BEHAVIOR Challenge 2026 β task 67 (boxing_food_after_dinner), 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 50-task generalist
behavior_50t_checkpoint(no 2025 specialist exists for task ids >= 50). - Training config:
pi_behavior_b1k_2026_task67(in the team pipeline repo'sbehavior-1k-solution/src/b1k/training/config.py). - Data: 2026 challenge demos, task 67 episodes 13400-13579 (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_task67/norm_stats.jsonβ normalization stats_CHECKPOINT_METADATA
Serve with scripts/serve_b1k.py --task-id 67 policy:checkpoint --policy.config pi_behavior_b1k_2026_task67 --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 67 (6 goal predicates: 2 plates countertop -> sink, 2 tacos -> tupperware, tupperware -> fridge; fridge closed)
Leaderboard score: q = 0.0833 (sum 0.83 over 301-310; no full completions)
| instance | 301 | 302 | 303 | 304 | 305 | 306 | 307 | 308 | 309 | 310 |
|---|---|---|---|---|---|---|---|---|---|---|
| q | 0 | 0 | 0 | 0 | .333 | .333 | .167 | 0 | 0 | 0 |
Credit comes from the plates-to-sink goals; the policy follows the demo order and often stalls reaching for the tupperware in a lower cabinet. The fridge-closed predicate is already true at reset, so it earns nothing.