Internal representation of sensory information for training autonomous robot

In this paper we report on our experiments in training an autonomous robot using a hierarchical neural network containing a topographical map in its hidden layer. The map topologically organizes the sensory information of the robot and propagates this information to the next layer that is trained in supervised manner. Through some physical experiments, we show that the order in the internal representation is important in supporting the success of the supervised learning of the robot to acquire a good strategy for operating in physical environments.