# 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.