BEHAVIOR Challenge 2026 β task 12 (preparing_lunch_box), 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 specialist
checkpoint_2, which already covered task 12 (2025 q 0.517 public / 0.367 hidden; 50 task slots, so its task-12 embedding is kept). - Training config:
pi_behavior_b1k_2026_task12(in the team pipeline repo'sbehavior-1k-solution/src/b1k/training/config.py). - Data: 2026 challenge demos, task 12 episodes 2400-2579 (last 20 held out), RGB cameras only.
- Recipe: 10,000 steps (final step is 9999), batch 16, peak lr 5e-5, 1x A100 80GB.
- Checkpoint is params-only (no optimizer state): it can be served or used as an init, not
--resumed.
Layout
params/β orbax/ocdbt parametersassets/b1k_2026_task12/norm_stats.jsonβ normalization stats_CHECKPOINT_METADATA
Serve with scripts/serve_b1k.py --task-id 12 policy:checkpoint --policy.config pi_behavior_b1k_2026_task12 --policy.dir <this dir>.
Evaluation (local; q = fraction of goal predicates newly satisfied, 1.0 on success)
Scored the 2026 leaderboard way: public instances 0-9 (ids 301-310), one rollout each, default 1.5x timeouts, task score = sum / 10.
Task 12 (6 goal predicates: apple halves, sandwich, cookie and tea bottle into the lunch box; fridge closed)
Leaderboard score: q = 0.3667 (sum 3.67 over 301-310; no full completions)
| instance | 301 | 302 | 303 | 304 | 305 | 306 | 307 | 308 | 309 | 310 |
|---|---|---|---|---|---|---|---|---|---|---|
| q | .333 | 0 | .667 | .5 | .667 | .667 | .333 | 0 | .5 | 0 |
Partial credit on 7 of 10 instances; the fridge-closed predicate is already true at reset, so it earns nothing.