Learning of 2D grasping strategies from box-based 3D object approximations

In this paper, we bridge and extend the approaches of 3D shape approximation and 2D grasping strategies. We begin by applying a shape decomposition to an object, i.e. its extracted 3D point data, using a flexible hierarchy of minimum volume bounding boxes. From this representation, we use the projections of points onto each of the valid faces as a basis for finding planar grasps. These grasp hypotheses are evaluated using a set of 2D and 3D heuristic quality measures. Finally on this set of quality measures, we use a neural network to learn good grasps and the relevance of each quality measure for a good grasp. We test and evaluate the algorithm in the GraspIt! simulator.