Learning behavior selection through interaction based on emotionally grounded symbol concept

In this paper, we propose a learning algorithm for action selection mechanism in the EGO architecture, which we proposed for autonomous behavior control of a humanoid robot. The concept of behavior value is introduced for action selection. The behavior value of each behavior module depends on external stimuli and internal states, and the behavior module with the higher behavior value is selected in the situation. We address the importance of learning the behavior value of each behavior. We describe how to compute behavior values for behavior modules through interaction with humans and environment. We implemented the learning algorithm on QRIO SDR-4X II, a small humanoid robot, and confirmed that for a given interaction driven behavior module, a high behavior value is obtained when interacting with a friendly user. The same tendency is obtained for a proper color painted ball for soccer play behavior module.