Claire Dune

Contact

Email : Claire.Dune@irisa.fr
Address : IRISA / INRIA Rennes
Campus Universitaire de Beaulieu
35042 Rennes cedex - France
Tel : +33 2 99 84 74 32
Fax : +33 2 99 84 71 71
Assistant : +33 2 99 84 22 52 (Céline Ammoniaux)

Much technology remains to be developed to assist people with disabilities. As the severity of the patient's handicap increases, the complexity of the devices he must use increases. This requires significant efforts to simplify the interactions between the patient and these devices. The goal of my Ph.D. thesis is to build a "One click grasping tool", and we propose a new approach toward an intuitive human machine interface for grasping objects. The considered objects are presented in an unstructured manner and have previously unknown shape. The user controls the grasping action with only one click on an image of the whole scene, thus giving the system the only information it needs to perform the grasping task autonomously.

The robotic system is an arm mounted either on a wheelchair or on a mobile platform. It is equipped with two cameras. One is fixed, e.g. on the top of the wheelchair (eye-to-hand) and the other one is mounted on the end effector of the robotic arm (eye-in-hand). The system is fully calibrated . We do not make any assumption about the objects shape or texture, and do not use an appearance database or a model. The only assumptions we make are that the objects are rigid and static in the reference frame. (see figure)

The first step towards an autonomous grasping tool is to set the camera poses so as to bring the object in every camera's field of view. We developed a method based on the epipolar geometry of the two camera system. The object is known to lie on a 3D line of view defined by the optical centre of the eye to hand camera and the "clicked" point. The mobile camera (controlled by virtual visual servoing) then scans this line searching for the object. The most likely position is identified with classical image matching techniques. At the end of the process, every camera field of view holds the object. The method has been tested and validated on a robotic cell with a wide range of daily life objects.

Once both cameras ''see'' the object, the question that remains is: How do we gather sufficient information to reliably grasp for an unmodelled and unknown object using only visual input? We assume that the objects can be grasped by aligning the gripper with their minimal dimension perpendicularly to their main axis and centred on it. Since an accurate 3D reconstruction of the object is not needed to complete such a grasping, we propose computing the best-fit quadric for the object using the contours extracted from several views taken by the mobile camera. The quadric representation holds the advantage of being compact and gives the information needed. We also propose a method to automatically compute the velocity of the mobile camera during the ''reconstruction'' step based on an information criterion. At the end of the process, the object pose is sufficiently known to be able to properly grasp it.