RoboRACE-Bench / README.md
Nabidnur's picture
RoboRACE: code, cached agent outputs, results, media, SO-101 renders
627f161 verified
|
Raw History Blame Contribute Delete
54.9 kB
---
pretty_name: RoboRACE-Bench
license: other
license_name: roborace-mixed
license_link: https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/LICENSES.md
language:
- en
task_categories:
- robotics
- visual-question-answering
- image-classification
tags:
- robotics
- multi-agent
- byzantine-robustness
- truth-inference
- vlm
- llm-agents
- so-101
- mujoco
- erqa
- planbench
- cosmos-reason
size_categories:
- 1K<n<10K
configs:
- config_name: so101_views
data_files:
- split: history
path: viewer/so101_views/history.parquet
- split: evaluation
path: viewer/so101_views/evaluation.parquet
- config_name: so101_vlm_scores
data_files: viewer/so101_vlm_scores/train.parquet
- config_name: panda_views
data_files:
- split: history
path: viewer/panda_views/history.parquet
- split: evaluation
path: viewer/panda_views/evaluation.parquet
- config_name: panda_vlm_scores
data_files: viewer/panda_vlm_scores/train.parquet
- config_name: so101_holdout_views
data_files:
- split: history
path: viewer/so101_holdout_views/history.parquet
- split: evaluation
path: viewer/so101_holdout_views/evaluation.parquet
- config_name: so101_holdout_vlm_scores
data_files: viewer/so101_holdout_vlm_scores/train.parquet
- config_name: panda_holdout_views
data_files:
- split: history
path: viewer/panda_holdout_views/history.parquet
- split: evaluation
path: viewer/panda_holdout_views/evaluation.parquet
- config_name: panda_holdout_vlm_scores
data_files: viewer/panda_holdout_vlm_scores/train.parquet
- config_name: trackD_maze_results
data_files: viewer/trackD_maze_results/train.parquet
- config_name: maze_llm_agents
data_files: viewer/maze_llm_agents/train.parquet
- config_name: erqa_cosmos_reason2_2b
data_files:
- split: px448
path: viewer/erqa_cosmos_reason2_2b/px448.parquet
- split: px768
path: viewer/erqa_cosmos_reason2_2b/px768.parquet
- config_name: erqa_hosted_agents
data_files:
- split: test
path: viewer/erqa_hosted_agents/test.parquet
- config_name: planbench_agents
data_files: viewer/planbench_agents/train.parquet
- config_name: trackC_llm_planners
data_files: viewer/trackC_llm_planners/train.parquet
- config_name: trackA0_procedural_scores
data_files: viewer/trackA0_procedural_scores/train.parquet
---
<div align="center">
# RoboRACE
### Byzantine-robust multi-agent robotics — when some of your robots, cameras or AI agents are wrong on purpose
**CPU-only · 2 public benchmarks (ERQA, PlanBench) + 5 RoboRACE tracks · two robot arms, a mobile-robot maze swarm, real VLMs and LLMs · two learned models on one design (RACE-Net-V, RACE-Net-M) · a hold-out evaluated after a recorded freeze · every number regenerated from cached outputs**
[Leaderboard](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/results/LEADERBOARD.md) · [Protocol](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/BENCHMARK.md) · [Models](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/MODELS.md) · [Plan](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/PLAN.md) · [Architecture](#architecture) · [Videos](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/tree/main/media/videos) · [Figures](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/tree/main/media/figures)
[🤗 Dataset · full work](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench) · [🤗 Model · RACE-Net-V](https://huggingface.co/Nabidnur/RoboRACE-Net) · [🤗 Model · RACE-Net-M](https://huggingface.co/Nabidnur/RoboRACE-Net-M) · [🤗 Space · showcase](https://huggingface.co/spaces/Nabidnur/RoboRACE) · [🤗 Space · liar playground](https://huggingface.co/spaces/Nabidnur/RoboRACE-Playground)
<img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/robot_rollouts.gif" width="960" alt="SO-101 arm rollouts seen by five camera agents; three cameras lie; majority vote flips, RACE-S stays correct">
<sub>A MuJoCo SO-101 arm (Menagerie model) runs pick-and-place; ground truth is the final physics state. Five camera agents judge success with
Cosmos-Reason2-2B on a CPU. Then 3 of 5 cameras invert their reports: majority vote flips, <b>RACE-S</b> learns negative weights for the liars without labels and stays correct.
The GIF is a 12-second highlight; the full 43-second film (3 episodes) is the <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/robot_rollouts.mp4">MP4 (1280×720, 30 fps)</a>.</sub>
</div>
> **This is the Hugging Face home of RoboRACE.** It holds the complete, detailed work: code, the cached outputs of every
> agent (VLMs, LLMs, simulated robots), per-cell benchmark results, leaderboards, figures, videos and the 1,100 SO-101
> camera renders for each of two robot arms. Code is mirrored on GitHub: [github.com/Hisernberg/robotics](https://github.com/Hisernberg/robotics). Showcase page: [Space · Nabidnur/RoboRACE](https://huggingface.co/spaces/Nabidnur/RoboRACE) · interactive: [Space · liar playground](https://huggingface.co/spaces/Nabidnur/RoboRACE-Playground) · model: [Nabidnur/RoboRACE-Net](https://huggingface.co/Nabidnur/RoboRACE-Net).
