🎆 2026.Mar.10, we have released the code and data now, please have a try!

🎆 2026.Feb.22, EmbodMocap has been accepted to CVPR2026, codes and data will be released soon.

For new users, follow this order:

1. Main Pipeline - Quick downloads, preview / visualization, running the pipeline, and step-by-step workflow notes

   • docs/embod_mocap.md

2. Installation - Set up the environment, core dependencies, and manual download references

   • docs/install.md

3. Visualization - Generate rendered videos or inspect scenes and motions interactively with Viser

   • docs/visualization.md

Notes:

• Compared to the paper version, the open-source release replaces PromptDA with LingbotDepth.
• fast is mainly for users who only care about mesh + motion for embodied tasks.
• standard is for users who also need RGBD/mask assets for training reconstruction models.
• We provide an interactive visualization tool based on Viser - give it a try!

Our Viser-based visualization tool allows you to interactively browse scenes, sequences, and SMPL motions in 3D:

Features:

• Switch between multiple scenes and sequences
• Interactive 3D viewing of scene mesh and SMPL motion
• Real-time camera trajectory visualization
• Frame-by-frame playback control

See docs/visualization.md for detailed usage.

If you find this project useful in your research, please consider citing us:

@inproceedings{wang2026embodmocap,
title = {EmbodMocap: In-the-Wild 4D Human-Scene Reconstruction for Embodied Agents.},
booktitle = {CVPR},
author = {Wang, Wenjia and Pan, Liang and Pi, Huaijin and Lou, Yuke and Ren, Xuqian and Wu, Yifan and Liao, Zhouyingcheng and Yang, Lei, Dabral, Rishabh and Theobalt, Christian and Komura, Taku},
year = {2026}
}

We acknowledge VGGT, TRAM, ViTPose, Lang-Segment-Anything, PromptDA, Lingbot-Depth, SAM, COLMAP for their awesome codes.

Feel free to contact me for other questions or cooperation: [email protected]