Learning a strategy for whole-arm grasping

Abstract

Traditionally, robot grasping has been approached in two separate phases: first, finding contact positions that yield optimal grasps and, then, moving the robot hand to these positions. This approach works well when the object's location is known exactly and the robot's control is perfect. However, in the presence of uncertainty, this approach often leads to failure, usually because the robot's gripper contacts the object and causes the object to move away from the grasp. To obtain reliable grasping in the presence of uncertainty, the robot needs to anticipate the possible motions of the object during grasping. Our approach is to compute a policy that specifies the robot's motions over a range of joint states of the object and gripper, taking into account the expected motion of the object when pushed by the gripper. We use methods from continuous-state reinforcement-learning to solve for these policies. We test our approach on the problem of whole-arm grasping for a PR2, where one or both arms, as well as the torso can all serve to create contacts.