# Efficient Embodied AI — Literature Database

This database covers the research scope of the CoRL 2026 workshop
[*Efficient Foundation Models for Real-Time Embodied AI*](https://efficient-embodied-ai.github.io/). That scope is making robot foundation models
(VLA policies, world-action models, emerging architectures) fast, cheap and safe enough to run on the robot.

- **2,232 unique items** (papers, preprints, tech reports, blogs, repos, benchmarks), mostly from 2024–2026.
- The items fall into nine topic clusters (C1–C9, defined below).
- The snapshot was taken on 2026-09-30.

## Files

| File | Contents |
|---|---|
| `literature_db.jsonl` | The full database, one JSON record per line. |
| `literature_db.csv` | The same data, flattened. List fields are joined with `; `. |
| `by_cluster/C1.md` … `C9.md` | Annotated bibliography per cluster, sorted by relevance, then year. Each opens with a cluster summary. |
| `efficiency_table.csv` | The 617 items that report efficiency numbers (model, params, hardware, metric, what it measures, benchmarks). |
| `overview.md` | Counts, technique taxonomy, top 30 must-reads, answers to the workshop's key questions, research gaps, items needing review. |
| `search_log.md` | Every query run and how the citation snowball was done. |
| `docs/` | The searchable website (GitHub Pages): https://salmitta.github.io/efficient-embodied-ai-literature/ |

## Website search

Everything runs in the visitor's browser. There is no server and no API key.

- **Keyword search** (MiniSearch): ranked, typo-tolerant matching over titles, TL;DRs, techniques, keywords, authors, venue and hardware.
- **Match by meaning** (on by default): the query is embedded in the browser with `Xenova/bge-small-en-v1.5` (about 34 MB, downloaded once). It is compared against the precomputed record embeddings in `docs/embeddings.bin`. Keyword and meaning scores are blended: short queries (names, acronyms) lean on keywords, longer descriptive queries lean on meaning.
- **Topic overview**: for any search or filter, a deterministic summary of the best-matching items: counts by year and cluster, top techniques, hardware mentioned, evidence shares, and reported efficiency numbers grouped by `metric_measures`.
- **Summarize this topic with AI** (optional): runs Qwen2.5 1.5B or 3B in the browser through [WebLLM](https://github.com/mlc-ai/web-llm) on WebGPU (about a 1 GB or 2 GB one-time download). It sees only the top 16 results and cites them as numbered links. It needs a recent desktop Chrome or Edge, or Safari with WebGPU. The output is AI-generated and should be checked against the cited items.

- **Propose research ideas** (optional, same in-browser model; the 3B model is recommended): proposes 3 testable ideas for the current query. Each idea has an idea, what it builds on, the gap it addresses, a first experiment (model, benchmark, hardware, metric type) and a risk. The ideas are grounded in:
  - the top 12 retrieved papers, which each idea must cite;
  - gap signals computed exactly from the matching set, such as the share with real-robot evaluation or onboard deployment, metric types never reported, clusters barely represented, and common benchmarks and hardware;
  - the workshop's five key questions.

  An optional constraints box (e.g. "must run on Jetson Orin NX; simulation only") steers the ideas. After generation, the page flags any idea that cites no retrieved paper, and shows the exact evidence the model was given. The ideas are starting points, not vetted proposals.

Opening `docs/index.html` directly from disk works for keyword search. Match by meaning needs the hosted site, or any local web server (`python3 -m http.server` inside `docs/`).

## Clusters

| | Cluster | Items |
|---|---|---|
| C1 | Efficient architectures for VLA policies and emerging designs | 733 |
| C2 | Compression for robot foundation models | 365 |
| C3 | Memory-efficient adaptation and fine-tuning | 453 |
| C4 | Benchmarking methodology and composability | 363 |
| C5 | Hardware-aware design and accelerators | 245 |
| C6 | Edge-cloud co-design and fleet systems | 218 |
| C7 | Safety under latency and timing variability | 539 |
| C8 | Efficient world models for planning and safety monitoring | 522 |
| C9 | People, venues and blogs (speaker/organizer and lab-blog sweep) | 260 |

An item can belong to several clusters, so the counts add up to more than the total.

## What `relevance` means

`relevance` (1–5) rates how central an item is to the workshop's question: making robot foundation models fast, cheap and safe
enough to deploy on the robot. It does not rate paper quality.

| Score | Meaning | Count |
|---|---|---|
| **5** | **Core.** Directly about the efficiency, real-time execution or deployability of robot or embodied foundation models, with substantive evidence. Also used for works the whole area builds on (e.g. π0, OpenVLA-OFT, RTC). | 248 |
| **4** | **Directly relevant.** An efficiency, latency, compression, adaptation, safety-under-latency or evaluation contribution for robot foundation models, narrower in scope or less established. | 529 |
| **3** | **Relevant.** Covers one of the following: <ul><li>an adjacent technique, or work where efficiency is a secondary angle</li><li>a foundational ML efficiency technique that robotics work builds on (AWQ, GPTQ, LoRA, FlashAttention, speculative decoding…)</li><li>safety, world-model or benchmark work relevant to the workshop questions</li></ul> | 953 |
| **2** | **Context.** A major robot foundation model, dataset or benchmark that the efficiency literature builds on or evaluates against (e.g. Open X-Embodiment, DROID), or background methods. | 497 |
| **1** | **Marginal.** Kept only for completeness. | 5 |

Caveats:
- Scores are the judgment of the subagent that recorded the item, from its abstract. They were not re-calibrated across agents. When several agents recorded the same item, the highest score was kept. Treat differences of one point as noise.
- For a core reading set, filter `relevance >= 4` (777 items). `relevance >= 3` gives 1,730 items.
- `foundational_technique: true` marks generic (non-robotics) ML techniques that are included because robotics work builds on them. It is independent of relevance.

