Abstract

Intelligence arises not only from seeing the world, but also from interacting with it. Through this continuous observation and interaction, humans develop internal world models that predict how their actions shape the environment, and rely on these models to guide everyday decisions. This talk explores how such models can be learned and utilized by egocentric 3D agents—AI systems that perceive from a first-person view, act through 3D body motions, and plan by simulating the consequences of their actions. Using EgoAgent (ICCV 2025) as an example, I will show how visual representation learning, motion prediction, and world modeling can be unified within a clean and mutually reinforcing framework. I will conclude by discussing future steps toward richer spatial understanding and more reliable autonomous agents.

Bio: Lu Chen is a Ph.D. student at the State Key Laboratory of CAD&CG, Zhejiang University, advised by Prof. Xiaowei Zhou. His research interests lie in computer vision and visualization, with a focus on world models and human-centric learning. He received his B.Eng. degree from Shandong University, where he was awarded the Taishan Honors College Dean’s Award for Excellence.