Grad Seminar: Safe and Robust Reinforcement Learning for Autonomous Vehicle Control

Abstract

Achieving safe and robust reinforcement learning (RL) for autonomous vehicle (AV) control requires a multi-faceted approach. This seminar presents Safe RL, where we integrate a high-fidelity simulation framework to enhance realism, apply transfer learning to improve data efficiency and training stability, and enforce safety using a Control Barrier Function (CBF) with a risk signal. We conduct stability analysis of our RL-CBF adaptive cruise control system, focusing on throttle and brake control under safety constraints. In Robust RL, we investigate adversarial threats in DRL-based AVs, analyzing existing attack strategies and introducing the novel Optimism Induction Attack (OIA), which manipulates state perception to bypass CBF safeguards. Our findings provide key insights into improving the safety, stability, and resilience of RL-driven AV systems.

Presenter

Saeedeh Lohrasbi, PhD candidate in Systems Design Engineering

Join online

Attending this seminar will count towards the graduate student seminar attendance milestone!