Unsupervised Learning of Interactive Behavior for HRI

In this paper, we present our efforts toward building interactive robots that can learn how to interact naturally with human partners in different environments and contexts. The main feature of our approach is that it relies completely on unsupervised learning and time series analysis techniques that allow the robot to build its own interaction protocol representation from the bottom up. The final controller of the robot learned this way is a hierarchy of either dynamical systems or probabilistic networks with complexity that is automatically adjusted to the interaction protocol to be learned. We report two examples of applying this technique to learn an explicit interaction protocol in a master-slave settings (guided navigation) and an implicit protocol in a teammate settings (a listener robot).