Abstract:

In this paper, we propose a novel kernel for the identification of the inverse dynamics of robotic manipulators based on Gaussian Process Regression. The proposed kernel, called Lagrangian Inspired Polynomial (LIP) kernel is based on two main ideas. First, instead of directly modeling the joint torques, we model as GPs the kinetic and potential energy of the system. To this aim, we prove a polynomial characterization of the kinetic and potential energy and we define a polynomial kernel that encodes this property. Second, we derive the GP prior on the joint torques by leveraging on the knowledge of Lagrange’s equations and by applying the properties of GPs under linear operators. We validated our method on a Franka Emika Panda robot, a 7 DOF cobot. The collected results show that the proposed method outperforms state-of-the-art black-box estimators based on Gaussian Processes in terms of prediction accuracy and generalization.