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.
Checkpoint sweep: task 38 spraying_for_bugs
Simulator rollouts of all 10 checkpoints (every 2,500 steps) of the Pi0.5 fine-tune jermo1120/b1k-task38-spraying-for-bugs (lr5e-6/) on BEHAVIOR-1K task spraying_for_bugs. step_24999 is the same weights as jermo1120/pi05-b1k-task38-spraying-for-bugs.
Instances: public_test indices 10–14 (instance ids 311–315), the non-leaderboard half of the split, 1 rollout each. The leaderboard runs (301–310) are in jermo1120/b1k-task38-spraying-for-bugs-official-eval.
Results (q_score per instance)
| checkpoint | 311 | 312 | 313 | 314 | 315 | mean q_score | full successes |
|---|---|---|---|---|---|---|---|
step_02500 |
0.5 | 0 | 0 | 0 | 0 | 0.10 | 0 |
step_05000 |
0 | 0 | 0.5 | 0.5 | 1.0 ✓ | 0.40 | 1 |
step_07500 |
0.5 | 0 | 0.5 | 0 | 1.0 ✓ | 0.40 | 1 |
step_10000 |
0 | 0 | 0 | 0 | 0.5 | 0.10 | 0 |
step_12500 |
0.5 | 0.5 | 0.5 | 0.5 | 0.5 | 0.50 | 0 |
step_15000 |
0 | 1.0 ✓ | 0.5 | 0 | 0.5 | 0.40 | 1 |
step_17500 |
0.5 | 0.5 | 0.5 | 0 | 0.5 | 0.40 | 0 |
step_20000 |
0.5 | 0.5 | 0 | 0.5 | 0 | 0.30 | 0 |
step_22500 |
0 | 0.5 | 0 | 0 | 0.5 | 0.20 | 0 |
step_24999 |
0.5 | 0 | 0.5 | 0.5 | 0.5 | 0.40 | 0 |
Every checkpoint covers one of the two plants on some instances. Full successes (both plants) came from step_05000, step_07500 and step_15000, one each. With one rollout per instance, differences between checkpoints are mostly noise.
Setup
| Evaluator | BEHAVIOR-1K v3.9.3-post2 (python -m omnigibson.eval.eval), OmniGibson 3.9.3, Isaac Sim 5.1 |
| Episode limit | 9,720 steps (evaluator default: 1.5× mean human-demo length) |
| Robot / wrapper | omnigibson/eval/r1pro.yaml (robot robot_r1), default wrapper (224×224 RGB + proprio) |
| Policy server | behavior-1k-solution ca556f7 + the model repo's changes.patch, config pi_behavior_b1k_fast, 20 flow steps, the winners' eval tricks on, served through serve_2026_adapter.py (2025→2026 observation adapter, task id 38) |
| Hardware | 1× RTX 5090 (simulator and policy on the same GPU) |
The model was trained on 2025-format data. serve_2026_adapter.py maps the 2026 evaluator's 61-dim proprio into the 2025 256-dim layout and sets the task id to 38 (the same id in both numberings). Base velocity is passed in the robot frame, while the 2025 training data used the world frame.
q_score gives partial credit. The task's only goal is "every potted plant (2) is covered in insectifuge", so per-instance scores are 0, 0.5 or 1.0 (success).
Files: per-rollout metrics in results/<checkpoint>/*.json and results/summary.csv; videos (head + both wrist cameras) in videos/<checkpoint>/; the server adapter and run script at the top level.
- Downloads last month
- 61