Games and Filters: A Road to Safe Intelligence

Watch a 49-minute seminar from MIT featuring Princeton Assistant Professor Jaime Fernández Fisac exploring how to develop safer autonomous systems and AI through game theory and control principles. Learn about game-theoretic reinforcement learning approaches that create robust safety filters for complex robotics challenges like legged locomotion and urban driving. Discover methods for reducing conservativeness while maintaining safety by considering how players' beliefs evolve during interactions. Examine emerging research on using generative AI's self-querying capabilities to better understand uncertainty, identify hazards, and predict action consequences. Gain insights from Fisac's experience as both an academic researcher and former Waymo scientist on developing trustworthy safety assurances for human-AI systems. The talk concludes with a forward-looking perspective on creating general safety filters to guide human-AI interactions toward beneficial outcomes.