RIL: Riemannian Incremental Learning of the Inertial Properties of the Robot Body Schema

Abstract

We transform classical robot inertial parameter identification into an online learning problem by integrating state-of-the-art gradient descent techniques and first-order principles from mechanics and differential geometry. Through this, incremental learning of fully physically feasible inertial properties without requiring any prior information is made possible. This is achieved using a version of Riemannian gradient descent equipped with experience replay that guarantees feasible parameter updates at all times during learning. Analysis of the method's performance are done on a virtual manipulator focusing on the influence that different measurement setups have on the estimation as well as on parameter feasibility and re-learning. Finally, we present experimental results on a real 7 DoF manipulator and evaluate the quality of the generated inverse dynamics torques and the corresponding model error.