Research On Grasping Algorithm Based On Optimized ResNet Structure

Robot technology is a relatively new modern technology that crosses traditional engineering boundaries,and its appearance has greatly improved social productivity.Grabbing is an essential function for robots to enter the real world.For humans,identifying and grasping objects is a relatively simple task.But for robots,identifying and grasping objects is a very challenging task,which involves target detection,grasping pose determination,robotic arm motion planning and control and so on.With the convolutional neural network(CNN)model showing excellent results in many application fields,more and more scholars began to pay attention to the research based on the convolutional neural network model.If the scale of the neural network model is too large,the network will not work or overfitting problems will occur,and the calculation cost will be too high,and the number of connections will be redundant.Therefore,the neural network model needs to be optimized.In this paper,a robot gripping pose detection algorithm based on optimized ResNet-50 structure is proposed,and experiments are conducted in a simulation environment to verify the algorithm.First of all,this article established a robot gripping system framework.The framework was divided into three parts: object perception,grip detection a nd grip execution.It introduced the acquisition method of depth image,the principle of distance measurement of depth camera,the principle of pinhole imaging,and the mapping relationship between the color image space and depth image space obtained by depth camera.A method for representing the position and pose of object grabbing was established,and the coordinate conversion relationship between the pixel coordinate system of the image and the robot base coordinate system was studied.Secondly,this paper introduced the basic unit in the neural network model,and built a grab pose detection algorithm based on ResNet-50 structure.The Cornell data set was used to train the neural network model,and the multi-object data set was used as the test data set to evaluate the neural network model.The rec ognition accuracy rate of the grab rectangle reached 91.7%.Then,this paper studied the optimization of ResNet neural network model.The control variable method was used to conduct experiments,and the effects of various optimization methods on ResNet neural network model training were compared,and an optimization scheme with good effect was finally obtained.The grabbing pose detection algorithm built in this article added such an optimization scheme,and the final recognition accuracy of the grab rectangle was 96.5%,which exceeded the previous accuracy of 91.7%.Finally,in order to conduct simulation experiments,this paper built a simulation environment in gazebo,an automatic grasping system based on the ROS robot software architecture,and used Move It! to plan and configure the robotic arm.In this paper,the hand-eye calibration and internal parameters of the depth camera was used to obtain the position and posture of the object in the robot's base coordinate system,and several gripping experiments were carried out in the simulation environment.The experimental results show that the robot gripping system constructed by the subject has good robustness,and that the constructed gr ab pose detection algorithm has good performance.