Robotic Grasp Detection Using Deep Learning And Geometry Model Of Soft Hand

Robotic grasping using a vision system is one of the most interesting areas in robotic.The goal of this research is to find out the graspable positions and orientations of the objects,and plan to grasp them.But it is by far not solved yet,especially grasping the objects in the dense cluster.This paper proposes a method using deep learning and geometry model of soft hand for detecting the graspable positions and orientations of the objects from clustered scenes,given a point cloud from a single depth camera placed in the middle of the robot.With the method this paper proposed,which is different with others,it takes the collision problem into consideration without any segmentation and recognition of the objects.Significantly,this paper also analyses the computational complexity and the spacial complexity of CNN(Convolution Neural Network,a kind of deep learning),which can provide some useful tips for engineer to design the deep learning models.Notably,the labeled training data set will be generated automatically with some criteria.As to say,these criteria,with which to make decision whether the grasp hypotheses ar e handles or not,will be instead of the trained deep learning model.Firstly,a geometry model of soft hand(more flexible than hard hand)will be designed for searching appreciate closing cube in the 3D point clouds.A closing cube is regarded as a grasp hypothesis,which should be whole contained in the geometry model of soft hand without any collisions.And it includes all the parameters of positions and orientations which can be used to robotic grasping.Secondly,we use a deep learning method(Mod-Le Net)to classify and rank these grasp hypotheses,so that we can find out the best handle.After the compare experiments with the algorithm of SVM and Mod-Lenet,we can see that,the number of the antipodal the Mod-Le Net found is less than the SVM one,but all the antipodals left are all best graspable.