Object Grasping Based On Convolution Neural Network

For the last few years,along with the applications of robots in various fields,institutions further focus on the robot research to inprove its performance indices.As one of the basic functions of robots,object grasping has become an important research direction In order to improve the success rate and accuracy of object grasping,many researchers studied the grasping rectangle of objects and select the best grasping rectangle.Referring to the characteristics of object grasping of human,a three-level serial convolution neural network(CNN)for object grasp detecting is proposed to realize high-accuracy grasp of unknown objects.The proposed three-level CNN can be described as follows.The first level network is designed to locate the position of objects roughly,which is prepared for the next level network to search and determine the position of the grasping rectangle.The second level network is designed to obtain the preselected grasping rectangles and meant to catch much less features with a quite smaller network,so as to find out the available object grasping rectangles and eliminate unavailable ones.The third level network is designed to reevaluate the preselected object grasping rectangles and meant to catch much more features with a quite larger network,so as to evaluate every single preselected object grasping rectangle accurately and obtain the best object grasping rectangle.The experimental results validate that the grasping accuracy of the three-level CNN increases by 6.1%compared to the single CNN.When the object grasping rectangle is obtained only using the RGB image,the accuracy of the object grasping rectangle is easily affected by the color and background of the object.To solve it,this dissertation modifies the structure of the three-level CNN,i.e.a particular CNN is added to the second level network of the three-level CNN to deal with the deep image,and the selecting algorithm of the best grasping rectangle is optimized to make up for the deficiency of the original three-level CNN.In the experiments.the accuracy of the object grasping rectangle increases by 2.6%compared to the previous one.A high-accuracy grasping operation is accomplished on the Youbot robot,whose grpper is used as the grasping actuator.When the object is grasped with the gripper,the size and structure of the grpper severely limits types of the objects.Thus it is difficult to grasp the objects with complex geometric shapes,such as heart-like,triangle and pentagram.The multi-finger dexterous hand with multiple degrees-of-freedom and multiple joints is suitable to grasp various geometric shape objects,but most of current approaches by using muti-finger dexterous hand requires 3D model of the object.With the absence of 3D models,it is difficult to grasp objects accurately.To solve these problems,this dissertation presents a two-level CNN for object grasping by using multi-fingered dexterous hand.The two-level CNN is used to obtain the distribution of fingers on the object when the object is grasped by the dexterous hand.After the experiments,the accuracy of the finger distribution obtained by using the two-level CNN is more than 95%.And with the absence of 3D model of the object,the two-level CNN realizes the high-accuracy grasping operation of unknown objects with the Shadow multi-finger dexterous hand.