Nabidnur commited on
Commit
ddeb297
·
verified ·
1 Parent(s): 0c91091

RoboRACE: code, cached agent outputs, results, media, SO-101 renders

Browse files
Files changed (2) hide show
  1. .gitattributes +0 -6
  2. README.md +219 -61
.gitattributes CHANGED
@@ -6,9 +6,3 @@
6
  *.webp binary
7
  *.mp4 binary
8
  *.pdf binary
9
- media/figures/racenet_scaling.png filter=lfs diff=lfs merge=lfs -text
10
- models/racenet/candidates/v1_kaggle/model.safetensors filter=lfs diff=lfs merge=lfs -text
11
- models/racenet/candidates/v1_local/model.safetensors filter=lfs diff=lfs merge=lfs -text
12
- models/racenet/candidates/v2_local/model.safetensors filter=lfs diff=lfs merge=lfs -text
13
- viewer/erqa_hosted_agents/test.parquet filter=lfs diff=lfs merge=lfs -text
14
- models/racenet_m/model.safetensors filter=lfs diff=lfs merge=lfs -text
 
6
  *.webp binary
7
  *.mp4 binary
8
  *.pdf binary
 
 
 
 
 
 
README.md CHANGED
@@ -40,6 +40,22 @@ configs:
40
  path: viewer/panda_views/evaluation.parquet
41
  - config_name: panda_vlm_scores
42
  data_files: viewer/panda_vlm_scores/train.parquet
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
43
  - config_name: trackD_maze_results
44
  data_files: viewer/trackD_maze_results/train.parquet
45
  - config_name: maze_llm_agents
@@ -113,6 +129,8 @@ The GIF is a 12-second highlight; the full 43-second film (3 episodes) is the <a
113
  | `data/so101/episodes.jsonl` | per-episode metadata: task, family (put / stack), scripted mode, **physics ground truth**, final object pose |
114
  | `results/so101/vlm_scores.jsonl` | Cosmos-Reason2-2B P(success) for every (episode, camera) — config **`so101_vlm_scores`** |
115
  | `data/panda/...`, `results/panda/vlm_scores.jsonl` | the same for the **Franka Panda** replication (220 episodes × 5 cameras) — configs **`panda_views`**, **`panda_vlm_scores`** |
 
 
116
  | `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`** |
117
  | `results/maze_llm/<agent>.jsonl` | LLM agents on text mazes: every attempt, validator feedback, repair count — config **`maze_llm_agents`** |
118
  | `models/racenet/` | RACE-Net weights (safetensors), config and training log; also at [Nabidnur/RoboRACE-Net](https://huggingface.co/Nabidnur/RoboRACE-Net) |
@@ -149,11 +167,12 @@ pip install -r requirements.txt && python scripts/run_benchmark.py --reuse-track
149
  | **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 |
150
  | **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 |
151
  | **PlanBench** Blocksworld | plan generation, 92 instances | official one-shot prompt | LLM-Modulo + swarm cascade: see [table](#planbench-blocksworld--the-validator-is-the-planner) |
152
- | **Track A** SO-101 camera swarm | success verification under 7 attacks | receiver alone 71.3 · majority 59.8 | **82.7** race_s_online (new) (rank 1) |
153
  | **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) |
154
- | **RACE-Net** (our model, 0.2 M params) | learned fusion, zero-shot on real camera swarms | RACE-S online 82.7 / 87.0 | **81.6** SO-101 / **81.8** Panda |
 
155
  | **Track B** 12-robot rendezvous | mission success, up to 5 Byzantine robots | W-MSR 82.1 | **94.3** RACE-C + W-MSR (rank 1) |
156
- | **Track D** maze swarm (6 mobile robots) | SPL under 6 attacks | beta-reputation 63.8 · majority 55.7 | **66.2** RACE map (rank 1) |
157
  | **Track C** LLM planners | exact-plan accuracy | zero-shot mean 88.1 | engineered mean **98.2** · flash-lite 69.0 → **92.9** |
158
 
159
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/overview.png" width="920"></p>
@@ -162,7 +181,7 @@ pip install -r requirements.txt && python scripts/run_benchmark.py --reuse-track
162
  [SO-101 camera swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/robot_rollouts.mp4) ·
163
  [Byzantine rendezvous](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/swarm_rendezvous.mp4) ·
164
  [PlanBench LLM-Modulo replay](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/planbench_modulo.mp4) ·
165
- [maze swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_swarm.mp4) ·
166
  [Franka Panda camera swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/panda_rollouts.mp4)
167
 
168
  ## The idea in one paragraph
@@ -186,6 +205,24 @@ clean-data gains), **RACE-C + W-MSR** (continuous robot consensus) and an **engi
186
  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).
187
  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>
188
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
189
  ---
190
 
191
  ## Part I — public benchmarks
@@ -271,14 +308,14 @@ near-miss placements and wrong objects. Each of 5 cameras is an agent (Cosmos-Re
271
 
