Not Dead Yet ? Is Physics-based Animation the Key to World Models?
David I. W. Levin - University of Toronto
Oct. 23, 2026, 3:30 p.m. - Oct. 23, 2026, 4:30 p.m.
ENGMC 11
Hosted by: Paul Kry
"World Model" has become a term synonymous with large-scale machine learning models whose output is a convincing 2D projection of a 3D world. The promise of dynamic, photorealistic, large-scale interactive worlds is compelling, but that promise and the excitement thereof obscure major technical challenges with a purely data-driven, machine learning approach. For most applications targeted by world models in the applied engineering and physical AI spaces, consistency of the model is as important as raw output quality. In this talk, I will try to convince you, via a tour of my own work and computer graphics methods past, that lessons learned from and algorithms developed for physics-based animation, in conjunction with modern machine learning methods, are key to unlocking physics simulation at scale and are a legitimate path to real, physically consistent dynamic 3D worlds.
David I. W. Levin is an Associate Professor of Computer Science at the University of Toronto and a Principal Research Scientist at NVIDIA. His research lies at the intersection of computer graphics, computational mechanics, and machine learning, with a focus on physics-based animation, elastodynamics, contact mechanics, and geometry processing. His lab develops high-performance numerical formulations—spanning mixed finite element methods, neural implicit simulation, and learned physical parameter estimation—to construct physically grounded, scalable world models for physical AI and robotics. His work is regularly published across premier venues in graphics, machine learning, and robotics, including ACM Transactions on Graphics (SIGGRAPH), ICLR, ICML, CVPR, and IROS.