## Efficiency numbers are not comparable across `metric_measures`

`reported_metric` holds the number as the paper reports it. `metric_measures` says what kind of quantity it is. **Numbers from different
`metric_measures` categories must not be compared or ranked against each other.** Even within one category, compare with care.

| `metric_measures` | What it is | Items | Watch out for |
|---|---|---|---|
| `closed-loop control` | Rate or latency of the full perception → policy → actuation loop while the robot (or simulator) runs | 37 | The closest to "real-time on the robot", but setups still differ (cameras, action horizon, network). |
| `open-loop chunk rate` | How fast the model produces action chunks or steps in isolation | 18 | A 50-action chunk at 5 Hz is not a 250 Hz controller. Execution, sensing and replanning are excluded. |
| `throughput` | Actions, tokens or requests per second, often batched or on a server | 69 | Batched throughput (e.g. at concurrency 8) says little about batch-1 latency on one robot. |
| `relative latency` | Speedup or latency reduction vs. the paper's own baseline (e.g. "2.3× faster") | 227 | The baselines, hardware and precision differ between papers. A 3× speedup in one paper and a 3× speedup in another are not the same thing. |
| `relative memory` | Memory or model-size reduction vs. a baseline | 34 | Weight memory, peak VRAM and unified-memory footprint are different quantities. |
| `energy` | Joules or watts per inference, action or episode | 17 | Methods differ: board power vs. GPU rail, idle-subtracted or not, internal telemetry vs. external meter. |
| `other` | FLOPs, parameter counts, task-success changes, mixed metrics | 215 | Heterogeneous. Read `reported_metric` itself. |

Further cautions:
- **Hardware is often missing.** Only 174 of the 617 items with numbers state the hardware. A latency with no hardware should not be compared with anything.
- **Numbers come from abstracts, pages and tables as the authors state them.** Nothing was re-measured, and ambiguous numbers were left out rather than guessed.
- **Faster per step is not always faster per task.** Several works show that per-step speedups do not always shorten tasks or preserve success (e.g. the Speedup Paradox, 2606.28529, and From Inference Efficiency to Embodied Efficiency, 2603.19131).

## Verification status and pending items

Every record was verified against a fetched source (arXiv, OpenReview, Semantic Scholar/OpenAlex abstract, proceedings or project page).
A random sample of 60 arXiv records matched arXiv titles exactly.

As of 2026-09-30:
- **1 item still unverified:** `s2-simulation-evaluating-robot-policies-asynchronous-real-time` ("Simulation for Evaluating Robot
  Policies Should Be Asynchronous and Real-Time"). Its title and authors come from Semantic Scholar, but no abstract could be retrieved.
  It is marked `verified: false`.
- **Venue spot-checks: done.** 22 venues that were originally filled in without a fetched source were re-checked (details in `overview.md`):
  - 20 confirmed;
  - 1 corrected (2310.17552: CoRL 2023 → ICRA 2024);
  - 1 reverted to "arXiv" (2502.04296: a CVPR 2025 claim that could not be confirmed).
- **Other venues:** most 2026 preprints are listed as "arXiv". Their acceptance status was not checked.

## Coverage notes

- **How the database was built:**
  - Nine topic subagents, 801 logged queries.
  - A citation snowball (Semantic Scholar citations and references of all 571 relevance ≥ 4 arXiv items; 2,574 candidate abstracts triaged).
  - Two targeted gap and recency sweeps.
  - Deduplication by arXiv ID → DOI → URL → normalized/fuzzy title.
- **Saturation:**
  - Search saturated for quantization, MoE/early exit, speculative decoding and continual learning.
  - It did not fully saturate for efficient world-action models and adaptive-compute VLAs, where new preprints appear weekly.
- **Not covered:** public workshop submissions. None were visible on OpenReview at snapshot time (the deadline is 2026-10-12).

## Rebuilding

```bash
python3 _work/merge.py          # raw subagent outputs in _work/raw, _work/raw2 -> literature_db.*, by_cluster/, efficiency_table.csv
python3 _work/build_reports.py  # -> overview.md, search_log.md
cd _work && npm install && node embed.mjs && cd ..   # -> docs/embeddings.bin (needs Node 22+; rerun whenever the database changes)
python3 _work/build_site.py     # -> docs/ (website; warns if the embeddings are out of date)
```

## Record schema

`id` (arXiv ID, DOI or slug), `title`, `authors`, `year`, `venue`, `type`, `url`, `pdf_url`, `code_url`, `clusters`, `keywords`,
`model_class`, `techniques`, `base_model`, `params`, `hardware`, `reported_metric`, `metric_measures`, `benchmarks`,
`real_robot_eval`, `onboard_deployment`, `tldr` (written by the curators, not copied from abstracts), `relevance`, `relevance_reason`,
`foundational_technique`, `verified`, `found_via`.