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.