Guest lectures
Prediction: Dr. Boris Ivanovic
Title: Behavior Prediction: A Nucleus of Modern AV Research
Abstract: Research on behavior prediction, the task of predicting the future motion of agents, has had an outsized impact on multiple aspects of autonomous vehicles (AVs). From direct improvements in online driving performance to deeper connections between AV stack modules to enabling closed-loop training and evaluation in simulation with intelligent reactive agents, behavior prediction has served as a nucleus for much of modern AV research. In this lecture, I will discuss recent advancements along each of these directions, covering modern approaches for behavior prediction, generalization to unseen environments, tighter integrations of AV stack components (towards end-to-end AV architectures), and methods for simulating the behaviors of agents. Finally, I will outline some open research problems in modeling human motion and their potential impacts on downstream driving performance.
Bio: Boris is currently a Senior Research Scientist and Manager in NVIDIA’s Autonomous Vehicle Research Group. His research interests include novel end-to-end AV architectures, sensor and traffic simulation, AI safety, and the thoughtful integration of foundation models in AV development. Prior to joining NVIDIA, he received his Ph.D. in Aeronautics and Astronautics under the supervision of Marco Pavone in 2021 and an M.S. in Computer Science in 2018, both from Stanford University. He received his B.A.Sc. in Engineering Science from the University of Toronto in 2016.
Planning: Dr. David Fridovich-Keil
Title: Auto-Encoding Bayesian Inverse Games
Abstract: This talk will present our recent work on inverting noncooperative games. In these problems, we aim to identify hidden parameters of agents’ objective functions and/or constraints based upon observations of their actions. Existing approaches are limited to reconstructing point estimates of these hidden parameters; in this work, however, we develop a variational inference framework for estimating the full Bayesian posterior. We will see how a motion planner can then make use of this additional information to generate safer and more efficient trajectories in human-robot interactions.
Bio: David Fridovich-Keil is an assistant professor at the University of Texas at Austin. David’s research spans optimal control, dynamic game theory, learning for control, and robot safety. While he has also worked on problems in distributed control, reinforcement learning, and active search, he is currently investigating the role of dynamic game theory in multi-agent interactive settings such as traffic. David’s work also focuses on the interplay between machine learning and classical ideas from robust, adaptive, and geometric control theory. David is the recipient of an NSF Graduate Research Fellowship and an NSF CAREER award.
Safe control: Dr. Sylvia Herbert
Title: Blending methods to generate safe controllers: combining data-driven control barrier function approximations with Hamilton-Jacobi reachability analysis
Abstract: While interviewing for faculty positions I was repeatedly asked how the tool I often use for generating safety filters for controllers (Hamilton-Jacobi reachability analysis) compares to the very popular control barrier function (CBF) method. I then spent the first couple years as a faculty member exploring the answer to this, supported by an ONR YIP. In this talk I will discuss: The relationship between these two methods How we can use reachability analysis to refine CBF approximations to provide safety guarantees How we can generalize these safety functions for hard-to-model dynamics (e.g. interaction behavior among pedestrians)
Bio: Sylvia Herbert joined the UCSD MAE department as an Assistant Professor after graduating with her PhD from UC Berkeley in 2020. She now has a research group of 7 PhD students along with several MS and undergraduate students. These students work on a range of projects, including:
- Connections between Hamilton-Jacobi reachability and CBFs (this talk) [links: 1, 2, 3, 4]
- More generally blending different computational methods and theory for HJ reachability [links: 1, 2]
- Koopman-Hopf reachability analysis for very efficient approximate reachable sets and control policies link
- Reasoning about safety when you are forced to interact with an uncertain stochastic environment (e.g. pushing around debris to reach a goal) link
- Control theoretic methods applied to learning frameworks (e.g. when fractal landscapes occur in policy gradient methods) link
- Safe reinforcement learning [1, 2] For more on these topics please see our group website. For questions/comments, Professor Herbert can be reached at sherbert@ucsd.edu