Behavior-based Task Learning by Demonstration for Mobile Manipulators

Task learning for behavior‐based mobile manipulation is formalized as a behavior recognition problem in this paper. The goal is to generate a behavior diagram through human demonstration, which can be used as a template for execution of a specified task. In order to extract significant and meaningful information from sensory data, various features are defined to convert the sensory data to the feature space. A generic task learning approach based on combined machine learning techniques is used to classify behaviors from the feature space. Not only can this approach recognize already‐known behaviors, it can also provide an easy method to add new behaviors to a system by combining new training datasets with the existing ones through retraining a support vector machine kernel (SVM) model. Common household applications, such as door‐opening and window‐cleaning tasks, are implemented for the evaluation of the proposed method.

Publication year:2013

Accessibility:Open