nvidia/LIBERO_LeRobot_v3

LIBERO LeRobot v3

Dataset Summary

nvidia/LIBERO_LeRobot_v3 is a LeRobotDataset v3.0 conversion of the LIBERO robot manipulation benchmark. LIBERO is designed for studying lifelong robot learning and knowledge transfer across language-conditioned manipulation tasks. This repository packages the LIBERO task suites as LeRobot-compatible datasets with Parquet state/action data, MP4 video observations, and LeRobot metadata.

The dataset is organized as five top-level suite folders:

• libero_spatial
• libero_object
• libero_goal
• libero_90
• libero_10

Each suite is stored independently in LeRobotDataset v3.0 format.

Dataset Details

Dataset Description

LIBERO, short for LIfelong learning BEnchmark on RObot manipulation, is a benchmark for studying lifelong learning in decision-making and robot manipulation. It focuses on knowledge transfer across tasks involving spatial relations, object manipulation, goal changes, and long-horizon behavior.

This version converts LIBERO demonstration data to the LeRobotDataset v3.0 standard. LeRobot v3.0 uses file-based Parquet and MP4 shards plus structured metadata, replacing per-episode files with chunked files that can contain multiple episodes.

This dataset is ready for commercial or non-commercial uses.

Dataset Owner(s)

NVIDIA Corporation

Dataset Creation Date

2026-05-26

Dataset Sources

• Repository: https://huggingface.co/datasets/nvidia/LIBERO_LeRobot_v3
• LIBERO project: https://libero-project.github.io/
• LIBERO paper: https://arxiv.org/abs/2306.03310
• LIBERO GitHub: https://github.com/Lifelong-Robot-Learning/LIBERO
• LeRobotDataset v3.0 documentation: https://huggingface.co/docs/lerobot/lerobot-dataset-v3

Version:

v1.0

License/Terms of Use

This dataset is released under the OpenMDW1.1

Intended Usage

This dataset is intended for research and development in:

• imitation learning
• robot policy learning
• vision-language-action model training
• lifelong robot learning
• knowledge transfer evaluation
• language-conditioned manipulation
• LeRobot-compatible data loading, streaming, and training workflows

Out-of-Scope Use

This dataset is not intended to be used as the sole validation source for safety-critical robot deployment. Policies trained on this dataset should be evaluated carefully before use in real-world physical systems.

Dataset Characterization

Data Collection Method

[Not Applicable]

Labeling Method

[Not Applicable]

Dataset Format

The repository contains one LeRobotDataset v3.0 dataset per suite folder. Each suite follows the LeRobot v3 layout:

<suite>/
├── data/
│   └── chunk-000/
│       └── file-*.parquet
├── meta/
│   ├── info.json
│   ├── stats.json
│   ├── tasks.parquet
│   └── episodes/
│       └── chunk-000/
│           └── file-*.parquet
└── videos/
    ├── observation.images.image/
    │   └── chunk-000/
    │       └── file-*.mp4
    └── observation.images.wrist_image/
        └── chunk-000/
            └── file-*.mp4

Dataset Quantification

The following values are taken from each suite's meta/info.json.

Suite	Episodes	Frames	Task IDs	FPS	Robot
libero_10	379	101,469	10	20	Franka
libero_90	3,921	569,249	73	20	Franka
libero_object	454	66,984	10	20	Franka
libero_spatial	432	52,970	10	20	Franka
libero_goal	428	52,042	10	20	Franka
Total	5,614	842,714	113	20	Franka

Note: the upstream LIBERO task suite naming includes LIBERO-90; this converted repository's libero_90/meta/info.json reports 73 task IDs.

Total Data Storage: 4.5 GB

Features

Common feature fields include:

• action: 7-dimensional robot action vector.
• observation.state: 8-dimensional robot state vector.
• observation.images.image: RGB video observation, 256 x 256, 20 FPS, AV1 MP4.
• observation.images.wrist_image: RGB wrist-camera video observation, 256 x 256, 20 FPS, AV1 MP4.
• timestamp, frame_index, episode_index, task_index, and index.

The libero_90 suite additionally includes:

• observation.states.ee_state
• observation.states.gripper_state
• observation.states.joint_state

File Format

• Tabular robot state/action data: Apache Parquet.
• Episode metadata: chunked Parquet under meta/episodes/.
• Task metadata: meta/tasks.parquet.
• Video observations: MP4 files under videos/.
• Dataset schema and statistics: meta/info.json and meta/stats.json.

Loading

Because this repository contains multiple LeRobot datasets under top-level suite folders, select the suite you want to use and point your local LeRobot workflow at that suite directory.

Example: download a single suite from the Hub.

from huggingface_hub import snapshot_download
from pathlib import Path

repo_dir = snapshot_download(
    repo_id="nvidia/LIBERO_LeRobot_v3",
    repo_type="dataset",
    allow_patterns=["libero_spatial/**"],
)

suite_dir = Path(repo_dir) / "libero_spatial"
print(suite_dir)

Use suite_dir as the local LeRobot dataset root in workflows that accept a local LeRobotDataset directory.

Dataset Creation

Source Data

The source data is derived from LIBERO, a benchmark for lifelong robot learning and language-conditioned robot manipulation. The LIBERO paper describes high-quality human-teleoperated demonstration data for the benchmark's tasks.

Conversion

This repository converts LIBERO data into LeRobotDataset v3.0. In LeRobot v3.0, data and video frames from multiple episodes are grouped into larger shard files, while metadata records episode boundaries, task IDs, feature schemas, statistics, and path templates.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Citation

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

@inproceedings{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},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023}
}

Please also cite or reference LeRobot if you use LeRobot tooling or dataset loaders.

References

• LIBERO project page: https://libero-project.github.io/
• LIBERO arXiv paper: https://arxiv.org/abs/2306.03310
• LIBERO GitHub repository: https://github.com/Lifelong-Robot-Learning/LIBERO
• Hugging Face LeRobot LIBERO documentation: https://huggingface.co/docs/lerobot/en/libero
• LeRobotDataset v3.0 documentation: https://huggingface.co/docs/lerobot/lerobot-dataset-v3
• Porting large datasets to LeRobot v3.0: https://huggingface.co/docs/lerobot/main/porting_datasets_v3