Simultaneous estimation of role and response strategy in human-robot role-reversal imitation learning

In this paper, we describe a novel imitation learning method which enables an autonomous robot to acquire response strategy and to estimate roles through human-robot realtime interaction. The robot becomes able to respond to human user's social action, e.g. bye-bye and shake hand, correctly. We constructed the learning method based on role reversal imitation which is found in human infants in developmental psychological researches. A probabilistic model is proposed which assumes that delayed reactions are stochastically generated by initiative actions. In an experiment, we show a robot hand became able to exhibit correct reaction and estimate whether another's action is an initiative action or a reaction.