yxma/React-lerobot

React — LeRobot v2.1 edition

LeRobot-format mirror of yxma/React — handheld bimanual tactile-visual manipulation (no robot arm). Loadable directly with the lerobot library.

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("yxma/React-lerobot")

Contents

• 36 episodes, 240,040 frames @ 30 fps, 2 tasks (motherboard, pushT)
• 5 video streams per frame (640×480, h264): observation.images.{view_left,view_middle,view_right,tactile_left,tactile_right}
• observation.state (14) = left+right GelSight 6-DoF poses (xyz+quat)
• observation.object_pose (7) + observation.object_pose_valid (object tracked only on motherboard; 0/NaN-free on pushT)
• observation.tactile (6) = L/R contact intensity/area/mixed
• action (14) = next-frame observation.state

Notes

• No depth here (LeRobot video is lossy h264). Lossless depth + per-task camera extrinsics live in the source repo yxma/React under data/<task>/depth/ and data/<task>/calibration/.
• Camera extrinsics/intrinsics are per-task (motherboard=May-12, pushT=June-26) — see the source repo.
• This is a handheld dataset: action is the next-frame pose (a convenience target), not a robot command.

License

CC-BY-4.0.