Research On Six-axis Robotic Grasping Detection Based On Deep Learning

With the development of social productivity and the rise of the wave of artificial intelligence,robotic arms have been widely used in various industries and have become an important part of the future intelligent manufacturing industry,of which visual grasping is a hot issue in field of robotic arms.Traditional industrial robotic arms require high requirements for accurate position and 3D model information of objects during operation,which is difficult to be applied to unstructured work scenarios.In order to improve the accuracy,autonomy and safety of the intelligent robotic arm in grasping objects,this study conducts a research on visual detection and bit pose estimation for the plane grasping task of the six-axis robotic arm in unknown scenes,which involves the modeling of the robotic arm system and the calibration of the visual system,the object detection grasping method based on convolutional networks and the pixel point grasping method,and considering the real-time nature of the grasping task,the two grasping The detection methods are improved considering the real-time nature of the grasping task.The main research and results of this study are as follows:(1)Calibration test and analysis of robotic arm vision systemStarting from the hardware and software of the robotic arm grasping system and the functional requirements that should be satisfied,the visual grasping platform of the robotic arm is built,and the modeling and calibration of the visual system are realized,including the calibration of the internal reference of the depth camera and the hand-eye calibration between the relative positions of the camera and the robotic arm,and a more accurate model of the visual system of the robotic arm and the corresponding relationship matrix is gained,which provides a basis for the subsequent research of the visual grasping method.(2)Research on robotic arm object grasping system based on the object detection taskBased on YOLO series object detection,real-time detection of target objects in the scene is achieved.The rotation-free planar grasping pose estimation combining object detection and depth images is investigated,and the pose information of the target object is obtained based on the bounding box and centroid depth information.Considering that the grasp detection task needs to balance detection speed and accuracy,and the object detection dataset in this study is small,the YOLO network is improved to increase the detection speed of the model while ensuring its detection accuracy.The results show that a deeper network should be used to process large-scale datasets,while small datasets should pay attention to the overfitting problem,and the object detection-based grasping method has the characteristics of simple operation and easy implementation.(3)A study of pixel-level grasping detection under planar grasping conditionsDue to the limitations of the rectangular frame grasping method,a pixel point representation-based planar grasping pose estimation method is studied,and a grasping detection network is designed based on the idea of semantic segmentation,which takes a360×360 single-channel depth image as input and outputs a grasping pose of the same size consisting of grasping point confidence,grasping width and grasping angle.In addition,the grasp detection network is improved to enhance the performance of the model.In addition to adding the attention mechanism,an image segmentation network is used instead of the main structure of the network,and the grasp detection experiments are conducted for objects in planar scenes using the directional rectangular frame grasp method and the improved pixel-level grasp method,respectively.The results show that the inference time of the pixellevel grasping network is still in the millisecond range after adding the attention mechanism,and the maximum grasping confidence,which indicates the grasping success rate,is improved by 7.2%,which is a good balance of speed and accuracy.Compared with the rectangular frame grasping method,the pixel-level grasping method has better prediction results,and finally the pixel-level grasping method is applied in the visual grasping platform to achieve the grasping operation of the target object.