A Behavior-based Approach for Task Learning on Mobile Manipulators

This paper describes a general method for task learning on behavior-based mobile manipulators. Primitive motions are defined as elementary behaviors in the system providing fast reactive interactions with the environment and the users. These behaviors are also used as basic units to construct the execution of a task. In this context, task learning is the process to find the mapping from the sensory data to the elementary behaviors. In this paper, we present a general method to extract the patterns caused by a certain behavior in the sensory data. and efficiently find the mapping from sensory data to behaviors in feature space. Machine learning techniques such as Decision Tree and Support Vector Machine (SVM) are used to select relevant features and classify patterns respectively. To verify this approach, we demonstrate two different tasks, Door-opening and Table-cleaning with our mobile manipulator. The robot can successfully execute these tasks in an unstructured environment, also with different objects.