and then fine-tuned on robot data to jointly predict actions and future frames for language-conditioned manipulation.", "relevance": 2, "relevance_reason": "context: widely used robot foundation model/baseline that efficiency work builds on or compares against", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2303.04137; cited by 2410.05273, 2410.08001"}
{"id": "2310.10639", "title": "Zero-Shot Robotic Manipulation with Pretrained Image-Editing Diffusion Models", "authors": ["Kevin Black", "Mitsuhiko Nakamoto", "P. Atreya", "H. Walke", "Chelsea Finn", "Aviral Kumar", "Sergey Levine"], "year": 2023, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2310.10639", "pdf_url": "https://arxiv.org/pdf/2310.10639", "code_url": "", "clusters": ["C8"], "keywords": ["image-editing diffusion", "subgoal generation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "SuSIE uses a pretrained image-editing diffusion model to imagine subgoal images that a low-level goal-conditioned policy then reaches.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 29 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cites 2301.04104, 2303.04137"}
{"id": "2202.02005", "title": "BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning", "authors": ["Eric Jang", "A. Irpan", "Mohi Khansari", "Daniel Kappler", "F. Ebert", "Corey Lynch", "S. Levine", "Chelsea Finn"], "year": 2022, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2202.02005", "pdf_url": "https://arxiv.org/pdf/2202.02005", "code_url": "", "clusters": ["C4"], "keywords": ["imitation learning", "zero-shot task generalization", "dataset"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Large-scale teleoperated imitation-learning study showing language/video-conditioned policies can generalize zero-shot to unseen manipulation tasks.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 31 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2112.03227, 2212.06817"}
{"id": "2208.06193", "title": "Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning", "authors": ["Zhen-Dong Wang", "Jonathan J. Hunt", "Ming-Yuan Zhou"], "year": 2022, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2208.06193", "pdf_url": "https://arxiv.org/pdf/2208.06193", "code_url": "", "clusters": ["C1"], "keywords": ["offline RL", "diffusion policy", "Diffusion-QL"], "model_class": "diffusion/flow policy", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Represents offline-RL policies as diffusion models combined with Q-learning, showing expressive policy classes help offline RL.", "relevance": 2, "relevance_reason": "Context: foundational diffusion-policy RL work (links=14).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2303.04137, 2406.01586"}
{"id": "2204.01691", "title": "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances", "authors": ["Brian Ichter", "Anthony Brohan", "Yevgen Chebotar", "Chelsea Finn", "Karol Hausman", "Alexander Herzog", "Daniel Ho", "Julian Ibarz", "A. Irpan", "Eric Jang", "Ryan C. Julian", "Dmitry Kalashnikov", "S. Levine", "Yao Lu", "Carolina Parada", "Kanishka Rao", "P. Sermanet", "Alexander T. Toshev", "Vincent Vanhoucke", "F. Xia", "Ted Xiao", "Peng Xu", "Mengyuan Yan", "Noah Brown", "Michael Ahn", "Omar Cortes", "Nicolas Sievers", "Clayton Tan", "Si-Chun Xu", "D. Reyes", "Jarek Rettinghouse", "Jornell Quiambao", "P. Pastor", "Linda Luu", "Kuang-Huei Lee", "Yuheng Kuang", "Sally Jesmonth", "N. Joshi", "Kyle Jeffrey", "Rosario M Jauregui Ruano", "Jasmine Hsu", "K. Gopalakrishnan", "Byron David", "Andy Zeng", "C. Fu"], "year": 2022, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2204.01691", "pdf_url": "https://arxiv.org/pdf/2204.01691", "code_url": "", "clusters": ["C1"], "keywords": ["LLM planning", "affordances", "SayCan"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Grounds LLM task planning in robot affordance value functions so a language model only proposes skills the robot can actually execute.", "relevance": 2, "relevance_reason": "Context: seminal robot foundation model system (links=44).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2212.06817, 2307.15818"}
{"id": "2205.11269", "title": "Dynamic Split Computing for Efficient Deep Edge Intelligence", "authors": ["Arian Bakhtiarnia", "Nemanja Milošević", "Qi Zhang", "Dragana Bajović", "Alexandros Iosifidis"], "year": 2022, "venue": "ICML 2022 DyNN Workshop", "type": "paper", "url": "https://arxiv.org/abs/2205.11269", "pdf_url": "https://arxiv.org/pdf/2205.11269", "code_url": "", "clusters": ["C6"], "keywords": ["dynamic split computing", "channel-aware split"], "model_class": "other", "techniques": ["dynamic split point selection"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Chooses the DNN split location at runtime according to channel data rate and server load, using natural bottleneck layers so no retraining is required.", "relevance": 2, "relevance_reason": "Generic dynamic split computing, precursor to network-aware VLA splitting.", "foundational_technique": true, "verified": true, "found_via": "search query"}
