PhysWorld

Robot Learning from a Physical World Model

Google DeepMind · USC · Stanford · Toyota Research Institute

TL;DR: PhysWorld unifies video generation and real-to-sim world modeling for zero-shot robotic manipulation.

From Video Generation to Robotic Manipulation

Video Generation

Robotic Manipulation

Real-to-Sim World Modeling from a Single Video

Learning with the Physical World Model

Citation

@inproceedings{physworld,
  title={Robot Learning from a Physical World Model},
  author = {Mao, Jiageng and He, Sicheng and Wu, Hao-Ning and You, Yang and Sun, Shuyang and Wang, Zhicheng and Bao, Yanan and Chen, Huizhong and Guibas, Leonidas and Guizilini, Vitor and Zhou, Howard and Wang, Yue},
  year={2025},
}