Shape-Primitive Based Object Recognition and Grasping

Abstract:
Grasping objects from unstructured piles is an important, but difficult task. We present a new framework to grasp objects composed of shape primitives like cylinders and spheres. For object recognition, we employ efficient shape primitive detection methods in 3D point clouds. Object models composed of such primitives are then found in the detected shapes with a probabilistic graph-matching technique. We implement object grasping based on the shape primitives in an efficient multi-stage process that successively prunes infeasible grasps in tests of increasing complexity. The final step is to plan collision-free reaching motions to execute the grasps. With our approach, our service robot can grasp object compounds from piles of objects, e. g., in transport boxes.