When: Thursday 17th of March, 1PM AEDT
Where: The talk will be partially presented at the Rose Street Seminar area (J04) and partially online via Zoom, RSVP here.
Speaker: Nicholas Barbara
Title: Learning stabilising feedback policies with controller augmentation: the road to online learning
Complex robotic systems are difficult to control across a wide range of tasks. Their restricted versatility motivates the need for learning-based controllers that allow a robot to adapt to new environments during operation. However, most online learning architectures do not have the stability guarantees required for use in safety-critical robotic systems. This seminar presents a promising new method for learning stabilising, nonlinear feedback controllers using controller augmentation, which we expect to be well-suited to the online learning problem. We combine the Youla-Kucera parameterisation with Recurrent Equilibrium Networks to construct a controller architecture with natural closed-loop stability guarantees for linear systems. Together with a reinforcement learning algorithm based on random search, we demonstrate that our architecture facilitates efficient learning and excellent controller performance on a range of partially-observed linear systems. We further discuss extensions to nonlinear systems and real-world experiments, with the aim of applying our methods to online learning in practical robotic systems.
Nicholas Barbara received his Bachelor’s degree in Aerospace Engineering (Honours), Physics, and Applied Mathematics at the University of Sydney in 2020. He is currently a Ph.D. student at the Australian Centre for Field Robotics working on robust reinforcement learning and its application to robotic systems. His research interests include optimal control, online learning, and complex systems.