{"id": "2211.17192", "title": "Fast Inference from Transformers via Speculative Decoding", "authors": ["Yaniv Leviathan", "Matan Kalman", "Yossi Matias"], "year": 2022, "venue": "International Conference on Machine Learning", "type": "paper", "url": "https://arxiv.org/abs/2211.17192", "pdf_url": "http://arxiv.org/pdf/2211.17192", "code_url": "", "clusters": ["C1"], "keywords": ["speculative decoding", "LLM inference"], "model_class": "other", "techniques": ["speculative decoding"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Introduces speculative decoding: a small draft model proposes several tokens that the large model verifies in parallel, giving exact-distribution speedups for autoregressive transformers.", "relevance": 2, "relevance_reason": "foundational technique used by parallel/speculative VLA decoding work", "foundational_technique": true, "verified": true, "found_via": "snowball round 2: cited by 2406.09246, 2408.17355"}
{"id": "2210.11691", "title": "FogROS G: Enabling Secure, Connected and Mobile Fog Robotics with Global Addressability", "authors": ["Kaiyuan Chen", "Jiachen Yuan", "Nikhil Jha", "Jeffrey Ichnowski", "John Kubiatowicz", "Ken Goldberg"], "year": 2022, "venue": "IROS 2022 Cloud Robotics Workshop", "type": "paper", "url": "https://arxiv.org/abs/2210.11691", "pdf_url": "https://arxiv.org/pdf/2210.11691", "code_url": "", "clusters": ["C6"], "keywords": ["fog robotics", "secure routing", "mobility"], "model_class": "other", "techniques": ["global data plane routing", "proxy design"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Secure routing layer giving fog robots persistent global names so they can roam across networks while offloading workloads to edge and cloud.", "relevance": 2, "relevance_reason": "Early FogROS networking layer; infrastructure only.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "2210.03370", "title": "GNM: A General Navigation Model to Drive Any Robot", "authors": ["Dhruv Shah", "Ajay Sridhar", "Arjun Bhorkar", "Noriaki Hirose", "Sergey Levine"], "year": 2022, "venue": "ICRA 2023", "type": "paper", "url": "https://arxiv.org/abs/2210.03370", "pdf_url": "https://arxiv.org/pdf/2210.03370", "code_url": "", "clusters": ["C9"], "keywords": ["cross-embodiment", "navigation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Cross-robot navigation policy trained on 60 h of data from 6 robots that deploys onboard to new platforms.", "relevance": 2, "relevance_reason": "Foundational small cross-embodiment policy.", "foundational_technique": false, "verified": true, "found_via": "author sweep: Dhruv Shah"}
{"id": "2210.06407", "title": "Interactive Language: Talking to Robots in Real Time", "authors": ["Corey Lynch", "Ayzaan Wahid", "Jonathan Tompson", "Tianli Ding", "James Betker", "Robert Baruch", "Travis Armstrong", "Peter R. Florence"], "year": 2022, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2210.06407", "pdf_url": "http://arxiv.org/pdf/2210.06407", "code_url": "", "clusters": ["C4"], "keywords": ["Language-Table", "real-time language interaction"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Real-time language-instructable robot framework with a ~600k-trajectory language-labeled dataset and benchmark.", "relevance": 2, "relevance_reason": "Context: Language-Table dataset/benchmark.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2307.15818, 2310.06114"}
{"id": "2204.10297", "title": "Learning to Fold Real Garments with One Arm: A Case Study in Cloud-Based Robotics Research", "authors": ["Ryan Hoque", "Kaushik Shivakumar", "Shrey Aeron", "Gabriel Deza", "Aditya Ganapathi", "Adrian Wong", "Johnny Lee", "Andy Zeng", "Vincent Vanhoucke", "Ken Goldberg"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2204.10297", "pdf_url": "https://arxiv.org/pdf/2204.10297", "code_url": "", "clusters": ["C4", "C6"], "keywords": ["cloud-based robotics research", "remote execution", "Reach platform"], "model_class": "other", "techniques": ["remote policy execution"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Uses a cloud platform for low-latency remote execution of fabric-folding policies on physical robots to benchmark algorithms.", "relevance": 2, "relevance_reason": "Cloud-hosted policy execution/benchmarking.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "2206.14244", "title": "Masked World Models for Visual Control", "authors": ["Younggyo Seo", "Danijar Hafner", "Hao Liu", "Fangchen Liu", "Stephen James", "Kimin Lee", "P. Abbeel"], "year": 2022, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2206.14244", "pdf_url": "http://arxiv.org/pdf/2206.14244", "code_url": "", "clusters": ["C8"], "keywords": ["model-based RL", "masked autoencoder"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Decouples visual representation learning (MAE) from dynamics learning in visual model-based RL.", "relevance": 2, "relevance_reason": "Context: world-model MBRL (links=10).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2406.01586, 2409.12514"}
{"id": "2205.06773", "title": "Optimized Partitioning and Priority Assignment of Real-Time Applications on Heterogeneous Platforms with Hardware Acceleration", "authors": ["Daniel Casini", "Paolo Pazzaglia", "Alessandro Biondi", "Marco Di Natale"], "year": 2022, "venue": "Journal of Systems Architecture, Volume 124, March 2022, 102416", "type": "paper", "url": "https://arxiv.org/abs/2205.06773", "pdf_url": "https://arxiv.org/pdf/2205.06773", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["heterogeneous platforms", "real-time"], "model_class": "other", "techniques": ["schedulability analysis", "partitioning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Framework to partition and prioritize real-time applications on GPU/FPGA-equipped embedded platforms with linearized schedulability analysis, inspired by WATERS 2019 ADAS case.", "relevance": 2, "relevance_reason": "Schedulability analysis for accelerator-based embedded inference.", "foundational_technique": false, "verified": true, "found_via": "gap check: chunk certification: arXiv abs:\"worst-case execution time\" AND (\"neural network\" OR DNN OR GPU)"}
