Towards socially adaptive robots: A novel method for real time recognition of human-robot interaction styles

Automatically detecting different styles of play in human-robot interaction is a key challenge towards adaptive robots, i.e. robots that are able to regulate the interactions and adapt to different interaction styles of the robot users. In this paper we present a novel algorithm for pattern recognition in human-robot interaction, the Cascaded Information Bottleneck Method. We apply it to real-time autonomous recognition of human-robot interaction styles. This method uses an information theoretic approach and enables to progressively extract relevant information from time series. It relies on a cascade of bottlenecks, the bottlenecks being trained one after the other according to the existing Agglomerative Information Bottleneck Algorithm. We show that a structure for the bottleneck states along the cascade emerges and we introduce a measure to extrapolate unseen data. We apply this method to real-time recognition of Human-Robot Interaction Styles by a robot in a detailed case study. The algorithm has been implemented for real interactions between humans and a real robot. We demonstrate that the algorithm, which is designed to operate real time, is capable of classifying interaction styles, with a good accuracy and a very acceptable delay. Our future work will evaluate this method in scenarios on robot-assisted therapy for children with autism.