| In view of the low level of intelligence in the application of traditional industrial robots in the manufacturing industry,it is impossible to meet the production and assembly requirements for customized,single-piece and small-batch products.There is an urgent need to develop intelligent industrial robots for automated assembly to achieve autonomous perception,decisionmaking and execution.Ability.This paper takes this as the research goal,based on 3D machine vision technology,to study the problem of parts recognition and positioning for robot assembly scenes;based on deep reinforcement learning algorithm,develop the autonomous shaft hole assembly ability of Sawyer robot,and improve the intelligence of robot assembly.The main research work of this paper is as follows.(1)Part recognition and localization based on 3D point cloud.The part recognition method based on point cloud global features realizes the cognitive demand for classification of each unknown part in the robot operation scene;the part recognition and localization method based on point cloud local features realizes the detection of specified specific target parts from multiple part visual point clouds and their localization.By task simplification,it can be used for the localization problem after point cloud global feature-based part recognition.Among them,in the point cloud global feature-based part recognition,the shape description component is increased for the problem that the VFH(Viewpoint Feature Histogram)global feature descriptor is not strong in describing the part shape and the recognition effect is sensitive to the shooting perspective,and a secondary recognition scheme is designed to improve its recognition accuracy.The 3D CAD model of the part is established,and the model point cloud feature database is constructed;the k-nearest neighbor search algorithm based on kd-tree is implemented to retrieve and recognize the visual point cloud of the unknown part in the library.Simulation experiments show that the initial recognition of parts with different morphology can achieve 100%correct recognition rate;when the initial recognition of similar parts in a special viewpoint generates confusion,the secondary recognition through the fusion of point clouds of dual viewpoints finally also achieves 100%correct recognition rate.Part recognition and positioning based on point cloud local features can be realized from the specific needs,based on point cloud SHOT(Signature of Histogram of Orientation)local feature matching and Hough voting method,to identify the required target parts from the scene visual point cloud Corresponding visual point cloud,and achieve precise positioning.Through simulation experiments,the target recognition and positioning method based on the local features of the point cloud can accurately identify the target shaft and hole from the scene visual point cloud of multiple parts,and the average root mean square error of registration positioning is 9.8805e-7m,the average absolute translation error and absolute rotation error are 4.9950e-4m and 2.528e-2rad respectively,and the algorithm takes an average of 2229ms,which meets the precision and real-time requirements of robot assembly.This method has an average root mean square error of 6.0267e-10m for shaft hole positioning after global feature recognition,and an average time-consuming of 739ms.The positioning accuracy and real-time performance are further improved.(2)Robot autonomous shaft hole assembly simulation based on deep reinforcement learning.Based on the robot operating system ROS(Robot Operating System),establish TD3(Twin Delayed Deep Deterministic Policy Gradient)deep reinforcement learning algorithm network structure and data interaction framework;based on the Gazebo simulation environment,establish the corresponding Sawyer robot shaft hole assembly reinforcement learning environment,design State space,action space,environment initialization function,design reward function based on the force and pose information of the assembly process.Using a shaft hole with a nominal diameter of 30mm,a fit gap of 0.1mm,and a chamfer of 1.5mm,the joint algorithm and the environment are used for shaft hole assembly training.After 1762 rounds of training,the model converges.Import shaft holes with different nominal diameters,clearances and chamfers into the simulation environment for simulation experiment verification.The results show that for shaft holes with nominal diameters of 20mm and 40mm,minimum clearance of 0.08mm,and chamfers of 1mm,the comprehensive success rate of the assembly simulation test can reach More than 96%,and the average time spent in the assembly process is less than 4.52s,which verifies that the robot autonomous axis hole assembly based on deep reinforcement learning algorithm has good robustness and meets the assembly real-time requirements.(3)Robot target recognition,positioning and autonomous shaft hole assembly experiments.Build the Sawyer robot assembly experiment scene,perform hand-eye calibration based on the KinectV2 camera,obtain on-site visual point clouds,and perform preprocessing operations such as point cloud filtering and segmentation.The part recognition and positioning experiment based on the three-dimensional point cloud was completed.For the shaft hole with a nominal diameter of 30mm,a fitting gap of 0.1mm,and a chamfer of 1.5mm,the automatic process from robot grabbing to assembly was realized. |
PDF Full Text Request