Research On Grasping Technology Of Service Robot Based On Deep Learning Object Detection

After decades of development,the robot grasping technology in the industrial field has made grade progress and extensive application.However,the service robots which can achieve precise grasping in daily life are still in the stage of laboratory research stage and have not yet widely used in daily life.As China begins to enter an aging stage,society's demand for service robots is increasing.It is one of the necessary conditions for a service robot to accurately grasp the daily necessities such as tea cups,drink bottles,etc.in the daily environment to be able to care for the disabled and the elderly.Compared with industrial robots,the working environment of service robots is more dynamic and complicated.The previous researchers mainly use the 3D model based on the object to grasp the robot,but it is difficult to meet the needs of the service robot to grasp the unknown object accurately.Therefore,this topic studies the grasping technology of service robots.Firstly,the grasping detection method based on deep learning is used to build grasping detection network of Grasp-YOLO,and determine the position and pose of the target to be grasped;then the robotic arm grasps according to the position and pose information.The main contents are as follows:Firstly,the basic network structure of convolutional neural networks is studied,and several typical convolutional neural networks are explored.The object detection method is studied,and the advantages and disadvantages of Faster R-CNN,YOLO,SSD are analyzed and compared through experiments.Experiments show that the YOLOv3 algorithm has a faster detection speed and a higher detection accuracy.YOLOv3 network is used as the backbone network for subsequent grasping detection.Secondly,we research and determine the robot's grasp parameter representation and grasping detection evaluation method,in-depth study of the YOLOv3 network,and improve the YOLOv3 network according to the characteristics of the service robot,and construct a grasp detection network model:Grasp-YOLO network,which uses the regression method to directly output the grasp parameters.A sample dataset of grasp detection based on living items is established,and the Grasp-YOLO network was trained on this sample dataset.The experimental show that the constructed method has a higher accuracy.Finally,the experiment platform of 6 DOF robotic arm grasping is built.Use the Microsoft Kinect v2 depth camera to collect the image of the target object,Grasp-YOLO network to process the image and output the grasp parameters,Move It! carries out the motion planning of the robotic arm.The Arduino UNO development board and the PCA9685 driver board drive the servo of the robotic arm to complete the grasping operation.The entire experimental platform uses the ROS system to connect all software and hardware.On this platform,a robot autonomous grasping experiment is carried out.Experiments show that the constructed grasping detection network can effectively improve the success rate of grasping,and verify the feasibility and effectiveness of the Grasp-YOLO network on the service robot autonomous grasping platform.