Self-initiated imitation learning. Discovering what to imitate

Abstract

Imitation learning is an important area in robotics and agents research because it provides an easy way for robot programming and also a bootstrapping technique for social learning. Available learning by imitation systems implicitly or explicitly assume that the boundaries of the actions to be imitated are set by the demonstrator and that the robot is in some imitation mode during the whole interaction session. A less researched area is self-initiated imitation in which the robot needs to decide for itself what to imitate from another imitatee that may not be actively involved in the demonstration process. In this paper, we propose a self-initiated imitation engine based on combining techniques from time-series analysis and causality discovery. The paper also reports a series of proof of concept experiments using simulated and real robots. These evaluations show that the proposed approach is capable of discovering important patterns of behavior during the interaction session and faithfully reproduces them.