DEEP REINFORCEMENT LEARNING FOR ADAPTIVE CONTROL IN ROBOTICS

Abstract

Adaptive control of robotic systems is challenging due to nonlinear system dynamics and time-varying parameters that govern the system’s behavior. Naturally, deep reinforcement learning (DRL) is a suitable choice for the control of non-linear systems that lack reliable physics-based models. However, DRL algorithms require extensive data to learn an optimal control policy. Furthermore, optimal policies struggle to generalize across unknown environments commonly encountered in real world tasks. In this work, I propose a set of DRL-based architectures ranging from policy blending to data-driven model predictive control (MPC). With these approaches, I achieve enhanced sample efficiency and successfully generalize across different parameterized environments. Given the augmented training times and widespread task generalization, I am able to complete robotic control tasks involving a parameter varying robotic arm as well as UAV flights subject to complex motor faults.