Multilevel part-based model for object manipulation

Currently, researches on robotic object manipulation mainly focus on basic actions like pick-and-place and fetch-and-carry, while sophisticated actions are much more useful in practice. In this article, we propose a multilevel part-based object model for complicated manipulations and an interaction-based sample collection method for efficient training. We implement our model with WAM Arm, Barret Hand and Kinect sensor in Robot Operating System (ROS) and test part-encoded actions on selected household objects. The result demonstrates that our model enables the robot arm to manipulate objects more dexterously.