Grasping with Intent to Place

We consider the problem of how to grasp an object given that we know its desired placement location. Instead of making path planning an intrinsic part of the grasp point evaluation algorithm, we attempt to evaluate the “goodness” of certain grasps by calculating metrics on the grasps and object point clouds themselves. This work aims to capture human intuitions about good grasps for certain placements and uses machine learning as a method of refining them. Introduction Grasping objects from stereo vision can be formulated as a supervised learning problem by learning the weights of 2D features on the disparity image. Using a maximum submatrix algorithm to find a rectangle that represents an oriented grasp point allows for an algorithm that runs in O(n), which is quite fast in practice and effective if there is a good grasping point distinguishable from the perspective of the camera. If not, the algorithm could be extended by taking pictures from multiple views and comparing the scores of the rectangles. However, from a personal robotics standpoint, there is more to consider when picking up an object than whether the robot could grip it at that point. Most of the time, the purpose of picking up an object is to put it down somewhere, often in a known terminal orientation. As humans, we naturally know how to do this without carefully planning the object’s movement. We consider features of the object -how much clearance we will have, how stable the grasp is, whether the object will slip out of our hands when we are manipulating it, but this not very difficult in most cases. For example, for a household robot, it will be important when placing things on shelves for there to be a clear method for removing the hand after making the placement. One simple rule to guarantee this is to have the robot always grasp the object such that its hand is between the robot and the object in the final configuration. We propose a machine learning algorithm that captures the spirit of such rules. Figure 1: If the intent is to flip the cup over, the grasp on the left will result in failure due to the gripper colliding with the ground. A better grasp is shown in the middle picture where the transformed position (right) is also a good grasp. Problem Motivation One motivation is to improve current grasps by removing bad grasps. To do this we combine 3D point cloud data in both the initial and final position of the object to calculate our features. This differs from current grasping algorithms because it uses the point cloud of the object in the final orientation as well. The same intuitions about grasping objects with limited information will be applied to tasks in personal robotics, to allow quick calculations to be done in cases where many alternatives are possible. As the grasping problem approaches a solution, placing will increase in importance. Problem Specification Given a stereo image, point cloud, rigid-body transformation from object initial to final position and a ranked list of grasps, re-rank the grasps using SVM-learned feature weights.