Learning feedforward controller for a mobile robot vehicle

Abstract

This paper describes the design and realisation of an on-line learning posetracking controller for a three-wheeled mobile robot vehicle. The controller consists of two components. The first is a constant-gain feedback component, designed on the basis of a second-order model. The second is a learning feedforward component, containing a single-layer neural network, that generates a control contribution on the basis of the desired trajectory of the vehicle. The neural network uses B-spline basis functions, enabling a computationally fast implementation and fast learning. The resulting control system is able to correct for errors due to parameter mismatches and classes of structural errors in the model used for the controller design. After sufficient learning, an existing static gain controller designed on the basis of an extensive model has been outperformed in terms of tracking accuracy.