{"id": "2209.05451", "title": "Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation", "authors": ["Mohit Shridhar", "Lucas Manuelli", "D. Fox"], "year": 2022, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2209.05451", "pdf_url": "https://arxiv.org/pdf/2209.05451", "code_url": "", "clusters": ["C1"], "keywords": ["voxel transformer", "multi-task", "6-DoF"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["RLBench"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Perceiver-based multi-task agent that encodes RGB-D voxels and language to predict next-best 6-DoF keyframe actions.", "relevance": 2, "relevance_reason": "Context: seminal RLBench multi-task policy baseline (links=19).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2212.06817, 2304.13705"}
{"id": "2105.10241", "title": "Predictive control barrier functions: Enhanced safety mechanisms for learning-based control", "authors": ["Kim P. Wabersich", "Melanie N. Zeilinger"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2105.10241", "pdf_url": "https://arxiv.org/pdf/2105.10241", "code_url": "", "clusters": ["C7"], "keywords": ["predictive control barrier functions", "safety filter", "feasibility"], "model_class": "other", "techniques": ["soft-constrained predictive CBF"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Adds an always-feasible soft-constrained predictive problem to predictive safety filters so they recover safety even when model assumptions are violated.", "relevance": 2, "relevance_reason": "Robust predictive filters for learned control.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"predictive safety filter\" learning-based"}
{"id": "2203.12601", "title": "R3M: A Universal Visual Representation for Robot Manipulation", "authors": ["Suraj Nair", "A. Rajeswaran", "Vikash Kumar", "Chelsea Finn", "Abhi Gupta"], "year": 2022, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2203.12601", "pdf_url": "https://arxiv.org/pdf/2203.12601", "code_url": "", "clusters": ["C1"], "keywords": ["visual representation", "Ego4D pretraining", "frozen encoder"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Pretrains a compact visual representation on human egocentric video for data-efficient downstream robot manipulation as a frozen perception module.", "relevance": 2, "relevance_reason": "Context: widely used pretrained robot visual encoder (links=24).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2303.04137, 2307.15818"}
{"id": "2203.07747", "title": "Real-Time Neural MPC: Deep Learning Model Predictive Control for Quadrotors and Agile Robotic Platforms", "authors": ["Tim Salzmann", "Elia Kaufmann", "Jon Arrizabalaga", "M. Pavone", "Davide Scaramuzza", "Markus Ryll"], "year": 2022, "venue": "IEEE RA-L 2022", "type": "paper", "url": "https://arxiv.org/abs/2203.07747", "pdf_url": "https://arxiv.org/pdf/2203.07747", "code_url": "", "clusters": ["C5", "C7"], "keywords": ["real-time neural MPC", "quadrotor", "embedded control"], "model_class": "other", "techniques": ["neural dynamics in real-time MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Framework to embed large neural network dynamics models into real-time MPC loops on embedded agile platforms such as quadrotors.", "relevance": 2, "relevance_reason": "Real-time learned control under timing constraints.", "foundational_technique": false, "verified": true, "found_via": "snowball: references of 2506.07339 (S2)"}
{"id": "2210.02317", "title": "Real-Time Reinforcement Learning for Vision-Based Robotics Utilizing Local and Remote Computers", "authors": ["Yan Wang", "Gautham Vasan", "A. Rupam Mahmood"], "year": 2022, "venue": "ICRA 2023", "type": "paper", "url": "https://arxiv.org/abs/2210.02317", "pdf_url": "https://arxiv.org/pdf/2210.02317", "code_url": "", "clusters": ["C6"], "keywords": ["local/remote computation split", "real-time RL", "wireless workstation"], "model_class": "other", "techniques": ["remote-local distributed learning (ReLoD)"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "Distributes SAC/PPO computation between a weak tethered computer and a wirelessly connected powerful workstation for real-time vision-based robot learning, quantifying the benefit.", "relevance": 2, "relevance_reason": "Local/remote compute split for robot learning.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: remote workstation robot policy"}
{"id": "2110.01110", "title": "Safe Control with Neural Network Dynamic Models", "authors": ["Tianhao Wei", "Changliu Liu"], "year": 2022, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2110.01110", "pdf_url": "https://arxiv.org/pdf/2110.01110", "code_url": "", "clusters": ["C7"], "keywords": ["MIND-SIS", "neural network dynamic models", "safety index synthesis"], "model_class": "other", "techniques": ["mixed-integer programming", "safety index synthesis"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Derives provably safe control laws for systems whose dynamics are neural networks via mixed-integer formulations and safety index synthesis.", "relevance": 2, "relevance_reason": "Foundational certified safe control for learned dynamics.", "foundational_technique": true, "verified": true, "found_via": "arXiv search: \"safety index\" synthesis safe control"}
