RichardCTian/b1k-2026-task10-ft1

BEHAVIOR Challenge 2026 — task 10 (set_up_a_coffee_station_in_your_kitchen), 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 checkpoint_1 (it already trained task 10; 50 task slots).
• Training config: pi_behavior_b1k_2026_task10 (in the team pipeline repo's behavior-1k-solution/src/b1k/training/config.py).
• Data: 2026 challenge demos, task 10 episodes 2000-2179 (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_task10/norm_stats.json — normalization stats
• _CHECKPOINT_METADATA

Serve with scripts/serve_b1k.py --task-id 10 policy:checkpoint --policy.config pi_behavior_b1k_2026_task10 --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 10 (6 goal predicates (ontop/nextto placements))

Leaderboard score: q = 0.2333 (sum 2.33 over 301-310; full completions: none)

instance	301	302	303	304	305	306	307	308	309	310
q	0.333	0	0.167	0.5	0.167	0.167	0.5	0.167	0	0.333

All partial credit, in sixths: 8 of 10 rollouts placed at least one item.