BEHAVIOR Challenge 2026 β task 30 (setting_the_fire), low-lr fine-tune (skill stages), step 2999
Selected checkpoint for the BEHAVIOR-1K 2026 submission.
Model
- Architecture: the 2025 BEHAVIOR Challenge winner's pi0.5-based policy (IliaLarchenko/behavior-1k-solution, ~3B params, task-ID embeddings, no language prompt).
- Init: the winner's task-specific
checkpoint_2, fine-tuned (not used unchanged). - Training config:
pi_behavior_b1k_2026_task30_lowlr(in our pipeline repo'sbehavior-1k-solution/src/b1k/training/config.py), with skill-annotation subtask stages. - Data: 2026 challenge demos, task 30 episodes 6000-6179 (180 demos), RGB cameras only.
- Recipe: 3000 steps from the winner's checkpoint_2, batch 16, cosine lr 5e-6 -> 1e-6 (warmup 100), skill stages. Frames inside the skill segments 'pick up from', 'turn on switch', 'ignite' and 'turn off switch' are repeated 3x per epoch. This is the final step 2999 (no screening; straight to the official run).
- Checkpoint is params-only (no optimizer state): it can be served or used as an init, not
--resumed.
Layout
params/β orbax/ocdbt parametersassets/β normalization stats_CHECKPOINT_METADATA
Serve with scripts/serve_b1k.py --task-id 30 policy:checkpoint --policy.config pi_behavior_b1k_2026_task30_lowlr --policy.dir <this dir>.
Evaluation (q = fraction of goal predicates newly satisfied)
Scored the 2026 leaderboard way: public instances 0-9 (ids 301-310), one rollout each, default 1.5x timeouts, task score = sum / 10.
Leaderboard score: q = 0.3625 (sum 3.6250 over 301-310; full completion: 308).
| instance | 301 | 302 | 303 | 304 | 305 | 306 | 307 | 308 | 309 | 310 |
|---|---|---|---|---|---|---|---|---|---|---|
| q | .375 | .375 | 0 | .125 | .500 | .250 | .250 | 1. | .500 | .250 |