272
  | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
273
  |---|---|---:|---:|---:|---:|---:|
274
- | **1** (1–1) | **race_s_online (new)** | **82.7** | 87.3 | 84.2 | 79.8 | 54.5 |
275
- | **2** (2–2) | **race_net (new, learned)** | **81.6** | 83.6 | 82.6 | 79.5 | 51.8 |
276
- | **3** (3–3) | **race_online (new)** | **77.7** | 83.6 | 79.7 | 73.9 | 49.1 |
277
- | **4** (4–5) | **race CGA (new)** | **77.1** | 83.6 | 79.7 | 72.1 | 39.1 |
278
- | **5** (4–6) | **race_s (new)** | **76.3** | 85.1 | 78.6 | 71.9 | 49.1 |
279
- | **6** (5–7) | race_online sign-only anchor (ablation) | **76.0** | 83.6 | 79.6 | 69.1 | 16.4 |
280
- | **7** (6–7) | race | **75.0** | 83.3 | 78.0 | 69.3 | 44.5 |
281
- | **8** (8–9) | receiver_only | **71.3** | 71.3 | 71.3 | 71.3 | 55.5 |
282
  | ref | REF stack_lr (labelled) | **82.6** | 92.7 | 84.9 | 78.1 | 48.2 |
283
 
284
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackA_vs_liars.png" width="980"></p>
@@ -300,13 +337,13 @@ physics-based labels were then generated with a **Franka Panda** (7-DOF, MuJoCo
300
  | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
301
  |---|---|---:|---:|---:|---:|---:|
302
  | **1** (1–1) | **race_s_online (new)** | **87.0** | 91.8 | 88.7 | 83.6 | 53.6 |
303
- | **2** (2–2) | **race_online (new)** | **83.2** | 90.0 | 85.9 | 78.0 | 49.1 |
304
- | **3** (3–4) | **race CGA (new)** | **82.7** | 90.0 | 85.9 | 76.6 | 34.5 |
305
- | **5** (4–7) | race_online sign-only anchor (ablation) | **81.8** | 90.0 | 85.9 | 74.2 | 5.5 |
306
- | **5** (3–6) | **race_net (new, learned)** | **81.8** | 81.8 | 82.6 | 80.5 | 55.5 |
307
- | **5** (3���6) | race | **81.7** | 94.2 | 86.8 | 71.9 | 5.5 |
308
- | **7** (6–7) | **race_s (new)** | **80.4** | 91.3 | 82.9 | 75.3 | 50.0 |
309
- | **8** (8–8) | receiver_only | **74.5** | 74.5 | 74.5 | 74.5 | 56.4 |
310
  | ref | REF stack_lr (labelled) | **84.9** | 96.4 | 87.0 | 80.7 | 50.0 |
311
 
312
  <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>
@@ -318,56 +355,119 @@ majority vote 48.8). The left panel plots every method's mean accuracy on the SO
318
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackAP_leaderboard.png" width="760"></p>
319
  </details>
320
 
321
- ### RACE-Net — our own fusion model
322
 
323
- RACE and RACE-S run a hand-specified EM on every window. **RACE-Net** is a 213,889-parameter axial set transformer that has
324
  *learned* the same receiver-anchored inference: it reads a whole window of unlabelled swarm reports, knows only which column is the receiver,
325
- and outputs P(success) per episode in one forward pass. It is trained **only on synthetic swarms** (no benchmark data, labels or agent outputs)
326
- and then applied unchanged to the real Cosmos-Reason2 camera swarms.
 
327
 
328
  <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>
329
 
330
- | mean accuracy over all attack cells | A0 procedural (development) | SO-101 (test) | Franka Panda (fresh test) |
331
  |---|---:|---:|---:|
332
  | majority vote | 68.5 | 59.8 | 62.2 |
333
  | Dawid–Skene (crowd-kit) | 69.4 | 60.1 | 65.0 |
334
  | RACE online | 90.3 | 77.7 | 83.2 |
335
  | RACE-S online | 91.2 | 82.7 | 87.0 |
336
- | **RACE-Net (learned, zero-shot)** | **90.4** | **81.6** | **81.8** |
 
337
  | *labelled stacker (reference)* | 84.8 | 82.6 | 84.9 |
338
 
339
- Liar-majority cells: RACE-Net 79.5 on SO-101 and 80.5 on the Panda
340
  (RACE-S online: 79.8 / 83.6).
341
  Weights, config and a standalone inference file are in [`models/racenet`](models/racenet) and on the Hub:
342
  [🤗 Nabidnur/RoboRACE-Net](https://huggingface.co/Nabidnur/RoboRACE-Net).
343
 
344
- **Checkpoints and selection.** Rule, fixed before the candidates were compared: highest mean accuracy on Track A0 (development); ties within 0.1 point go to the higher synthetic accuracy (v2 generator). Test-track columns were computed after
345
- the choice and are shown for transparency.
346
-
347
- | checkpoint | generator | steps | synthetic acc. (v1 / v2) | A0 dev (selects) | SO-101 | Panda | Panda, no attack |
348
- |---|---|---:|---:|---:|---:|---:|---:|
349
- | **v1 · local CPU · 6k steps (released)** | v1 | 6,000 | 93.4 / 96.8 | **90.4** | 81.6 | 81.8 | 81.8 |
350
- | v1 · Kaggle CPU · 15k steps | v1 | 15,000 | 93.9 / 97.2 | **90.2** | 80.9 | 82.3 | 84.5 |
351
- | v2 · local CPU · 7k steps | v2 | 7,000 | 92.6 / 96.9 | **90.2** | 82.6 | 83.5 | 88.5 |
352
 
353
- Generator v2 widens agent skill (receivers down to near chance, near-perfect peers) and trains on swarms of up to 12 + 3 agents. It was designed after the v1 results on the test tracks and the scaling study were known, so for v2 checkpoints the SO-101 and Panda columns are **not** untouched test results; the v1 rows are.
 
