A vision-based 3D grasp planning approach with one single image

Many vision-based 3D grasping methods detect the stable grasping points based on several images. In their works, the full or partial 3D shape models are usually built for guiding the manipulations of the multi-DOF grippers. However, the uncertainties of detection and reconstruction of the contact points could lead to an unstable grasp. In this paper, the problem of vision-based 3D grasping from a single image of object is investigated, and the "attractive regions" proposed in our previous works [9] is utilized to analyze the 3D grasping process. (1) It is proven that a form closure grasp can be achieved if the state of 3D object is in the attractive region formed in the 3D grasping process. (2) Meanwhile, the initial contact regions, from which the 3D form-closure grasp can be achieved, are computed from the contour of the object. (3) A practical grasping system with four-finger gripper is designed, and a vision-based grasp planning approach is formulated. Finally, an experiment is conducted to show the efficiency of the proposed method.