Research On Intelligent Learning Method Of Robot Grasping Process Based On Vision

With the continuous development of robotics technology and the intelligent demand of social production and life,robots have been widely used in various fields of society,from small children’s toys to large multi-axis robotic arms installed on the space station.Grasping is the basic and important function of the application range of the robot.To enable the robot to grasp intelligently,a good grasping detection result is the prerequisite for the smooth progress of the grasping task.However,in the unstructured environment with various types of objects,unstable placement,slanting,and dense arrangement,the robot’s target recognition and grasping ability will be reduced.In this paper,the object detection and grasp pose estimation methods based on deep learning were studied,and the multi-object recognition and grasp methods of robots were realized in the scene of diverse objects,disordered,and dense placement,and the robot grasp experiments were carried out in the relevant scenes.The main research work is as follows:(1)The overall framework of the robot visual grasping system is designed.The internal parameter calibration of binocular camera of visual perception system and hand-eye calibration of manipulator and camera are completed.(2)Based on the target detection network,the Retinanet changes the characteristics of the kanetah layer,replaces FPN of the original network with PANet,and adds the attention mechanism CBAM to the network,so that the network can better recognize and locate objects in dense scenes.In order to effectively separate the object from the background when the object’s aspect ratio is large and there is inclination and dense arrangement,the prediction of Angle is added in the prediction part of this paper,so that the horizontal anchor frame becomes the rotating anchor frame.By introducing KFIoU,which solves the boundary problem and square-like problem caused by the regression of rotating box,KFIoU can also make the measurement and loss consistent.Based on the self-built object dense scene data set,the multiobject recognition and location test experiment is carried out,and the test accuracy reaches91.8%,which is better than the original network,and verifies the effectiveness of the improved model.(3)The robot grasping detection method is studied.The Oriented Arrow Representation Model(OAR)was introduced to represent the grasping configuration of objects.In order to avoid Angle conflict and simplify labeling,the OAR model of the same position was merged into the Adaptive Grasping Attribute Model(AGA).AFFGA-Net pixel level grab pose estimation network was introduced,and some modifications were made on the basis of AFFGA-Net.Resnet50 network was used to replace Resnet100 of the original network to speed up the running rate of the network.Secondly,the hole rate of the mixed cavity convolutional layer of the network was set to 3.Avoid low information utilization and local feature loss due to large hole rate.The Cornell data set was also re-labeled,and the network model was trained and tested with the re-labeled data set,and the capture detection accuracy of 95.3% was obtained.(4)Build a robot grasping experiment platform,combine the target detection network with the grasping detection network,and complete the grasping algorithm design.A single object grasping experiment and a multi-object grasping experiment are designed,and the performance of the algorithm and the robot grasping reliability are tested and verified.The object detection and grasping detection method proposed in this topic can effectively realize object identification,positioning and pose estimation under the scene of object density,arbitrary inclination and random posture,which helps to improve the intelligent level of visual robot grasping and provides certain value for robot application in assembly,sorting and service and other related fields.