Vision planning for grasping of unknown 3-dimensional objects using a two-fingered parallel-jawed gripper.

摘要

Vision planners used for grasping attempt to either recognize or reconstruct the object to be grasped. Recognition systems, while needing only a few images, have difficulty with unknown objects, and reconstruction systems, while being able to grasp unknown objects, require several images to construct a CAD level model of the object. By incorporating the grasp planner with the view planner, a system is developed that can grasp unknown three-dimensional objects while generally requiring only a few images. In the worst case, the required number of views is the same as in view planning for reconstruction.; A system using off-the-shelf components without relying upon structured light, lasers, or a stereo camera system is proposed. The system is modeled for a basic five degree-of-freedom MICROBOT TeachMover robot equipped with a black-and-white pinhole camera mounted on a compliant two-fingered parallel-jawed gripper. Three-dimensional images are obtained from projecting points in the workspace back into the image space as the robot moves relative to the desired object. The surface of the object is assumed to be matte and free from specular reflections and shadows. As the surface of the object is being modeled, a grasp plan is formulated for an antipodal grasp. The planned grasp is to be stable enough to resist gravity and to maximize the chance of gripping the object successfully given errors in the system. During the planning, a constant density and hardness of the object is assumed since these parameters are not known a priori and cannot be determined visually. The position for the next best view is determined by the view location that will most likely acquire the missing information required to a stable grasp. Only grasps for the general task of transporting, sorting, removal, or reorientation for further inspection; rather than for fine manipulation, are considered.; The minimal view criteria for grasping an object are given and it is shown that any shape can be held by a compliant two-fingered parallel jawed gripper. Using the view criteria and view planning system it is possible to predict the number of required views necessary to plan a stable grasp of a given object.