RoboRACE-Bench / docs /MODELS.md
Nabidnur's picture
RoboRACE: code, cached agent outputs, results, media, SO-101 renders (part 8)
f5c9954 verified
|
Raw History Blame Contribute Delete
4.56 kB
# The RACE-Net family: one design, two models
Both models in this repository answer the same question in two domains: *a robot receives reports from peers, some of
whom lie in a structured way; what is true?* They share one design, taken from RACE and AIP in the *Liars Are Information*
program, and they differ only in what an "item" is.
| | **RACE-Net-V** (verification) | **RACE-Net-M** (map) |
|---|---|---|
| Task | pick-and-place: did the arm succeed? | maze: which cells are walls? |
| Item | one episode in the recent window | one map cell |
| Peer report | a score in [0, 1] from a camera + VLM | a claim *free* / *wall* with a time stamp |
| The receiver's anchor | its own (noisy) score for every episode | its own exact sensing of some cells |
| Code | `src/robotics/swarm/racenet.py` | `src/robotics/swarm/racenet_m.py` |
| Size | 213,889 parameters | 44,837 parameters |
| Trained on | synthetic swarms (`sample_batch`) | simulated maze swarms with random two-coin liars (`sample_channels`) |
| Replaces | RACE / RACE-S (EM per window) | `race_map` (Bayesian model averaging) |
## The shared design
1. **Anchor.** The receiver's own evidence is the only thing assumed honest. Nothing is voted on and no labels are used.
This breaks the symmetry that makes majority vote and unanchored crowd models (Dawid–Skene, GLAD, MACE) fail when
liars are the majority.
2. **In-context channel estimation.** Each peer's *channel* (how its reports relate to the truth) is estimated from the
current unlabelled window or map by comparing the peer with the anchor. RACE-Net-V does this with attention across
episodes inside each agent's column; RACE-Net-M does it with counts of claim × anchor, pooled over cells.
3. **A swarm view.** Attention across peers lets the estimate for one peer depend on the others (who else is reliable,
who repeats whom).
4. **Signed use.** The read-out gives evidence with a sign: a reliable peer is trusted, a consistent liar is **inverted
and used**, a random peer is ignored. Reputation systems (beta-reputation, EigenTrust) can only discard.
5. **Sets, not sequences.** No positional encodings anywhere: both models are permutation-equivariant over peers and
over items, so they run on any swarm size, window length or map size.
6. **Synthetic training, zero-shot use.** Neither model sees benchmark data. RACE-Net-V is trained on a generator of
synthetic swarms; RACE-Net-M on random two-coin liars with change points, never on the benchmark's named attacks.
RACE-Net-M adds **iterative refinement**: after a first pass, each peer is also checked against the belief implied by
all the *other* peers on cells the receiver has not sensed (leave-one-out), for three rounds. A robot can therefore
verify a peer it has never overlapped with, through a peer it already trusts.
## System-level pieces built on the models
- **RACE-Pool** (`fusion.race_pool`, pick-and-place). An AIP-style pool of three receiver-anchored experts that fail in
different ways: RACE-Net-V (learned), RACE-S (score-level EM) and RACE (vote-level EM). Each expert's log-odds are
centred on its own threshold, scaled to unit spread inside the window and summed. The expert set and the equal weights
were fixed on the development track (A0).
- **Reachability repair** (`trackD.repair_field`, maze). The task guarantees that the goal can be reached. When the
fused map says it cannot, the standard planner drops every peer and plans on the robot's own observations. The repair
keeps the fused map and un-believes only the claimed walls that enclose the goal. It is evaluated on top of every
fusion rule (ablation table in the leaderboard); it helps against the colluding ring and can hurt a rule whose map is
already polluted with false walls.
## Protocol
| | RACE-Net-V | RACE-Net-M |
|---|---|---|
| Training data | `racenet.sample_batch` (generators v1–v3) | maze seeds 5000+ |
| Validation | held-out synthetic swarms | maze seeds 9000+ (cell accuracy, next to the Bayesian rule) |
| Development (selection, tuning) | Track A0 | maze seeds 0–39 (`scripts/tune_trackD.py`) |
| Test | SO-101, Franka Panda, and a hold-out generated after the freeze | maze seeds 1000–1019 |
Things that were tried and did not work are kept in the record: an always-on risk-aware planner (worse than optimistic
planning on development mazes), a minimum-violation fallback (no gain: the robots detoured through unexplored cells),
and the v2 generator for RACE-Net-V (better with weak receivers, worse on large swarms).