You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

BEHAVIOR-1K 170k policy rollouts

LeRobot v3.0 dataset containing 1,000 policy rollout episodes from checkpoint step 170,000: 100 tasks × public-test instance IDs 0–9 (scene instance IDs 301–310).

Successes and failures are included together: 215 successes, 785 failures. Mean final q-score across all 1,000 episodes is 0.37890160117660115 (0.379 rounded).

Contents

  • data/: 14,889,690 frames of actions, observations, rewards, and termination flags.
  • videos/: head, left-wrist, and right-wrist RGB streams at 30 FPS.
  • meta/: LeRobot metadata, task definitions, episode indexing, and statistics.
  • json_bundles/: 100 per-task result bundles containing episode outcomes and metrics.
  • episode_index_map.json: maps package IDs in the result bundles to LeRobot episode indices.

Robot type: R1Pro. Action dimension: 23. State dimension: 61. Source actions and states were preserved exactly, without an additional state transform. Converted episode coverage, source-array equality, video frame counts, and video time ranges were verified before publication.

These are policy rollouts on public-test instances, not expert demonstrations. Training on these episodes means those same instances cannot serve as an independent evaluation set. Result bundles contain episode metadata and metrics, not simulator replay/state files. Internal execution paths and infrastructure metadata have been omitted from the public bundles.

Evaluation

See EVALUATION_README.md for endpoint URLs, the evaluation command, and the WSS support patch.

Downloads last month
307