 
 
 
 
 
 
 
354
 
355
- The 15k-step run was trained in a Kaggle CPU notebook (4 vCPU, AMD EPYC 7B12; 130 min); its log is in `results/kaggle/racenet_training.json`. All three checkpoints are in `models/racenet/candidates/`.
356
 
357
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/racenet_scaling.png" width="980"></p>
358
 
359
  **Scaling study** (`scripts/racenet_analysis.py`; an independent synthetic generator, not the training distribution; 20 seeds per point):
360
 
361
- - **Liar majority.** With 3 of 5 agents lying: RACE-Net 79.6, RACE-S online 77.6, RACE online 76.8,
362
- majority 45.5. With 15 of 16 lying: 83.2 / 83.4 / 78.7 / 21.9.
363
- - **Short windows.** With only 8 unlabelled episodes: RACE-Net 67.5, RACE-S online 72.8, receiver alone 70.0;
364
  no label-free method beats the receiver before about 32 episodes.
365
- - **Large swarms.** At 24 agents (larger than any training swarm): RACE-Net 89.1, RACE online 93.4,
366
  RACE-S online 93.2.
367
- - **Weak receiver (open weakness).** If the receiver is only 55% accurate and nobody lies, RACE-Net (57.1) and RACE-S online
368
- (54.4) stay near the receiver's own accuracy, while majority reaches 81.8 and discrete RACE 82.4.
369
- Score-level anchoring trusts a weak receiver too much.
370
- - **v2 checkpoint (post hoc).** Training on a wider skill range helps with weak receivers (59.8 at 55% and 82.5 at 70% receiver accuracy, against 57.1 and 77.4 for the released model) but costs accuracy on large swarms (80.8 against 89.1 at 24 agents). It is not uniformly better, which agrees with the development track not selecting it.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
371
 
372
  ### Track B — rendezvous with Byzantine robots
373
 
@@ -412,27 +512,76 @@ step on its fused belief map. Up to four robots corrupt their broadcasts (false
412
  that walls off the goal, sleepers). Score = SPL (success weighted by path length) of the honest robots.
413
 
414
  <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);
415
- RACE map fusion 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
416
  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>
417
 
 
 
 
 
418
  | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
419
  |---|---|---:|---:|---:|---:|---:|
420
- | **1** (1–1) | **race_map** | **66.2** | 76.8 | 69.7 | 61.8 | 0.0 |
421
- | **2** (2–2) | reputation | **63.8** | 76.8 | 67.9 | 58.5 | 0.0 |
422
- | **3** (3–3) | solo | **61.6** | 61.5 | 62.0 | 61.2 | 13.9 |
423
- | **4** (4–4) | trust_all | **56.7** | 76.8 | 59.9 | 51.7 | 0.0 |
424
- | **5** (5–5) | majority | **55.7** | 76.8 | 61.5 | 48.1 | 0.0 |
 
 
 
 
 
425
  | ref | REF oracle | **71.3** | 76.8 | 74.5 | 67.5 | 0.0 |
426
 
427
- Honest robots reaching the goal in time: RACE map 94.2%, beta-reputation 91.9%,
428
- no communication 95.2%, naive merging 83.8%, majority 82.1%
429
- (oracle that knows the liars: 98.2%).
 
 
 
 
430
 
431
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackD_vs_liars.png" width="980"></p>
432
 
433
- `race_map` infers each peer's two-coin channel (how reliable are its "wall" claims, how reliable its "free" claims) by Bayesian model
434
- averaging on cells the receiver sensed itself, and adds each claim as a signed log-likelihood ratio. Beta-reputation (Jøsang & Ismail, 2002)
435
- uses the same verified evidence but can only *discard* a liar. Weak spot: many wall-inventing liars (0.0 worst cell).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
436
 
437
  <details><summary><b>More Track D figures</b></summary>
438
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackD_leaderboard.png" width="640"></p>
@@ -478,13 +627,19 @@ the models are newer, hosted agents were run on quota-limited subsets with a bri
478
  | Dawid–Skene implemented correctly | ≥90% decision agreement with the crowd-kit reference implementation on identical data |
479
  | RACE-S calibration is rank-invariant | a cubic distortion of one agent changes ≤2% of decisions |
480
  | SO-101 labels are physical | labels read from the simulator after settling; replay of a cached seed must reproduce its label |
481
- | 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; RACE-Net trained on synthetic swarms only |
482
  | findings replicate on a new robot | all methods frozen before the Franka Panda data was generated (Track A-Panda) |
483
- | RACE-Net is a set function | output unchanged (1e-6) under any permutation of peers or episodes (`tests/test_racenet.py`) |
484
  | maze checker is exact | all 40 BFS-shortest paths validate with their optimal length (`tests/test_maze.py`) |
 
 
 
 
 
 
485
 
486
 
487
- `pytest` runs 57 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.
488
 
489
  ## Reproduce (CPU)
490
 
@@ -530,13 +685,15 @@ src/robotics/vlm/ cosmos_reason.py (logit P(yes) verifier)
530
  scripts/ run_benchmark · train_racenet · select_racenet · racenet_analysis · make_figures · make_videos · make_gifs · make_architecture · make_plotnn · make_readme ·
531
  publish_hf + make_space + make_playground (Hugging Face dataset, model, Spaces) · agents
532
  results/ LEADERBOARD.md · leaderboard.json · cached agent outputs
 