## Videos
<table><tr>
<td width="50%"><video controls muted loop playsinline width="100%" src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/robot_rollouts.mp4"></video><br>
<sub><b>SO-101 camera swarm</b>: three physics rollouts, five Cosmos-Reason2 camera agents, honest swarm vs 3-of-5 liars.</sub></td>
<td width="50%"><video controls muted loop playsinline width="100%" src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/swarm_rendezvous.mp4"></video><br>
<sub><b>Byzantine rendezvous</b>: 12 robots, 4 colluding liars — DeGroot vs W-MSR vs RACE-C + W-MSR.</sub></td>
</tr><tr>
<td width="50%"><video controls muted loop playsinline width="100%" src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/planbench_modulo.mp4"></video><br>
<sub><b>PlanBench LLM-Modulo</b>: a real cached LLM plan is rejected by the exact simulator and repaired.</sub></td>
<td width="50%"><video controls muted loop playsinline width="100%" src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_swarm.mp4"></video><br>
<sub><b>Maze swarm (Track D)</b>: three of six robots invert their map reports — majority map vs RACE map.</sub></td>
</tr><tr>
<td width="50%"><video controls muted loop playsinline width="100%" src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/panda_rollouts.mp4"></video><br>
<sub><b>Franka Panda camera swarm</b>: the cross-embodiment replication.</sub></td>
<td width="50%"><video controls muted loop playsinline width="100%" src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/camera_swarm.mp4"></video><br>
<sub><b>Track A0</b> (first iteration): procedural scenes judged by the same camera swarm.</sub></td>
</tr></table>
## What is in this repository
| path | content |
|---|---|
| `data/so101/{hist,eval}/<seed>/<camera>.png` | 220 SO-101 episodes × 5 cameras (448×448), with per-split `metadata.jsonl` |
| `viewer/<config>/*.parquet` | Parquet copies behind the dataset viewer; **`so101_views`** embeds every image with its task, physics label, failure mode and Cosmos score |
| `data/so101/episodes.jsonl` | per-episode metadata: task, family (put / stack), scripted mode, **physics ground truth**, final object pose |
| `results/so101/vlm_scores.jsonl` | Cosmos-Reason2-2B P(success) for every (episode, camera) — config **`so101_vlm_scores`** |
| `data/panda/...`, `results/panda/vlm_scores.jsonl` | the same for the **Franka Panda** replication (220 episodes × 5 cameras) — configs **`panda_views`**, **`panda_vlm_scores`** |
| `data/*_holdout/...`, `results/*_holdout/vlm_scores.jsonl`, `results/holdout/FREEZE.json` | the **hold-out**: 170 new episodes per arm, evaluated only after the recorded freeze commit — configs **`so101_holdout_views`**, **`panda_holdout_views`** and the matching `*_vlm_scores` |
| `models/racenet`, `models/racenet_m` | weights of both models (**RACE-Net-V**, **RACE-Net-M**) and every candidate checkpoint |
| `results/trackD/rows.json` | Track D maze swarm: SPL and success of every fusion rule in every (maze, attack, liar-count) cell — config **`trackD_maze_results`** |
| `results/maze_llm/<agent>.jsonl` | LLM agents on text mazes: every attempt, validator feedback, repair count — config **`maze_llm_agents`** |
| `models/racenet/` | RACE-Net weights (safetensors), config and training log; also at [Nabidnur/RoboRACE-Net](https://huggingface.co/Nabidnur/RoboRACE-Net) |
| `results/kaggle/` | independent replication run on Kaggle CPU (log + comparison) |
| `results/erqa/cosmos_r2_2b*.jsonl` | ERQA: full option distributions for all 400 questions at ≤448² and ≤768² px — config **`erqa_cosmos_reason2_2b`** |
| `results/erqa/<agent>.jsonl` | ERQA answers + raw replies of hosted VLM agents (Gemini, Qwen, Gemma, Gemini Robotics-ER) — config **`erqa_hosted_agents`** |
| `results/planbench/<agent>.jsonl` | PlanBench Blocksworld: every attempt, parsed plan, validator verdict, repair count — config **`planbench_agents`** |
| `results/trackC/answers.jsonl` | Track C: zero-shot / engineered / saboteur answers of 4 LLMs — config **`trackC_llm_planners`** |
| `results/*/rows.json`, `results/leaderboard.json`, `results/LEADERBOARD.md` | per-cell results of every method under every attack and receiver; bootstrap-ranked leaderboards |
| `media/` | figures, architecture diagrams (Raschka style + PlotNeuralNet), videos (MP4 + GIF) |
| `src/`, `scripts/`, `tests/`, `docs/` | the full code base, protocol ([BENCHMARK](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/BENCHMARK.md)) and research plan ([PLAN](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/PLAN.md)) |
```python
from datasets import load_dataset
views = load_dataset("Nabidnur/RoboRACE-Bench", "so101_views", split="evaluation") # images + labels
scores = load_dataset("Nabidnur/RoboRACE-Bench", "so101_vlm_scores", split="train") # Cosmos-Reason2 P(success) per camera
erqa = load_dataset("Nabidnur/RoboRACE-Bench", "erqa_cosmos_reason2_2b", split="px448")
plans = load_dataset("Nabidnur/RoboRACE-Bench", "planbench_agents", split="train")
```
Re-run every number on a CPU from this repository:
```bash
huggingface-cli download Nabidnur/RoboRACE-Bench --repo-type dataset --local-dir RoboRACE-Bench && cd RoboRACE-Bench
pip install -r requirements.txt && python scripts/run_benchmark.py --reuse-tracks && python scripts/make_figures.py
```
---
## Headline results
| benchmark | what is measured | strongest baseline | **ours** |
|---|---|---|---|
| **ERQA** (Gemini Robotics team, 400 Q) | reproduce Cosmos-Reason2-2B on CPU | NVIDIA reports 37.75 | **33.8** (95% CI 29.3–38.5) ✓ inside |
| **ERQA** multi-agent | 4 VLM agents, 283 shared Q | best single agent 54.1 | best swarm **54.8** · robot's own model 37.1 → **43.8** with CGA |
| **PlanBench** Blocksworld | plan generation, 92 instances | official one-shot prompt | LLM-Modulo + swarm cascade: see [table](#planbench-blocksworld--the-validator-is-the-planner) |
| **Track A** SO-101 camera swarm | success verification under 7 attacks | receiver alone 71.3 · majority 59.8 | **83.4** race_net (new, learned) (rank 1) |
| **Track A-Panda** cross-embodiment | same protocol on a Franka Panda, methods frozen | receiver alone 74.5 · majority 62.2 | **87.0** race_s_online (new) |
| **RACE-Net-V** (our model, 0.2 M params) | learned fusion, zero-shot on real camera swarms | RACE-S online 82.7 / 87.0 | **83.4** SO-101 (rank 1) / **84.2** Panda · RACE-Pool **82.0** / **85.9** |
| **Hold-out** (new episodes, both arms, evaluated after the freeze) | same protocol, nothing tuned on it | receiver alone 72.7 / 77.1 · majority 62.3 / 64.2 | RACE-S online **82.6** / **85.4** · RACE-Pool **81.2** / **86.7** · RACE-Net-V 79.0 / 84.3 |
| **Track B** 12-robot rendezvous | mission success, up to 5 Byzantine robots | W-MSR 82.1 | **94.3** RACE-C + W-MSR (rank 1) |
| **Track D** maze swarm · **RACE-Net-M** (our model, 45 K params) | SPL under 6 attacks, 10 fusion rules | W-MSR trimming 64.1 · beta-reputation 63.8 · EigenTrust 63.0 | **70.2** RACE-Net-M (rank 1; oracle 71.3) · hand-written RACE map 66.2 |
| **Track C** LLM planners | exact-plan accuracy | zero-shot mean 88.1 | engineered mean **98.2** · flash-lite 69.0 → **92.9** |
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/overview.png" width="920"></p>
**Videos** (GIF previews play inline below; full-quality 1280×720 MP4s):
[SO-101 camera swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/robot_rollouts.mp4) ·
[Byzantine rendezvous](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/swarm_rendezvous.mp4) ·
[PlanBench LLM-Modulo replay](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/planbench_modulo.mp4) ·
[maze swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_swarm.mp4) · [maze, wall-inventing liars](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_false_walls.mp4) ·
[Franka Panda camera swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/panda_rollouts.mp4)
## The idea in one paragraph
A robot team fuses evidence from many agents. Some of them are broken, biased, colluding or adversarial. Majority vote and classical
crowdsourcing (Dawid–Skene, GLAD, MACE) fail once liars are the majority, because the latent-truth model becomes symmetric. **RACE**
(from the *Liars Are Information* program) breaks the symmetry with an **anchor**: each receiver knows it is honest. Peers then get *signed*
channel weights: positive means trust, negative means **invert**, near zero means discard. A consistent liar carries information. This repo adds
**RACE-S** (score level), **online** variants (against sleepers), **confidence-gated anchoring (CGA)** (a measured compromise between robustness and
clean-data gains), **RACE-C + W-MSR** (continuous robot consensus), an **engineered / LLM-Modulo** planning layer, and two **learned** versions of the idea:
**RACE-Net-V** for camera swarms and **RACE-Net-M** for map fusion, plus **RACE-Pool**, an AIP-style pool of anchored experts.
