---
license: apache-2.0
pretty_name: LIBERO-3D
task_categories:
- robotics
tags:
- robotics
- imitation-learning
- reinforcement-learning
- embodied-ai
- computer-vision
- depth
- point-cloud
- npz
size_categories:
- 100K<n<1M
---

# LIBERO-3D

LIBERO-3D is a processed LIBERO demonstration dataset packaged for 3D-aware robot learning. It keeps the original LIBERO task suites while converting each trajectory into per-episode `.npz` files with RGB, depth, camera parameters, segmentation, states, and actions.

This repository is intended as a storage and distribution repo for dataset files. The Hugging Face dataset viewer is not expected to render these files directly because the payload is stored as binary `.npz` episodes rather than tabular data.

## Contents

The repository is organized into four suites:

```text
libero_goal/
libero_object/
libero_spatial/
libero_10/
README.md
```

Each suite contains 10 tasks. Each task contains 50 episodes. In total this repo contains 40 tasks and 2000 episodes.

Each task directory follows this structure:

```text
<suite>/<task_name>/
  metadata.json
  _SUCCESS
  episode_000/data.npz
  episode_001/data.npz
  ...
  episode_049/data.npz
```

## Data Format

Each `data.npz` episode stores trajectory-aligned arrays. The exact keys can be checked from the accompanying `metadata.json`. The processed files include at least the following modalities:

- `state`: robot state, shape `(T, 8)`
- `actions`: control actions, shape `(T, 7)`
- `agentview_rgb`: static camera RGB frames, shape `(T, 256, 256, 3)`
- `agentview_depth`: static camera depth frames, shape `(T, 256, 256)`
- `wrist_rgb`: wrist camera RGB frames, shape `(T, 256, 256, 3)`
- `wrist_depth`: wrist camera depth frames, shape `(T, 256, 256)`
- `wrist_intrinsic`: wrist camera intrinsics, shape `(T, 3, 3)`
- `wrist_extrinsic`: wrist camera extrinsics, shape `(T, 4, 4)`
- `wrist_segmentation`: wrist segmentation maps, shape `(T, 256, 256, 2)`

The task-level `metadata.json` also records:

- suite name and task name
- original source demo path
- image resolution
- state semantics
- static and dynamic camera calibration fields
- per-episode step counts and tensor shapes

## Suite Summary

- `libero_goal`: 10 goal-conditioned manipulation tasks
- `libero_object`: 10 object-centric pick-and-place tasks
- `libero_spatial`: 10 spatial-relation manipulation tasks
- `libero_10`: 10 long-horizon tasks from the LIBERO-10 suite

## Source

This dataset is derived from the LIBERO benchmark datasets and reorganized into processed per-episode `.npz` files for downstream 3D learning pipelines.

- Official LIBERO repository: https://github.com/Lifelong-Robot-Learning/LIBERO
- Official Hugging Face dataset release for the original demo files: https://huggingface.co/datasets/yifengzhu-hf/LIBERO-datasets

## Usage Notes

- This repo is for file hosting and dataset download, not for in-browser preview.
- Keep the directory names stable if downstream code expects suite names such as `libero_goal` or `libero_10`.
- Do not upload macOS metadata files such as `._*` or `.DS_Store`.
- Large-file upload tooling such as `huggingface_hub` or `hf upload-large-folder` is recommended.

## License

This processed release follows the upstream LIBERO dataset license labeling used in the official Hugging Face dataset repository: `apache-2.0`.