The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Simulator evaluation of jermo1120/b1k-task37-clean-a-trumpet
Rollouts of the fine-tuned Pi0.5 policy jermo1120/b1k-task37-clean-a-trumpet (checkpoint lr5e-6/step_09999) on BEHAVIOR-1K task clean_a_trumpet, run in the official 2026 BEHAVIOR evaluator.
Status: complete.
Results
| Rollouts finished | 10 / 10 |
| Successes | 0 |
| Mean q_score | 0.000 |
| instance_id | success | q_score | steps |
|---|---|---|---|
| 301 | no | 0.000 | 7961 |
| 302 | no | 0.000 | 7961 |
| 303 | no | 0.000 | 7961 |
| 304 | no | 0.000 | 7961 |
| 305 | no | 0.000 | 7961 |
| 306 | no | 0.000 | 7961 |
| 307 | no | 0.000 | 7961 |
| 308 | no | 0.000 | 7961 |
| 309 | no | 0.000 | 7961 |
| 310 | no | 0.000 | 7961 |
Per-rollout metrics are in results/*.json and results/summary.csv. Each rollout has a video in videos/ (head camera plus both wrist cameras) and a simulator log in logs/.
Instance 309: the simulator physics blew up after the robot pushed into a door early in the episode, and it was launched into the sky (base travel about 22.7 km). Its distance metrics are meaningless; the rollout still counts as a failure.
Diagnostics
All 0.0 so far. diagnostics/ has single-rollout tests of the base-velocity frame (zeroed, rotated back to world frame) and of the original checkpoint_3 in the same setup. The original checkpoint_3 also scores 0, which points to the 2025→2026 evaluator adapter rather than the fine-tune. See diagnostics/README.md.
Setup
| Evaluator | BEHAVIOR-1K v3.9.3-post2 (python -m omnigibson.eval.eval), OmniGibson 3.9.3, Isaac Sim 5.1 |
| Split | public_test, instance indices 0–9, 1 rollout each, default seed |
| Robot config | omnigibson/eval/r1pro.yaml |
| Episode limit | 7,960 steps (evaluator default: 1.5× mean human-demo length) |
| Policy server | scripts/serve_2026_adapter.py, behavior-1k-solution ca556f7, config pi_behavior_b1k_fast, 20 flow steps, the winners' eval tricks on |
| Hardware | 1× RTX 5090 (simulator and policy on the same GPU) |
Caveats
The model was trained on 2025-format data, so the policy server adapts the 2026 evaluator's observations (see the docstring in scripts/serve_2026_adapter.py):
- The 61-dim 2026 proprio is scattered into the 2025 256-dim layout. Fields the model doesn't read stay zero.
- The task ID is overridden to the 2025 index (37).
- Base velocity is passed in the robot's local frame, but the training data used the world frame. Converting it would need the robot's global yaw, which is privileged. This may hurt base motion.
In instance 301 the robot picked up the scrub brush and reached the trumpet, but timed out without cleaning it.
- Downloads last month
- 658