Sensho/libero-lerobot-v2

LIBERO (LeRobot v2 + MP4)

A re-encoding of physical-intelligence/libero into LeRobot v2.0 with MP4 videos instead of per-frame PNGs, prepared for use with the NVIDIA DreamZero data pipeline.

What changed from physical-intelligence/libero

• Per-frame PNG images encoded into per-episode H.264 MP4 (yuv420p, CRF 18) for decord-compatible video reading.
• Camera keys renamed from image / wrist_image to observation.images.agentview / observation.images.wrist (matching DreamZero's observation.images.* convention).
• Added a per-frame annotation.task string column with the task description.
• info.json features for cameras switched from dtype: "image" to dtype: "video".

State, action, episode counts, and task definitions are unchanged from the source.

Dataset summary

• Robot: Franka Emika Panda (single arm)
• Episodes: 1,693 across 4 LIBERO suites (libero-spatial, libero-object, libero-goal, libero-10)
• Total frames: 273,465
• Tasks: 40
• FPS: 10
• State (8-dim): [eef_pos×3, eef_axis_angle×3, gripper_qpos×2]
• Action (7-dim): [delta_pos×3, delta_axis_angle×3, gripper×1]
• Cameras: agentview (third-person) + wrist (eye-in-hand), 256×256 RGB

Usage (DreamZero post-training)

hf download <YOUR_USER>/libero-lerobot-mp4 \
    --repo-type dataset --local-dir ./data/libero_lerobot/all

# Subset to N demos/task
python subset_lerobot_episodes.py \
    --src ./data/libero_lerobot/all \
    --dst ./data/libero_lerobot/all_10demos \
    --demos-per-task 10 --seed 42

# Add GEAR metadata (in DreamZero repo)
python scripts/data/convert_lerobot_to_gear.py \
    --dataset-path ./data/libero_lerobot/all_10demos \
    --embodiment-tag libero_sim \
    --state-keys '{"eef_pos": [0, 3], "eef_axis_angle": [3, 6], "gripper_qpos": [6, 8]}' \
    --action-keys '{"delta_eef_pos": [0, 3], "delta_eef_axis_angle": [3, 6], "gripper": [6, 7]}' \
    --relative-action-keys delta_eef_pos delta_eef_axis_angle \
    --task-key annotation.task --action-horizon 24 --force

Source attribution

This dataset is a derivative of:

• physical-intelligence/libero — itself a LeRobot conversion of openvla/modified_libero_rlds.
• The original LIBERO benchmark by Liu et al. 2023.

License

CC-BY 4.0, inherited from the source datasets.

Citation

@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={arXiv preprint arXiv:2306.03310},
  year={2023}
}