## Architecture
<table><tr>
<td width="52%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/cosmos_reason2_verifier.png"></td>
<td width="48%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/race_s.png"><br><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/roborace_system.png"></td>
</tr></table>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/plotnn/cosmos_verifier_plotnn.png" width="900"><br>
<sub>Same verifier as a <a href="https://github.com/HarisIqbal88/PlotNeuralNet">PlotNeuralNet</a> tensor diagram (one 448×448 view → 196 image tokens → P(yes)).
Exact numbers come from the checkpoint: 2.44 B parameters (LLM decoder 1,409 M, embeddings 311 M, untied head 311 M, ViT 302 M, merger + DeepStack 101 M).
Vector version: <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/plotnn/cosmos_verifier_plotnn.pdf">PDF</a> (LaTeX source alongside). Diagrams are drawn by <code>scripts/make_architecture.py</code> in the style of Sebastian Raschka's LLM architecture figures.</sub></p>
### Our two models: one design
<table><tr>
<td width="50%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/racenet.png"></td>
<td width="50%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/racenet_m.png"></td>
</tr></table>
| | **RACE-Net-V** · pick-and-place | **RACE-Net-M** · maze |
|---|---|---|
| question | did the arm succeed? (camera swarm) | which cells are walls? (robot swarm) |
| receiver's anchor | its own score for every episode | its own exact sensing of some cells |
| structure | axial set transformer over agents × episodes | anchored channel encoder → peer transformer → signed read-out, 3 refinement rounds |
| shared design | anchor on the receiver · estimate each peer's channel in context · **invert** consistent liars · sets, not sequences · synthetic training only |
| size · training | 213,889 parameters · synthetic swarms | 44,837 parameters · simulated mazes with random liars |
Both come from RACE and AIP in the *Liars Are Information* program: RACE supplies the anchored, signed channel estimation; AIP supplies the pooling of
experts (RACE-Pool). The full description is in [docs/MODELS.md](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/MODELS.md).
---
## Part I — public benchmarks
### ERQA — embodied reasoning
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/erqa_results.png" width="980"></p>
**Reproduction.** Cosmos-Reason2-2B on 4 CPU cores (bf16, ≤448² px, option letter = argmax of next-token logits): **33.8%**
(95% CI 29.3–38.5, n = 400, 1.3 s per question). NVIDIA reports
37.75%, which is inside our CI. Multi-view questions are the weakest category (16.2%).
**Resolution ablation (≤768² px):** 35.5% overall. Only 31.2% of questions contain an image larger than 448², and 19 of 400 answers change. Multi-view rises from 16.2% to 24.3% (4 gained vs 1 lost of 37; exact McNemar p = 0.38). The direction suggests resolution matters for multi-view matching, but the effect is not significant at n = 37, and most of the weakness remains a capability limit of the 2B model.
**Multi-agent swarm** (cosmos_r2_2b, gemini-3.1-flash-lite, gemini-3.5-flash-lite, qwen3.8-27b@cerebras; 283 questions answered by every agent, seeded random order):
| system | accuracy | 95% CI | category-reweighted |
|---|---:|---|---:|
| ⊕ Dawid-Skene (crowd-kit) | **54.8** | 48.9–60.5 | 54.5 |
| Qwen3.8-27B @Cerebras | **54.1** | 48.2–59.8 | 53.7 |
| ⊕ RACE · receiver = Gemini-3.1-flash-lite | **53.0** | 47.2–58.7 | 52.9 |
| ⊕ RACE · receiver = Gemini-3.5-flash-lite | **53.0** | 47.2–58.7 | 52.9 |
| ⊕ RACE · receiver = Qwen3.8-27B @Cerebras | **53.0** | 47.2–58.7 | 52.9 |
| ⊕ RACE sign-only anchor · receiver = Gemini-3.1-flash-lite | **53.0** | 47.2–58.7 | 52.9 |
| ⊕ RACE sign-only anchor · receiver = Gemini-3.5-flash-lite | **53.0** | 47.2–58.7 | 52.9 |
| ⊕ RACE sign-only anchor · receiver = Qwen3.8-27B @Cerebras | **53.0** | 47.2–58.7 | 52.9 |
| ⊕ RACE sign-only anchor · receiver = Cosmos-Reason2-2B (CPU, ours) | **52.3** | 46.5–58.0 | 52.1 |
| Gemini-3.5-flash-lite | **51.9** | 46.1–57.7 | 51.4 |
| ⊕ plurality | **51.2** | 45.4–57.0 | 50.8 |
| Gemini-3.1-flash-lite | **50.2** | 44.4–56.0 | 50.0 |
| ⊕ RACE CGA · receiver = Cosmos-Reason2-2B (CPU, ours) | **43.8** | 38.2–49.6 | 43.3 |
| Cosmos-Reason2-2B (CPU, ours) | **37.1** | 31.7–42.9 | 36.5 |
| ⊕ RACE · receiver = Cosmos-Reason2-2B (CPU, ours) | **37.1** | 31.7–42.9 | 36.5 |
Gemini Robotics-ER 2, the SOTA reference, could only be queried on 46 shared questions before its free-tier daily quota ran out. On that subset it scores 82.6%.
<details><summary><b>Cosmos-Reason2-2B as a success verifier</b> (calibration study on held-out scenes)</summary>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/verifier_calibration.png" width="820"><br>
<sub>The model's P(yes) is biased toward "yes": the naive 0.5 threshold accepts almost everything. A threshold chosen on a calibration split
lifts held-out accuracy from 56% to 83% (AUROC 0.925); <code>scripts/eval_verifier.py</code>.</sub></p>
</details>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/erqa_pareto_cga.png" width="560"></p>
**Robustness vs usefulness.** With the robot's own CPU model as the receiver, the strong RACE anchor is the only rule whose worst
attacked cell never drops below the robot alone (37.1 vs 37.1), but it gains nothing on clean data.
The sign-only anchor and Dawid–Skene gain a lot on clean data (52.3 / 53.0)
and collapse under coherent liars (worst cell 15.2).
**CGA** sits in between: clean 43.8 and worst cell 31.1.
On an earlier 136-question subset its worst cell matched the receiver. On the full 220-question set and on Track A
(39.1 vs 55.5) it does **not**, so we report CGA as a compromise point, not a dominating method.
