Abstract:

In this paper, we analyze the implementation of feedback linearization control scheme based on full data-driven inverse dynamics models. We made no use of physical models in the definition of the inverse dynamics, that was learned entirely from previously recorded data via Gaussian Process Regression (GPR). The resulting controller was tested on a simulated manipulator with 7 degrees of freedom (dof), to solve a trajectory tracking problem. Different kernel functions were tested, in particular, we analyzed the performance obtained by Squared Exponential (SE) kernel and the recently introduced Geometrically Inspired Polynomial (GIP) kernel. Results show that GIP obtains better tracking precision and is more robust w.r.t. the presence of an initial tracking errors. On the contrary, poor generalization properties of SE kernel deeply undermine control performance when the robot is located far from the poses seen during training.