Competence-aware Planning and Control

Hamidreza Modares
Assistant Professor
Michigan State University

Abstract: Existing motion planning approaches typically separate the high-level planning from the low-level control. That is, the high-level planner does not account for the competence of the low-level controller in following its plans, given uncertainties on the system’s dynamics model and its surrounding environments. Even though closed-loop motion planners account for the existence of kinematically feasible or kinodynamically feasible trajectories, they typically assume the availability of high-fidelity robot dynamics or a rich data set and do not account for the destructive conflicts that might arise between the robot’s safety and performance requirements.
While it is natural to leverage reinforcement learning (RL) to deal with uncertainties, imposing safety constraints on RL during design time is a daunting challenge. In this talk, safe RL algorithms will be discussed that can co-learn performance and safety specifications to proactively resolve conflicts. To reduce the data requirements for learning, a physics-informed learning approach will learn control policies that conform with both current data samples as well as the available prior knowledge. A sampling-based planner is then designed to account for the RL’s competence in delivering acceptable behaviors.

Bio: Hamidreza Modares received a B.S. degree from the University of Tehran, Tehran, Iran, in 2004, an M.S. degree from the Shahrood University of Technology, Shahrood, Iran, in 2006, and the Ph.D. degree from the University of Texas at Arlington, Arlington, TX, USA, in 2015, all in Electrical Engineering. He is currently an Associate Professor in the Department of Mechanical Engineering at Michigan State University. Prior to joining Michigan State University, he was an Assistant professor in the Department of Electrical Engineering at Missouri University of Science and Technology. His current research interests include reinforcement learning, safe control, machine learning in control, distributed control of multi-agent systems, and robotics. During the past five years, he has served as an Associate Editor for IEEE Transactions on Neural Networks and Learning Systems, Neurocompiting, and IEEE Transactions on Systems, Man, and Cybernetics: Systems.