Reinforcement learning for adaptive gain in sliding mode controller: applications in robot arms

This paper presents a novel method for providing adaptive gains for Sliding Mode Control (SMC), based on reinforcement learning (RL). SMC is a non-linear robust control method capable of reference tracking, even in the presence of uncertainties. However, the main drawback of SMC is ‘chattering’, which leads to high frequency, low-amplitude oscillations of the input signal. This effect may be reduced by applying an adaptive gain, computed using RL. The RL agent is designed based on Deep Q-Network and is trained on a set of fixed reference joint angles on a two-link robot. 600 episodes are used to train the RL agent for each joint, with each episode lasting 10 seconds. Simulation results show improved performance in terms of convergence and error reduction, compared to two fixed gain SMCs. Preliminary experimental results obtained using a single motor, single link setup show that the RL-SMC method is capable of controlling the joint angle with 0.5° precision, with significant reduction in chattering compared to the fixed gain case. This is the first time that Double DQN has been used for adaptive gain selection in SMC.