Learning Inverse Dynamics by Gaussian Process Regression Under the Multi-task Learning Framework

In this chapter, dedicated to Dit-Yan’s mentor and friend George Bekey on the occasion of his 80th birthday, we investigate for the first time the feasibility of applying the multi-task learning (or called transfer learning) approach to the learning of inverse dynamics. Due to the difficulties of modeling the dynamics completely and accurately and solving the dynamics equations analytically to obtain the control variables, the machine learning approach has been regarded as a viable alternative to the robotic control problem. In particular, we learn the inverse model from measured data as a regression problem and solve it using a nonparametric Bayesian kernel approach called Gaussian process regression (GPR). Instead of solving the regression tasks for different degrees of freedom (DOFs) separately and independently, the central thesis of this work is that modeling the inter-task dependencies explicitly and allowing adaptive transfer of knowledge between different tasks can make the learning problem much easier. Specifically, based on data from a 7-DOF robot arm, we demonstrate that the learning accuracy can often be significantly increased when the multi-task learning approach is adopted.