Abstract
We have developed neuro-robot system as a suitable small-scaled model system for embodied brain information processing. Neuro-robot behavior is generated from spatiotemporal pattern of electrical activity evoked by an input from the outer world. Various algorithms can be proposed for the interface between neuronal activity and the robot behavior. In this study, we developed neuro-robot with Self-Organization Map (SOM), which generates robot behaviors simultaneously without interruptions for teacher learning. The system has an e-puck robot body, equipped with IR sensors. Electrical stimulation is applied according to the value of IR sensors. Evoked responses are measured by 64 electrodes, and a feature vector of the spatiotemporal pattern is generated. The input layer with 64 nodes has been provided to input feature vector. A two-dimensional output layer with 10 x 10 nodes has been provided. The function of SOM is dimension reduction. To restrict the winner against a particular input, the winner for that input should be fixed at a seed node and the input should be applied repeatedly during initial learning process. We confirmed that this seeding approach could perform a successive collision avoidance behavior.
