Games and Filters: A Road to Safe Intelligence
Speaker
Jaime Fernández Fisac
Princeton University
Host
Anoopkumar Sonar
MIT CSAIL
Abstract:
Despite their growing sophistication, autonomous systems still struggle to operate safely in uncertain, open-world situations—as highlighted by public skepticism toward early automated driving technologies. Meanwhile, the excitement around generative AI has been tempered by concerns about potential harms from poorly understood human–AI interactions, where existing guardrails often obscure rather than remove underlying pitfalls. Comprehensive safety assurances remain elusive in both domains—but could insights from one catalyze breakthroughs in the other?
This talk will demonstrate how bridging AI’s learned representations and control theory’s safety principles lays a strong common foundation for certifiable intelligent systems. First, we will explore how game-theoretic reinforcement learning synthesizes robust safety filters with clear-cut guarantees for robotics problems beyond the reach of model-based methods, from legged locomotion to urban driving. Next, we will discuss the value of closing the safety–learning loop by accounting for players’ evolving beliefs during interactions, reducing conservativeness without compromising safety. Finally, we will review early evidence that generative AI systems can use introspective self-queries to refine situational uncertainty, identify novel hazards, and anticipate the future consequences of their actions on users, with strong implications on AI alignment. The talk will end with a vision for general human–AI safety filters that monitor interactions and proactively steer them towards safe and beneficial outcomes.
Bio:
Jaime Fernández Fisac is an Assistant Professor of Electrical and Computer Engineering at Princeton University, where he directs the Safe Robotics Laboratory and co-directs Princeton AI4ALL. His research integrates control systems, game theory, and artificial intelligence to equip robots with transparent safety assurances that users and the public can trust. Before joining Princeton, he was a Research Scientist at Waymo, where he pioneered new approaches to interaction planning that continue to shape how autonomous vehicles share the road today. He is also the co-founder of Vault Robotics, a startup developing agile delivery robots that work alongside human drivers. Prof. Fisac holds an Engineering Degree from Universidad Politécnica de Madrid, a Master’s in Aeronautics from Cranfield University, and a Ph.D. in Electrical Engineering and Computer Sciences from the University of California, Berkeley. His work has been featured in The Wall Street Journal and WIRED, and recognized with the Google Research Scholar Award and the NSF CAREER Award.
Despite their growing sophistication, autonomous systems still struggle to operate safely in uncertain, open-world situations—as highlighted by public skepticism toward early automated driving technologies. Meanwhile, the excitement around generative AI has been tempered by concerns about potential harms from poorly understood human–AI interactions, where existing guardrails often obscure rather than remove underlying pitfalls. Comprehensive safety assurances remain elusive in both domains—but could insights from one catalyze breakthroughs in the other?
This talk will demonstrate how bridging AI’s learned representations and control theory’s safety principles lays a strong common foundation for certifiable intelligent systems. First, we will explore how game-theoretic reinforcement learning synthesizes robust safety filters with clear-cut guarantees for robotics problems beyond the reach of model-based methods, from legged locomotion to urban driving. Next, we will discuss the value of closing the safety–learning loop by accounting for players’ evolving beliefs during interactions, reducing conservativeness without compromising safety. Finally, we will review early evidence that generative AI systems can use introspective self-queries to refine situational uncertainty, identify novel hazards, and anticipate the future consequences of their actions on users, with strong implications on AI alignment. The talk will end with a vision for general human–AI safety filters that monitor interactions and proactively steer them towards safe and beneficial outcomes.
Bio:
Jaime Fernández Fisac is an Assistant Professor of Electrical and Computer Engineering at Princeton University, where he directs the Safe Robotics Laboratory and co-directs Princeton AI4ALL. His research integrates control systems, game theory, and artificial intelligence to equip robots with transparent safety assurances that users and the public can trust. Before joining Princeton, he was a Research Scientist at Waymo, where he pioneered new approaches to interaction planning that continue to shape how autonomous vehicles share the road today. He is also the co-founder of Vault Robotics, a startup developing agile delivery robots that work alongside human drivers. Prof. Fisac holds an Engineering Degree from Universidad Politécnica de Madrid, a Master’s in Aeronautics from Cranfield University, and a Ph.D. in Electrical Engineering and Computer Sciences from the University of California, Berkeley. His work has been featured in The Wall Street Journal and WIRED, and recognized with the Google Research Scholar Award and the NSF CAREER Award.