GT-111/libero_v3_eef

LIBERO-90 v3 EEF (LeRobot v3.0)

LIBERO-90 benchmark dataset converted to LeRobot v3.0 format with lossless H.264 video (CRF=0, yuv444p).

Format

• LeRobot v3.0 — single dataset with parquet data files + MP4 video
• 8D EEF: xyz + quaternion (wxyz, scalar-first) + gripper
• State: absolute pose (axis-angle → quaternion)
• Action: delta position + delta quaternion (rotvec → quaternion) + absolute gripper
• 2 cameras: front view (observation.images.image) + wrist view (observation.images.wrist_image)
• 40 language-conditioned manipulation tasks, 1,693 episodes, ~273K frames
• Raw values (no normalization); use meta/norm_stats.json for statistics

Statistics

Metric	Value
Episodes	1,693
Total frames	273,465
Tasks	40
Robot	Franka Panda
FPS	10
Camera resolution	256×256

Loading

from lerobot.datasets import LeRobotDataset

ds = LeRobotDataset("GT-111/libero_v3_eef")
# ds["observation.state"]  → absolute EEF pose [x, y, z, qw, qx, qy, qz, gripper]
# ds["action"]             → delta position + absolute quat [dx, dy, dz, dqw, dqx, dqy, dqz, gripper]

Acknowledgments

Original dataset: Physical Intelligence (CC-BY-4.0).

@article{liu2023libero,
  title={LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
  author={Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2023}
}