Research On Robot Grabbing Control Algorithm Based On Visual Pose Estimation

With the rapid development of “Industry 4.0” research projects,traditional industrial robots have been unable to meet the needs of modern production.In order to provide sufficient power for the sustainable development of industry,the intelligent development of industrial robots has become an urgent demand in China.In recent years,the achievements in the field of target detection have become increasingly rich.Using the object detection algorithm to control the robot and guide the robot to make the current action in time according to the surrounding environment has become the mainstream trend of machine intelligence.However,there are still some fundamental challenges,such as image blurring caused by motion and sharp position differences between objects,as well as control errors caused by long system delays.In addition,the shape features of the object in detection are mainly presented by two-dimensional images,which greatly reduces the shape information of the target.For the grasping robot,the two-dimensional image information does not provide sufficient shape feature information for the grasping robot.To address these issues,this thesis proposes a novel robot visual servo control pipeline based on deep detection network and spatial pose estimation.Firstly,we build a robot platform to realize the automatic grasping function of the object.The grasping robot uses the vision sensor to collect the video stream,analyze and reason on it,and transmits the visual data processing information to the control side.Secondly,we introduce a deep learning-based detection network called YOLOv3-tiny to perform accurate and real-time moving object detection.Thirdly,we design an advanced key point attitude estimation method to provide sufficient object 3D information for the robot to grasp.Finally,we use a double loop-based cascade PID algorithm to drive the robot system precisely.After the vision algorithm drive the robot to the specified position,the grasping mechanism works to realize the automatic grasping of the object.Extensive comparison and simulation experiments demonstrate the effectiveness on object detection and object pose estimation of designed control system.