Self-organization based visuo-motor coordination for a real camera and manipulator system

In the paper, we present a new method for the coordination of a six-joint manipulator whose controller is directly connected to a stereo camera system. We try to avoid the calibration processes that generally have to be implemented very precisely. This is achieved by decomposing the whole nonlinear transformation into local linear transformations and then making the linearization map recorded by the manipulator controller. The linearization is carried out by a learning process based on a Kohonen-style self-organization network. We use several strategies to reduce the huge computational costs that are required in real implementation of the learning algorithm. Results of the experiments are reported to show the properties of this method and its applicability in real situations