### PlanBench Blocksworld — the validator is the planner
<table><tr><td width="55%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/planbench.png"></td>
<td width="45%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/planbench_modulo.gif"><br><sub>A real cached LLM plan: the simulator rejects step 11 ("the red block is not clear");
the error message is fed back, and the repaired plan reaches the goal. <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/planbench_modulo.mp4">MP4</a></sub></td></tr></table>
Our exact Blocksworld simulator accepts **all 500 official ground-truth plans** (`tests/test_blocksworld.py`), so it can replace VAL on this domain.
Results on 92 instances that every listed agent answered:
| agent / strategy | one-shot (official prompt) | **LLM-Modulo** (≤3 repairs) | gain | LLM calls / task |
|---|---:|---:|---:|---:|
| gemini-3.1-flash-lite | 69.6 | **95.7** | +26.1 | 1.47 |
| gpt-oss-120b@groq | 91.3 | **100.0** | +8.7 | 1.10 |
| qwen3.8-27b@cerebras | 97.8 | **100.0** | +2.2 | 1.02 |
| qwen3.8-27b@groq | 43.5 | **72.8** | +29.3 | 2.26 |
| swarm cascade-Modulo (cheap → expensive) | – | **100.0** | | 1.02 |
| swarm verify-select (one-shot, 4 agents) | – | **100.0** | | 4.00 |
| swarm plurality (one-shot, 4 agents) | – | **90.2** | | 4.00 |
The same Qwen3.8-27B weights behave very differently on two providers (Groq vs Cerebras). That is a reproducibility hazard worth reporting in its own right.
---
## Part II — RoboRACE-Bench
### Track A — SO-101 camera swarm
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/so101_gallery.png" width="900"></p>
220 physics rollouts (110 unlabelled history, 110 evaluation, balanced). Failures are real physics: missed grasps, drops in transit,
near-miss placements and wrong objects. Each of 5 cameras is an agent (Cosmos-Reason2-2B). Every camera takes a turn as the receiver.
| rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
|---|---|---:|---:|---:|---:|---:|
| **1** (1–1) | **race_net (new, learned)** | **83.4** | 88.2 | 85.1 | 80.0 | 53.6 |
| **2** (2–2) | **race_s_online (new)** | **82.7** | 87.3 | 84.2 | 79.8 | 54.5 |
| **3** (3–3) | **race_pool (new)** | **82.0** | 86.2 | 83.8 | 78.7 | 54.5 |
| **4** (4–4) | **race_online (new)** | **77.7** | 83.6 | 79.7 | 73.9 | 49.1 |
| **5** (5–6) | **race CGA (new)** | **77.1** | 83.6 | 79.7 | 72.1 | 39.1 |
| **6** (5–7) | **race_s (new)** | **76.3** | 85.1 | 78.6 | 71.9 | 49.1 |
| **7** (6–8) | race_online sign-only anchor (ablation) | **76.0** | 83.6 | 79.6 | 69.1 | 16.4 |
| **8** (7–8) | race | **75.0** | 83.3 | 78.0 | 69.3 | 44.5 |
| ref | REF stack_lr (labelled) | **82.6** | 92.7 | 84.9 | 78.1 | 48.2 |
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackA_vs_liars.png" width="980"></p>
<details><summary><b>More Track A figures</b>: per-camera quality, method leaderboard, accuracy by attack</summary>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackA_cameras.png" width="900"></p>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackA_leaderboard.png" width="720"></p>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackA_heatmap.png" width="820"></p>
</details>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/race_s_weights.jpg" width="600"><br><sub>RACE-S signed weights, learned without labels: inverted cameras get <b>negative</b> weight (their reports are reused), noisy or stuck cameras get ≈0.</sub></p>
### Track A-Panda — does it replicate on a different robot?
Every method and hyper-parameter was frozen (and pushed) before this data existed. The same tasks, cameras, failure modes and
physics-based labels were then generated with a **Franka Panda** (7-DOF, MuJoCo Menagerie): 220 new episodes, 1,100 new Cosmos-Reason2 views.
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/panda_rollouts.gif" width="960"><br><sub>Franka Panda rollouts with the camera swarm's verdicts. <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/panda_rollouts.mp4">MP4</a></sub></p>
| rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
|---|---|---:|---:|---:|---:|---:|
| **1** (1–1) | **race_s_online (new)** | **87.0** | 91.8 | 88.7 | 83.6 | 53.6 |
| **2** (2–2) | **race_pool (new)** | **85.9** | 90.5 | 87.7 | 82.4 | 53.6 |
| **3** (3–3) | **race_net (new, learned)** | **84.2** | 86.0 | 85.6 | 81.5 | 54.5 |
| **4** (4–4) | **race_online (new)** | **83.2** | 90.0 | 85.9 | 78.0 | 49.1 |
| **5** (5–6) | **race CGA (new)** | **82.7** | 90.0 | 85.9 | 76.6 | 34.5 |
| **6** (6–8) | race_online sign-only anchor (ablation) | **81.8** | 90.0 | 85.9 | 74.2 | 5.5 |
| **7** (5–7) | race | **81.7** | 94.2 | 86.8 | 71.9 | 5.5 |
| **8** (7–8) | **race_s (new)** | **80.4** | 91.3 | 82.9 | 75.3 | 50.0 |
| ref | REF stack_lr (labelled) | **84.9** | 96.4 | 87.0 | 80.7 | 50.0 |
<table><tr><td width="50%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/cross_embodiment.png"></td><td width="50%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/panda_gallery.png"></td></tr></table>
The best method on the Panda is **race_s_online (new)** (87.0; liar-majority cells 83.6 vs
majority vote 48.8). The left panel plots every method's mean accuracy on the SO-101 against the Panda.
<details><summary><b>Full Track A-Panda bar chart (all methods)</b></summary>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackAP_leaderboard.png" width="760"></p>
</details>
### RACE-Net-V — the learned verification model
RACE and RACE-S run a hand-specified EM on every window. **RACE-Net-V** is a 214,145-parameter axial set transformer that has
*learned* the same receiver-anchored inference: it reads a whole window of unlabelled swarm reports, knows only which column is the receiver,
and outputs P(success) for every episode. It is trained **only on synthetic swarms** (no benchmark data, labels or agent outputs)
and then applied unchanged to the real Cosmos-Reason2 camera swarms. The released checkpoint is the EM-refinement variant: it also receives the RACE-S and RACE estimates of the window and learns to refine them.
**RACE-Pool** is the AIP-style pool of RACE-Net-V, RACE-S and RACE ([design](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/MODELS.md)).
