Stable Model-based Control with Gaussian Process Regression for Robot Manipulators
For robotics practitioners, this provides a stable, nonparametric method to improve tracking precision without requiring precise parametric models.
The paper proposes a computed-torque control law using Gaussian Process regression to compensate for unmodeled dynamics in robot manipulators, guaranteeing stochastically bounded tracking error and asymptotic stability. In simulations and experiments, it outperforms classical computed-torque approaches in tracking precision.
Computed-torque control requires a very precise dynamical model of the robot for compensating the manipulator dynamics. This allows reduction of the controller's feedback gains resulting in disturbance attenuation and other advantages. Finding precise models for manipulators is often difficult with parametric approaches, e.g. in the presence of complex friction or flexible links. Therefore, we propose a novel computed-torque control law which consists of a PD feedback and a dynamic feed forward compensation part with Gaussian Processes. For this purpose, the nonparametric Gaussian Process regression infers the difference between an estimated and the true dynamics. In contrast to other approaches, we can guarantee that the tracking error is stochastically bounded. Furthermore, if the number of training points tends to infinity, the tracking error is asymptotically stable in the large. In simulation and with an experiment, we demonstrate the applicability of the proposed control law and that it outperforms classical computed-torque approaches in terms of tracking precision.