Visually guided object grasping

{ In this paper we describe a method for aligning a robot gripper | or any other end eeector | with an object. An example of such a gripper/object alignment is grasping. The task consists of, rst computing an alignment condition, and second servoing the robot such that it moves and reaches the desired position. A single camera is used to provide the visual feedback necessary to estimate the location of the object to be grasped, to determine the gripper/object alignment condition, and to dynamically control the robot's motion. The original contributions of this paper are the following. Since the camera is not mounted onto the robot it is crucial to express the alignment condition such that it does not depend on the intrinsic and extrinsic camera parameters. Therefore we developp a method for expressing the alignment condition (the relative location of the gripper with respect to the object) such that it is projective invariant, i.e., it is view invariant and it does not require a calibrated camera. The central issue of any image-based servoing method is the estimation of the image Jacobian. This Jacobian relates the 3-D velocity eld of a moving object to the image velocity eld. In the past, the exact estimation of this Jacobian has been avoided because of the lack of a fast and robust method to estimate the pose of a 3-D object with respect to a camera. We discuss the advantage of using an exact image Jacobian with respect to the dynamic behaviour of the servoing process and we describe a new method for computing the pose of an object with respect to the camera that observes this object. This new pose computation method avoids the non-linear structure of the perspective camera model by using a rst-order approximation of perspective | paraperspective. After just a few iterations, the paraperspective pose converges to perspective pose. From an experimental point of view, we show the interest of exact versus approximate Jacobian estimation. Finally, it is important to stress the fact that, in all the visual tasks that are described below, the camera is either not calibrated or poorly calibrated.