CSE DSI Machine Learning Seminar with Sylvia Herbert (UCSD)
Constructing Neural Safety Filters for Autonomous Systems
Safety filters are attractive as a modular framework for “robustifying” an autonomous system, and have become increasingly popular in an era of black-box nominal control policies. However, constructing valid safety filters is a challenge for many real-world high-dimensional autonomous systems. Learning-based approaches to designing safety filters are one possible solution, though they suffer from learning errors and are particularly sensitive to distribution shifts in online deployment—an aspect that is not ideal for safety.
In this talk, I will highlight two of the approaches we have taken to improve outcomes for neural safety filters:
- For lower-dimensional systems (e.g. 4D-6D), we have introduced tools for refining neural safety filters after training. This takes advantage of the scalability of learning-based approaches with the safety guarantees of control-theoretic methods. Additionally, refinement allows for online adaptation in the face of novel information.
- For higher-dimensional systems (7D-50D+) we show how semi-supervised learning using techniques from applied math and control theory can better guide the learning process. This work was recently nominated for best paper at the upcoming Learning for Dynamics and Control (L4DC) conference. Additionally, by parameterizing environmental conditions, we can show adaptability online.
Sylvia Herbert is an Assistant Professor of Mechanical and Aerospace Engineering at the University of California, San Diego. Her research focus is to enable efficient and safe decision-making in robots and other complex autonomous systems while reasoning about uncertainty in real-world environments and human interactions. These techniques are backed by both rigorous theory and physical testing on robotic platforms.
Prior to joining UCSD, Professor Herbert received her PhD in Electrical Engineering from UC Berkeley, where she studied with Professor Claire Tomlin on safe and efficient control of autonomous systems. Before that, she earned her BS/MS at Drexel University in Mechanical Engineering. She is the recipient of the ONR Young Investigator Award, 2023 IROS Robocup Best Paper Award, 2025 Learning for Dynamics and Control (L4DC) Best Paper Nomination, Hellman Fellowship, UCSD JSOE Early Career Faculty Award, UC Berkeley Chancellor’s Fellowship, NSF GRFP, UC Berkeley Outstanding Graduate Student Instructor Award, and the Berkeley EECS Demetri Angelakos Memorial Achievement Award for Altruism.
On Zoom, but can be viewed in Keller 3-180.