Modelling of robot dynamics based on a multi-dimensional RBF-like neural network

Abstract

The modelling of robot manipulator dynamics by means of a neural architecture is presented. Such a model able to generate a decoupling and linearising feedback in the control system of the robot. In a structured model approach, a RBF-like neural network (radial basis function NN) is used to represent and adapt all model parameters with their dependences on the joint positions. The neural network is hierarchically organised to reach optimal adjustment to the common structural knowledge about the identification problem. A fixed, grid based neuron placement, together with application of polynomial basis functions is utilised favourably for a very effective recursive implementation. In this way, a neural network based online identification of a dynamic model is enabled for a complete industrial 6 joint robot with reasonable effort and good results.