Design of a Tracking Controller Based on Machine Learning

Abstract

Tracking control of multibody systems is a challenging task requiring detailed modeling and control expertise. Especially in the case of closed-loop mechanisms, inverse kinematics as part of the controller may become a game stopper due to the extensive calculations required for solving nonlinear equations and inverting complicated functions. The procedure introduced in this paper substitutes such advanced human expertise by artificial intelligence through the utilization of surrogates, which may be trained from data obtained by classical simulation. The necessary steps are demonstrated along a parallel mechanism called λ-robot. Based on its mechanical model, the workspace is investigated, which is required to set proper initial conditions for generating data covering the used operation space of the robot. Based on these data, artificial neural networks are trained as surrogates for inverse kinematics and inverse dynamics. They provide forward control information such that the remaining error behavior is governed by a linear ordinary differential equation, which allows applying a linear quadratic regulator (LQR) from linear control theory. An additional feedback loop of the tracking error accounts for model uncertainties. Simulation results validate the applicability of the proposed concept.