Determining final grasp pose of robot end effector after traversing to pre-grasp pose

A robot uses a two-step vision process to precisely grasp an object. First, it uses an initial camera to get a rough idea of where the object is and moves its gripper to a "pre-grasp" position nearby. Then, a camera on the gripper itself takes a closer look, comparing what it sees to a library of known visual features and their associated grasping instructions to determine the exact final position for the gripper. The claims narrow this to a specific method where the final grasp pose strictly conforms to a candidate pose derived from this second vision step.