# Research & Build Plan — Byzantine-Robust Multi-Agent Robotics **Thesis:** robot teams increasingly depend on many agents: cameras, other robots, VLM judges and LLM planners. Some of them will be wrong in *structured* ways (biased, stuck, colluding, adversarial). The *Liars Are Information* line (RACE, AIP, AIP-AD) showed that structured liars can be estimated without labels, anchored on the receiver's own honesty, and then **inverted or de-biased and used**, not merely discarded. This project carries that idea into robotics. It measures the result on a public, rankable, CPU-runnable benchmark, and it puts engineering ahead of scale: cheap models plus good plumbing should beat the strongest single model. Status: ☑ done · ◐ in progress · ☐ planned. Numbers live in [`results/LEADERBOARD.md`](../results/LEADERBOARD.md) and are regenerated by `scripts/run_benchmark.py`. This plan never hard-codes them. --- ## 1. Inventory: what we build on | Asset | Role here | |---|---| | **RACE v3.1** (`Nabidnur/liars-are-information`, `gemma4-race-pack`) | anchored, clone-aware latent-truth estimator; vendored as `src/robotics/swarm/race.py` | | **AIP / AIP-AD** (`Nabidnur/AIP-AD`) | dissent poolers, Agent Calibration Layer (ACL), GRPO arbiter; ACL + `aip_mt` reused in Track A | | E8/E10 live-swarm methodology | cached honest + saboteur answers, pre-registered hypotheses → Track C design | | `Sim2Reason-Data`, TRACEA, MAGNet | candidate sources for future tracks (physics-reasoning QA, trace explanations, multi-agent graphs) | | **Cosmos-Reason2-2B** (gated, accessible) | VLM success verifier on CPU (~1.4 s per view, bf16) | | Groq / Cerebras / Gemini free tiers | LLM planner agents (OpenRouter free tier persistently rate-limited; SambaNova needs a payment method) | | MuJoCo 3 + EGL software rendering | procedural tabletop scenes, 5-camera rig, no GPU needed | ## 2. Phase A — CPU research sprint ☑ | # | Item | Status | |---|---|---| | A1 | Multi-provider LLM client with fallback, live-verified model defaults | ☑ | | A2 | Skill-constrained planner, canonical plan form, procedural task generator (12 kinds incl. traps) | ☑ | | A3 | Cosmos-Reason2 verifier (one-pass P(yes) from logits; optional CoT rationale) + calibration study | ☑ held-out AUROC 0.925, 83% acc (56% at the naive 0.5 threshold) | | A4 | **RoboRACE-Bench**: 3 tracks, 10 attacks, 20+ methods, receiver rotation, bootstrap-rank leaderboard ([spec](BENCHMARK.md)) | ☑ | | A5 | New methods: RACE-S (score-level RACE), online RACE / RACE-S, RACE-C+W-MSR (online isolation), engineered planner (validator + repair) | ☑ #1 in Tracks A and B | | A6 | Systems study: engineered small-model swarm vs best single model (McNemar-tested) | ☑ | | A7 | Figures (12) and videos (camera swarm, robot swarm) generated from results | ☑ `media/` | ### Pre-registered hypotheses for the sprint Written before the final runs. **Verdicts:** H1 ✅, H2 ✅, H3 ✅, H4 ✅ (offline variants; the online variants fix it), H5 ◐ (engineering helps 3 of 4 models; the best zero-shot model already scored 100%). Details are in the README. - **H1 (Track A):** under liar-majority cells, anchored methods (RACE, RACE-S) beat majority, mean and Dawid–Skene by ≥ 10 points of mean accuracy. - **H2 (Track A):** with no attack, RACE-S is within 3 points of the labelled stacker (label-free ≈ supervised). - **H3 (Track B):** RACE-C+W-MSR has the best mean mission success across all cells, and beats W-MSR on `collude` and `two_faced` at F = 4 of 12. - **H4 (Track B, expected failure):** history-fitted defences are hurt by `sleeper` attacks more than robust statistics are. - **H5 (Track C):** an engineered planner beats the same model zero-shot. The best engineered system beats the best zero-shot single model on the 48-task suite. ## 3. Phase A2 — research-grade iteration (this release) ☑ | # | Item | Status | |---|---|---| | B1 | **Public benchmark: ERQA.** Cosmos-Reason2-2B reproduction on CPU (reported 37.75% must lie in our CI), resolution ablation, multi-agent swarm with hosted VLMs, letter-level attacks, SOTA reference subset | ☑ | | B2 | **Public benchmark: PlanBench Blocksworld.** Exact simulator validated on all 500 official GT plans; official one-shot vs LLM-Modulo; swarm strategies (plurality / verify-select / cascade) | ☑ | | B3 | **Track A on a real arm model:** SO-101 (MuJoCo Menagerie), IK pick-and-place, physics ground truth, 4 failure modes, 220 episodes × 5 cameras | ☑ | | B4 | **Validation against references and theory:** crowd-kit DS/GLAD/MACE baselines and an agreement test, the W-MSR safety theorem, rank-invariance, ERQA reproduction | ☑ | | B5 | **Confidence-gated anchoring (CGA)** plus the sign-only anchor ablation (robustness ↔ usefulness trade-off) | ☑ Proposed post hoc on a 136-question ERQA subset, where its worst cell matched the receiver. On the full 220-question set and on Track A that floor does **not** hold, so CGA is reported as a compromise point. | | B6 | Architecture diagrams (Raschka style + PlotNeuralNet) from the real checkpoint config; 1280×720 videos (robot rollouts, rendezvous, LLM-Modulo) | ☑ | | B7 | README generated from `results/leaderboard.json` (no