533
  media/ figures · videos (mp4 + gif) · architecture (Raschka-style + PlotNeuralNet)
534
  ```
535
 
536
  ## Limitations
537
  - 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.
538
  - 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.
539
- - RACE-Net is trained on a hand-written simulator of swarms; where real agent errors differ from it (see the comparison table), the EM-based RACE-S can be better.
 
540
  - Online variants fit on the current unlabelled evaluation window; a deployed robot would use a sliding window.
541
  - 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.
542
  - PlanBench uses a seeded 100-instance subset. Track C saturates after engineering.
@@ -545,7 +702,8 @@ media/ figures · videos (mp4 + gif) · architecture (Raschka
545
  ## Provenance & citation
546
  RACE / *Liars Are Information* ([public demo Space](https://huggingface.co/spaces/Nabidnur/liars-are-information-demo); the dataset `Nabidnur/liars-are-information` is private) ·
547
  AIP-AD ([`Nabidnur/AIP-AD`](https://huggingface.co/Nabidnur/AIP-AD)) · ERQA (Gemini Robotics team) · PlanBench (Valmeekam et al., NeurIPS 2023 D&B) ·
548
- LLM-Modulo (Kambhampati et al., ICML 2024) · W-MSR (LeBlanc et al., IEEE JSAC 2013) · Dawid & Skene (1979) · crowd-kit (Toloka) ·
 
549
  Cosmos-Reason2 (NVIDIA) · SO-101 model (MuJoCo Menagerie) · PlotNeuralNet (H. Iqbal) · diagram style after S. Raschka.
550
 
551
  ```bibtex
 
40
  path: viewer/panda_views/evaluation.parquet
41
  - config_name: panda_vlm_scores
42
  data_files: viewer/panda_vlm_scores/train.parquet
43
+ - config_name: so101_holdout_views
44
+ data_files:
45
+ - split: history
46
+ path: viewer/so101_holdout_views/history.parquet
47
+ - split: evaluation
48
+ path: viewer/so101_holdout_views/evaluation.parquet
49
+ - config_name: so101_holdout_vlm_scores
50
+ data_files: viewer/so101_holdout_vlm_scores/train.parquet
51
+ - config_name: panda_holdout_views
52
+ data_files:
53
+ - split: history
54
+ path: viewer/panda_holdout_views/history.parquet
55
+ - split: evaluation
56
+ path: viewer/panda_holdout_views/evaluation.parquet
57
+ - config_name: panda_holdout_vlm_scores
58
+ data_files: viewer/panda_holdout_vlm_scores/train.parquet
59
  - config_name: trackD_maze_results
60
  data_files: viewer/trackD_maze_results/train.parquet
61
  - config_name: maze_llm_agents
 
129
  | `data/so101/episodes.jsonl` | per-episode metadata: task, family (put / stack), scripted mode, **physics ground truth**, final object pose |
130
  | `results/so101/vlm_scores.jsonl` | Cosmos-Reason2-2B P(success) for every (episode, camera) — config **`so101_vlm_scores`** |
131
  | `data/panda/...`, `results/panda/vlm_scores.jsonl` | the same for the **Franka Panda** replication (220 episodes × 5 cameras) — configs **`panda_views`**, **`panda_vlm_scores`** |
132
+ | `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` |
133
+ | `models/racenet`, `models/racenet_m` | weights of both models (**RACE-Net-V**, **RACE-Net-M**) and every candidate checkpoint |
134
  | `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`** |
135
  | `results/maze_llm/<agent>.jsonl` | LLM agents on text mazes: every attempt, validator feedback, repair count — config **`maze_llm_agents`** |
136
  | `models/racenet/` | RACE-Net weights (safetensors), config and training log; also at [Nabidnur/RoboRACE-Net](https://huggingface.co/Nabidnur/RoboRACE-Net) |
 
167
  | **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 |
168
  | **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 |
169
  | **PlanBench** Blocksworld | plan generation, 92 instances | official one-shot prompt | LLM-Modulo + swarm cascade: see [table](#planbench-blocksworld--the-validator-is-the-planner) |
170
+ | **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) |
171
  | **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) |
172
+ | **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** |
173
+ | **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 |
174
  | **Track B** 12-robot rendezvous | mission success, up to 5 Byzantine robots | W-MSR 82.1 | **94.3** RACE-C + W-MSR (rank 1) |
175
+ | **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 |
176
  | **Track C** LLM planners | exact-plan accuracy | zero-shot mean 88.1 | engineered mean **98.2** · flash-lite 69.0 → **92.9** |
177
 
178
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/overview.png" width="920"></p>
 
181
  [SO-101 camera swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/robot_rollouts.mp4) ·
182
  [Byzantine rendezvous](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/swarm_rendezvous.mp4) ·
183
  [PlanBench LLM-Modulo replay](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/planbench_modulo.mp4) ·
184
+ [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) ·
185
  [Franka Panda camera swarm](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/panda_rollouts.mp4)
186
 
187
  ## The idea in one paragraph
 
205
  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).
206
  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>
207
 
208
+ ### Our two models: one design
209
+
210
+ <table><tr>
211
+ <td width="50%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/racenet.png"></td>
212
+ <td width="50%"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/architecture/racenet_m.png"></td>
213
+ </tr></table>
214
+
215
+ | | **RACE-Net-V** · pick-and-place | **RACE-Net-M** · maze |
216
+ |---|---|---|
217
+ | question | did the arm succeed? (camera swarm) | which cells are walls? (robot swarm) |
218
+ | receiver's anchor | its own score for every episode | its own exact sensing of some cells |
219
+ | structure | axial set transformer over agents × episodes | anchored channel encoder → peer transformer → signed read-out, 3 refinement rounds |
220
+ | shared design | anchor on the receiver · estimate each peer's channel in context · **invert** consistent liars · sets, not sequences · synthetic training only |
221
+ | size · training | 213,889 parameters · synthetic swarms | 44,837 parameters · simulated mazes with random liars |
222
+
223
+ Both come from RACE and AIP in the *Liars Are Information* program: RACE supplies the anchored, signed channel estimation; AIP supplies the pooling of
224
+ experts (RACE-Pool). The full description is in [docs/MODELS.md](https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/blob/main/docs/MODELS.md).
225
+
226
  ---
227
 
