Learning Where to Look: Autonomous Development of Gaze Behavior for Natural Human-Robot Interaction

Gaze is one of the most important nonverbal behaviors in regulating humanhuman interactions in close encounters. Several approaches have been proposed to design robots that can exhibit human-like natural gazing behavior. Most of the available techniques though are based on careful hard-coding of behavioral rules that are usually extracted from analysis of human-human interaction corpora or theories of gaze behavior in humans. This approach has two major disadvantages. First, it is very hard to pin down these behavioral rules specially taking into account cultural effects, and the effect of power distribution on gaze behavior. Second, this approach generates systems that are not generalizable to other nonverbal behaviors. In this paper, we report the development and evaluation of an unsupervised gaze controller that tries to alleviate these two weaknesses of engineered gazing behavior. The system utilizes two major learning mechanisms. First, it learns a set of recurrent gaze patterns called basic interactive acts using a motif discovery algorithm. A hierarchical controller to activate these acts as needed during the interaction is then learned resulting in a grounded controller. The system was implemented and evaluated in comparison with a reactive gaze controller that was previously shown to provide human-like gazing behavior. The proposed system was shown to be superior in terms of naturalness, human-likeness and comfort of the human partner.