Research On Stacking Object Grasping Method Based On Deep Learning

With the increasingly serious problem of aging in China,society's demand for household service robots is becoming more and more urgent.However,the current robotic vision grasping technology for unstructured environments is not yet mature,which limits the popularity of service robots.Therefore,studying how robots achieve autonomous grasping in complex scenarios has become one of the core technologies of intelligent robots.This article focuses on the common scenes of stacked objects in daily life,combines with the theory and technology of deep learning,and focuses on the research and realization of robot grasping sequence recognition and grasping area detection.The main research of this article is as follows:(1)Regarding the problem of low accuracy of current manipulation relationship inference algorithms,a new manipulation relationship inference algorithm based on VLMRN neural network model is proposed.This algorithm first uses a single-order and multi-layer object detector SSD to obtain object image position information;then,improves the situation where only visual features are used as inference clues,extracts the position features of object pairs as supplements for visual features;finally,experimental verification is performed on the VMRD dataset,the results show that the algorithm effectively improves the accuracy of manipulation relationship of identifying object pairs and enhances the model's inferential capability.(2)In terms of the problem that the current robot target grasping area detection algorithm cannot take into account both detection accuracy and real-time,a new real-time robot target grasping area detection algorithm based on SE-Retina Grasp neural network model is proposed.This algorithm first extracts the position and angle of the grasp frame based on the first-order target detection model Retina Net;secondly,it fully integrates the feature information of the high and low layers to enhance the detection performance of the small grasp frame;finally,experimental verification is performed on the Cornell dataset,the results shows that the algorithm improves the efficiency of grasping detection while achieving higher detection accuracy,and meets the requirements of real-time detection.(3)Combining the two algorithms mentioned above,a grasping method for stacked object scenes is realized.The method first generates the manipulation relationship tree by using the results of the manipulation relationship inference algorithm;and uses the deep post-order traversal algorithm to obtain the grasping order of the current scene;secondly,intercepts the current object image area as the input of the grasping area detection algorithm to obtain the grasping area of the grasped object;then,the robot performs grasping according to the coordinate-transformed image grasping position;finally,an experimental verification is carried out on the simulation environment,and the results come out the method which can grasp the target items in the correct order.