Talk Abstract: Rapid advances in content capture and generative AI are dramatically simplifying the process of creating rich virtual worlds. But a virtual world is lifeless without compelling actors. The ability to populate a world with agents that can perform highly skilled behaviors or engage in complex problem-solving activities stands to have broad impact: it will increase the richness of simulated worlds and improve their utility as proxies for real-world situations. In this talk, I’ll focus on recent efforts creating virtual characters that perform skilled tasks in virtual environments. I’ll describe methods that learn skilled behaviors from demonstrations of large video datasets (e.g., hitting a tennis ball into the court, strategic positioning in competitive games) as well as discuss the design of new world-simulation platforms specifically architected for training AI agents orders of magnitude faster than what is possible with prior off-the-shelf game engine frameworks.

Speaker Biography: Kayvon Fatahalian is an Associate Professor in the Computer Science Department at Stanford University. Kayvon’s current research focuses on the design of systems for high-performance graphics/simulation, the analysis of video at scale, and AI-assisted content creation.