Gbf Network Architectures for Robot Vision

A versatile robot manipulator is based on techniques of computer vision and neural network learning. For grasping objects four principal tasks have to be done in a cycle | detect the desired object and the grasping ngers in the images, evaluate the spatial relationship wrt. grasping stability, choose a more stable grasping pose (if possible), and move the manipulator to it. In this work we focus on two subproblems thereof, specifying the hand{eye coordination and evaluating the grasping situation. These procedures are based on (nonlinear) functions which are not known a priori and therefore have to be learned. We uniformly approximate the required functions by means of networks of Gaussian Basis Functions (GBF networks). Modifying the number of GBFs and/or the accessory size of the gaus-sian support changes the quality of the learned function. We use GBF networks both to specify the hand{eye coordination and to recognize the grasping situation and show how various network conngurations innuence the accuracy and therefore the usefulness of the function approximations. All experiments are carried out in real world environments using an industrial robot manipulator and the computer vision system KHOROS.