Continuous Control of a Robot Manipulator Using Deep Deterministic Policy Gradient

摘要

Deep reinforcement learning (DRL) addresses the problems that previously limited the performance of RL algorithms while working with high-dimensional state and action spaces. In this paper, we explore the deep deterministic policy gradient (DDPG) algorithm that operates over continuous action spaces. The application of reference tracking for a two-link robot manipulator (TLRM) in uncertain environments is considered. The TLRM is subjected to uncertainties such as frictional forces and external torque disturbances. In the simulation study, we compare the performance of our RL-based controller with the well-known proportional-derivative (PD) controller. Results indicate a considerable improvement in the mean square error (MSE) and variance accounted for (VAF) metrics when the RL-based controller is utilized.