Research On Two-finger Manipulator Grasp Synthesis Based On Convolutional Neural Network

The rise of deep learning technology has driven the growth of the intelligent robot market,making the interaction between robots and the physical world become a hot research topic.Based on the autonomous grasping of the manipulator,this paper studies the solution of using convolution neural network to realize the stable grasping point planning of the object surface.In this paper,scene depth image is used as input information,and a two-step grasp planning scheme of “sample first,then predicting” is adopted.Firstly,the Laplace method is used to extract the edge pixels of the body in the deep image,and the opposing-plantar method is used to generate the grabbing space.Then,based on importance sampling method,candidate grasping sets are obtained from grasping space.Finally,the trained grasping predictive convolution neural network model is used to predict the confidence of each grasping success of candidate sets,and the maximum of them is taken as the planning result to guide the robot to complete grasping.In order to apply convolution neural network to grasp prediction,a model based on convolution neural network is established,where the grasp prediction problem is abstracted as a classification problem in deep learning.The output of network is the confidence level of a grasp point which belongs to the successful grasp.This paper describes a grasp by its central coordinates and direction of the two-fingered manipulator.Depth images of 32×32 pixels taken as the center are used as input information of the network.The cross-entropy distance between the output of the network and the label data is used as the cost function of the training of the control network,so that the network can extract the advanced features of the depth image to learn the relationship between the depth image and the grasping category,and realize the grasp prediction.In the actual environment,due to the existence of systematic errors such as camera assembly and motion clearance of manipulator arm,it is necessary to introduce a kind of grasping index which is robust to system errors in the calculation of data set label.The grasping model of two-fingered manipulator and the method of force closure analysis are studied in this paper.The probability of force closure is introduced as the basis of training data label calculation.When a grasping force closure probability is greater than the threshold,the label will be labeled,and vice versa.In order to generate candidate grasp prediction sets,this paper studies the method of object boundary extraction based on Laplace operator,which is the basis of constructing the space of target grasping points.Meanwhile,we determines the strategy of candidate grasp sampling based on importance sampling and optimizes the probability density function of high quality grabbing point distribution by cross-entropy method.In this way,the candidate grasping sets can contain the grasping points with high confidence in the success of grasping in the space of the metatarsal grasping points to the greatest extent.