hand-typed numbers) | ☑ | **Next:** make the anchor adaptive with a statistical test (e.g. a receiver-vs-peer agreement test with a controlled false-alarm rate) instead of a confidence gate, and pre-register it on a fresh benchmark not yet inspected (candidates: RoboVQA, Where2Place, the Cosmos physical-AI bench). ## 3b. Phase A3 — breadth, replication and our own model ☑ | # | Item | Status | |---|---|---| | C1 | **Track D: maze swarm** with a new robot type (mobile robots), 6 attacks, SPL metric; `race_map` vs beta-reputation, majority, naive merging | ☑ rank 1 on 500 cells | | C2 | **Track A-Panda:** cross-embodiment replication with every method frozen beforehand | ☑ | | C3 | **RACE-Net:** our own learned fusion model (axial set transformer, synthetic training only), zero-shot on real camera swarms; dev-only checkpoint selection; scaling study (liar fraction, window, swarm size, receiver skill) | ☑ weights on the Hub. It matches the EM methods but does not beat RACE-S online. Open weakness: a weak receiver. | | C4 | **LLM agents in text mazes** with exact-validator repair | ☑ | | C5 | **Published-number context** for ERQA and PlanBench, copied from the papers' tables with sources | ☑ | | C6 | **Kaggle CPU:** RACE-Net training run and an independent replication of the leaderboard | ☑ | | C7 | Coverage: hosted ERQA and PlanBench agents re-run after quota resets | ☑ | | C8 | Hugging Face: dataset (both robots), model repo, showcase Space, interactive playground Space | ☑ | ## 3c. Phase A4 — one design, two trained models ☑ | # | Item | Status | |---|---|---| | D1 | **RACE-Net-M:** learned map fusion (anchored channel encoder, peer transformer, signed read-out, 3 refinement rounds), trained on random two-coin liars | ☑ rank 1 on Track D | | D2 | Track D literature baselines: W-MSR-style trimming, EigenTrust, Dawid–Skene | ☑ | | D3 | Reachability repair (planner) with an ablation on every fusion rule | ☑ helps only on an accurate map | | D4 | **RACE-Net-V v3:** dev-informed generator, EM-refinement variant, larger model; candidates trained on Kaggle CPU | ☑ | | D5 | **RACE-Pool:** AIP-style pool of three anchored experts | ☑ | | D6 | Development data for the arms (history splits), fresh hold-out with a recorded freeze commit | ☑ | | D7 | Paired significance tests in the leaderboard; scaling studies for both models | ☑ | | D8 | Kaggle CPU: independent replication of the leaderboard, training runs, maze studies | ☑ | ## 4. Phase B — strengthen the science (next, still CPU) ☐ 0. **Harder Track C:** Track C saturates after engineering (every engineered receiver is ~98%). Add longer-horizon tasks, partial observability and ambiguous references so planner fusion becomes informative. 1. **Real camera frames.** Track A now uses an SO-101 physics simulation; the next step is real LeRobot SO-101 / LIBERO episode videos (public HF datasets) for Track A, judging the final frame of each episode against the dataset's own success labels. 2. **Heterogeneous verifier agents:** add Gemini vision and a second open VLM next to Cosmos, so agent errors are less correlated. 3. **GRPO arbiter** (from AIP-AD) over the Track A pooling strategies as a learned meta-method. 4. **Adaptive attackers:** gate-aware liars that lie only when the receiver is uncertain (the hardest RACE attack in the original study). 5. **Sleeper defence for RACE-S:** online re-estimation with change-point detection (RACE `C6_ONLINE`). 6. Publish the benchmark: HF dataset (cached agent outputs + scenes) plus a Space with the leaderboard. Submissions are fusion rules run against the cached outputs. ## 5. Phase C — embodied (needs a GPU) ☐ These items carry over from the original plan and are unchanged in substance. 1. Reproduce GR00T N1.7 LIBERO numbers (`scripts/eval_libero.sh`). Target: within ±3 points of the reported 96.5% average. 2. RL experts in MuJoCo Playground (`SO101PickCube`) produce demonstrations. Then fine-tune GR00T N1.7 / SmolVLA (`scripts/train_groot_so101.sh`, `scripts/train_smolvla_kaggle.sh`). 3. **Close the loop:** the planner swarm decomposes the task, GR00T executes, the camera-swarm verifier (RACE-S) judges, and the robot retries or replans on failure. Measure long-horizon success with and without the robust verifier (this was RQ2 of the original plan). 4. Multi-robot: two SO-101 arms share verdicts, and one arm's camera is deliberately degraded. ## 6. Evaluation rules (all phases) - Every number comes from a script and lands in `results/`. Generated docs are never hand-edited. - Unlabelled history and evaluation splits are separate. Thresholds and channel estimates use history only. - Report the 95% Wilson CI for accuracies, bootstrap rank intervals for leaderboards, and exact McNemar tests for paired system comparisons. - Report failures as prominently as wins (see H4). ## 7. Keys & compute Keys are read from environment variables only (`.env.example`). They are never written to the repo or logs. CPU budget for the full sprint: about 2 h, dominated by 1,000 Cosmos-Reason2 view scorings and about 600 rate-limited LLM calls. Re-running the leaderboard from caches takes a few minutes.