Reinforcement Learning Based Control for Uncertain Robotic Manipulator Trajectory Tracking

摘要

This paper investigates the trajectory tracking control method for a robotic manipulator with uncertainties. A compound controller combining a traditional control law with deep reinforcement learning is developed to improve tracking accuracy and adaptability. The model-based control method is able to increase sampling efficiency for learning control strategy. The introduction of deep reinforcement learning based on soft actor-critic structure and Lyapunov function constrain enables the system to compensate unknown uncertainties and remain stable. Eventually, a 3-DOF manipulator is used to show the effectiveness of the proposed controller. Comparative simulation results demonstrate that the compound controller acquires higher tracking accuracy than the pure model-based control method.