TianxingChen/RoboTwin2.0, revision 3dc3b798668feb99ac61cc9086d84cbcc3d79186, files dataset/<task>/<set>.zip with the sets arx-x5_clean_50.zip, arx-x5_randomized_500.zip; aloha-agilex_clean_50.zip, aloha-agilex_randomized_500.zip; ur5_clean_50.zip, ur5_randomized_500.zip; franka_clean_50.zip, franka_randomized_500.zip; piper_clean_50.zip, piper_randomized_500.zip.
RoboTwin 2.0 is the work of Tianxing Chen et al. (RoboTwin-Platform/RoboTwin). Cite their paper when you use this data.

License

MIT, as the source; the notice is in LICENSE. The files here were changed from the source as listed below.

Changes from the source

Format: LeRobot v3.0 (parquet tables, one video stream per camera), 17 fps label (the source rate is 16.67 Hz; no frame was dropped or added).
observation.state and action: 16 values, [x y z qw qx qy qz gripper] for the left arm, then the right arm. Absolute end-effector poses in metres and unit quaternions (w first). The state is the source's recorded end-effector pose; the action is the pose of the next frame's joint targets, computed with the embodiment's kinematics. A constant offset per arm is added to the positions (each arm's base at a common place); the axes are the source's. Gripper: 0 closed .. 1 open.
Video: head camera (cam_high) and the two wrist cameras (cam_left_wrist, cam_right_wrist), resized from 320x240 to 224x168 (area filter, same 4:3 shape), AV1 (crf 26).
Language: one instruction per episode, taken from the source's instruction list for that episode.
Other source fields (joint values, point clouds, camera parameters, the third-person view) are not included.
999 episodes named by the source's set sizes are not in its files.
35 source episodes are not included (a joint command far from the recorded motion, or no instruction file).