{"id": "2210.03094", "title": "VIMA: General Robot Manipulation with Multimodal Prompts", "authors": ["Yunfan Jiang", "Agrim Gupta", "Zichen Zhang", "Guanzhi Wang", "Yongqiang Dou", "Yanjun Chen", "Li Fei-Fei", "Anima Anandkumar", "Yuke Zhu", "Linxi (Jim) Fan"], "year": 2022, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2210.03094", "pdf_url": "http://arxiv.org/pdf/2210.03094", "code_url": "", "clusters": ["C1", "C4"], "keywords": ["multimodal prompts", "VIMA-Bench"], "model_class": "VLA", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["VIMA-Bench"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Transformer agent that takes interleaved text/image prompts for many manipulation tasks, with an accompanying benchmark.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2212.06817, 2307.15818"}
{"id": "2204.06252", "title": "What Matters in Language Conditioned Robotic Imitation Learning Over Unstructured Data", "authors": ["Oier Mees", "Lukás Hermann", "Wolfram Burgard"], "year": 2022, "venue": "IEEE Robotics and Automation Letters", "type": "paper", "url": "https://arxiv.org/abs/2204.06252", "pdf_url": "https://arxiv.org/pdf/2204.06252", "code_url": "", "clusters": ["C4"], "keywords": ["CALVIN", "language-conditioned imitation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["CALVIN"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Systematic study of design choices for language-conditioned imitation from unstructured play data, benchmarked on CALVIN.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 12 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2307.15818, 2405.12213"}
{"id": "2109.13396", "title": "Bridge Data: Boosting Generalization of Robotic Skills with Cross-Domain Datasets", "authors": ["F. Ebert", "Yanlai Yang", "Karl Schmeckpeper", "Bernadette Bucher", "G. Georgakis", "Kostas Daniilidis", "Chelsea Finn", "S. Levine"], "year": 2021, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2109.13396", "pdf_url": "https://doi.org/10.15607/rss.2022.xviii.063", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "cross-domain", "WidowX"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Multi-domain, multi-task robot dataset on a low-cost arm showing that pooling data across domains boosts generalization of new skills.", "relevance": 2, "relevance_reason": "Context: foundational dataset in OXE/BridgeData lineage (links=27).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2212.06817, 2304.13705"}
{"id": "2112.13222", "title": "Edge Robotics: Edge-Computing-Accelerated Multi-Robot Simultaneous Localization and Mapping", "authors": ["Peng Huang", "Liekang Zeng", "Xu Chen", "Ke Luo", "Zhi Zhou", "Shuai Yu"], "year": 2021, "venue": "IEEE Internet of Things Journal", "type": "paper", "url": "https://arxiv.org/abs/2112.13222", "pdf_url": "https://arxiv.org/pdf/2112.13222", "code_url": "", "clusters": ["C6"], "keywords": ["edge robotics", "multi-robot SLAM", "robot-edge-cloud"], "model_class": "other", "techniques": ["hierarchical edge/cloud map fusion"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "RecSLAM sends robots' raw data to edge servers for real-time map fusion and to the cloud for global merging, after measurements show pure cloud offload is communication-bound.", "relevance": 2, "relevance_reason": "Robot-edge-cloud hierarchy (pre-foundation-model).", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:edge-cloud AND abs:robot"}
{"id": "2108.10470", "title": "Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning", "authors": ["Viktor Makoviychuk", "Lukasz Wawrzyniak", "Yunrong Guo", "Michelle Lu", "Kier Storey", "M. Macklin", "David Hoeller", "N. Rudin", "Arthur Allshire", "Ankur Handa", "Gavriel State"], "year": 2021, "venue": "NeurIPS Datasets and Benchmarks", "type": "paper", "url": "https://arxiv.org/abs/2108.10470", "pdf_url": "https://arxiv.org/pdf/2108.10470", "code_url": "", "clusters": ["C4"], "keywords": ["GPU simulation", "Isaac Gym"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "GPU-resident physics simulator that trains policies end-to-end on GPU without CPU bottlenecks.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2405.05941, 2409.00588"}
{"id": "2107.14483", "title": "ManiSkill: Generalizable Manipulation Skill Benchmark with Large-Scale Demonstrations", "authors": ["Tong-Zhou Mu", "Z. Ling", "Fan-Bo Xiang", "Derek Yang", "Xuanlin Li", "Stone Tao", "Zhiao Huang", "Zhiwei Jia", "Hao Su"], "year": 2021, "venue": "NeurIPS Datasets and Benchmarks", "type": "paper", "url": "https://arxiv.org/abs/2107.14483", "pdf_url": "https://arxiv.org/pdf/2107.14483", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "manipulation", "simulation"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["ManiSkill"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "ManiSkill: simulation benchmark for generalizable manipulation skills with large-scale demonstrations.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2306.03310, 2405.05941"}
