Rasearch On Mechanical Arm Grasping Detection Using Deep Learning

As a new form of robots,intelligent robots are expected to replace humans in time-consuming,high-intensity and dangerous tasks in unknown environments.Grasping technology of the robotic arm has made significant progress and applications in the face of a structured environment and a single known target.However,in many work scenes,the environment is unknown and the work target forms are different.It's susceptible to changes in ambient light,object occlusion,etc.,which requires the robot gripping system to have high autonomy and robustness.In this paper,we use the powerful knowledge transfer and nonlinear fitting capabilities of deep learning technology to study the application of convolutional neural network in robotic arm grasping detection.Based on the convolutional neural network technology,two autonomous detection methods for the robotic arm grasping position and posture are designed,and the process of robotic arm autonomous grasping facing unknown environment and strange targets is completed.The main contents include the following:Firstly,the robot gripping scene and grasping system flow are designed.The representation of the grasping position and posture of the end gripper on the camera image is discussed.Secondly,the convolutional neural network technology is used to complete the end-to-end detection of the grasping pose of the gripper.A convolutional network model is designed to obtain the appropriate center position of object grabbing.Then,based on the grabbing center position,a grabbing angle and gripper opening width are obtained using a heuristic search method.Trained and tested on Cornell Dataset and got 87.5% grasp detection accuracy.Thirdly,for the end-to-end method that takes the entire image as input,the grasp detection is disturbed by background information and irrelevant object information.An object detection model for grabble objects is trained.The object detection process gives the approximate location of the object to be captured in the working scene,which not only helps reduce the impact of the background and irrelevant objects in the scene on the capture detection,but also reduces the image range of the capture analysis.And the object detection gives the semantic labels of each object in the scene,which is helpful for the planning of complex and orderly crawling tasks;Finally,using the output features of the object detection in the previous work,a two-stage rapid grasp detection method is designed.Considering the trade-off between accuracy and speed of grasp detection,combined with the output of the object detection model and the area analysis method,a two-level grasp detection convolutional neural network is proposed.Training and testing on the Cornell Dataset resulted in 92.8% detection accuracy and 1.1169 s detection speed.While achieving a high grab success rate,it also takes into account the real-time requirements of the robot grab system.