GT-111/robotwin_v3_eef

RobotWin v3 EEF (LeRobot v3.0)

RobotWin 50-task bimanual benchmark, converted to LeRobot v3.0 with lossless stream-copied H.264 video.

Format

• Per-task LeRobot v3.0 datasets — each task is a subdirectory under clean/ or aug/
• 16D bimanual EEF: left arm [xyz, qw, qx, qy, qz, gripper] + right arm [xyz, qw, qx, qy, qz, gripper]
• Quaternion: wxyz, scalar-first
• State: absolute pose. Action: absolute position + absolute quaternion + gripper
• 3 cameras: cam_high + cam_left_wrist + cam_right_wrist, 640×480
• 2 subsets: clean (50 demos/task) and aug (500 demos/task), 100 total datasets
• ~32K episodes, ~6M total frames
• Raw values; use {subset}/{task}/meta/norm_stats.json for per-task stats

Loading

from lerobot.datasets import LeRobotDataset

ds = LeRobotDataset("GT-111/robotwin_v3_eef/clean/adjust_bottle-demo_clean_collect_200-50")

Acknowledgments

Original dataset: robbyant/robotwin-clean-and-aug-lerobot (CC BY-NC-SA 4.0).