Manipulation planning of similar objects by part correspondence

Many innovative ideas in robotics have been inspired by neuroscience and, in particular, by the investigation of how intelligence and perception work. In this paper we explore an approach for semantic robot grasping, which combines programming by demonstration, automatic 3D shape segmentation and manipulation planning by parts. Neuro-psychology studies have evidenced the influence of shape decomposition for human perception of objects. In accordance to these findings a robot manipulation system is presented which is capable of learning and planning manipulation tasks for similar objects. The proposed approach allows a robot to perform intelligent grasping tasks by modeling the topology of an object. Manipulation tasks are demonstrated in virtual reality using a data glove. Results show that 3D shape segmentation enables both object retrieval and part-based grasping according to the affordances of an object.