RichardCTian/b1k-2026-task40-ft1

BEHAVIOR Challenge 2026 — task 40 (make_microwave_popcorn), 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-40 specialist checkpoint_4 (2025 public-leaderboard q 0.90 on this task; 50 task slots).
• Training config: pi_behavior_b1k_2026_task40 (in the team pipeline repo's behavior-1k-solution/src/b1k/training/config.py).
• Data: 2026 challenge demos, task 40 episodes 8000-8179 (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/ — orbax/ocdbt parameters
• assets/b1k_2026_task40/norm_stats.json — normalization stats
• _CHECKPOINT_METADATA

Serve with scripts/serve_b1k.py --task-id 40 policy:checkpoint --policy.config pi_behavior_b1k_2026_task40 --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 40 (2 goal predicates (popcorn cooked + in a bowl))

Leaderboard score: q = 0.4000 (sum 4.00 over 301-310; full completions: 301, 302, 307, 308)

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

Every rollout scored 1.0 or 0: the two goals land together or not at all.