Grasp Detection Based On 2D Images

Grabbership is an important skill for both service and industrial robots.Combining stable gripping on arbitrary objects has always been a major challenge for the robotics field.The robotics field involves visual perception and geometric analysis of the gripping objects.Existing grip synthesis methods can be roughly divided into three categories:analysis methods,data-driven methods,and object representation methods.The data-driven approach attempts to learn crawling by identifying target objects and matching them against a database 3D model in order to transfer pre-defined or pre-synthesized grabs to target objects.This is based on a strong assumption that there is a similar object model in the database so that meaningful crawling can be performed.However,for an object model that is not in the database,valid prior knowledge cannot be transferred from the database.For this situation,purely geometrical analysis will be more useful.Existing analytical methods search for force or shape closure principles by listing all possible analytical methods.The object representation method uses a similar process,first approaching the target object through the original shape in order to more easily test the grab closure.Both of these methods are aimed at finding physically feasible grabs.However,such synthetic grabs are generally not semantically reasonable.Grabber algorithm is very important for the application of service robots.Calculating a stable grab has always been a challenging problem.Many aspects such as the geometry of the target object,the geometry and kinematics of the robotic arm,and the contact between the manipulator and the object should be considered.In this paper,we mainly study the capture of three-dimensional objects in a real scene.The input is a single 2D image.We first process the input image using the Graph Cut algorithm,and divide the real scene 2D image containing the target object to obtain the 2D shape of the target object.Then,based on the hypothesis of convexity and symmetry,we find a vertical axis for each part of the object to reflect the convexity of the 2D shape obtained by maximizing each part on this axis.After completing the 2D shape,we will get each 3D model of the completed 2D segment through a sweep algorithm and represent it as a union of generalized cylinders.And use the Laplacian smoothing algorithm to generate a smooth three-dimensional model.Based on the finally obtained symmetric 3D model,we propose a pure geometric analysis method to realize the grab analysis of this three-dimensional model,and synthesize physical and semantically reasonable two-point grabs.Our main contribution lies in the selection of two geometric standards that are reasonably captured.These two criteria are derived from observations of human nature's crawling activities.The observed result is that humans tend to grab a target object at two points of the object with similar local geometries,while the two grabbing points are in different global structures.This often leads to physically stable and psychologically comfortable grasping.The final physics experiment also proved that our method produced relatively good experimental results.