Neural net applications in advanced robot control systems

Abstract

The paper outlines DLR's present approaches to neural net based robot control in different areas within the frame of the BMFT- funded project NERES. The investigations aim mainly at the improvement of robot dynamics based on joint level control for a new generation of torque-controlled light-weight robots on one side, and on learning and self-improvement of sensory feedback involving the robot's environment on the other side. For these latter investigations force-torque (tactile) feedback is taken into account as well as nontactile sensorfusion of TV-images and range finder information. Tabular knowledge-bases, backpropagation techniques in multilayer nets as well as Kalman filter algorithms are the main technique applied to these feedback-type problems. In addition to these control-oriented topics, world-modelling by feeding sensory data range into Kohonen-nets has been initiated recently.