Neural Network Based Friction Compensation for Joints in Robotic Motion Control

Friction exits in robot manipulators in static or dynamic form at different stages, it functions between contacting surfaces of the joints, impairing servo-loop tracking performance. To address this problem, in this paper, we propose a modified neural network (NN) structure under the framework of multivariable time series forecasting, which compromises the calculate efficiency and accuracy. In this structure, both historical data and instantaneous data are utilized as inputs, and the torques needed to compensate for the effects of friction of each joint are the targets this structure need to forecast. Besides, both the tracking errors and the prediction errors are applied to update NN parameters. Furthermore, an autocorrelation parameter is introduced to modify the inputs and outputs of the NN structure, dealing with auto correlated errors for time series forecasting. The NN structure is tested on a real robot manipulator for a specific task trajectory. To solve the limitation of actual data, a conventional friction model with noise is applied to simulate the real friction, thus producing adequate training data to pre-train NN model, and the final NN model is fine-tuned using real task data on basis of the pre-trained model. Experiments show that the predicted compensating torques applied to robot by adding a feedforward control to the current command are always close to the desired torques, and tracking errors of joints are effectively reduced on the task trajectory.