Research On Workpieces Unordered Sorting Method Based On Deep Learning 6D Pose Estimation

Driven by the "Made in China 2025" development strategy,industrial robots are gradually becoming intelligent and becoming an important part of modern manufacturing industry.Sorting task is one of the important applications of robots.How to realize the accurate sorting of randomly placed workpieces by robots has become a hot research issue.The existing methods mainly focus on the sorting of planar disordered objects.For the sorting of piled disordered objects,most of the methods adopted are based on template or voting method,which is not high precision,large amount of calculation and difficult to meet the industrial requirements.In this paper,a disordered sorting method based on 6D pose estimation of deep learning is proposed to realize automatic sorting of disordered placed workpieces,which has certain research significance and practical value.The main research work of this paper is as follows:(1)Aiming at the problem that 2D object detection method is difficult to obtain 6D position information of workpieces,a new method of workpiece location based on improved G2L network was proposed.Firstly,the segmentation network QueryInst is used to replace the YOLOv3 module in the original network to solve the problem of poor segmentation effect of point cloud when workpieces are stacked.Then,the channel attention mechanism SE is embedded in the G2L network translation positioning point cloud feature extraction module to improve the feature extraction capability of the network.Finally,a 3D model feature extraction network module is designed to complement the features and improve the prediction accuracy of the network in the case of object occlusion.The experimental results showed that the improved network average ADD increased by 0.62%on the LINEMOD data set,12.54%on the Linemod-occluded data set and 2.07%on the self-made workpiece data set,which proved the effectiveness of the improvement.(2)In order to solve the problem that the robot may collide during sorting and moving,an improved RRT*algorithm and collision detection model are proposed to plan the path.Firstly,the forward and inverse kinematics of the robot is analyzed,then the improved RRT*algorithm is used to explore the path,and then the path is optimized.Finally,the collision detection model is established and the angles of each joint of the robot are solved to obtain a collision-free and optimal robot motion path.Experimental results show that the improved RRT*algorithm is superior to both RRT and RRT*algorithms.Search success improved by an average of about 27%in simple 3D environments and about 32%in complex 3D environments.Meanwhile,experiments show that the proposed algorithm can plan a collision-free path for ABB IRB1200 robot.(3)A complete platform of disordered sorting system was built,and several sorting experiments were carried out in scenes without stacking and with stacking,and the experimental results were analyzed to verify the feasibility of the designed scheme and system,which can successfully sort the disordered placed workpieces.