Using proprioceptive information for the development of robot body representations

Abstract

As part of the attempt to improve a robot's flexibility and adaptation by adopting biologically inspired developmental methods, we trained a multilayer perceptron model (MLP) to develop body representations of a humanoid robot using proprioceptive and motor information. The information used were the left arm joint positions, the motor commands and the electric currents applied to these joints. By babbling its left arm, that is by executing a self-exploration behaviour, the robot gathered sensorimotor information for training the model. Once having learned the relation between these different modalities, the model can be used for running predictive processes. We present our first training results and discuss further research possibilities.