Design Of Robotic Grabbing System Based On Point Cloud Deep Learning

With the continuous advancement of China’s "Smart Manufacturing 2025" strategy,traditional manufacturing is facing transformation and upgrading.Manufacturing robots are increasingly doing industrial robots.However,in actual production applications,industrial robots can usually only perform actions according to pre-planned paths through teaching and programming,and lack the ability to sense the surrounding environment.Although 2D vision systems can improve the industrial robot’s perception to a certain extent,Still can not meet the needs of intelligent manufacturing.In recent years,deep learning has been used more and more in the industry with its powerful fitting ability,especially in the field of computer vision.This article studies the application of robotic gripping parts.A robotic grabbing system based on an infrared binocular structured light camera was designed and tested.The first chapter introduces the research background and significance of this article,then analyzes the development status of robot grabbing technology,and finally discusses the current research status of object pose detection combined with deep learning.Then summarize the research status to highlight the necessity of the work in this paper.The second chapter analyzes the system requirements and establishes the basic framework of the robot grasping system.The working principle of several 3D cameras is analyzed,and an industrial infrared binocular structured light camera is selected according to requirements,and the camera parameters are given.Then it was determined that the UR10 robot was selected as the robot of this subject,and the corresponding aero-mechanical gripper was configured.The third chapter studies the point cloud data acquisition and point cloud preprocessing.First analyze the monocular and binocular imaging models of the camera,and perform monocular and binocular calibration on the camera.After converting the data obtained by the calibrated camera into a point cloud,the pre-processing issues such as point cloud filtering,segmentation,and downsampling are studied.The filtering mainly removes outliers in the point cloud;segmentation is mainly used to separate the part data and the background data of the workbench.The workbench background plane is calculated according to the random sampling consistency algorithm,and the plane is used as a boundary to obtain the part point cloud.Point cloud data.Downsampling is mainly used to reduce the point cloud density and improve the subsequent calculation speed while retaining the key information of the part point cloud as much as possible.This article implements a voxel-based point cloud downsampling algorithm to downsample the point cloud data.In Chapter 4,an point-to-point robot grabbing detection model based on the Point Net model is designed.First,an overall point cloud deep learning network framework based on the Point Net model is proposed,and then the principles of the Point Net network model are analyzed in detail,including the STN model and the perceptron model.The point cloud data is out of order and multiperspective characteristics are targeted for design.The characteristics of point cloud features are extracted to build a robot grab detection algorithm for this paper.In Chapter 5,a three-dimensional vision-guided robotic grabbing system including a PC,an infrared binocular structured light camera,a robot,a pneumatic gripper,and a control cabinet is built based on ROS under Linux,and the production of a point cloud data set is explained.Finally,the robot gripping system is tested.Chapter 6 summarizes and prospects the design work of the robot grabbing system based on point cloud deep learning.