228
  ## Part I — public benchmarks
 
308
 
309
  | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
310
  |---|---|---:|---:|---:|---:|---:|
311
+ | **1** (1–1) | **race_net (new, learned)** | **83.4** | 88.2 | 85.1 | 80.0 | 53.6 |
312
+ | **2** (2–2) | **race_s_online (new)** | **82.7** | 87.3 | 84.2 | 79.8 | 54.5 |
313
+ | **3** (3–3) | **race_pool (new)** | **82.0** | 86.2 | 83.8 | 78.7 | 54.5 |
314
+ | **4** (4–4) | **race_online (new)** | **77.7** | 83.6 | 79.7 | 73.9 | 49.1 |
315
+ | **5** (5–6) | **race CGA (new)** | **77.1** | 83.6 | 79.7 | 72.1 | 39.1 |
316
+ | **6** (5–7) | **race_s (new)** | **76.3** | 85.1 | 78.6 | 71.9 | 49.1 |
317
+ | **7** (6–8) | race_online sign-only anchor (ablation) | **76.0** | 83.6 | 79.6 | 69.1 | 16.4 |
318
+ | **8** (7–8) | race | **75.0** | 83.3 | 78.0 | 69.3 | 44.5 |
319
  | ref | REF stack_lr (labelled) | **82.6** | 92.7 | 84.9 | 78.1 | 48.2 |
320
 
321
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackA_vs_liars.png" width="980"></p>
 
337
  | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
338
  |---|---|---:|---:|---:|---:|---:|
339
  | **1** (1–1) | **race_s_online (new)** | **87.0** | 91.8 | 88.7 | 83.6 | 53.6 |
340
+ | **2** (2–2) | **race_pool (new)** | **85.9** | 90.5 | 87.7 | 82.4 | 53.6 |
341
+ | **3** (3–3) | **race_net (new, learned)** | **84.2** | 86.0 | 85.6 | 81.5 | 54.5 |
342
+ | **4** (4–4) | **race_online (new)** | **83.2** | 90.0 | 85.9 | 78.0 | 49.1 |
343
+ | **5** (5–6) | **race CGA (new)** | **82.7** | 90.0 | 85.9 | 76.6 | 34.5 |
344
+ | **6** (6–8) | race_online sign-only anchor (ablation) | **81.8** | 90.0 | 85.9 | 74.2 | 5.5 |
345
+ | **7** (5–7) | race | **81.7** | 94.2 | 86.8 | 71.9 | 5.5 |
346
+ | **8** (7–8) | **race_s (new)** | **80.4** | 91.3 | 82.9 | 75.3 | 50.0 |
347
  | ref | REF stack_lr (labelled) | **84.9** | 96.4 | 87.0 | 80.7 | 50.0 |
348
 
349
  <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>
 
355
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackAP_leaderboard.png" width="760"></p>
356
  </details>
357
 
358
+ ### RACE-Net-V — the learned verification model
359
 
360
+ 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
361
  *learned* the same receiver-anchored inference: it reads a whole window of unlabelled swarm reports, knows only which column is the receiver,
362
+ and outputs P(success) for every episode. It is trained **only on synthetic swarms** (no benchmark data, labels or agent outputs)
363
+ 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.
364
+ **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)).
365
 
366
  <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>
367
 
368
+ | mean accuracy over all attack cells | A0 procedural (development) | SO-101 (test) | Franka Panda (test) |
369
  |---|---:|---:|---:|
370
  | majority vote | 68.5 | 59.8 | 62.2 |
371
  | Dawid–Skene (crowd-kit) | 69.4 | 60.1 | 65.0 |
372
  | RACE online | 90.3 | 77.7 | 83.2 |
373
  | RACE-S online | 91.2 | 82.7 | 87.0 |
374
+ | **RACE-Net-V (learned, zero-shot)** | **91.5** | **83.4** | **84.2** |
375
+ | **RACE-Pool (RACE-Net-V + RACE-S + RACE)** | **92.4** | **82.0** | **85.9** |
376
  | *labelled stacker (reference)* | 84.8 | 82.6 | 84.9 |
377
 
378
+ Liar-majority cells: RACE-Net-V 80.0 on SO-101 and 81.5 on the Panda
379
  (RACE-S online: 79.8 / 83.6).
380
  Weights, config and a standalone inference file are in [`models/racenet`](models/racenet) and on the Hub:
381
  [🤗 Nabidnur/RoboRACE-Net](https://huggingface.co/Nabidnur/RoboRACE-Net).
382
 
383
+ **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
384
+ data; no label-free method ever used their labels, so they serve as development data for the two arms, and the evaluation
385
+ splits stay test sets. Test columns were computed after the choice and are shown for transparency. On the same development
386
+ sets RACE-S online scores 84.9 and RACE online 83.8.
 
