Actor-Critic Neural Network Based Finite-time Control for Uncertain Robotic Systems

This paper investigates reinforcement learning (RL) based finite-time control (FTC) of uncertain robotic systems. The proposed methodology consists of a terminal sliding mode based finite-time controller and an Actor-Critic (AC)-based RL loop that adjusts the output of the neural network. The terminal sliding mode controller is designed to ensure calculable settling time, as compared to conventional asymptotic stability. The AC-based RL loop uses recursive least square technique to update the critic network and policy gradient algorithm to estimate the parameters of actor network. We show that the AC is beneficial to improve robustness of terminal sliding mode controller both in approaching stage and near equilibrium. The performance of proposed controller is compared to that with only terminal sliding mode controller. The simulation results show that proposed controller outperforms pure terminal sliding mode controller, and that AC is a successful supplement to FTC.