On Evolving a Robust Force-Tracking Neural Network-Based Impedance Controller

Abstract :

This paper describes the use of evolutionary techniques to design a robust forcetracking impedance controller. Current state-of-the-art approaches start the analysis and design of the properties of the impedance controller from a manually-given set of impedance parameters, since no well-defined methodology has been yet presented to obtain them. In this work, an impedance controller represented as an artificial neural network (ANN) is described, whose optimal parameters are obtained in a simple way by means of evolutionary techniques. Furthermore, the controller is generalised and provided with force tracking capabilities through an on-line parameter estimator that dynamically computes the weights of the ANN-based impedance controller based on the current force reference. The resulting controller presents robustness against uncertainties both on the robot and/or the environmental model. The performance of the controller has been evaluated on a range of experiments using a model of a two-link robotic arm and a non-linear model of the environment. The results evidenced a great performance on force-tracking tasks as well as particular robustness against parametric uncertainties.