Autonomous development of gaze control for natural human-robot interaction

Gaze behavior is one of the most important nonverbal behaviors during human-human close encounters. For this reason, many researchers in natural human-robot interaction focus on developing robots that can achieve human-like gaze behavior. Many approaches have been proposed to achieve this natural gaze behavior based on accurate analysis of human behaviors during natural interactions. One limitation of most available approaches is that the behavior is hardwired to the robot and learning techniques are used only, if ever, for adjusting the parameters of the behavior. In this paper we propose and evaluate a different approach in which the robot learns natural gaze behavior by watching natural interactions between humans. The proposed approach uses the LiEICA architecture developed by the authors and is completely unsupervised which leads to grounded behavior. We compare the resulting gaze controller with a state-of-the-art gaze controller that achieved human-like behavior and show that the proposed approach leads to a more natural gaze behavior based on subjective evaluations of subjects.