Behavior-based Task Learning by Demonstration on Mobile Manipulation

Users should have the opportunity to teach new tasks on their robot to fit the specific needs in its own environment. Behaviors are good building blocks for complex tasks because they are in charge of a specific control objective and have an intuitive name. After encapsulating controls within a behavior, the task learning problem for behavior-based systems becomes a behavior recognition problem. Behavior-based task learning can be focused on either behavior segmentation, behavior recognition or behavior cooperation. The behavior diagram is expressed by an execution matrix, which consists of only numerical information referring to the parameters of the behaviors or transitions. The parameter values are determined by the behavior classification.A feature space is introduced to extract more meaningful information from sensory data. Geometric features, such as the plane feature or the line feature, detect whether the movement of points fulfills a certain geometric relationship. A combined machine learning approach, which consists of decision tree and support vector machine, is proposed for behavior-based task learning. The decision tree is used to select relevant features from the feature pool, and support vector machine is applied to find the mapping from the feature space to behaviors. This way, corresponding behaviors can be recognized during demonstration. This method also has the ability to handle the diversity of robot configurations and users. Post-processing techniques are proposed to improve the classifi-cation results.