Abstract: Large language and generative models have achieved remarkable success by scaling with internet-scale data. In contrast, physical AI, agents that perceive and act in the real world, still lags far behind. Today, both academia and industry largely pursue scalability in physical AI by collecting ever-growing amounts of action–video-paired data and training large networks, such as VLA models. In my view, this approach is heading in the wrong direction. The central challenge is not to scale data for imitation, but to build a reality world in computation, a structured, interactive training ground where agents can perceive, act, and learn.
In this talk, I will discuss how to construct reality world simulators that can effectively scale AI learning in the physical world. I will introduce a three-pronged recipe: 1) Real-world simulation: generating controllable, interactive digital cousins of real environments from monocular videos. 2) Scene-aware human modeling: bringing realistic dynamics into simulated spaces to enhance the safety and social compliance of trained robots. 3) Hybrid world simulation: combining the strengths of graphics-based simulators and neural world models into a unified framework. Finally, I will share several future directions that I believe can move us closer to generalizable physical AI in the real world.
Speaker Bio.: Dr. Wayne Wu is a postdoctoral researcher at UCLA, where he works with Prof. Bolei Zhou. He received his Ph.D. in Computer Science and Technology from Tsinghua University in June 2022 and was previously a visiting Ph.D. student at Nanyang Technological University. He also spent several years in industry, where he led products that reached more than 10 million end users worldwide.
His research lies at the intersection of computer vision, computer graphics, and robotics. He focuses on developing infrastructure to scale up robot learning, specifically building reality world simulators for generalizable physical AI. He was honored with the UCLA Chancellor’s Award for Postdoctoral Research in 2025. One of his works was selected as a Best Paper Candidate at CVPR 2023, and five others were selected as Oral or Spotlight presentations at top AI conferences. He serves as an Area Chair at CVPR 2026.