Robotic constant-force grinding control with a press-and-release model and model-based reinforcement learning

Abstract

When a workpiece is ground by a robot, the issue often arises that the grinding force signal can easily suffer overshoot during the impact stage and instability during the processing stage. In this paper, a force control algorithm for use in the impact and processing stages of robotic constant-force grinding is proposed based on a press-and-release model and model-based reinforcement learning. For the impact stage, a press-and-release model for compensating for the robot deformation is established to enable the indirect control of the magnitude of the force signal by regulating the ratio of the pressing and release times to prevent overshoot. In the processing stage, model-based reinforcement learning is applied to quickly obtain the optimal processing parameters. Through iterative experiments, model-based reinforcement learning is used to update the model and then continue to search for the optimal processing parameters until the normal force reaches the desired state. Experimental results show that force control based on the proposed algorithm converges fast; in the impact and processing stages, the normal force converges to the set range after 4 and 3 iterations, respectively. The normal force is more stable than it is under position control. Moreover, the surface roughness Ra of the workpiece is reduced by 30.68%.