Neural Network Based Friction Compensation of Motion Control on A 6-DoF Robot Manipulator

Friction is the force resisting relative motions between two bodies, it causes cycle oscillations and impairs closedloop tracking performance for robot manipulators. To address this problem, in this paper, we propose a neural network (NN) structure with attention mechanism, which compromises the calculate efficiency and accuracy. In this structure, both instantaneous data and feature abstracted from historical data are utilized as inputs, and the torques needed to compensate for the effects of friction of each joint are the targets. Besides, the proposed NN method are contrasted with XGBoost method and convolutional neural network method to demonstrate that NN can achieve better approximation performance in offline experiment. Training data are collected by traversing through the working velocity of robot joints on the task trajectory, based on which the best model will be produced and is deployed on servo controller of a 6-degree-of-freedom (DoF) robot manipulator. Finally, the NN structure is tested on a real robot manipulator for online validation, the robot manipulator is programmed to run on the same trajectory with random speeds for a period of time. Online experiment results show that the predicted compensating torques applied to robot by adding feedforward control to the current command are always closer to the desired torques, and tracking errors of joints are effectively reduced.