BEHAVIOR Challenge 2026 — task 1 (picking_up_trash), fine-tune ft1, step 9999
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(it already trained task 1; 50 task slots). - Training config:
pi_behavior_b1k_2026_task1(in our pipeline repo'sbehavior-1k-solution/src/b1k/training/config.py). - Data: 2026 challenge demos, task 1 episodes 200-379 (last 20 held out), RGB cameras only.
- Recipe: 10,000 steps (final step is 9999), batch 16, peak lr 5e-5, 1x A100 80GB. Final training loss 0.186.
- Checkpoint is params-only (no optimizer state): it can be served or used as an init, not
--resumed.
Layout
params/— orbax/ocdbt parametersassets/b1k_2026_task1/norm_stats.json— normalization stats_CHECKPOINT_METADATA
Serve with scripts/serve_b1k.py --task-id 1 policy:checkpoint --policy.config pi_behavior_b1k_2026_task1 --policy.dir <this dir>.
Evaluation (local; 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.
Task 1 (3 goal predicates (soda cans in the ashcan))
Leaderboard score: q = 0.5667 (sum 5.67 over 301-310; full completions: 302, 306)
| instance | 301 | 302 | 303 | 304 | 305 | 306 | 307 | 308 | 309 | 310 |
|---|---|---|---|---|---|---|---|---|---|---|
| q | 0.667 | 1. | 0.333 | 0.667 | 0 | 1. | 0 | 0.667 | 0.667 | 0.667 |