Incremental gesture discovery for interactive robots

Human Robot Interaction using natural interactive modalities like gesture and verbal communication channels is becoming an important research direction because of the increase in social applications of robotics. Nevertheless, the naturalness of the interaction is usually restricted by the fact that in most cases the set of gestures or verbal commands that can be used is limited and predefined by the designer. Recently, the authors proposed a three stages development/learning process that allows the robot to learn natural interaction protocols with no predefined gestures or verbal commands. The most critical stage in this system is the first stage in which the robot discovers recurrent gestures using constrained motif discovery (CMD). Most available CMD algorithms are either batch algorithms with the exception of MCInc which is not suitable for the task at hand due to their reliance on random sampling from the whole data set. In this paper we propose a novel incremental CMD algorithm that does not have this restriction. The paper also evaluates the effectiveness of using the action stream as an additional source to constrain the discovery process.