| Assembly is an essential part of industrial manufacturing,and the workload of the pro duct assembly loop accounts for 20%-70% of the entire product manufacturing workload,s o it is significant to improve the automation and intelligence level of the assembly process.In recent years,collaborative robot assembly has been introduced as a new generation of pr oduction mode into the intelligent manufacturing system,and the combination of artificial i ntelligence and robotics has become the key to solve the problem of human-machine collab orative assembly.This thesis focuses on the theme of collaborative human-machine assemb ly in dynamic scenarios,and investigates the target recognition,flexible grasping,robot ass embly skill learning,robot dynamic obstacle avoidance and path planning in the robot asse mbly process to provide theoretical and technical support for the application of collaborativ e human-machine assembly in industry.The paper relies on the National Natural Science F oundation of China and the Natural Science Foundation of Guizhou Province,and the main research contents and innovation points are as follows:(1)A Mask R-CNN-based target detection model is proposed for the recognition and s egmentation of assembly parts based on low accuracy,slow recognition speed and inability of part segmentation of robot targets in complex assembly environments.The model optim izes the backbone network of Mask R-CNN,adds the SE module in the feature extraction n etwork to improve the part segmentation accuracy,and conducts comparison experiments w ith the traditional algorithm based on the VOC tes-dev public dataset,and the experimental results show that the model improves the accuracy rate and recognition time by 12.15% and0.115 s,respectively.in order to verify the model’s effectiveness in assembly part in order t o verify the effectiveness of the model in assembly part detection,the group built its own as sembly part dataset-RAData,and introduced data augmentation and migration learning met hods to enhance the learning efficiency of the model on RAData while increasing the numb er of samples and using the prior knowledge in the source model.The experimental results show that the Mask R-CNN model improves the F1 value and recognition time by 10.64%and 0.145 s,respectively,compared with the traditional model,which can effectively realize the recognition and segmentation of assembly parts in assembly scenarios and provide sup port for subsequent tasks such as robot grasping and robot assembly.(2)To address the problems of many types of parts and variable part position patterns i n the assembly environment,which easily lead to low robot grasping accuracy and poor gra sping flexibility:(1)A robot grasping method based on Canny operator contour detection th eory is proposed to achieve reliable robot grasping of multiple types and irregularly placed parts.To further improve the grasping accuracy,a B*Gaus modeling method is proposed,a nd in the robot grasping comparison experiment,the grasping accuracy reaches 94% when using the B*Gaus model,which is 7% better than the traditional Gaus method.(2)A 3D poi nt cloud-based robot grasp detection model(MR-GPD)is proposed,which combines the ta rget segmentation theory of Mask R-CNN network into the robot grasp pose detection algor ithm(GPD),and constructs a robot grasp candidate screening mechanism by designing the MR-GPD network structure and model loss function to achieve flexible grasp of contact an d stacked parts in assembly scenarios.To verify the effectiveness of the MR-GPD model,a comparative experiment of robot grasping under unstructured assembly scenarios is design ed.The experimental results show that the grasping success rate of the MR-GPD model rea ches 95.86%,which is 23.28% higher than that of the traditional GPD method,providing ef fective grasping of assembly parts.(3)Aiming at the problems of complex assembly process,difficulty in learning robot a ssembly skills and low efficiency,an interactive robot reinforcement learning(RLFI)frame work is proposed.This framework is based on HMM model and Actor Critical learning mo del.By building a strategy network,designing a fuzzy logic reward and punishment functio n,and establishing an optimization mechanism based on prior feedback,the robot can main tain efficient and stable learning efficiency.Simulation results show that the proposed fram ework has excellent performance in robot technology learning.In order to verify the effecti veness of the proposed RLFI framework in the actual man-machine collaborative assembly,a mechanical product assembly scene composed of five sub tasks was designed.Under the RLFI framework,the robot learned the skills of part grabbing,assembly sequence,assembl y path and assembly action,and effectively completed the scheduled assembly tasks on the built assembly physical platform.(4)In order to solve the problems that robots can not perceive the environment inform ation,lack the ability of dynamic obstacle avoidance and path planning in random and unce rtain assembly scenes:(1)A constraint based robot anti-collision strategy is proposed.This strategy uses the nearest neighbor node(KNN)algorithm to cluster the 3D point cloud ima ges of the robot assembly scene,calculate the shortest distance between the obstacle and th e robot,and control the joint movement of the robot arm through the constraint frame to ac hieve obstacle avoidance.(2)In order to solve the problem that the Basic RRT algorithm is inefficient when it faces dynamic obstacles,an improved RRT path planning algorithm is p roposed and verified in Gazebo simulation experiment.Finally,through robot kinematics m odeling and depth space calibration,a physical experiment platform for robot assembly dyn amic obstacle avoidance is built,and several groups of robot assembly obstacle avoidance e xperiments are designed.The experimental results show that the robot can realize dynamic obstacle avoidance in different assembly scenes,and generate curvature smooth paths,whic h further improves the assembly ability of the robot in actual assembly scenes. |
PDF Full Text Request