Inertially-safe Grasping of Novel Objects

Abstract

We address the problem of grasping novel objects in a way that minimizes the torques that the hand is required to produce. The problem of grasping new objects has received a lot of attention in the past few years [1], [2], [3], [4], most of which has focus on finding grasps for which the shape of the gripper locally matches the shape of the object. While selecting a grip that makes it difficult for the object to escape is clearly important, it also seems natural that the inertial parameters of objects need to be taken into account. In this paper, we present a planner that suggests grasps that respect both of these criteria. Our planner searches for grasping points where the gripper nicely fits to the object, with a strong preference for grasps that are near the center of mass of the object. In other words, our planner finds grasping solutions that fulfill constraints imposed by both local shape and forces acted by the object. Our planner allows the robot to avoid situations that can be potentially dangerous for the robot itself, for example trying to grasp a large object by one of its extremities and risk damaging the robot’s hand. The main challenges of planning grasps for novel objects is that we only have an incomplete model of the object available. If we assume that we take a single snapshot of the object from a 3D sensor, it then is an involved task to find suitable finger placements of the robotic hand on the unknown backside surface of the object. It is also equally hard to infer the position of the object’s center of mass when only one side of the object is visible. We overcome the finger placement problem by using a part-based grasp planner, by which previously-learned prototypical parts [2] are fitted to the incomplete object snapshot, yielding a gripper pose that aligns the fingers to the object’s surface. We then infer the object’s center of mass from vision. The object’s center of mass is set at the center of gravity of the visible side. While this approach is simple, it already allows us to produce useful behaviors, as shown in the next section. In future work, we envision learning the mapping between a partial view and the object’s mass and center of mass. We believe that important clues for predicting the mass of an object can be obtained from vision. The object’s