# MOSAIC: Multi-Source Human Motion Capture and Unitree G1 Retargeted Motions for Humanoid Motion Tracking and Teleoperation (BAAI)

## Facts

- Provider: Beijing Academy of Artificial Intelligence (BAAI) Humanoid team
- License: CDLA-Permissive-2.0 (https://huggingface.co/datasets/BAAI-Humanoid/MOSAIC_Dataset)
- Size: 5,290 files
- Format: NumPy (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/BAAI-Humanoid/MOSAIC_Dataset
- Access: public, readable without an account
- Category: Robotics
- Source: https://arxiv.org/abs/2602.08594
- Page: https://scigantic.com/archive/c1e3d72d-c8de-59b2-8786-30da3e4b952f
- Keywords: Robotics, robot learning, NumPy, Hugging Face

## About

The built-in motion dataset of MOSAIC, a BAAI method for bridging the sim-to-real gap in generalist humanoid motion tracking and teleoperation with rapid residual adaptation. It holds human motions in an AMASS-style format and the same motions retargeted to the Unitree G1 humanoid robot as NPZ files, from three sources: optical motion capture (Vicon), inertial motion capture (IO-AI, with raw MCAP recordings and a conversion script), and motions generated by the GENMO model from text prompts; the same prompts guided the optical capture and are released in motion_prompt/. A small amount of stage-II adaptor training data is added: about 30 minutes of PICO VR teleoperation data and 30 minutes of Noitom inertial mocap. Folders are human/ (optical_mocap, inertial_mocap, generated_genmo by activity such as walk, dance and daily activities) and G1/ with matching retargeted clips. About 4,300 NPZ, 880 PyTorch .pt and 59 MCAP files, about 46 GB (measured 2026-11-21); read with numpy and torch. Released under CDLA-Permissive-2.0 by BAAI (external tools keep their own terms); cite MOSAIC, arXiv:2602.08594.

## Load the data

```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="BAAI-Humanoid/MOSAIC_Dataset", filename="G1/adaptor_data/noitom_inertial_mocap/mocap_motion_bosheng_20260129_232616_smoothed_0_499.npz", repo_type="dataset")
print(path)
```

## Data structure

File types in sampled files: NumPy

Example file: `G1/adaptor_data/noitom_inertial_mocap/mocap_motion_bosheng_20260129_232616_smoothed_0_499.npz`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/c1e3d72d-c8de-59b2-8786-30da3e4b952f
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/c1e3d72d-c8de-59b2-8786-30da3e4b952f
- Over MCP: `get_schema_card` with id `c1e3d72d-c8de-59b2-8786-30da3e4b952f` at https://api.scigantic.com/mcp

## Related topics

- [Robotics datasets](https://scigantic.com/topics/robotics)
- [NumPy datasets](https://scigantic.com/topics/numpy)
- [Robot learning datasets](https://scigantic.com/topics/robot-learning)
