Reinforcement-Learning-Based Robust Force Control for Compliant Grinding via Inverse Hysteresis Compensation

Traditional robotic grinding is prone to damage workpiece surface due to the high unmatched stiffness of manipulators and the difficulty in measuring actual contact force (ACF). Clearly, new robotic grinding combined with passive compliance devices (PCDs) driven by pneumatic actuators (PACs) definitely have wider applications. However, external disturbances and inherent complex hysteretic nonlinearities in PACs may severely degrade grinding precision. Therefore, it is a challenging issue for compliance systems to diminish hysteretic nonlinearities and maintain desired grinding force through robust control while reducing control efforts and improving response performance. To this end, this article introduces a PAC-driven PCD to make it easier to realize high-precision force control than the control of complex manipulators. For overcoming the difficulty in measuring ACF, we propose a novel control framework, where the hysteresis of the PAC is excluded from closed control loop, and its inverse compensator is utilized to accurately plan the control objective. Importantly, a reinforcement-learning-based robust controller is designed to realize the planned control objective. To the best of our knowledge, after elaborately developing the inverse hysteresis compensator under the proposed framework, this article, for the first time, presents an effective method to simultaneously realize disturbance suppression, control effort optimization, and error elimination for passive compliance systems. Finally, hardware experiments are carried out to verify the effectiveness and robustness of the proposed method.