ON AN ADAPTABLE CONNECTIONIST ROBOT CONTROL-SYSTEM

Living organisms seem to handle complex tasks by using basic reflexes as building blocks, from which larger units of behavior are assembled: some compositional method for learning to reach new goals by combinin g familiar action sequences into more complex new actions is necessary to overcome scaling problems associated with non-compositional algori thms. Artificial Neural Networks (ANN) seem to offer a good approach t o test the hypotheses mentioned above. In order to integrate the conce pts and models from both the fields of motor control and ANN we propos e a new connectionist neural network that matches a biological model f or the neural assemblies: the paper shows a control scheme similar to that observed in a spinal animal. In this paper we discuss how repetit ive input signals to the ANN will be transformed into the appropriate sequences of commands, coded in a task space, to control a robot arm.