A Reinforcement-Learning Approach to Control Robotic Manipulator Based on Improved DDPG

摘要

One of the exciting development in the recent decades has been the capacity to teach robots using Rein-forcement Learning (RL) techniques to execute certain tasks. Deep Deterministic Policy Gradient (DDPG) is one of those RL techniques. This paper proposes an adaptive robust controller based on an improved DDPG algorithm for position and velocity control of an n-link robotic manipulator, which is nonlinear and uncertain. The designed controller takes care of model nonlinearities, uncertainties and also time-varying external disturbances. The controller is based on multiple actor networks and a critic network. A reward function is proposed in order to guarantee a stable and effective learning of the proposed DDPG agent. Finally, numerical simulation are performed using two-link robot manipulator as an example. The outcomes demonstrate the robustness, adaptability and trajectory tracking accuracy of this control approach. Neural Lyapunov stability is also shown for the proposed controller.