About the Talk “Occlusion Games: Strategically maintaining safety in partially occluded multiagent environments”: Robot motion planning in dense multiagent environments—such as autonomous driving—often requires handling the possibility of previously undetected agents emerging from occluded regions and threatening the robot’s safety. Computing collision-free motion strategies among unknown agents potentially lurking beyond the robot’s field of view is challenging for real-time decision-making: probabilistic predictions can be blindsided by low-probability but high-criticality events, whereas worst-case planning methods that compute the forward-reachable set (under specified control bounds) of these potential unseen agents often lead to prohibitively conservative motion and even deadlock situations. In this talk, we will flesh out the fundamental structure of this decision problem, characterizing it as a dynamic pursuit-evasion game that conveniently admits an exact decomposition into two phases: first, the evader must prevent the hidden pursuer from sneaking up on it undetected; if/when the pursuer is detected by the evader’s sensors, the game transitions to a traditional (perfect state feedback information) pursuit-evasion phase. This analysis yields a simple but powerful insight: the robot can safely traverse its partially occluded environment if and only if any hidden agent is always detected before it becomes unavoidable. We introduce the new notion of forward-hidden set, a smaller forward-reachable set that only contains the future states that a potential unseen agent could reach without first being detected. This enables a tractable methodology to compute real-time motion plans that account for the robot’s future ability to detect and avoid any currently unseen agents, bringing together game theory and active perception. Our simulation results in CARLA demonstrate assertive and invariably collision-free autonomous vehicle motion in various driving scenarios, including truck overtaking and intersection traversal. We will end by discussing possible extensions of the framework to a broader family of operational design domains.
About the Speaker: Jaime Fernández Fisac is an Assistant Professor of Electrical and Computer Engineering at Princeton University, where he directs the Safe Robotics Laboratory http://saferobotics.princeton.edu/. His research combines control theory with machine learning and artificial intelligence to enable robots to operate safely in the physical world and the human space. Prior to joining the Princeton faculty, he was a Research Scientist at Waymo (formerly Google’s Self-Driving Car project) from 2019 to 2020, working on autonomous vehicle safety and interaction. Jaime received his Engineering Diploma from the Universidad Politécnica de Madrid, Spain in 2012, his M.Sc. in Aeronautics from Cranfield University, U.K. in 2013, and his Ph.D. in Electrical Engineering and Computer Sciences from the University of California, Berkeley in 2019. He is a recipient of the La Caixa Foundation Fellowship, the Leon O. Chua Award, and the Google Research Scholar Award.