LeRobot v3.0 dataset containing 1,000 policy rollout episodes, covering 100 tasks × 10 public-test instances (public indices 0–9, scene instance IDs 301–310), with 14,718,807 frames.

Both successful and unsuccessful task outcomes are included: 308 successes and 692 failures. Mean final q-score is 0.5142542971; every task contributes 10 episodes and has equal weight in this mean.

Data format

Robot: R1Pro.
Actions: 23-dimensional raw action vectors.
State: 61-dimensional proprioceptive state vectors.
RGB videos: head (720 × 720), left wrist (480 × 480), and right wrist (480 × 480), at 30 FPS, encoded with HEVC.
Rewards: zero before the final transition; the final transition contains the evaluator's final q-score.
Episode endings: termination and truncation flags are retained.

Actions and states retain their source values without additional smoothing or normalization during conversion. For hybrid rollouts with T actions and T+1 observations, the trailing observation is omitted so the dataset contains T aligned rows.

Files

data/: Parquet files containing actions, observations, rewards, and termination flags.
videos/: three RGB streams per episode.
meta/: LeRobot task definitions, episode indexing, feature descriptions, and statistics.
json_bundles/: 100 per-task JSON files, each containing 10 episodes with outcomes, evaluator metrics, and source provenance. Public instance IDs are 0–9; original evaluator instance IDs are retained in metrics where present. These are metadata bundles, not simulator replay/state files.
episode_index_map.json: maps bundle package IDs to LeRobot episode indices.

Evaluation service

We provide up to 160 parallel policy servers, with one rollout per server at a time. See EVALUATION_README.md for endpoint URLs, connection requirements, and evaluation commands, and policy_scheme.patch for WSS support.

Downloads last month: 67

Total file size:

74.5 GB