 
 
 
387
 
388
+ | 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 |
389
+ |---|---|---:|---:|---:|---:|---:|---:|---:|
390
+ | **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 |
391
+ | v3 large · Kaggle · 9k | v3 · plain | 9,000 | 589,729 | 97.9 | 83.8 | 85.6 | 83.6 / 84.1 | 82.2 / 85.9 |
392
+ | 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 |
393
+ | v2 · local · 7k | v2 · plain | 7,000 | 213,889 | 96.7 | 83.4 | 85.6 | 82.6 / 83.5 | 81.9 / 86.1 |
394
+ | 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 |
395
+ | v1 · local · 6k | v1 · plain | 6,000 | 213,889 | 97.6 | 82.2 | 85.0 | 81.6 / 81.8 | 81.7 / 85.5 |
396
+ | v1 · Kaggle · 15k | v1 · plain | 15,000 | 213,889 | 97.8 | 82.0 | 85.0 | 80.9 / 82.3 | 81.4 / 85.4 |
397
 
398
+ Checkpoints marked Kaggle were trained in Kaggle CPU notebooks (4 vCPU; logs in `results/kaggle/`). All checkpoints are in `models/racenet/candidates/`.
399
 
400
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/racenet_scaling.png" width="980"></p>
401
 
402
  **Scaling study** (`scripts/racenet_analysis.py`; an independent synthetic generator, not the training distribution; 20 seeds per point):
403
 
404
+ - **Liar majority.** With 3 of 5 agents lying: RACE-Net-V 80.1, RACE-S online 77.6, RACE online 76.8,
405
+ majority 45.5. With 15 of 16 lying: 86.7 / 83.4 / 78.7 / 21.9.
406
+ - **Short windows.** With only 8 unlabelled episodes: RACE-Net-V 71.7, RACE-S online 72.8, receiver alone 70.0;
407
  no label-free method beats the receiver before about 32 episodes.
408
+ - **Large swarms.** At 24 agents (larger than any training swarm): RACE-Net-V 93.8, RACE online 93.4,
409
  RACE-S online 93.2.
410
+ - **Weak receiver (open weakness).** If the receiver is only 55% accurate and nobody lies, RACE-Net-V reaches 65.8 and RACE-S online
411
+ 54.4, while majority reaches 81.8 and discrete RACE 82.4. Score-level anchoring trusts a weak
412
+ receiver too much; from a 70% receiver upwards the gap is gone.
413
+ - **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).
414
+ - **RACE-Pool** on the same sweeps: 79.2 with 3 of 5 lying, 94.9 at 24 agents, 68.2 with a 55% receiver.
415
+
416
+ ### Hold-out — a test nothing was tuned on
417
+
418
+ Both earlier test tracks were looked at while RACE-Net-V was being developed. For a clean final number, **1,700 new camera
419
+ views** were generated with new seeds for both arms (60 history + 110 evaluation episodes each). The views were scored by the frozen Cosmos-Reason2
420
+ pipeline, and no fusion method was run on them until commit [`5b2bc8e`](https://github.com/Hisernberg/robotics/commit/5b2bc8e89da83f8893f2bf99e962b9312ded18b4) had frozen every model,
421
+ rule and threshold (`results/holdout/FREEZE.json`; `run_benchmark.py` refuses to evaluate the hold-out without that record).
422
+
423
+ <table><tr><td width="50%">
424
+
425
+ **SO-101 hold-out**
426
+
427
+ | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
428
+ |---|---|---:|---:|---:|---:|---:|
429
+ | **1** (1–1) | **race_s_online (new)** | **82.6** | 86.7 | 84.1 | 79.7 | 53.6 |
430
+ | **2** (2–2) | **race_pool (new)** | **81.2** | 84.9 | 82.9 | 78.0 | 52.7 |
431
+ | **3** (3–3) | **race_online (new)** | **79.9** | 86.5 | 82.1 | 75.6 | 52.7 |
432
+ | **4** (4–5) | **race CGA (new)** | **79.4** | 86.5 | 82.1 | 74.3 | 37.3 |
433
+ | **5** (4–6) | **race_net (new, learned)** | **79.0** | 77.1 | 80.0 | 77.2 | 55.5 |
434
+ | **6** (5–6) | race_online sign-only anchor (ablation) | **78.4** | 86.5 | 82.2 | 71.3 | 11.8 |
435
+ | **7** (7–7) | race | **74.3** | 80.0 | 76.8 | 69.7 | 33.6 |
436
+ | ref | REF stack_lr (labelled) | **82.6** | 93.6 | 85.3 | 77.5 | 45.5 |
437
+
438
+ </td><td width="50%">
439
+
440
+ **Franka Panda hold-out**
441
+
442
+ | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
443
+ |---|---|---:|---:|---:|---:|---:|
444
+ | **1** (1–1) | **race_pool (new)** | **86.7** | 90.9 | 88.7 | 83.1 | 56.4 |
445
+ | **2** (2–3) | **race_online (new)** | **85.9** | 92.7 | 89.0 | 80.1 | 51.8 |
446
+ | **3** (2–4) | **race_s_online (new)** | **85.4** | 89.1 | 87.0 | 82.5 | 58.2 |
447
+ | **4** (3–5) | **race CGA (new)** | **84.8** | 92.7 | 88.9 | 77.1 | 35.5 |
448
+ | **5** (4–6) | **race_net (new, learned)** | **84.3** | 87.5 | 85.7 | 81.6 | 54.5 |
449
+ | **6** (5–7) | race_online sign-only anchor (ablation) | **83.4** | 92.7 | 88.9 | 73.0 | 6.4 |
450
+ | **7** (6–7) | race | **82.1** | 89.5 | 85.3 | 76.1 | 48.2 |
451
+ | ref | REF stack_lr (labelled) | **82.6** | 93.6 | 84.7 | 78.5 | 50.0 |
452
+
453
+ </td></tr></table>
454
+
455
+ | hold-out | comparison | difference | 95% CI (cells) | better / worse |
456
+ |---|---|---:|---|---:|
457
+ | SO-101 | race_pool − race_s_online | -1.4 | -1.8 … -1.0 | 110 / 199 |
458
+ | SO-101 | race_net − race_s_online | -3.6 | -4.2 … -3.0 | 65 / 254 |
459
+ | SO-101 | race_net − race_online | -0.9 | -1.6 … -0.2 | 148 / 191 |
460
+ | Panda | race_pool − race_s_online | +1.3 | +1.0 … +1.5 | 270 / 49 |
461
+ | Panda | race_net − race_s_online | -1.2 | -1.5 … -0.8 | 80 / 225 |
462
+ | Panda | race_net − race_online | -1.6 | -2.3 … -0.9 | 81 / 252 |
463
+
464
+ **Reading.** On data nothing was tuned on, the receiver-anchored family keeps its margin over every literature baseline
465
+ (majority 62.3 / 64.2, Dawid–Skene (crowd-kit) 62.3 / 68.8, receiver alone
466
+ 72.7 / 77.1). Among our own methods the EM-based ones are ahead: RACE-S online is rank 1 on SO-101
467
+ and RACE-Pool is rank 1 on the Panda. The learned **RACE-Net-V is rank 5 and 5**
468
+ (79.0 / 84.3); its first place on the SO-101 test track (83.4) does **not** replicate here. For camera swarms
469
+ 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
470
+ learning the inference from simulated swarms and fitting it on the window at hand.
471
 
