Probabilistic Models of Object Geometry for Grasp Planning

Robot manipulators generally rely on complete knowledge of object geometry in order to plan motions and compute successful grasps. However, manipulating real-world objects poses a substantial modelling challenge. New instances of known object classes may vary from learned models. Objects that are not perfectly rigid may appear in new configurations that do not match any of the known geometries. In this paper we describe an algorithm for learning generative probabilistic models of object geometry for the purposes of manipulation; these models capture both non-rigid deformations of known objects and variability of objects within a known class. Given a single image of partially occluded objects, the model can be used to recognize objects based on the visible portion of each object contour, and then estimate the complete geometry of the object to allow grasp planning. We provide two main contributions: a probabilistic model of shape geometry and a graphical model for performing correspondence between shape descriptions. We show examples of learned models from image data and demonstrate how the learned models can be used by a manipulation planner to grasp objects in cluttered visual scenes.