<table><tr><td width="42%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/racenet.png"></td><td width="58%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/racenet.png"></td></tr></table>
| mean accuracy over all attack cells | A0 procedural (development) | SO-101 (test) | Franka Panda (test) |
|---|---:|---:|---:|
| majority vote | 68.5 | 59.8 | 62.2 |
| Dawid–Skene (crowd-kit) | 69.4 | 60.1 | 65.0 |
| RACE online | 90.3 | 77.7 | 83.2 |
| RACE-S online | 91.2 | 82.7 | 87.0 |
| **RACE-Net-V (learned, zero-shot)** | **91.5** | **83.4** | **84.2** |
| **RACE-Pool (RACE-Net-V + RACE-S + RACE)** | **92.4** | **82.0** | **85.9** |
| *labelled stacker (reference)* | 84.8 | 82.6 | 84.9 |
Liar-majority cells: RACE-Net-V 80.0 on SO-101 and 81.5 on the Panda
(RACE-S online: 79.8 / 83.6).
Weights, config and a standalone inference file are in [`models/racenet`](models/racenet) and on the Hub:
[🤗 Nabidnur/RoboRACE-Net](https://huggingface.co/Nabidnur/RoboRACE-Net).
**Checkpoints and selection.** Rule: highest mean accuracy of RACE-Net-V over three development sets (A0 evaluation window, SO-101 history split, Panda history split; equal weight); ties within 0.1 point go to the higher synthetic accuracy. The history splits are the benchmark's unlabelled calibration
data; no label-free method ever used their labels, so they serve as development data for the two arms, and the evaluation
splits stay test sets. Test columns were computed after the choice and are shown for transparency. On the same development
sets RACE-S online scores 84.9 and RACE online 83.8.
| checkpoint | generator · type | steps | params | synthetic (v3) | dev: RACE-Net-V (selects) | dev: RACE-Pool | test SO-101 / Panda: RACE-Net-V | test: RACE-Pool |
|---|---|---:|---:|---:|---:|---:|---:|---:|
| **v3 refine · local · 6k (released)** | v3 · EM-refinement | 6,000 | 214,145 | 97.8 | **84.0** | 85.6 | 83.4 / 84.2 | 82.0 / 85.9 |
| v3 large · Kaggle · 9k | v3 · plain | 9,000 | 589,729 | 97.9 | 83.8 | 85.6 | 83.6 / 84.1 | 82.2 / 85.9 |
| v3 · Kaggle · 15k · seed 1 | v3 · plain | 15,000 | 213,889 | 97.9 | 83.7 | 85.5 | 83.6 / 83.6 | 82.0 / 85.9 |
| v2 · local · 7k | v2 · plain | 7,000 | 213,889 | 96.7 | 83.4 | 85.6 | 82.6 / 83.5 | 81.9 / 86.1 |
| v3 · Kaggle · 15k · seed 0 | v3 · plain | 15,000 | 213,889 | 97.9 | 83.3 | 85.6 | 83.8 / 84.0 | 81.9 / 86.1 |
| v1 · local · 6k | v1 · plain | 6,000 | 213,889 | 97.6 | 82.2 | 85.0 | 81.6 / 81.8 | 81.7 / 85.5 |
| v1 · Kaggle · 15k | v1 · plain | 15,000 | 213,889 | 97.8 | 82.0 | 85.0 | 80.9 / 82.3 | 81.4 / 85.4 |
Checkpoints marked Kaggle were trained in Kaggle CPU notebooks (4 vCPU; logs in `results/kaggle/`). All checkpoints are in `models/racenet/candidates/`.
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/racenet_scaling.png" width="980"></p>
**Scaling study** (`scripts/racenet_analysis.py`; an independent synthetic generator, not the training distribution; 20 seeds per point):
- **Liar majority.** With 3 of 5 agents lying: RACE-Net-V 80.1, RACE-S online 77.6, RACE online 76.8,
majority 45.5. With 15 of 16 lying: 86.7 / 83.4 / 78.7 / 21.9.
- **Short windows.** With only 8 unlabelled episodes: RACE-Net-V 71.7, RACE-S online 72.8, receiver alone 70.0;
no label-free method beats the receiver before about 32 episodes.
- **Large swarms.** At 24 agents (larger than any training swarm): RACE-Net-V 93.8, RACE online 93.4,
RACE-S online 93.2.
- **Weak receiver (open weakness).** If the receiver is only 55% accurate and nobody lies, RACE-Net-V reaches 65.8 and RACE-S online
54.4, while majority reaches 81.8 and discrete RACE 82.4. Score-level anchoring trusts a weak
receiver too much; from a 70% receiver upwards the gap is gone.
- **Against the first release.** Weak receivers: 65.8 at 55% and 87.3 at 70% receiver accuracy (first release 57.1 and 77.4). 24-agent swarms: 93.8 (first release 89.1).
- **RACE-Pool** on the same sweeps: 79.2 with 3 of 5 lying, 94.9 at 24 agents, 68.2 with a 55% receiver.
### Hold-out — a test nothing was tuned on
Both earlier test tracks were looked at while RACE-Net-V was being developed. For a clean final number, **1,700 new camera
views** were generated with new seeds for both arms (60 history + 110 evaluation episodes each). The views were scored by the frozen Cosmos-Reason2
pipeline, and no fusion method was run on them until commit [`5b2bc8e`](https://github.com/Hisernberg/robotics/commit/5b2bc8e89da83f8893f2bf99e962b9312ded18b4) had frozen every model,
rule and threshold (`results/holdout/FREEZE.json`; `run_benchmark.py` refuses to evaluate the hold-out without that record).
