Grasping of unknown objects on a planar surface using a single depth image

In this study, we present a novel method for grasping of an unknown object on a planar surface. Given a single depth image, the planar surface and the object are extracted by employing Random Sample Consensus. Then, the principal axis of the object is approximated by means of Principal Component Analysis. The gripper of a robotic arm approaches the object in a perpendicular direction to the plane, and grasps it in an orientation determined by the normal of the plane and the obtained principal axis of the object. In this method, no 3D shape model or off-line learning is required. In order to demonstrate the efficacy of our method, we developed a grasping system using a real robotic arm and an inexpensive depth camera. The system had a 100% success rate when grasping 18 unknown household objects including a marker pen, a pencil, an eraser, a tennis ball, a Rubik's Cube, a T-shirt and a AAA battery. The results of this study imply that the system can grasp a wide range of unknown household objects and that our method is valuable for grasping of objects on a planar surface.