Research On Flexible Assembly Of Robot Based On Visual Localization And Force-pose Image Learning

Assembly is important in the production of products,Robot assembly methods based on instruction or off-line programming are difficult to meet the complex assembly tasks and ensure the flexibility of assembly.By installing visual sensors and designing control algorithms,the automaticity of robot assembly can be greatly improved.However,problems such as large visual location error,poor flexibility of assembly and workpiece shielding still exist.In this case,force control can be introduced.In order to solve these problems,this paper combines deep learning technology to complete flexible assembly of robot by learning force/ torque and pose information.The main research contents of the paper are as follows:(1)A flexible robot assembly method based on vision and force-pose image learning was proposed: Based on the visual information to get workpiece pose,complete the alignment between the workpiece;In view of the visual error caused low alignment accuracy,difficult to guarantee the assembly quality and flexibility,the flexible assembly method based on force-pose images learning was proposed,the force/ torque and pose data in contact assembly process are converted into force-pose images,and motion labels were set,the complex force-pose coupling relationship in assembly was learned by image features,and then assembly was completed according to the classification results of force-pose images in contact assembly process,and assembly flexibility was ensured.(2)In order to obtain force-pose image samples,a sample collection process including motion direction determination algorithm,force-pose image generation rules and difficult/easy sample balance method was proposed.The force/torque information and the corresponding pose information in the assembly process of different initial poses were collected for 3000 times,a force-pose image dataset consisting of 127,741 images and their corresponding motion labels was generated,then different classification models are used,the classification loss and accuracy of different models are analyzed and the optimal model is selected as the final assembly motion direction classification model for contact assembly experimental,the experimental results show that the model can correct the alignment error of workpiece and complete contact assembly successfully,the success rate can reach 96.7%.(3)A vision-based workpiece location algorithm is studied,to solve the location and alignment problems between workpiece.Firstly,the rgb data and depth data are read by depth camera,the mathematical principle of camera imaging is derived,and the internal parameters and distortion of the camera are calibrated,second,the camera was fixed at the end of the manipulator to complete the hand-eye calibration,to obtain the hand-eye transformation matrix,then the depth image calibration was performed,finally,the trained SSD model and the marker detection algorithm were used to obtain the alignment pose of the workpiece.(4)Staubli Tx90 manipulator,force sensor and depth camera were used as experimental equipment,and the robot flexible assembly platform system is built based on ROS.The assembly between RJ45 connector as the research object,the workpiece alignment was achieved based on the workpiece visual location algorithm,then the force-pose image is generated in real time during the assembly process,as the input of the assembly motion direction classification model to complete the contact assembly.Experimental results show that the method can control the contact force within 30 N and the contact torque within 0.15N·m during the assembly process,and the flexible assembly can be successfully completed.