Design And Development Of Industrial Robot System Based On Deep Learning

In the production and manufacturing process,industrial robots can greatly save labor costs and efficiently and reliably complete given tasks.Currently,robots mostly use traditional algorithms to perceive the surrounding environment and complete specified actions through manual teaching.With the normalization of the epidemic situation,traditional manufacturing is gradually transforming into intelligent and automated,which requires robot systems to have stable and reliable environmental perception capabilities and to be competent in various complex and changing real environments.Based on RGB and RGB-D visual sensors,this thesis constructs a robot grasping system platform,creates a corresponding dataset,completes the estimation of the position and posture of the object,and develops a corresponding grasping system.The main works are as follows:(1)The calibration problem of visual sensors and robot systems is studied.Firstly,the imaging principle of the camera is analyzed,and the single target calibration of the camera is understood.Secondly,the hand-eye calibration method of the robot is studied,and finally,the internal and external parameters of the camera and the transformation matrix from the camera coordinate system to the robot base coordinate system are obtained through experiments.(2)A robot system that simultaneously estimates the 6D pose information of multiple target objects from a single view and realizes accurate grasping is proposed.In view of the problems such as weak texture,multiple types of target objects,complex grasping environment,and difficult dataset production in object grasping applications,high-quality training data can be obtained through synthetic dataset without manual annotation,which greatly reduces the workload of annotation.Using the UR5 robot,grasping experiments were conducted in a real scene,which has high practicality.(3)A new method for estimating the 6D pose from a single RGB image is proposed.In view of the poor performance of existing methods in occlusion and truncation,the network uses RANSAC voting to predict the invisible key point pixels on the object through visible pixels on the object.The method was evaluated on public datasets,and a dataset was created using real objects.Stable grasping was achieved using this method.(4)A precise and efficient sorting method for robot grasping in cluttered and occluded environments is designed.The system is based on RGB images and depth images.Firstly,the labels and surface center points of all objects to be grasped are estimated.Secondly,suitable objects to be picked up are selected based on depth and exposed area.The pick-up point is calculated through hand-eye calibration,and the pick-up direction is calculated by fitting the object’s depth point cloud.Finally,the robot arm is controlled to achieve six degrees of freedom sorting.