Scaling 4D World Models for Physical AI: Balancing Physics and Data

abstract:

4D world models seek to simulate how objects interact across space and time, providing a foundation for Physical AI. Two complementary paradigms have emerged. Physics-based simulators offer accuracy and interpretability but remain computationally expensive and difficult to adapt to the real world. Purely data-driven AI approaches promise unprecedented scalability, yet foundation models capable of faithfully modeling the physical world remain elusive.

In this talk, I argue that these approaches are more alike than they appear. Ultimately, all simulation methods are data-driven—the key difference lies in whether knowledge is encoded through manually derived physical laws or learned directly from data. From this perspective, I will review the evolution of data-driven physics simulation and discuss how combining physics with modern AI enables scalable, accurate, and adaptable 4D world models. Finally, I will showcase applications of this unified approach in computer animation, robotics, advanced manufacturing, and protein engineering.

bio:

Peter Yichen Chen is the Director of the UBC PhysAI Lab, where he is an Assistant Professor of Computer Science and Mechanical Engineering and a Green College Leading Scholar. Prior to joining UBC, he was a researcher at Meta and MIT. His interdisciplinary work spans computer graphics, machine learning, scientific computing, and robotics. His research has been recognized with multiple SIGGRAPH Paper Awards and, most recently, the SIGGRAPH/Eurographics SCA Early Career Researcher Award. It has also received research support from NVIDIA, Meta, and Amazon.