<table><tr><td width="50%">
**SO-101 hold-out**
| rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
|---|---|---:|---:|---:|---:|---:|
| **1** (1–1) | **race_s_online (new)** | **82.6** | 86.7 | 84.1 | 79.7 | 53.6 |
| **2** (2–2) | **race_pool (new)** | **81.2** | 84.9 | 82.9 | 78.0 | 52.7 |
| **3** (3–3) | **race_online (new)** | **79.9** | 86.5 | 82.1 | 75.6 | 52.7 |
| **4** (4–5) | **race CGA (new)** | **79.4** | 86.5 | 82.1 | 74.3 | 37.3 |
| **5** (4–6) | **race_net (new, learned)** | **79.0** | 77.1 | 80.0 | 77.2 | 55.5 |
| **6** (5–6) | race_online sign-only anchor (ablation) | **78.4** | 86.5 | 82.2 | 71.3 | 11.8 |
| **7** (7–7) | race | **74.3** | 80.0 | 76.8 | 69.7 | 33.6 |
| ref | REF stack_lr (labelled) | **82.6** | 93.6 | 85.3 | 77.5 | 45.5 |
</td><td width="50%">
**Franka Panda hold-out**
| rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
|---|---|---:|---:|---:|---:|---:|
| **1** (1–1) | **race_pool (new)** | **86.7** | 90.9 | 88.7 | 83.1 | 56.4 |
| **2** (2–3) | **race_online (new)** | **85.9** | 92.7 | 89.0 | 80.1 | 51.8 |
| **3** (2–4) | **race_s_online (new)** | **85.4** | 89.1 | 87.0 | 82.5 | 58.2 |
| **4** (3–5) | **race CGA (new)** | **84.8** | 92.7 | 88.9 | 77.1 | 35.5 |
| **5** (4–6) | **race_net (new, learned)** | **84.3** | 87.5 | 85.7 | 81.6 | 54.5 |
| **6** (5–7) | race_online sign-only anchor (ablation) | **83.4** | 92.7 | 88.9 | 73.0 | 6.4 |
| **7** (6–7) | race | **82.1** | 89.5 | 85.3 | 76.1 | 48.2 |
| ref | REF stack_lr (labelled) | **82.6** | 93.6 | 84.7 | 78.5 | 50.0 |
</td></tr></table>
| hold-out | comparison | difference | 95% CI (cells) | better / worse |
|---|---|---:|---|---:|
| SO-101 | race_pool − race_s_online | -1.4 | -1.8 … -1.0 | 110 / 199 |
| SO-101 | race_net − race_s_online | -3.6 | -4.2 … -3.0 | 65 / 254 |
| SO-101 | race_net − race_online | -0.9 | -1.6 … -0.2 | 148 / 191 |
| Panda | race_pool − race_s_online | +1.3 | +1.0 … +1.5 | 270 / 49 |
| Panda | race_net − race_s_online | -1.2 | -1.5 … -0.8 | 80 / 225 |
| Panda | race_net − race_online | -1.6 | -2.3 … -0.9 | 81 / 252 |
**Reading.** On data nothing was tuned on, the receiver-anchored family keeps its margin over every literature baseline
(majority 62.3 / 64.2, Dawid–Skene (crowd-kit) 62.3 / 68.8, receiver alone
72.7 / 77.1). Among our own methods the EM-based ones are ahead: RACE-S online is rank 1 on SO-101
and RACE-Pool is rank 1 on the Panda. The learned **RACE-Net-V is rank 5 and 5**
(79.0 / 84.3); its first place on the SO-101 test track (83.4) does **not** replicate here. For camera swarms
the recommendation is therefore RACE-S online or RACE-Pool. RACE-Net-V is within 1 to 4 points of them, which is the honest size of the gap between
learning the inference from simulated swarms and fitting it on the window at hand.
### Track B — rendezvous with Byzantine robots
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/swarm_rendezvous.gif" width="960"><br><sub><a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/swarm_rendezvous.mp4">MP4</a> · same robots, same lies, three consensus rules; red dashed = link isolated online by RACE-C.</sub></p>
| rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
|---|---|---:|---:|---:|---:|---:|
| **1** (1–1) | **race_c+wmsr** | **94.3** | 100.0 | 94.1 | – | 0.0 |
| **2** (2–3) | geomedian | **87.1** | 100.0 | 86.7 | – | 0.0 |
| **3** (3–5) | median | **84.3** | 100.0 | 83.8 | – | 0.0 |
| **4** (3–5) | trimmed | **83.7** | 100.0 | 83.2 | – | 0.0 |
| **5** (3–6) | wmsr | **82.1** | 100.0 | 81.5 | – | 12.5 |
| **6** (2–6) | **race_c** | **72.9** | 82.5 | 72.6 | – | 0.0 |
| **7** (7–7) | mean | **3.2** | 100.0 | 0.0 | – | 0.0 |
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackB_vs_liars.png" width="980"></p>
<details><summary><b>More Track B figures</b>: method leaderboard, mission success by attack</summary>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackB_leaderboard.png" width="640"></p>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackB_heatmap.png" width="720"></p>
</details>
### Track C — LLM planners: same models, better plumbing
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackC_systems.png" width="760"></p>
<details><summary><b>Track C swarm leaderboard</b> (real LLM saboteurs, every agent as receiver)</summary>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackC_leaderboard.png" width="720"></p>
<sub>After engineering, every receiver is ~98% accurate on its own, so fusion can only cost accuracy; RACE costs the least. Full table: <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/results/LEADERBOARD.md">leaderboard</a>.</sub>
</details>
<details><summary><b>Track A0</b> — the first iteration (procedural tabletop scenes, kept for continuity)</summary>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/camera_swarm.gif" width="720"><br>
<sub>Procedural MuJoCo scenes judged by the same five-camera Cosmos swarm; superseded by the SO-101 physics rollouts above.
<a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/camera_swarm.mp4">MP4</a> · results in the <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/results/LEADERBOARD.md">leaderboard</a> (Track A0).</sub></p>
</details>
### Track D — maze exploration with lying robots
Six mobile robots must each reach a goal in an unknown maze. Each senses its 8 neighbouring cells, broadcasts what it sensed, and re-plans every
step on its fused belief map. Up to four robots corrupt their broadcasts (false walls, false openings, full inversion, random, a colluding bloc
that walls off the goal, sleepers). Score = SPL (success weighted by path length) of the honest robots.
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_swarm.gif" width="960"><br><sub>Three of six robots invert their reports. Majority map fusion believes the false walls (red);
RACE-Net-M verifies each peer against cells the robot sensed itself and <b>inverts</b> the liars. This is the maze with the largest gap between the
two rules; averages are in the table. <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_swarm.mp4">MP4</a></sub></p>
**RACE-Net-M** is the learned map-fusion model ([design](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/MODELS.md)): 44,837 parameters, trained on simulated maze swarms with *random* two-coin liars.
The benchmark's attack names never appear in its training, and the colluding ring attack is outside that family. It is compared with six rules from the
literature and with our hand-written `race_map` on 20 fresh mazes × 25 conditions.
| rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
|---|---|---:|---:|---:|---:|---:|
| **1** (1–1) | **race_net_m + reach-repair** | **70.2** | 76.8 | 73.1 | 66.8 | 0.0 |
| **2** (2–2) | **race_net_m** | **69.8** | 76.8 | 72.6 | 66.4 | 0.0 |
| **3** (3–3) | **race_map** | **66.2** | 76.8 | 69.7 | 61.8 | 0.0 |
| **4** (4–5) | wmsr_trim | **64.1** | 76.8 | 65.6 | 61.5 | 13.9 |
| **5** (4–6) | reputation | **63.8** | 76.8 | 67.9 | 58.5 | 0.0 |
| **6** (5–7) | eigentrust | **63.0** | 75.0 | 66.8 | 58.2 | 0.0 |
| **7** (6–8) | dawid_skene | **62.0** | 76.3 | 67.4 | 55.5 | 0.0 |
| **8** (7–8) | solo | **61.6** | 61.5 | 62.0 | 61.2 | 13.9 |
| **9** (9–9) | trust_all | **56.7** | 76.8 | 59.9 | 51.7 | 0.0 |
| **10** (10–10) | majority | **55.7** | 76.8 | 61.5 | 48.1 | 0.0 |
| ref | REF oracle | **71.3** | 76.8 | 74.5 | 67.5 | 0.0 |
- **Rank 1.** RACE-Net-M scores 69.8 (70.2 with reachability repair), against 66.2 for the hand-written `race_map` and 64.1 for the best
literature baseline (wmsr_trim). The oracle that knows who lies scores 71.3.
