BEHAVIOR Challenge 2026 — task 32 (wash_a_baseball_cap), fine-tune ft1, step 40000
Selected checkpoint for the team's BEHAVIOR-1K 2026 submission (task owner: Richard).
Model
- Architecture: the 2025 BEHAVIOR Challenge winner's pi0.5-based policy (IliaLarchenko/behavior-1k-solution, ~3B params, task-ID embeddings).
- Init: the winner's task-group specialist
checkpoint_3(IliaLarchenko/behavior_submission), 50 task slots; task 32 uses slot 32. - Training config:
pi_behavior_b1k_2026_task32— skill-annotation subtask stages, 2026 challenge demos for task 32 (episodes 6400-6579; last 20 held out), RGB cameras only, 50,000 steps, batch 16, peak lr 5e-5, cosine decay. - Checkpoint selection: steps 10000/20000/30000/40000/49999 screened on test instances 311-315 (10000 and 49999 dropped: first three all zero; 20000 and 30000 mean 0.10; 40000 mean 0.40).
- Params-only checkpoint (no optimizer state):
params/(orbax/ocdbt) +assets/(normalization stats, FAST tokenizer). - Serve with
serve_b1k.py --task-id 32 policy:checkpoint --policy.config pi_behavior_b1k_2026_task32 --policy.dir <this repo>.
Official public instances 301-310 — q = 0.350 (sum 3.5/10)
| instance | 301 | 302 | 303 | 304 | 305 | 306 | 307 | 308 | 309 | 310 |
|---|---|---|---|---|---|---|---|---|---|---|
| q | 0 | 0 | 0 | 1.0 | 0 | 1.0 | 0 | 0.5 | 0.5 | 0.5 |
Test instances (selection only): 311=0, 312=0, 313=1.0, 314=0, 315=1.0.
Evaluation: BEHAVIOR-1K @ b1979916e (v3.9.2 series), RGBDFullResWrapper, r1pro.yaml, one rollout per instance.
Credit
Built on the 2025 winner's code and checkpoint (IliaLarchenko, Apache-2.0).