{"id": "10.1145/3445814.3446746", "title": "Robomorphic computing: a design methodology for domain-specific accelerators parameterized by robot morphology", "authors": ["Sabrina M. Neuman", "Brian Plancher", "Thomas Bourgeat", "Thierry Tambe", "S. Devadas", "V. Reddi"], "year": 2021, "venue": "ASPLOS 2021", "type": "paper", "url": "https://doi.org/10.1145/3445814.3446746", "pdf_url": "https://arxiv.org/pdf/DOI:10.1145/3445814.3446746", "code_url": "", "clusters": ["C5"], "keywords": ["domain-specific accelerator", "robot dynamics", "FPGA/ASIC"], "model_class": "other", "techniques": ["morphology-parameterized accelerator generation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Seminal methodology that turns a robot's kinematic topology into a custom accelerator for dynamics gradients used in motion planning and control.", "relevance": 2, "relevance_reason": "Foundational robotics accelerator work (pre-foundation-model).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2608.04428 (references)"}
{"id": "2108.11898", "title": "Supervised Compression for Resource-Constrained Edge Computing Systems", "authors": ["Yoshitomo Matsubara", "Ruihan Yang", "Marco Levorato", "Stephan Mandt"], "year": 2021, "venue": "WACV 2022", "type": "paper", "url": "https://arxiv.org/abs/2108.11898", "pdf_url": "https://arxiv.org/pdf/2108.11898", "code_url": "", "clusters": ["C6"], "keywords": ["supervised compression", "split computing", "stochastic bottleneck"], "model_class": "other", "techniques": ["learned feature compression", "distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": true, "tldr": "Compresses split-point features with a distilled stochastic bottleneck and learned prior, beating naive image compression for device-to-server offloading.", "relevance": 2, "relevance_reason": "Foundational split-computing feature compression.", "foundational_technique": true, "verified": true, "found_via": "arXiv API: split computing robot"}
{"id": "2111.11038", "title": "Time-Critical Tasks Implementation in MEC based Multi-Robot Cooperation Systems", "authors": ["Rui Yin", "Yineng Shen", "Huawei Zhu", "Xianfu Chen", "Celimuge Wu"], "year": 2021, "venue": "China Communications", "type": "paper", "url": "https://arxiv.org/abs/2111.11038", "pdf_url": "https://arxiv.org/pdf/2111.11038", "code_url": "", "clusters": ["C6"], "keywords": ["MEC", "multi-robot cooperation", "energy", "time-critical"], "model_class": "other", "techniques": ["joint computation/communication resource optimization"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "A master robot acts as an MEC server for sensing robots, with schemes optimizing computation and communication resources to meet deadlines while balancing energy.", "relevance": 2, "relevance_reason": "Robot-as-edge-server resource allocation.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: edge server robot policy"}
{"id": "2108.03298", "title": "What Matters in Learning from Offline Human Demonstrations for Robot Manipulation", "authors": ["A. Mandlekar", "Danfei Xu", "Josiah Wong", "Soroush Nasiriany", "Chen Wang", "Rohun Kulkarni", "Li Fei-Fei", "S. Savarese", "Yuke Zhu", "Roberto Mart'in-Mart'in"], "year": 2021, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/2108.03298", "pdf_url": "https://arxiv.org/pdf/2108.03298", "code_url": "", "clusters": ["C4"], "keywords": ["robomimic", "offline imitation learning", "benchmark"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["robomimic"], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Large empirical study of offline imitation/RL algorithms on simulated and real multi-stage manipulation with human demos, released as the robomimic benchmark.", "relevance": 2, "relevance_reason": "Context: robomimic benchmark used throughout diffusion-policy efficiency work.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2303.04137, 2304.13705"}
{"id": "2005.07648", "title": "Language Conditioned Imitation Learning Over Unstructured Data", "authors": ["Corey Lynch", "P. Sermanet"], "year": 2020, "venue": "Robotics: Science and Systems Conference", "type": "paper", "url": "https://arxiv.org/abs/2005.07648", "pdf_url": "https://doi.org/10.15607/rss.2021.xvii.047", "code_url": "", "clusters": ["C1"], "keywords": ["language-conditioned imitation", "play data"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Learns language-conditioned policies from unstructured play data with a small fraction of language labels.", "relevance": 2, "relevance_reason": "Context: early language-conditioned policy work (links=12).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2112.03227, 2212.06817"}
{"id": "2010.02193", "title": "Mastering Atari with Discrete World Models", "authors": ["Danijar Hafner", "T. Lillicrap", "Mohammad Norouzi", "Jimmy Ba"], "year": 2020, "venue": "International Conference on Learning Representations", "type": "paper", "url": "https://arxiv.org/abs/2010.02193", "pdf_url": "https://arxiv.org/pdf/2010.02193", "code_url": "", "clusters": ["C8"], "keywords": ["DreamerV2", "discrete latents", "MBRL"], "model_class": "world model", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "DreamerV2 learns behaviours purely inside a discrete-latent world model and reaches human-level Atari performance; a foundational latent world-model reference.", "relevance": 2, "relevance_reason": "context: major robot FM/dataset/benchmark, 23 core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2112.11490, 2203.04955"}
