Research On Grasp Detection Of Industrial Robot Based On Deep Learning

Robot grasping is an important part of intelligent robots,and is widely used in many fields,such as manufacturing,logistics,services,food,and medical treatment,etc.With the rapid development of robotics and artificial intelligence,the intelligent robot grasping has become a hot spot for applications and research in the field of academia and industry.At present,the combined mode of AI + robotic arm provides robots with a variety of grasping solutions,such as visual servo guidance,reinforcement learning grasping,grasping pose generation,6D pose estimation,etc.However,In the multi-object sorting scenario,there is a situation of partial occlusion and stacking,the accuracy,real-time performance and generalization ability of algorithms still need to be improved now.In order to solve the above problems,this paper takes the grasp detection problem as the research object,proposes a grasp detection network based on Faster-RCNN,builds an intelligent robot grasp system,and conducts relevant experimental research to verify the effectiveness of the network and the system.Firstly,for the multi-object sorting work scene,the framework of the robotic grasping system based on Kinect was established,and the structure and characteristics of the system control execution unit,such as the UR5 robot arm and gripper,were introduced in detail;and the construction of the Kinect vision system was mainly studied.The mathematical principles of RGBD camera calibration,image matching and hand-eye calibration are analyzed,and related experiments are performed to calibrate the camera parameters and determine the relative coordinate transformation between the robot arm and the camera.Secondly,this paper discusses the technical foundation of convolutional neural network in detail,and establish the theoretical basis for the subsequent multi-task convolutional neural network.Aiming at the requirements of multi-category and multi-object grasp detection tasks,a mathematical model of object grasping was established,a network dataset containing public data and self-made data was constructed,and an improved Faster-RCNN-based grasp detection network was proposed.This paper adds the rotation anchor boxes and a new grasp detection task branch.The network inputs the preprocessed RGB image,and outputs the object class and the rectangular grasp parameter.Experiment results show that the network has good performance under the public evaluation metrics.Finally,this paper establishes a software framework of grasp system,which includes a hardware layer,an algorithm layer,and a user interface layer.In order to achieve the simulation experiment platform under the ROS,this project uses Solid Works to export the URDF model of each component of the system,configures Move It! trajectory planning library and move?group node,and configures Gazebo simulation environment and ros-control control middleware,establishes the connection between simulation,controller and transmission,and builds xacro modeling file,and finally achieves the joint simulation of ROS and Gazebo;In addition,this paper performs object detection and virtual grasp experiment,in the aspect of detection experiment,the network's prediction accuracy is 90.6%,and the detection speed is 8.22FPS;in the aspect of virtual grasp experiment,the average detection accuracy is 90.5%,and the average grasp success rate is 83.2%.The effectiveness of the grasp system was verified through experiments.