PUB - Publikationen an der Universität Bielefeld

For a social robot, the ability of learning tasks via human demonstration is very crucial. But most current approaches suffer from either the demanding of the huge amount of labeled training data needed, or the limited recognition cabability caused by the domain-specific modeling. It would be amazing when the robot could start the task learning with only a few labeled demonstrations and optimize the models by observation without human teaching. This paper puts forward a semi-supervised incremental strategy for the robot to learn the manipulation tasks performed by the user. Based on the set of prelearned object-specific manipulative primitives, the task model is extended from a basic Markov model, that can be initialized with few labeled data, and updated continously when new unlabeled data is available. Furthermore, the system also has the capability to detect the unseen tasks during the learning process and to build up new task models from them. Thus, the strategy enables the robot not only to elaborate the known tasks incrementally, but also to extend the task knowledge incrementally. The results of experiments in an office environment show the applicability of this approach.