Research On Object Grasping Of Two-finger Robot Based On Deep Learning

It is an indispensable interaction between robot and the real world for robot to grasp objects.For humans,grasping objects can be done without thinking,but it is difficult for robots to accurately and flexibly grasp an object.To achieve the same dexterity and interactivity as human action,robots still have many mechanical and computational problems to be overcome.In addition,robotic grasping is a good example of a minimal study that can support many other related studies.In this paper,the selection of grasping position on rigid objects(such as tools,household articles,packaging articles and industrial parts)for the two-finger manipulator is studied by combining the object grasping of manipulator and deep learning.Due to data noise and object occlusion,it is difficult to accurately infer the physical properties of the shape,pose,material,mass,and position of the contact point between the manipulator and the object,this makes it difficult for the robot to grasp various objects.In this paper,a deep neural network trained by a large number of human grasping tag data is used to grasp a variety of different objects.The depth image of the object is acquired by the Kinect2.0 camera,and a number of candidate grasp positions are generated according to the ForceClosure conditions.The image of the object is cut with the appropriate size according to the center of the grasp point and the grasp direction,and then the trained neural network is used to evaluate the grasp robustness of each pair of grasp points,and the optimal grasp point is selected by the output of the deep learning network.For some limitations of the original grasp method,such as when there is a small gap inside the object,the robot grasps is easily to collide with the object,leading to failed grasp results.At this time,the original method cannot select a reliable grasping point position.We proposes to select the peripheral contour grasping.Firstly,the narrow area inside the object is filled by applying closed operation in morphology,and then the grasping point is re-selected,which solves the problem of easily collision between robot hand and object.Mean while,selecting grasp point on the outer contour of the object can greatly reduces the number of grasp point pairs,which greatly improves the efficiency of program.In order to realize the selective grasping of objects in multi-object scenes,this paper gets the bounding box of the target object by extracting SIFT feature from the color image,and selects the grasp positions in the region of interest to achieve the successful grasping of specific object.The experimental results of simulation and real grasping of two-fingered manipulator show that the method of selecting grasping position based on depth learning can successfully grasp different objects with arbitrary posture in the field of vision,which can meet the requirement of manipulator accurate and flexible grasping of objects.