Identification-Assisted Sim-to-Real Reinforcement Learning for Adaptive Sliding Mode Control of Industrial Robotic Manipulators

Abstract

This paper presents an identification-assisted sim-to-real reinforcement learning framework for adaptive sliding mode control of n-degree-of-freedom (n-DOF) industrial robotic manipulators. Within this framework, a reinforcement-learning-based adaptive sliding mode control with time-delay estimation (ASMCTDERL) algorithm is proposed, in which a Deep Deterministic Policy Gradient (DDPG)-trained Actor network generates online adjustment factors for the SMCTDE controller parameters. To reduce the mismatch between the simulation environment and the physical robotic system, the dynamic parameters of the experimental manipulator are identified from input–output data using the least-squares method. The identified parameters are then incorporated into the MATLAB/Simulink model to construct an identified simulation environment for DDPG training. After training, the trained Actor network is transferred to the physical robotic system to adaptively tune the SMCTDE controller during operation. The stability of the baseline SMCTDE controller is analyzed using Lyapunov theory under bounded time-delay estimation error, while the reinforcement-learning-based adaptation is constrained within admissible parameter ranges to preserve stable controller operation. The proposed framework is experimentally validated on a 6-DOF industrial robotic manipulator under payload variation and time-varying motor efficiency conditions. Simulation and experimental results show that the proposed ASMCTDERL controller achieves smaller tracking errors and lower RMSE values than conventional SMC and fixed-gain SMCTDE controllers. These results demonstrate the effectiveness of the proposed identification-assisted sim-to-real reinforcement learning framework for adaptive controller parameter tuning in industrial robotic manipulators.