- **Goal reached in time:** RACE-Net-M + repair 97.8%, RACE-Net-M 97.4%, `race_map` 94.2%,
W-MSR trimming 96.5%, beta-reputation 91.9%, majority 82.1% (oracle 98.2%).
- **Wall-inventing liars**, the weak spot of `race_map` (59.5), are handled: RACE-Net-M 71.7.
- **Still open:** the colluding ring (66.2 with repair, oracle 71.0). The liars are honest everywhere
else, so their channel looks reliable until a robot stands next to the goal.
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackD_vs_liars.png" width="980"></p>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_false_walls.gif" width="960"><br><sub>Three of six robots invent walls. Beta-reputation believes dozens of them;
RACE-Net-M learns that these peers' wall reports are worthless while their "free" reports are true, and keeps only the latter.
<a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_false_walls.mp4">MP4</a></sub></p>
Belief quality on held-out mazes (accuracy on cells a peer reported and the receiver has not sensed; 6,111 snapshots):
| model | all | 3 of 5 peers lie | 5 of 5 peers lie |
|---|---:|---:|---:|
| `race_map` (hand-written Bayesian rule) | 90.1 | 82.7 | 72.9 |
| rounds = 1 (no refinement) | 91.8 | 85.7 | 76.1 |
| rounds = 2 | 92.2 | 86.4 | 77.4 |
| rounds = 3 (released) | 92.3 | 86.7 | 77.6 |
| rounds = 3, second seed | 92.3 | 86.7 | 77.6 |
**Other swarm and map sizes.** The model was trained on 6 robots in 12×12-cell mazes. Unchanged, with half of the robots lying
(10 new mazes × 5 attacks per row; computed in a Kaggle CPU notebook):
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/racenet_m_scaling.png" width="960"></p>
| setting | liars | beta-reputation | W-MSR trimming | race_map | RACE-Net-M + repair | oracle |
|---|---:|---:|---:|---:|---:|---:|
| 4 robots · 8×8 | 2 | 74.4 | 72.0 | 73.3 | **73.6** | 76.1 |
| 6 robots · 12×12 (training size) | 3 | 61.1 | 54.0 | 63.6 | **67.7** | 66.9 |
| 10 robots · 12×12 | 5 | 70.7 | 60.5 | 70.6 | **76.7** | 77.5 |
| 8 robots · 16×16 | 4 | 57.5 | 59.4 | 62.5 | **69.5** | 67.7 |
In small mazes (8×8) every rule is within a few points of the others: there is little to gain from peers. The gap grows with the map and the swarm.
**Reachability repair** (planner ablation, same 500 cells). When the fused map says the goal cannot be reached, the standard planner drops every peer;
the repair un-believes only the claimed walls that enclose the goal. It helps every rule against the colluding ring, but it hurts rules whose map is
already polluted with false walls, so it only pays off on top of an accurate fusion model:
| fusion rule | standard fallback | + reach-repair | collude: standard → repair | false_walls: standard → repair |
|---|---:|---:|---:|---:|
| majority | 55.7 | 49.2 | 63.2 → 66.2 | 57.6 → 45.5 |
| eigentrust | 63.0 | 61.0 | 62.6 → 65.0 | 53.8 → 42.2 |
| reputation | 63.8 | 61.8 | 62.7 → 66.2 | 57.0 → 44.8 |
| race_map | 66.2 | 65.6 | 63.2 → 66.2 | 59.5 → 54.3 |
| race_net_m | 69.8 | 70.2 | 63.2 → 66.2 | 71.7 → 71.2 |
<details><summary><b>More Track D figures</b></summary>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackD_leaderboard.png" width="640"></p>
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackD_heatmap.png" width="720"></p>
</details>
### LLM agents in text mazes
The same mazes as ASCII text: the agent answers with a move string and an exact simulator returns the first illegal move (LLM-Modulo, ≤3 repairs).
All 40 shortest paths validate (`tests/test_maze.py`). Free-tier quotas cut some agents short, so `n` differs per agent: compare one-shot with Modulo within a row, not rows with each other.
<table><tr><td width="55%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/maze_llm.png"></td><td width="45%">
| agent | n | one-shot | Modulo | 4×4 | 6×6 | calls |
|---|---:|---:|---:|---:|---:|---:|
| gpt-oss-120b@groq | 24 | 100.0 | **100.0** | 100.0 (n=17) | 100.0 (n=7) | 1.00 |
| qwen3.8-27b@groq | 16 | 93.8 | **100.0** | 100.0 (n=13) | 100.0 (n=3) | 1.06 |
| qwen3.8-27b@cerebras | 40 | 72.5 | **87.5** | 100.0 (n=20) | 75.0 (n=20) | 1.65 |
| gemini-3.1-flash-lite | 26 | 11.5 | **38.5** | 62.5 (n=16) | 0.0 (n=10) | 3.38 |
</td></tr></table>
---
## Part III — how this compares with published numbers
<p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/published_context.png" width="1000"></p>
Grey bars are copied from the cited papers' own tables (ERQA: Gemini Robotics report, Feb 2025, and the NVIDIA Cosmos-Reason2 model card;
PlanBench: Valmeekam et al. 2024, 600 instances). Teal bars are measured in this repo. They are **context, not same-condition comparisons**:
the models are newer, hosted agents were run on quota-limited subsets with a brief-reasoning prompt, and our PlanBench numbers use a seeded
100-instance subset. The full tables with sources and confidence intervals are in the [leaderboard](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/results/LEADERBOARD.md#part-iii--context-published-numbers-next-to-ours).
---
## Validation — how we know the numbers are right
| claim | evidence |
|---|---|
| ERQA pipeline is faithful | reported 37.75% lies inside our 95% CI for Cosmos-Reason2-2B |
| Blocksworld validator ≡ VAL here | accepts 500 / 500 official ground-truth plans; rejects ≥95% of swapped or truncated plans |
| W-MSR implemented correctly | honest states never leave the honest hull (W-MSR safety theorem), 5 attacks × F = 1..3 |
| Dawid–Skene implemented correctly | ≥90% decision agreement with the crowd-kit reference implementation on identical data |
| RACE-S calibration is rank-invariant | a cubic distortion of one agent changes ≤2% of decisions |
| SO-101 labels are physical | labels read from the simulator after settling; replay of a cached seed must reproduce its label |
| no test-set tuning | RACE-S hyper-parameters fixed on synthetic data; CGA / sign-only reported as ablations on *every* benchmark; Track D tuned on separate development mazes; both RACE-Net models trained on simulated swarms only and selected on development data |
| findings replicate on a new robot | all methods frozen before the Franka Panda data was generated (Track A-Panda) |
| RACE-Net-V and RACE-Net-M are set functions | outputs unchanged under any permutation of peers or items (`tests/test_racenet.py`, `tests/test_racenet_m.py`) |
| maze checker is exact | all 40 BFS-shortest paths validate with their optimal length (`tests/test_maze.py`) |
| hold-out was not peeked at | `results/holdout/FREEZE.json` records the commit that froze every model and rule; `run_benchmark.py` skips the hold-out without it |
| RACE-Net-M uses liars instead of discarding them | a full inverter alone recovers the map, a coin-flipper adds no evidence (`tests/test_racenet_m.py`) |
| literature baselines are faithful | EigenTrust drops the inverter, trimming removes F votes per cell, Dawid–Skene is right up to a global flip (`tests/test_racenet_m.py`) |
| differences are not noise | paired bootstrap intervals (over mazes for Track D) and exact sign tests in the [leaderboard](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/results/LEADERBOARD.md) |
| maze data hygiene | five disjoint sets of maze seeds: training, validation, development, benchmark, scaling study ([protocol](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/BENCHMARK.md)) |
| independent replication on Kaggle CPU | a Kaggle notebook (4 vCPU, Intel(R) Xeon(R) CPU @ 2.20GHz) re-ran the tests (60 passed, 4 skipped) and recomputed every track, the hold-out included, from the public Hugging Face copy (revision f4fdc380) in 215 min: 1,864 leaderboard values compared, 1,862 identical, 2 differ by at most 0.00007 (one third-party baseline (crowd-kit GLAD, an iterative optimiser) on the SO-101 hold-out) |
`pytest` runs 64 tests. See [docs/BENCHMARK.md](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/BENCHMARK.md) for the full protocol and threats to validity.
