Model Learning of Robot Inverse Dynamics based on Self-Organizing Map Gaussian Process Regression

Abstract

An accurate inverse dynamics model of robotic systems is key in robotic applications. Analytic inverse dynamics models suffer from uncertainties of physical parameters estimation, some complex frictions and actuator dynamics. In such cases, developing machine learning algorithm to approximate the inverse dynamics model from collected data become a hot research field. In this paper, a novel approach based on Self-Organizing Map Gaussian Process Regression (SOM-GPR) is proposed for modeling inverse dynamics of a robotic arm, which is a combination of the SOM and GPR. In our method, we use the SOM neural network to divide the global GP model into multiple small-sized local GP models. The prediction for a test point is performed by the nearest local GP model. With our approach, we can achieve the accurate modeling of robot inverse dynamics models. The effectiveness of our proposed algorithm has been verified in the experiment. The result of the experiment shows that the combination of two techniques we propose has a better performance over other extensions of GPR applied to learn the inverse dynamics model.