{"id": "2009.12293", "title": "robosuite: A Modular Simulation Framework and Benchmark for Robot Learning", "authors": ["Yu-Ke Zhu", "Josiah Wong", "A. Mandlekar", "Roberto Mart'in-Mart'in", "Abhishek Joshi", "Kevin Lin", "Abhiram Maddukuri", "Soroush Nasiriany", "Yi-Feng Zhu"], "year": 2020, "venue": "arXiv.org", "type": "preprint", "url": "https://arxiv.org/abs/2009.12293", "pdf_url": "https://arxiv.org/pdf/2009.12293", "code_url": "", "clusters": ["C4"], "keywords": ["simulation", "MuJoCo"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Modular MuJoCo-based simulation framework with standard manipulation benchmark tasks.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2306.03310, 2310.17552"}
{"id": "2012.06739", "title": "Sampling Training Data for Continual Learning Between Robots and the Cloud", "authors": ["Sandeep Chinchali", "Evgenya Pergament", "Manabu Nakanoya", "Eyal Cidon", "Edward Zhang", "Dinesh Bharadia", "Marco Pavone", "Sachin Katti"], "year": 2020, "venue": "ISER 2020", "type": "paper", "url": "https://arxiv.org/abs/2012.06739", "pdf_url": "https://arxiv.org/pdf/2012.06739", "code_url": "", "clusters": ["C6"], "keywords": ["fleet data sampling", "continual learning", "robot-cloud"], "model_class": "other", "techniques": ["onboard rare-event sampling"], "base_model": "", "params": "", "hardware": "", "reported_metric": "65.7-81.3% reduction in cloud storage, annotation and compute time", "metric_measures": "other", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": true, "tldr": "HarvestNet samples only rare, useful events on the robot for cloud retraining, cutting fleet-to-cloud data and compute costs.", "relevance": 2, "relevance_reason": "Fleet-cloud data cost reduction.", "foundational_technique": false, "verified": true, "found_via": "snowball from 1902.05703"}
{"id": "10.1007/978-3-030-36150-1_24", "title": "Edge-Enabled Autonomous Navigation and Computer Vision as a Service: A Study on Mobile Robot's Onboard Energy Consumption and Computing Requirements", "authors": ["Jens Lambrecht", "Eugen Funk"], "year": 2019, "venue": "ROBOT 2019 (Advances in Intelligent Systems and Computing)", "type": "paper", "url": "https://doi.org/10.1007/978-3-030-36150-1_24", "pdf_url": "", "code_url": "", "clusters": ["C6"], "keywords": ["edge offloading", "onboard energy", "mobile robot"], "model_class": "other", "techniques": ["edge offloading"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Measures how offloading a mobile robot's navigation and vision to edge services (software-as-a-service) changes its onboard energy use and compute requirements.", "relevance": 2, "relevance_reason": "verified 2026-09-30 by orchestrator: Springer abstract fetched. Early onboard-vs-edge energy evidence for C6 TCO questions.", "foundational_technique": false, "verified": true, "found_via": "gap check: Q2: OpenAlex title_and_abstract.search: energy consumption computing mobile robots"}
{"id": "1908.04935", "title": "Fog Robotics: A Summary, Challenges and Future Scope", "authors": ["Siva Leela Krishna Chand Gudi", "Benjamin Johnston", "Mary-Anne Williams"], "year": 2019, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1908.04935", "pdf_url": "https://arxiv.org/pdf/1908.04935", "code_url": "", "clusters": ["C6"], "keywords": ["fog robotics", "survey"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Short overview of fog robotics motivations (bandwidth, backhaul congestion, latency) and challenges for cloud-assisted robots.", "relevance": 2, "relevance_reason": "Background on fog robotics.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:cloud robotics AND abs:latency"}
{"id": "1910.10897", "title": "Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning", "authors": ["Tian-He Yu", "Deirdre Quillen", "Zhan-Peng He", "Ryan C. Julian", "Avnish Narayan", "Hayden Shively", "Adithya Bellathur", "Karol Hausman", "Chelsea Finn", "S. Levine"], "year": 2019, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/1910.10897", "pdf_url": "https://arxiv.org/pdf/1910.10897", "code_url": "", "clusters": ["C4"], "keywords": ["benchmark", "meta-RL", "multi-task"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Meta-World"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Simulated benchmark of 50 distinct manipulation tasks for multi-task and meta-RL evaluation, widely reused for evaluating fast policies.", "relevance": 2, "relevance_reason": "context: major robot foundation model/dataset/benchmark with many core links", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2112.03227, 2203.04955"}
{"id": "1910.11215", "title": "RoboNet: Large-Scale Multi-Robot Learning", "authors": ["S. Dasari", "F. Ebert", "Stephen Tian", "Suraj Nair", "Bernadette Bucher", "Karl Schmeckpeper", "Siddharth Singh", "S. Levine", "Chelsea Finn"], "year": 2019, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/1910.11215", "pdf_url": "https://arxiv.org/pdf/1910.11215", "code_url": "", "clusters": ["C4"], "keywords": ["dataset", "multi-robot"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Early large-scale multi-robot video/action dataset for pretraining visual manipulation models.", "relevance": 2, "relevance_reason": "Major robot foundation model/dataset/benchmark widely used as a baseline or data source by the efficiency literature (context).", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2212.06817, 2307.15818"}
