MOSAIC Model
Project Page | Paper | Code | Dataset | Model
This repository releases deployment-ready ONNX models for MOSAIC, introduced in:
In MOSAIC, a general motion tracker is trained in simulation, and interface-specific adaptation is handled via a lightweight residual adaptor that injects action-space corrections while preserving the general policy’s capabilities.
It includes models:
- gmt.onnx: the general motion tracking policy
- noitom_teleop.onnx: the adaptor policy for teleoperation using Noitom inertial mocap suit
- pico_teleop.onnx: the adaptor policy for teleoperation using PICO VR device
How to download
Option A: Download a single ONNX file
from huggingface_hub import hf_hub_download
onnx_path = hf_hub_download(
repo_id="BAAI-Humanoid/MOSAIC_Model",
filename="pico_teleop.onnx", # or "noitom_teleop.onnx"
)
print("Downloaded to:", onnx_path)
Option B: Download all files in this model repo
from huggingface_hub import snapshot_download
local_dir = snapshot_download(
repo_id="BAAI-Humanoid/MOSAIC_Model",
)
print("Downloaded to:", local_dir)
Usage
For constructing the correct deployable observations and mapping model outputs to Unitree G1 control targets, please use the official MOSAIC codebase and RobotBridge Deployment framework.
Citation
If you use these models for your research, please cite our paper:
@article{sun2026mosaic,
title = {MOSAIC: Bridging the Sim-to-Real Gap in Generalist Humanoid Motion Tracking and Teleoperation with Rapid Residual Adaptation},
author = {Zhenguo Sun and Bo-Sheng Huang and Yibo Peng and Xukun Li and Jingyu Ma and Yu Sun and Zhe Li and Haojun Jiang and Biao Gao and Zhenshan Bing and Xinlong Wang and Alois Knoll},
journal = {arXiv preprint arXiv:2602.08594},
year = {2026}
}
License
This dataset is released under Apache-2.0.