Signature based task description and perception for motion trajectory priented Robot Learning

Robot learning by demonstration (LbD) has been a natural learning paradigm through which robots can learn the underlying tasks from human´s demonstrations. In most LbD related work, the demonstration normally just includes simple and short-term actions like gestures and grasps. Little attention was paid on the demonstration learning of long-term motions. In this paper, we propose to study the motion trajectory oriented LbD, in which we emphasize that free form 3-D motion trajectories are informative and compact to characterize most long-term human motions. In our view, the demonstrations need to be well modeled in a form that is convenient and effective to be learned and understood by robots. To this end, we introduce a novel signature descriptor to describe the space motion trajectories extracted from human motions. The signature has rich descriptive invariants that enable a robot to learn generalized task knowledge. Based on the signature, the motion perception is also explored which is an effective interface to improve the learning interactions. An experiment about robot learning of human sign language was conducted to verify the signature´s effectiveness and illustrate its applications.