Semi-supervised Incremental Learning of Manipulative Tasks

摘要

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, or the limited recognition cabability caused by very domain-specific modeling. This paper puts forward a semi-supervised incremental strategy for the robot to learn the manipulative tasks performed by the user. The task models are extended Markov models, taking a set of pre-learned object-specific manipulative primitives as basic states. They can be initialized with few labeled data, and updated continously when new unlabeled data is available. Furthermore, the system also has the capability to reject unlabeled observation as unseen tasks and detect a new task model from a group of them. Thus, using this strategy, the robot only needs human teaching at every beginning, then elaborate the learned tasks, and even extend task knowledge by its own observation. The experimental results in an office environment show the applicability of this approach. 1