Rena-Tian/b1k-2026-task69-ft1

BEHAVIOR Challenge 2026 — task 69 (vacuuming_floors), openpi pi0.5 3-task model, step 49999

The checkpoint submitted for task 69 in our BEHAVIOR-1K 2026 entry.

Model

• Architecture: openpi pi0.5 (language-conditioned), not the 2025 winner's task-ID policy used for our other tasks.
• One model trained jointly on three 2026 tasks: 34 hanging_pictures, 63 installing_smoke_detectors, 69 vacuuming_floors (openpi config pi05_b1k_3task_calib, run run1, 50k steps; final step 49999).
• Data: 2026 challenge demos for those three tasks.
• Checkpoint is params-only here (params/ + assets/ normalization stats); the optimizer state was left out, so it can be served or used as an init but not --resumed.

Layout

• params/ — orbax/ocdbt parameters
• assets/b1k_3task_calib/norm_stats.json — normalization stats
• _CHECKPOINT_METADATA

Evaluation (q = fraction of goal predicates newly satisfied)

2026 leaderboard protocol: public instances 0-9 (ids 301-310), one rollout each, default 1.5x timeouts, task score = sum / 10.

Leaderboard score: q = 0.000 (all 10 rollouts hit the 3618-step limit; no success).

Task 69's goal is a single literal, not covered(floor, dust), and the simulator counts the floor as covered while even one dust particle remains, so every rollout scores 0 or 1 (no partial credit).

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