Enhancing the Safety and Reliability of Robot Learning

Tea Talk, November 14, 2025 Machine learning has become an increasingly popular approach to robot motion planning, as it can generate plans without the need for extensive manual modeling. However, despite its potential, it remains difficult to verify the correctness of learned models. In robotics, even minor planning errors can lead to hardware damage or pose safety risks to nearby humans. Our work investigates ways to improve the reliability of data-driven methods in on physical robotic systems. In this talk, I will present our research on (1) learning policies with stability guarantees, (2) out-of-distribution detection for sim-to-real transfer, and (3) learning viability for gait selection.