Simulation and Experimental Validation of Reinforcement Learning-Based Adaptive Sliding Mode Control for a Flexible-Joint Robotic Manipulator

Abstract

In this work, an adaptive sliding mode controller (ASMC) for a flexible-joint robotic manipulator using reinforcement learning is presented and validated by experiments and simulations. The control strategy has been designed to be implemented on a practical platform, which has been designed and fabricated as a four-degree-of-freedom robotic manipulator equipped with a flexible joint. RL-ASMC combines reinforcement learning and Adaptive Sliding Mode Control (ASMC) to adapt the parameters of the controller and improve the position tracking performance under different operating conditions. The non-linear robotic system and control algorithms were simulated in MATLAB/Simulink and real-time implementation of the control system was programmed in LabVIEW/ NI myRIO for the fabricated Robotic Platform. The same external disturbances, no load and payload were applied to ASMC and RL-ASMC. The key performance measures used were angular position error, settling time, the maximum estimated control torque, and disturbance rejection. Experimental results demonstrated that the performance of RL-ASMC is superior to the standalone ASMC, which decreases the angular position error about 71%, settling time about 16%, and the maximum estimated control torque about 10%, respectively. Moreover, the simulated and experimental results were found to be in good agreement, with the average deviations of 4.8% and 4.3% for ASMC and RL-ASMC at the reference angles of 135° and 45°, respectively. The results show the effectiveness, stability, and usability of RL-ASMC for the robotic manipulators with flexible joints.