## Reproduce (CPU)
```bash
bash scripts/fetch_assets.sh # SO-101 (MuJoCo Menagerie), PlotNeuralNet, ERQA, PlanBench
```
```bash
python -m venv .venv && .venv/bin/pip install --index-url https://download.pytorch.org/whl/cpu torch torchvision
```
```bash
.venv/bin/pip install -r requirements.txt
```
```bash
MUJOCO_GL=egl .venv/bin/python -m pytest -q
```
```bash
MUJOCO_GL=egl .venv/bin/python scripts/run_benchmark.py && .venv/bin/python scripts/make_figures.py && .venv/bin/python scripts/make_readme.py
```
Videos (need the SO-101 data and `ffmpeg`):
```bash
MUJOCO_GL=egl .venv/bin/python scripts/make_videos.py && .venv/bin/python scripts/make_gifs.py
```
RACE-Net (training about 40 min on 4 cores, then dev-only selection and the scaling study):
```bash
.venv/bin/python scripts/train_racenet.py --steps 6000 --gen v1 --out models/racenet/candidates/mine
```
```bash
.venv/bin/python scripts/select_racenet.py "mine=models/racenet/candidates/mine" "v1 local=models/racenet/candidates/v1_local" && .venv/bin/python scripts/racenet_analysis.py
```
Regenerating agent outputs needs `HF_TOKEN` (access to the gated `nvidia/Cosmos-Reason2-2B`) and LLM keys as environment variables
(`.env.example`). The scripts are `erqa_cosmos.py`, `erqa_api_agents.py`, `planbench_agents.py`, `so101_generate.py` and `so101_score_views.py`.
## Repository map
```
src/robotics/swarm/ race.py (RACE v3.1) · fusion.py (ACL, rank-Gauss, AIP, DS, RACE, RACE-S, CGA) · racenet.py (RACE-Net) · crowd.py · attacks.py
src/robotics/benchmarks/ erqa.py · blocksworld.py (exact simulator + validator) · maze_text.py
src/robotics/bench/ trackA/B/C/D · erqa_swarm · planbench_eval · leaderboard (bootstrap ranks)
src/robotics/sim/ so101_env.py · panda_env.py (IK pick-and-place, physics ground truth) · tabletop.py
models/racenet/ RACE-Net weights (safetensors), config, training log, selection.json, candidates/ (3 checkpoints)
src/robotics/vlm/ cosmos_reason.py (logit P(yes) verifier)
scripts/ run_benchmark · train_racenet · select_racenet · racenet_analysis · make_figures · make_videos · make_gifs · make_architecture · make_plotnn · make_readme ·
publish_hf + make_space + make_playground (Hugging Face dataset, model, Spaces) · agents
results/ LEADERBOARD.md · leaderboard.json · cached agent outputs
kaggle/ the two Kaggle CPU notebooks (RACE-Net training, independent replication)
media/ figures · videos (mp4 + gif) · architecture (Raschka-style + PlotNeuralNet)
```
## Limitations
- Track A renders are simulated. The SO-101 and Panda meshes are real, but lighting and textures are synthetic, and all cameras share one VLM.
- Track D is a grid world with exact local sensing; the 3-D view in the video is a rendering of the same run, not a physics simulation.
- RACE-Net-V is trained on a hand-written simulator of swarms. On the hold-out it is 1 to 4 points behind our EM methods (RACE-S online, RACE-Pool); its first place on the SO-101 test track did not replicate. A weak receiver (below about 60% accuracy) is still an open weakness of every score-level method here.
- RACE-Net-M is evaluated in one simulator (grid mazes, exact local sensing). A bloc that lies only around the goal still costs about 5 SPL points against the oracle.
- Online variants fit on the current unlabelled evaluation window; a deployed robot would use a sliding window.
- ERQA swarm numbers cover the questions that every hosted agent answered within free-tier quotas. Groq Qwen and Gemma-4 were throttled to near-zero coverage and are excluded.
- PlanBench uses a seeded 100-instance subset. Track C saturates after engineering.
- CGA was proposed after inspecting ERQA. It is reported on every benchmark with fixed settings, but it deserves confirmation on a fresh benchmark.
## Provenance & citation
RACE / *Liars Are Information* ([public demo Space](https://huggingface.co/spaces/Nabidnur/liars-are-information-demo); the dataset `Nabidnur/liars-are-information` is private) ·
AIP-AD ([`Nabidnur/AIP-AD`](https://huggingface.co/Nabidnur/AIP-AD)) · ERQA (Gemini Robotics team) · PlanBench (Valmeekam et al., NeurIPS 2023 D&B) ·
LLM-Modulo (Kambhampati et al., ICML 2024) · W-MSR (LeBlanc et al., IEEE JSAC 2013) · EigenTrust (Kamvar et al., WWW 2003) ·
beta-reputation (Jøsang & Ismail, 2002) · SPL (Anderson et al., 2018) · Dawid & Skene (1979) · crowd-kit (Toloka) ·
Cosmos-Reason2 (NVIDIA) · SO-101 model (MuJoCo Menagerie) · PlotNeuralNet (H. Iqbal) · diagram style after S. Raschka.
```bibtex
@misc{roborace2026,
title = {RoboRACE: Byzantine-Robust Multi-Agent Robotics with Receiver-Anchored Channel Estimation},
author = {Nabidnur},
year = {2026},
note = {https://github.com/Hisernberg/robotics}
}
```
<sub>This card is generated by <code>scripts/publish_hf.py</code> from the GitHub README, which is generated by <code>scripts/make_readme.py</code> from <code>results/leaderboard.json</code>.</sub>