April 29, 2022, zoom / ESB 2001

Optimal control has become the standard framework for developing autonomous decision making and control algorithms used in engineering, economics, and operations research. Despite their widespread use, these algorithms are sensitive to variations in the underlying dynamical model – whether used explicitly or not – resulting in suboptimal or unstable performance. This talk will present a new approach that stably combines online learning with optimal control for general uncertain nonlinear dynamical systems. Central to the approach is online adjustment of the learning rate, a key mechanism that allows one to safely combine learning with optimal feedback policies computed for a family of dynamical models. The implications of the developed framework in reinforcement learning, sim-to-real, transfer learning, and general optimal control will also be discussed.

Brett Lopez is an Assistant Professor in the Mechanical and Aerospace Engineering Department at UCLA. Prior to joining UCLA, he was a Postdoctoral Scholar at the NASA Jet Propulsion Laboratory in the Robotic Aerial Mobility Group where he led a team of engineers and researchers designing the next generation of autonomous aerial robots for the DARPA Subterranean Challenge. He obtained his Ph.D. (2019) and S.M. (2016) from MIT with a specialization in robotics and controls. He obtained his B.S. (2014) in Aerospace Engineering from UCLA where he received the Aerospace Engineering Outstanding Bachelor of Science award. Dr. Lopez’s research focuses on mobile robotics, control theory, and learning, with a particular emphasis on developing algorithms that possess strong performance guarantees despite the presence of various real-world uncertainties.