---
pretty_name: OMG-Data
task_categories:
- robotics
tags:
- LeRobot
- humanoid
- motion-generation
---

# OMG-Data

Official LeRobotDataset v3.0 release for [OMG](https://github.com/Tsinghua-MARS-Lab/OMG).
It contains Unitree G1 reference motion with text, audio, human reference conditioning.

## Dataset

- Episodes: 798181
- Frames: 121171939
- FPS: 30
- Robot: Unitree G1, 29 actuated joints

- `train`: episodes 0:636865
- `val`: episodes 636865:716840
- `test`: episodes 716840:798181

`observation.state` is the 36D G1 `qpos` at frame `t`. `action` is the target
36D G1 `qpos` at frame `t+1`; it is a reference-motion target, not a low-level
motor command. Text conditions use the standard LeRobot task table. Optional
aligned conditions are stored in `omg.audio.feature` and
`omg.humanref.motion`; their per-frame masks are `omg.condition.has_audio` and
`omg.condition.has_humanref`.

## Loading

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("THU-MARS/OMG-Data")
sample = dataset[0]
```

OMG training converts these source episodes into fixed windows and the 125D
motion representation. See the repository data documentation for direct and
materialized training commands.

## Terms

OMG-Data aggregates motions derived from multiple research datasets. Original
source licenses and usage terms continue to apply to their corresponding
motions; users must review those terms before redistribution or commercial use.