Gaussian Process Based Disturbance Compensation for an Inverted Pendulum

Abstract This paper considers an inverted pendulum with velocity input to demonstrate the validity of a new method for estimating unknown disturbances such as nonlinear friction and damping. With only one run of a given swing-up strategy, the input and output data of the system are collected. They are the basis for a Gaussian process modelling that is used to estimate the unknown disturbance term. The learned Gaussian process model can be used subsequently to predict this disturbance and serve for an online disturbance compensation. Simulation results and a comparison with a classical observer-based disturbance compensation indicate the benefits of the proposed approach.