This repo is the official implementation of Video2Act: A Dual-System Video Diffusion Policy with Robotic Spatio-Motional Modeling.

Pipeline

News

• [2025.12.2] Code is coming soon! Stay tuned for updates.

Overview

Video2Act is a dual-system video diffusion policy framework that combines robotic spatio-motional modeling for advanced manipulation tasks. The framework integrates:

• Spatial and Motion Processing: Spatial and Motion feature extraction from VDM
• Policy Learning: Diffusion-based transformer architecture for action generation
• RoboTwin Integration: Evaluation

Getting Started

Code is coming soon!

We are currently finalizing the codebase and will release:

• Complete training and evaluation code
• RoboTwin evaluation tools
• Detailed documentation and tutorials

Citation

If you find this work useful, please consider citing:

@misc{jia2025video2actdualsystemvideodiffusion,
      title={Video2Act: A Dual-System Video Diffusion Policy with Robotic Spatio-Motional Modeling}, 
      author={Yueru Jia and Jiaming Liu and Shengbang Liu and Rui Zhou and Wanhe Yu and Yuyang Yan and Xiaowei Chi and Yandong Guo and Boxin Shi and Shanghang Zhang},
      year={2025},
      eprint={2512.03044},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2512.03044}, 
}

License

This project will be released under the MIT License.