Rena-Tian/b1k-2026-task45-ft1

BEHAVIOR Challenge 2026 — task 45 (cook_hot_dogs), fine-tune ft1 (skill stages), step 6499

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.
• Training config: pi_behavior_b1k_2026_task45_r13k5 (in our pipeline repo's behavior-1k-solution/src/b1k/training/config.py), with skill-annotation subtask stages.
• Data: 2026 challenge demos, episodes 9000-9179, RGB cameras only.
• Recipe: 20k-step schedule (batch 16, cosine lr from 5e-5) run as warm restarts after two orbax save deadlocks: steps 0-13500 with config pi_behavior_b1k_2026_task45, then 6500 more steps from global step 13500 (lr 1.773e-5 -> 5e-6, config pi_behavior_b1k_2026_task45_r13k5). Local step 6499 = global step 19999. This is step 6499 (selected by screening on instances 311-315).
• Checkpoint is params-only (no optimizer state): it can be served or used as an init, not --resumed.

Layout

• params/ — orbax/ocdbt parameters
• assets/ — normalization stats
• _CHECKPOINT_METADATA

Serve with scripts/serve_b1k.py --task-id 45 policy:checkpoint --policy.config pi_behavior_b1k_2026_task45_r13k5 --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 = 0.6000 (sum 6.0000 over 301-310; full completion: 301, 302, 306, 307, 308, 310).

instance	301	302	303	304	305	306	307	308	309	310
q	1.	1.	0	0	0	1.	1.	1.	0	1.