{"id": "1810.07286", "title": "At Human Speed: Deep Reinforcement Learning with Action Delay", "authors": ["Vlad Firoiu", "Tina W. Ju", "Joshua B. Tenenbaum"], "year": 2018, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1810.07286", "pdf_url": "https://arxiv.org/pdf/1810.07286", "code_url": "", "clusters": ["C7"], "keywords": ["action delay", "predictive model", "reaction time"], "model_class": "world model", "techniques": ["predictive model to undo delay"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Shows deep RL degrades under human-level action delay and restores performance with a learned predictive model that compensates for the delay.", "relevance": 2, "relevance_reason": "Early delay compensation via prediction.", "foundational_technique": true, "verified": true, "found_via": "snowball: references of 2506.07339 (S2)"}
{"id": "1809.09260", "title": "Low Precision Policy Distillation with Application to Low-Power, Real-time Sensation-Cognition-Action Loop with Neuromorphic Computing", "authors": ["Jeffrey L. McKinstry", "D. Barch", "Deepika Bablani", "M. DeBole", "S. K. Esser", "J. Kusnitz", "J. Arthur", "D. Modha"], "year": 2018, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1809.09260", "pdf_url": "https://arxiv.org/pdf/1809.09260", "code_url": "", "clusters": ["C2"], "keywords": ["low precision", "policy distillation", "neuromorphic"], "model_class": "other", "techniques": ["low-precision distillation"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": ["Atari"], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Distills a full-precision RL policy into a low-precision network running real-time on neuromorphic hardware.", "relevance": 2, "relevance_reason": "Early low-precision policy distillation (pre-2022).", "foundational_technique": false, "verified": true, "found_via": "snowball from 2412.01034 (references)"}
{"id": "1811.02790", "title": "ROBOTURK: A Crowdsourcing Platform for Robotic Skill Learning through Imitation", "authors": ["A. Mandlekar", "Yuke Zhu", "Animesh Garg", "Jonathan Booher", "Max Spero", "Albert Tung", "Julian Gao", "John Emmons", "Anchit Gupta", "Emre Orbay", "S. Savarese", "Li Fei-Fei"], "year": 2018, "venue": "Conference on Robot Learning", "type": "paper", "url": "https://arxiv.org/abs/1811.02790", "pdf_url": "https://arxiv.org/pdf/1811.02790", "code_url": "", "clusters": ["C4"], "keywords": ["crowdsourced teleoperation", "dataset"], "model_class": "other", "techniques": [], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Crowdsourcing platform for 6-DoF teleoperated demonstration collection via smartphones.", "relevance": 2, "relevance_reason": "Context: data-collection platform underlying robomimic-style datasets.", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2310.06114, 2405.12213"}
{"id": "2018.00047", "title": "RoboX: An End-to-End Solution to Accelerate Autonomous Control in Robotics", "authors": ["Jacob Sacks", "Divya Mahajan", "R. C. Lawson", "H. Esmaeilzadeh"], "year": 2018, "venue": "ISCA 2018", "type": "paper", "url": "https://doi.org/10.1109/ISCA.2018.00047", "pdf_url": "https://arxiv.org/pdf/DOI:10.1109/ISCA.2018.00047", "code_url": "", "clusters": ["C5"], "keywords": ["MPC accelerator", "low power", "robotics"], "model_class": "other", "techniques": ["domain-specific accelerator for model predictive control"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Early accelerator and programming stack for MPC-based motion planning and control under low power budgets.", "relevance": 2, "relevance_reason": "Seminal robotics accelerator; classical control rather than FMs.", "foundational_technique": false, "verified": true, "found_via": "snowball from 2608.04428 (references)"}
{"id": "1812.00568", "title": "Visual Foresight: Model-Based Deep Reinforcement Learning for Vision-Based Robotic Control", "authors": ["F. Ebert", "Chelsea Finn", "S. Dasari", "Annie Xie", "Alex X. Lee", "S. Levine"], "year": 2018, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1812.00568", "pdf_url": "https://arxiv.org/pdf/1812.00568", "code_url": "", "clusters": ["C8"], "keywords": ["video prediction", "model-based RL"], "model_class": "world model", "techniques": ["video prediction", "MPC"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": null, "tldr": "Visual Foresight learns action-conditioned video prediction from self-supervised robot data and plans with it via model-predictive control for vision-based manipulation.", "relevance": 2, "relevance_reason": "context: seminal video-prediction world model for robot planning", "foundational_technique": false, "verified": true, "found_via": "snowball round 2: cited by 2203.04955, 2405.15223"}
{"id": "s2-control-delay-rl-memoryless-iros-2010", "title": "Control delay in Reinforcement Learning for real-time dynamic systems: A memoryless approach", "authors": ["E. Schuitema", "L. Buşoniu", "Robert Babuška", "P. Jonker"], "year": 2010, "venue": "IROS 2010", "type": "paper", "url": "https://doi.org/10.1109/IROS.2010.5650345", "pdf_url": "", "code_url": "", "clusters": ["C7"], "keywords": ["control delay", "real-time RL"], "model_class": "other", "techniques": ["delay-aware RL"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": null, "onboard_deployment": null, "tldr": "Proposes temporal-difference learning algorithms that account for control delay (the gap between sensing and acting) without augmenting the state, tested in simulation and on a real robot system.", "relevance": 2, "relevance_reason": "verified 2026-09-30 by orchestrator: abstract via OpenAlex (DOI 10.1109/IROS.2010.5650345). Seminal delay-aware RL cited by RTC.", "foundational_technique": true, "verified": true, "found_via": "snowball: references of 2506.07339 (S2)"}
