Behavior network acquisition in multisensor space for whole-body humanoid

This paper presents a design and the development of a robot system, which has the ability to acquire a behavior description by network representation called StateNet. In the StateNet, arcs represent whole-body motions of a robot, and nodes represent robot states, or multi-sensor body images. Also, there is another network where each node has attentions to the sensors. The system uses stored sensor information to determine attentions. This autonomous acquisition has diffuse nodes and lacks arcs. To solve these problems, this paper proposes a method to integrate nodes with clustering method and to create arcs by generating robot´s motions using GA-based (genetic algorithm) learning method. Finally, we show an experiment with a small whole-body humanoid.