GT-111/libero-90-v3

LIBERO-90 (LeRobot v3)

This is a LeRobot v3 format conversion of the LIBERO-90 benchmark dataset, with canonical 7D end-effector (EE) command-space action representation for multi-dataset robot learning.

Original Dataset

LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning

Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, Peter Stone. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. NeurIPS 2023 (Datasets and Benchmarks Track).

• Original Source: https://huggingface.co/datasets/physical-intelligence/libero
• Project Website: https://libero-project.github.io
• arXiv: https://arxiv.org/abs/2306.03310

LIBERO is a benchmark for lifelong robot learning with 130 procedurally generated manipulation tasks across 4 task suites. The dataset provides high-quality human-teleoperated demonstrations using a Franka Emika Panda robot arm.

License

This dataset is released under CC-BY-4.0, consistent with the original LIBERO dataset license.

Conversion Details

What we changed

1. Canonical EE Command-Space Actions: LIBERO actions are treated as absolute EE pose commands (position delta + quaternion + continuous gripper) rather than joint-space velocities. The 7D action vector is:

   • [dx, dy, dz, droll, dpitch, dyaw, gripper]

2. Gripper Canonicalization: Raw LIBERO stores gripper state as a 2-element vector (finger positions). We canonicalized this to a single continuous [0, 1] gripper openness value.

3. LeRobot v3 Format: Converted to the latest LeRobot v3 dataset layout with sharded video storage (MP4) and Parquet-based frame data.

4. Task-Based Prompts: Each frame is annotated with its task description for language-conditioned training.

What we preserved

• Both camera views: front camera (observation.images.image) and wrist camera (observation.images.wrist_image)
• Original episode boundaries and task assignments
• Frame-level timestamps

Dataset Structure

├── data/
│   └── chunk-*/file-*.parquet  # Frame data (7D action + 8D state + image paths)
├── videos/
│   └── observation.images.image/chunk-*/file-*.mp4
│   └── observation.images.wrist_image/chunk-*/file-*.mp4
├── meta/
│   ├── info.json                # Dataset metadata
│   ├── tasks.json               # Task vocabulary
│   ├── episodes.jsonl           # Episode index
│   └── stats.json               # Normalization statistics
└── norm_stats.json              # Precomputed quantile normalization stats

Statistics

Metric	Value
Total Episodes	3,921
Total Frames	569,249
Robot Type	Franka Emika Panda
State Dim	8
Action Dim	7

Usage

from lerobot.datasets import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="GT-111/libero-90-v3",
)

For training with the LWM-VLA / OpenPI framework:

from openpi.training.config import MultiDatasetPretrainDatasetSpec

MultiDatasetPretrainDatasetSpec(
    repo_id="GT-111/libero-90-v3",
    dataset_type="libero",
    weight=0.1,
)

Citation

If you use this dataset, please cite the original LIBERO paper:

@inproceedings{liu2024libero,
  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},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  volume={36},
  year={2024},
  note={Datasets and Benchmarks Track}
}

Version History

• v3.0 (current): LeRobot v3 conversion with canonical 7D EE command-space actions