BEHAVIOR Challenge 2026 — task 64 (setting_the_table), low-lr fine-tune (skill stages), step 29999

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 2025 winner's 50-task generalist behavior_50t_checkpoint (task 64 is a new 2026 task with no winner specialist), fine-tuned in two stages (below).
Training config: pi_behavior_b1k_2026_task64_cont (in our pipeline repo's behavior-1k-solution/src/b1k/training/config.py), with skill-annotation subtask stages.
Data: 2026 challenge demos, task 64 episodes 12800-12979 (180 demos), RGB cameras only.
Recipe: two stages: (1) 20k steps from the winner's 50-task generalist (behavior_50t_checkpoint), batch 16, peak lr 5e-5, skill stages (config pi_behavior_b1k_2026_task64); (2) +30k steps continuing from that step 19999 with a fresh warmup + cosine schedule, peak lr 5e-5, batch 16 (config pi_behavior_b1k_2026_task64_cont). This is the final step 29999 (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.

Serve with scripts/serve_b1k.py --task-id 64 policy:checkpoint --policy.config pi_behavior_b1k_2026_task64_cont --policy.dir <this dir>.

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 = 0.0375 (sum 0.3750 over 301-310; full completion: none).

instance	301	302	303	304	305	306	307	308	309	310
q	0	0	.125	0	.125	0	0	0	0	.125

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview