Template-Based Exploration of Grasp Selection

Autonomous robotic grasping is one of the pre-requisites for personal robots to become useful when assisting humans in households. Seamlessly easy for humans, it still remains a very challenging task for robots. The key problem of robotic grasping is to automatically choose an appropriate grasp configuration given an object as perceived by the sensors of the robot. An algorithm that autonomously provides promising grasp hypotheses has to be able to generalize over the large variations in size and geometry of everyday objects (for examples see Fig. 1). Object model-free approaches have been proposed that directly operate on point clouds provided by 3D sensors (for example a stereo camera system, or the Microsoft Kinect). Hsiao et. al. developed an algorithm that searches among feasible top and side grasps to maximize the amount of object mass between the finger tips of the robot gripper [1]. This and similar approaches use a fixed heuristic for grasp computation and thus lack the ability to adapt and improve the ranking of grasp hypotheses based on previous grasp executions. In [2] success rate of grasps is learned by trial-and-error from local descriptors extracted from 2d images. However, local 2d features for grasping are often not descriptive enough to select a 6d grasp configuration. In this paper, we propose a model free grasp selection algorithm that generates grasps for a wide variety of differently shaped objects and is able to improve from experience. In addition the repertoire of grasps can be extended by kinesthetic teaching. We propose an object part representation, the grasp heightmap, which (i) better generalizes over local features and (ii) also represents holistic features (see Fig. 2). Further, it is not restricted to specific hands, but performs well on different robots. Our approach is based on the simple assumption that similar objects can be grasped with similar grasp configurations. For example, a pen can be grasped