Robot learning in analog neural hardware

Abstract

This paper describes a mobile robot that learned local maneuvers according to the principles of classical and operant conditioning, which were performed in an analog neural hardware implementation. The neurons were equipped with fixed gain and Hebbian inputs, each of which was low-pass filtered for short term memory effects. A wheelchair equipped with sonar and tactile sensors was used as a mobile robot that was able to steer autonomously through narrow doorways after learning an obstacle avoidance task. The system presented here performed operant conditioning in analog hardware controlling a physical mobile robot, which, to our knowledge, was not shown before.