Wavelet neural network approach for control of non-contact and contact robotic tasks

In this paper, some basic ideas of wavelet approximation theory is analyzed and applied for intelligent control of manipulation robots in noncontact and contact tasks. In the first part of analysis, the wavelet neural network is applied as feedforward part of learning decentralized control algorithm for robotic nonconstant tasks. Two different approximation strategies are proposed: one where wavelet inputs are robot nominal internal robot coordinates, velocities and accelerations and other where as additional network inputs real robot internal positions and velocities are included. As second part of analysis wavelet networks are applied for classification of unknown dynamic characteristic of robot environment and learning of robot dynamic model for robot compliance control tasks. The applied method is based on application of wavelet network for classification of force sensor data through process of off-line training.