472
  ### Track B — rendezvous with Byzantine robots
473
 
 
512
  that walls off the goal, sleepers). Score = SPL (success weighted by path length) of the honest robots.
513
 
514
  <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);
515
+ 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
516
  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>
517
 
518
+ **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.
519
+ 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
520
+ literature and with our hand-written `race_map` on 20 fresh mazes × 25 conditions.
521
+
522
  | rank (90% CI) | method | mean | clean | liars < 50% | liars ≥ 50% | worst cell |
523
  |---|---|---:|---:|---:|---:|---:|
524
+ | **1** (1–1) | **race_net_m + reach-repair** | **70.2** | 76.8 | 73.1 | 66.8 | 0.0 |
525
+ | **2** (2–2) | **race_net_m** | **69.8** | 76.8 | 72.6 | 66.4 | 0.0 |
526
+ | **3** (3–3) | **race_map** | **66.2** | 76.8 | 69.7 | 61.8 | 0.0 |
527
+ | **4** (4–5) | wmsr_trim | **64.1** | 76.8 | 65.6 | 61.5 | 13.9 |
528
+ | **5** (4–6) | reputation | **63.8** | 76.8 | 67.9 | 58.5 | 0.0 |
529
+ | **6** (5–7) | eigentrust | **63.0** | 75.0 | 66.8 | 58.2 | 0.0 |
530
+ | **7** (6–8) | dawid_skene | **62.0** | 76.3 | 67.4 | 55.5 | 0.0 |
531
+ | **8** (7–8) | solo | **61.6** | 61.5 | 62.0 | 61.2 | 13.9 |
532
+ | **9** (9–9) | trust_all | **56.7** | 76.8 | 59.9 | 51.7 | 0.0 |
533
+ | **10** (10–10) | majority | **55.7** | 76.8 | 61.5 | 48.1 | 0.0 |
534
  | ref | REF oracle | **71.3** | 76.8 | 74.5 | 67.5 | 0.0 |
535
 
536
+ - **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
537
+ literature baseline (wmsr_trim). The oracle that knows who lies scores 71.3.
538
+ - **Goal reached in time:** RACE-Net-M + repair 97.8%, RACE-Net-M 97.4%, `race_map` 94.2%,
539
+ W-MSR trimming 96.5%, beta-reputation 91.9%, majority 82.1% (oracle 98.2%).
540
+ - **Wall-inventing liars**, the weak spot of `race_map` (59.5), are handled: RACE-Net-M 71.7.
541
+ - **Still open:** the colluding ring (66.2 with repair, oracle 71.0). The liars are honest everywhere
542
+ else, so their channel looks reliable until a robot stands next to the goal.
543
 
544
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackD_vs_liars.png" width="980"></p>
545
 
