Mechanical and Civil Engineering Seminar

Title: Safety Filters for Autonomous Systems: When They Misbehave and How to Fix Them

Abstract: Control Barrier Function (CBF)-based safety filters have become ubiquitous in modern safety-critical control, and are being deployed across an expanding range of autonomous systems, ranging from humanoid robots, self-driving vehicles, and aerospace systems. Their appeal is clear: given any nominal controller, possibly precomputed to optimize some performance metric, a safety filter provides a principled, minimally invasive correction that provably keeps the system within a desired safe set. Their simplicity, generality, and modularity has driven their rapid and widespread adoption. Despite this widespread use, in this talk we argue that their dynamical behavior is still far from being well-understood. We provide a variety of examples showing how closed-loop dynamics induced by CBF-based safety filters are far richer (and possibly far more dangerous) than their design intent suggests. Those include examples with unbounded trajectories, limit cycles, and undesired equilibria that can even be locally stable. These pathologies are not edge cases: we show that they arise for broad classes of nominal controllers and safe set geometries, including convex ones. Fortunately, for an important class of systems and safe sets (primarily linear dynamics paired with affine or quadratic CBFs) we identify concrete design principles that provably preclude these undesirable behaviors. We hope these results serve as a foundation for extending such guarantees to more general classes of systems, with the ultimate goal of providing safety filters with more rigorous stability and performance guarantees.

Bio: Pol Mestres received the bachelor's degree in mathematics and the bachelor's degree in engineering physics from the Universitat Politècnica de Catalunya, Barcelona, Spain, in 2020, and the master's and Ph.D. degrees in mechanical engineering in 2021 and 2025 respectively from the University of California, San Diego, La Jolla, CA, USA. He is currently a postdoctoral scholar at the California Institute of Technology. His research interests include safety-critical control, motion planning, and reinforcement learning.