| # BEHAVIOR 2026 Self-Evaluation | |
| This dataset contains three independent evaluation rounds: `round0/`, `round1/`, and `round2/`. Each round covers 100 tasks with instances 301-310, totaling 1,000 episodes. Results are reported separately for each round. | |
| Files are named `{task}_{instance}_{rollout}`. | |
| | Directory | Successful episodes | Macro Q-score | Video size | | |
| | --- | ---: | ---: | ---: | | |
| | `round0/` | 398 | 0.604701 | 29.50 GB | | |
| | `round1/` | 416 | 0.617207 | 29.32 GB | | |
| | `round2/` | 414 | 0.618579 | 29.06 GB | | |
| Each round contains `metrics/`, `videos/`, `robot/r1pro.yaml`, `wrapper/default_wrapper.py`, and `selection_index.json`. | |
| Macro Q-score is calculated by averaging the 10 instance Q-scores within each task, then averaging the 100 task means equally. | |