546
+ <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;
547
+ RACE-Net-M learns that these peers' wall reports are worthless while their "free" reports are true, and keeps only the latter.
548
+ <a href="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/videos/maze_false_walls.mp4">MP4</a></sub></p>
549
+
550
+ Belief quality on held-out mazes (accuracy on cells a peer reported and the receiver has not sensed; 6,111 snapshots):
551
+
552
+ | model | all | 3 of 5 peers lie | 5 of 5 peers lie |
553
+ |---|---:|---:|---:|
554
+ | `race_map` (hand-written Bayesian rule) | 90.1 | 82.7 | 72.9 |
555
+ | rounds = 1 (no refinement) | 91.8 | 85.7 | 76.1 |
556
+ | rounds = 2 | 92.2 | 86.4 | 77.4 |
557
+ | rounds = 3 (released) | 92.3 | 86.7 | 77.6 |
558
+ | rounds = 3, second seed | 92.3 | 86.7 | 77.6 |
559
+
560
+ **Other swarm and map sizes.** The model was trained on 6 robots in 12×12-cell mazes. Unchanged, with half of the robots lying
561
+ (10 new mazes × 5 attacks per row; computed in a Kaggle CPU notebook):
562
+
563
+ <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/racenet_m_scaling.png" width="960"></p>
564
+
565
+ | setting | liars | beta-reputation | W-MSR trimming | race_map | RACE-Net-M + repair | oracle |
566
+ |---|---:|---:|---:|---:|---:|---:|
567
+ | 4 robots · 8×8 | 2 | 74.4 | 72.0 | 73.3 | **73.6** | 76.1 |
568
+ | 6 robots · 12×12 (training size) | 3 | 61.1 | 54.0 | 63.6 | **67.7** | 66.9 |
569
+ | 10 robots · 12×12 | 5 | 70.7 | 60.5 | 70.6 | **76.7** | 77.5 |
570
+ | 8 robots · 16×16 | 4 | 57.5 | 59.4 | 62.5 | **69.5** | 67.7 |
571
+
572
+ 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.
573
+
574
+ **Reachability repair** (planner ablation, same 500 cells). When the fused map says the goal cannot be reached, the standard planner drops every peer;
575
+ 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
576
+ already polluted with false walls, so it only pays off on top of an accurate fusion model:
577
+
578
+ | fusion rule | standard fallback | + reach-repair | collude: standard → repair | false_walls: standard → repair |
579
+ |---|---:|---:|---:|---:|
580
+ | majority | 55.7 | 49.2 | 63.2 → 66.2 | 57.6 → 45.5 |
581
+ | eigentrust | 63.0 | 61.0 | 62.6 → 65.0 | 53.8 → 42.2 |
582
+ | reputation | 63.8 | 61.8 | 62.7 → 66.2 | 57.0 → 44.8 |
583
+ | race_map | 66.2 | 65.6 | 63.2 → 66.2 | 59.5 → 54.3 |
584
+ | race_net_m | 69.8 | 70.2 | 63.2 → 66.2 | 71.7 → 71.2 |
585
 
586
  <details><summary><b>More Track D figures</b></summary>
587
  <p align="center"><img src="https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/media/figures/trackD_leaderboard.png" width="640"></p>
 
627
  | Dawid–Skene implemented correctly | ≥90% decision agreement with the crowd-kit reference implementation on identical data |
628
  | RACE-S calibration is rank-invariant | a cubic distortion of one agent changes ≤2% of decisions |
629
  | SO-101 labels are physical | labels read from the simulator after settling; replay of a cached seed must reproduce its label |
630
+ | 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 |
631
  | findings replicate on a new robot | all methods frozen before the Franka Panda data was generated (Track A-Panda) |
632
+ | 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`) |
633
  | maze checker is exact | all 40 BFS-shortest paths validate with their optimal length (`tests/test_maze.py`) |
634
+ | 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 |
635
+ | 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`) |
636
+ | 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`) |
637
+ | 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) |
638
+ | 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)) |
639
+ | independent replication on Kaggle CPU | a Kaggle notebook (4 vCPU, Intel(R) Xeon(R) CPU @ 2.20GHz) re-ran the tests (53 passed, 4 skipped in 66.52s (0:01:06)) and recomputed the leaderboard from the public Hugging Face copy (Nabidnur/RoboRACE-Bench as of 2026-11-20 16:24 UTC): 1,217 leaderboard values compared, max abs difference 0.0000, 109.2 min |
640
 
641
 
642
+ `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.
643
 
644
  ## Reproduce (CPU)
645
 
 
685
  scripts/ run_benchmark · train_racenet · select_racenet · racenet_analysis · make_figures · make_videos · make_gifs · make_architecture · make_plotnn · make_readme ·
686
  publish_hf + make_space + make_playground (Hugging Face dataset, model, Spaces) · agents
687
  results/ LEADERBOARD.md · leaderboard.json · cached agent outputs
688
+ kaggle/ the two Kaggle CPU notebooks (RACE-Net training, independent replication)
689
  media/ figures · videos (mp4 + gif) · architecture (Raschka-style + PlotNeuralNet)
690
  ```
691
 
692
  ## Limitations
693
  - 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.
694
  - 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.
695
+ - 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.
696
+ - 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.
697
  - Online variants fit on the current unlabelled evaluation window; a deployed robot would use a sliding window.
698
  - 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.
699
  - PlanBench uses a seeded 100-instance subset. Track C saturates after engineering.
 
702
  ## Provenance & citation
703
  RACE / *Liars Are Information* ([public demo Space](https://huggingface.co/spaces/Nabidnur/liars-are-information-demo); the dataset `Nabidnur/liars-are-information` is private) ·
704
  AIP-AD ([`Nabidnur/AIP-AD`](https://huggingface.co/Nabidnur/AIP-AD)) · ERQA (Gemini Robotics team) · PlanBench (Valmeekam et al., NeurIPS 2023 D&B) ·
705
+ LLM-Modulo (Kambhampati et al., ICML 2024) · W-MSR (LeBlanc et al., IEEE JSAC 2013) · EigenTrust (Kamvar et al., WWW 2003) ·
706
+ beta-reputation (Jøsang & Ismail, 2002) · SPL (Anderson et al., 2018) · Dawid & Skene (1979) · crowd-kit (Toloka) ·
707
  Cosmos-Reason2 (NVIDIA) · SO-101 model (MuJoCo Menagerie) · PlotNeuralNet (H. Iqbal) · diagram style after S. Raschka.
708
 
709
  ```bibtex