README.md

---
license: cc-by-4.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- manipulation
- libero
- simulation
- depth
---

# LIBERO with depth (MuJoCo 3.3.2 re-render)

The four standard [LIBERO](https://libero-project.github.io/) benchmark suites,
re-rendered under **MuJoCo 3.3.2** and republished in **LeRobot v2.1** format with
**per-camera metric depth** added alongside the usual RGB. `no_noops` means idle
frames have been stripped.

## Contents

This repo holds **four independent LeRobot datasets**, one per suite:

| suite | episodes | frames | tasks |
|---|---|---|---|
| `libero_10_no_noops_lerobot` | 388 | 104,160 | 10 |
| `libero_goal_no_noops_lerobot` | 433 | 52,895 | 10 |
| `libero_object_no_noops_lerobot` | 457 | 67,309 | 10 |
| `libero_spatial_no_noops_lerobot` | 433 | 53,128 | 10 |
| **total** | **1711** | **277,492** | **40** |

Each subdirectory is a self-contained LeRobot dataset (`meta/`, `data/`, `videos/`).

## Format

| | |
|---|---|
| Robot | `franka` |
| fps | 20 |
| RGB | 512×512, 2 cameras, `.mp4` |
| Depth | 512×512, 2 cameras, FFV1 `.mkv`, uint16 millimetres |
| State / action | 8-D / 7-D |
| LeRobot version | `v2.1` |

Depth decoding is identical to the YAM datasets in this org — see e.g.
[`flex-pi/self_repair_gripper_bc`](flex-pi/self_repair_gripper_bc). Depth is declared as
`dtype: "depth_video"`, so the stock LeRobot loader ignores it.

## Loading

Because this repo contains four datasets, point the loader at a subdirectory rather
than the repo root:

```python
from huggingface_hub import snapshot_download
p = snapshot_download("flex-pi/libero_mujoco3.3.2_depth", repo_type="dataset",
                      allow_patterns="libero_spatial_no_noops_lerobot/*")
```

## Citation

Please cite the original LIBERO benchmark in addition to the Flex-π project.