BEHAVIOR Challenge 2026 β task 25 (clearing_food_from_table_into_fridge), low-lr fine-tune (skill stages), step 1999
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_1, fine-tuned (not used unchanged). - Training config:
pi_behavior_b1k_2026_task25_lowlr(in our pipeline repo'sbehavior-1k-solution/src/b1k/training/config.py), with skill-annotation subtask stages. - Data: 2026 challenge demos, task 25 episodes 5000-5179 (180 demos), RGB cameras only.
- Recipe: 2000 steps from the winner's checkpoint_1, batch 16, cosine lr 5e-6 -> 1e-6 (warmup 100), skill stages. This is the final step 1999 (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 25 policy:checkpoint --policy.config pi_behavior_b1k_2026_task25_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 = 0.1200 (sum 1.2000 over 301-310; full completion: none).
| instance | 301 | 302 | 303 | 304 | 305 | 306 | 307 | 308 | 309 | 310 |
|---|---|---|---|---|---|---|---|---|---|---|
| q | 0 | 0 | .200 | .200 | 0 | .200 | .200 | 0 | .400 | 0 |