Object Grasping Area Detection For Manipulator Desktop Operation

In the context of the leapfrog development of the global economy,human production mode is too inefficient,long duration,high risk,high error rate,vulnerable to environmental impact and other defects,intelligent robots applied in manufacturing,machinery and other industries emerged at the historic moment.The robotic arm grasping is the basic operation of moving and transporting the target object when the robot performs complex tasks,and the grasping detection is the core problem of the manipulator intelligent grasping system.In order to improve the recognition performance and action accuracy of the robot and enhance the interaction ability between the manipulator and the surrounding environment,this paper makes an in-depth study on the detection of the 2D plane grasping area of the manipulator facing the desktop.Using computer vision technology,a grasping detection method with high accuracy is proposed.The research work mainly includes:(1)When detecting two-dimensional plane target objects,in order to reduce the influence of target and background or irrelevant objects,reduce the analysis range of object grasping position in the picture and facilitate the determination of target object category information during orderly grasping,a target detection model based on yolov3 is proposed.This paper improves the loss function on the basic structure of yolov3,optimizes the size of anchors box,and uses the data enhancement method to expand the data set for training.Compared with the unoptimized yolov3 model,the better average percision and detection speed are obtained,and the improved detection model is used as the recognition model before grasping pose detection.(2)When detecting the grasping pose of two-dimensional plane targets,in order to improve the accurate estimation of grasping angle,a grasping detection method based on circular smooth label(CSL)is proposed.Firstly,the network model is improved and the angle processing module is added.Secondly,the boundary discontinuity caused by the One-Hot coding method based on regression and classification in the design of angle loss function is analyzed,also in order to improve the error tolerance of the model to adjacent angles of ground truth,CSL method is introduced to improve the loss function.Finally,the training is carried out on two capture data sets.The experimental results show that this method further improves the detection accuracy compared with other grasing detection algorithms.(3)Using UR5 robot arm and RGB-D camera,an "eye to hand" robot arm autonomous object grasping system is built,which verifies the performance of the detection method proposed in this paper applied to the robot arm automatic grasping platform.Firstly,the hand-eye calibration is used.Secondly,a series of grasping experiments on daily necessities are carried out on the platform.The success rate of the system is 83%,which shows that the detection algorithm proposed in this paper is feasible to grasp unknown objects in real scenes.