From spreading of behavior to dyadic interaction - A robot learns what to imitate

Imitation learning is a promising way to learn new behavior in robotic multiagent systems and in human-robot interaction. However, imitating agents should be able to decide autonomously which behavior, observed in others, is interesting to copy. This paper shows a method for extraction of meaningful chunks of information from a continuous sequence of observed actions by using a simple recurrent network (Elman Net). Results show that, independently of the high level of task-specific noise, Elman nets can be used for learning through prediction a reoccurring action patterns, observed in another robotic agent. We conclude that this primarily robot to robot interaction study can be generalized to human-robot interaction and show how we use these results for recognizing emotional behaviors in human-robot interaction scenarios. The limitations of the proposed approach and the future directions are discussed. C © 2010 Wiley Periodicals, Inc.