Learning interaction protocols by mimicking understanding and reproducing human interactive behavior

Understanding human interactive behavior is a key technology required for future robots. To achieve this goal, the robot should be able to recognize key patterns in human-human interactions. Moreover, the robot should be able to generate similar behaviors during its interaction with human partners. In this paper, an unsupervised system is proposed that allows the robot to build a generative model of the interaction protocol using interaction records. A system is evaluated in a guided navigation task and is shown to successfully learn the underlying interaction protocol.