Research On Generative Grasp Detection Method In Unstructured Environment

Grasp detection in unstructured environments refers to the use of computer vision techniques to detect the grasp position and pose of various objects in unseen environments,in order to complete subsequent grasping tasks.The generative approach,which uses fully convolutional networks to fit grasp poses,is one of the main methods of grasp detection.Unreasonable grasp point selection and inaccurate grasp direction are the main reasons for grasp task failure.With the generative grasp detection method as the background,this paper aims to study methods to improve grasp success rate from two aspects: grasp point and grasp direction.This has important implications for the application of intelligent robots in factory and home environments.The main work of this thesis is as follows:(1)A 4Do F grasp detection method based on angle map continuity and quality map gaussianization is proposed to solve the problems of fragmented angle labeling and difficulty in locating optimal grasping points in grasping images using the 4Do F grasping method.First,the grasp angle map is split into multiple submaps according to angle values,and the morphology method is used to filter out fragmented labeling in the submaps to improve the continuity of the grasp angle map.Then,Gaussian function is used to optimize the grasping quality map,and the center of the annotation region as the optimal grasping position to improve the positioning accuracy of the optimal grasping point.Finally,a AGGDN grasp detection network is built using Res Ne Xt modules to fuse grasp point and grasp direction attention maps.Experimental results show that the detection accuracy of AGGDN on the Jacquard and Cornell datasets is 94.4% and 95.5%,respectively,and the single detection time is 11 ms.AGGDN can improve the grasping position and pose detection ability of complex objects,and has good real-time performance(2)A 6Do F grasp detection method based on semantic instance reconstruction fusion is proposed to solve the problems of distinguishing multiple adjacent objects and fitting high-dimensional poses in 6Do F grasp methods.First,a regularization constraint based on multi-task learning is used to add a semantic instance reconstruction branch to complete implicit 3D reconstruction of the foreground and predict the central coordinates of each foreground point belonging to the instance by voting,in order to distinguish adjacent objects.Then,a high-dimensional pose dimensionality reduction learning method is proposed,using two orthogonal unit vectors to decompose the 3D rotation matrix,to improve the accuracy of pose learning.Finally,a SIRGN grasp detection network is built based on the Conv ONet 3D reconstruction network structure.Experimental results show that the grasp success rate of SIRGN in Packed and Pile scenes is 89.5% and 78.1%,respectively.SIRGN has a better fitting effect on grasping points and grasping directions,which can improve the robustness of grasping.(3)A adaptive grasping system is designed and implemented to address the problem of fixed workspace for traditional 6Do F grasping detection methods.A real environment robot grasping system is built to conduct grasp experiments in a real environment using AGGDN and SIRGN.The results show that both of depth map and TSDF data have small domain gap in simulation and real environments,and the proposed method can be directly transferred to real environment.By combining AGGDN and SIRGN,an adaptive grasping system is established for non-fixed workspace.The AGGDN method,which has a wider search range,is used to detect graspable areas and calculate workspace coordinates,and SIRGN is employed to detect the grasp pose and complete the grasp.The adaptive grasping experiment results show that the robot can find the object to be grasped more flexibly and use SIRGN to complete the grasping task efficiently.