Self-Organized Exploration and Automatic Sensor Integration From the Homeokinetic Principle

Abstract

Starting from the homeokinetic principle introduced earlier the present paper presents simple learning rules for the neurons of a closed loop robot controller. These learning rules are shown in a simple application case to realize a self-learning autonomous robot which can survive in a sufficiently simple world without any further external help. In particular we demonstrate that sensors are automatically integrated according to their response strength as soon as they deliver a signal to the controller. Moreover the system also can deal with the problem of a rapid change in the properties of the sensors. The basic effect observed is that the learning rule drives the robot in the sensorimotor loop into an explorative mode of behavior which however is sensitive to the reactions of the environment by way of the model error. From a dynamic systems point of view we have a closed loop control system with a pitchfork or a Hopf bifurcation (if the learning of the threshold is included) and the effect of the learning is to drive the system to a regime slightly above the bifurcation point where such systems are known to be particularly sensitive.