{"id": "pi0-on-a-budget-repo", "title": "pi0-on-a-budget", "authors": ["imjbassi"], "year": null, "venue": "GitHub", "type": "repo", "url": "https://github.com/imjbassi/pi0-on-a-budget", "pdf_url": "", "code_url": "", "clusters": ["C3"], "keywords": ["consumer GPU", "LoRA", "pi0-FAST", "12GB"], "model_class": "VLA", "techniques": ["LoRA", "frozen image encoder"], "base_model": "pi0-FAST", "params": "", "hardware": "RTX 4070 12GB", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Community attempt to LoRA-fine-tune pi0-FAST on a 12GB GPU by also freezing the image encoder; no training measurements yet.", "relevance": 2, "relevance_reason": "Practical memory-constrained fine-tuning notes; early-stage.", "foundational_technique": false, "verified": true, "found_via": "WebSearch: pi0 fine-tuning openpi LoRA recipe Physical Intelligence blog"}
{"id": "2607.10811", "title": "Distributed Agent System: Fault-Tolerant Collaboration Among Embodied Agents", "authors": ["Kai Yu", "Lu Chen", "Hanqi Li"], "year": 2026, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2607.10811", "pdf_url": "https://arxiv.org/pdf/2607.10811", "code_url": "", "clusters": ["C6", "C7"], "keywords": ["device-edge-cloud", "fault tolerance", "embodied agents"], "model_class": "other", "techniques": ["two-layer fault tolerance", "semi-formal agent protocols"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Device-edge-cloud framework framing reliability of heterogeneous embodied LLM agents as system-level fault tolerance.", "relevance": 1, "relevance_reason": "Conceptual device-edge-cloud agent framework.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: device-edge robot"}
{"id": "2505.03587", "title": "Meta-reasoning Using Attention Maps and Its Applications in Cloud Robotics", "authors": ["Adrian Lendinez", "Renxi Qiu", "Lanfranco Zanzi", "Dayou Li"], "year": 2025, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/2505.03587", "pdf_url": "https://arxiv.org/pdf/2505.03587", "code_url": "", "clusters": ["C6"], "keywords": ["meta-reasoning", "cloud robots"], "model_class": "other", "techniques": ["attention-map meta-reasoning"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Meta-reasoning framework using semantic attention maps, demonstrated on cloud-connected robots in real-world unexpected situations.", "relevance": 1, "relevance_reason": "Cloud robotics application, not efficiency-focused.", "foundational_technique": false, "verified": true, "found_via": "snowball from 1902.05703"}
{"id": "2210.03936", "title": "Cloud Native Robotic Applications with GPU Sharing on Kubernetes", "authors": ["Giovanni Toffetti", "Leonardo Militano", "Seán Murphy", "Remo Maurer", "Mark Straub"], "year": 2022, "venue": "IROS 2022 Cloud Robotics Workshop", "type": "paper", "url": "https://arxiv.org/abs/2210.03936", "pdf_url": "https://arxiv.org/pdf/2210.03936", "code_url": "", "clusters": ["C6"], "keywords": ["Kubernetes", "GPU sharing", "cloud native robotics"], "model_class": "other", "techniques": ["container orchestration", "GPU sharing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Experience report on running robotic applications on a Kubernetes cluster with real robots, highlighting networking and GPU-sharing shortcomings for deep-learning workloads.", "relevance": 1, "relevance_reason": "Background on shared GPU infrastructure for robots.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: ti:cloud AND ti:robot"}
{"id": "1901.03297", "title": "Robotics Evolution: from Remote Brain to Cloud", "authors": ["Alaa F. Sheta", "Nazeeh Ghatasheh", "Hossam Faris", "Ali Rodan"], "year": 2019, "venue": "arXiv", "type": "preprint", "url": "https://arxiv.org/abs/1901.03297", "pdf_url": "https://arxiv.org/pdf/1901.03297", "code_url": "", "clusters": ["C6"], "keywords": ["remote brain", "cloud robotics", "history"], "model_class": "other", "techniques": ["survey"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": false, "onboard_deployment": false, "tldr": "Overview tracing robotics from remote-brain designs to cloud robotics and the resulting compute-offload opportunities.", "relevance": 1, "relevance_reason": "Historical background only.", "foundational_technique": false, "verified": true, "found_via": "search query"}
{"id": "1811.05578", "title": "Fog Robotics for Efficient, Fluent and Robust Human-Robot Interaction", "authors": ["Siva Leela Krishna Chand Gudi", "Suman Ojha", "Benjamin Johnston", "Jesse Clark", "Mary-Anne Williams"], "year": 2018, "venue": "IEEE NCA 2018", "type": "paper", "url": "https://arxiv.org/abs/1811.05578", "pdf_url": "https://arxiv.org/pdf/1811.05578", "code_url": "", "clusters": ["C6"], "keywords": ["fog robotics", "HRI", "latency"], "model_class": "other", "techniques": ["fog computing"], "base_model": "", "params": "", "hardware": "", "reported_metric": "", "metric_measures": "", "benchmarks": [], "real_robot_eval": true, "onboard_deployment": false, "tldr": "Early fog-robotics architecture reducing cloud latency for human-robot interaction by processing at nearby fog nodes.", "relevance": 1, "relevance_reason": "Early background.", "foundational_technique": false, "verified": true, "found_via": "arXiv API: abs:fog robotics"}