Abstract
The development of joint attention related actions, such as gaze following and gaze alternation, is one of mysteries in infant development. Previous synthetic studies have proposed learning methods for gaze following without any explicit instructions as first step to understand such development. However, a robot was given a priori knowledge about which pair of sensory information and action should be associated. In this paper, we propose a learning mechanism that automatically and iteratively acquires social behavior by detecting and reproducing the causality inherent in interaction with a caregiver without such knowledge. The measurement of causality based on transfer entropy [1] is used to detect appropriate pairs of variables for acquiring social actions. The reproduction of the detected causality promotes other causality. In the computer simulation of human-robot interaction, we examine what kinds of behavior related to joint attention can be acquired sequentially by changing the behaviors of caregiver agent. The result indicates the actions are acquired in similar order to infant development.
