Episodes Preview R2006 Visualizer
20 episodes · 15 fps · 3 cameras · 960×600 h264

Pick wood block — R2006

LeRobot v3 dataset built from 13 R2006 MCAP recordings supplied for the task Pick up the wood block.

  • 6,349 frames, 20 episodes, 15 Hz (about 423.3 seconds of exported observations).
  • Robot: R2006; 24 joint positions in the source joint order.
  • Cameras: head, hips, wrist_R; RGB, 960 × 600, H.264.
  • One video file per camera per episode; no independently encoded stream concatenation.
  • Extra MCAP topics are excluded. Original MCAP files are not included.

Action meaning

action[t] = observation.state[t + 1] on the resampled 15 Hz timeline. These are derived absolute joint-position targets, not recorded control commands and not joint deltas. Joint units are preserved from the source. The prediction horizon is one frame (approximately 66.67 ms). The last observed state of each valid interval is used only as the final target. Actions never cross a recording or gap boundary.

The metadata lists all joint names in observation.state.names and action.names. Controller semantics and units must match these features when training or replaying. No GR00T embodiment configuration or success annotations are supplied.

Synchronization and repeated images

All modalities use MCAP log timestamps. Joint positions use the nearest sample within 100 ms. Images use the most recent earlier image, repeated until a new one arrives, with maximum age 6 seconds. A 15 Hz video therefore does not imply 15 newly captured images per second. Only the common time range of the selected streams is retained. Gaps split recordings into episodes; valid intervals shorter than one second are omitted. The 20 episodes are temporal segments, not necessarily 20 complete successful demonstrations.

Camera Median image age Maximum image age Repeated frames
head 0.511 s 3.836 s 92.9%
hips 0.350 s 2.484 s 89.9%
wrist_R 0.443 s 3.185 s 92.1%

synchronization.parquet records source filename, episode/frame indices, grid timestamps, source state/action timestamps, and source image timestamp and age for every exported frame. conversion_report.json includes source SHA-256 hashes, episode provenance, timing limits, joint names, and camera timing statistics.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    "andrejarden/pick-wood-block",
    video_backend="pyav",
)
frame = dataset[0]
state = frame["observation.state"]
action = frame["action"]
image = frame["observation.images.head"]

Authenticate with an account that can access this private repository before loading. Written using LeRobot 0.4.4 (dataset format v3).

Validation

validation.json records the checks performed: all state/action rows against source measurements, next-state targets within episodes, every image timestamp and age, every decoded video frame and presentation timestamp, and native LeRobot loading at the start, middle and end of every episode.

Downloads last month
157