Learning grasps with topographic features. (August 2015)

• Record Type: Journal Article
• Authors: Fischinger, David; Weiss, Astrid; Vincze, Markus
• Is Part Of: International journal of robotics research. Volume 34:Number 9(2015:Aug.)
• Page Start: 1167
• Page End: 1194
• Publication Date: 2015-08
• Subjects: Grasping in cluttered scenes -- topographic features -- machine learning -- height accumulated features; Robots -- Periodicals; Robots, Industrial -- Periodicals; 629.89205
• DOI: 10.1177/0278364915577105
• Languages: English
• ISSNs: 0278-3649

Abstract

We present a system for grasping unknown objects, even from piles or cluttered scenes, given a point cloud. Our method is based on the topography of a given scene and abstracts grasp-relevant structures to enable machine learning techniques for grasping tasks. We describe how Height Accumulated Features (HAF) and their extension, Symmetry Height Accumulated Features, extract grasp relevant local shapes. We investigate grasp quality using an F -score metric. We demonstrate the gain and the expressive power of HAF by comparing its trained classifier with one that resulted from training on simple height grids. An efficient way to calculate HAF is presented. We describe how the trained grasp classifier is used to explore the whole grasp space and introduce a heuristic to find the most robust grasp. We show how to use our approach to adapt the gripper opening width before grasping. In robotic experiments we demonstrate different aspects of our system on three robot platforms: a Schunk seven-degree-of-freedom arm, a PR2 and a Kuka LWR arm. We perform tasks to grasp single objects, autonomously unload a box and clear the table. Thereby we show that our approach is easily adaptable and robust with respect to different manipulators. As part of the experiments we compare our algorithm with a state-of-the-art method and show significant improvements. Concrete examples are used to illustrate the benefit of our approach compared with established grasp approaches. Finally, we show advantages of the symbiosis between our approach and object recognition.