Research On Industrial Robot's Graspling Method Based On Depth Learning And Binocular Vision

With the progress of science and the continuous improvement of intelligent manufacturing requirements,intelligent industrial robotic arms play an increasingly important role in the industry.Intelligent robotic arms which integrate advanced technologies such as mechanics,sensors,artificial intelligence and automation can enhance productivity,improve product quality and save a lot of manpower.In recent years,the number of intelligent robotic arm is increasing rapidly.The complex application scenarios raise a new challenge for the industrial robotic arms to improve the stability,accuracy and flexibility.What's more,because of the development of computer vision and artificial intelligence,the machine vision and the artificial intelligence play more important role in the intelligent industrial robot applications.Vision sensors are a critical component of intelligent robotic arms,it can measure the environment changes and provide visual information in the scene without contacting.The widely application of vision sensors makes the robotic arms can be flexibly plan and controlled according to the feedback in the real time.Robotic grasp as an important part of intelligent robot research covers many techniques such as grasping perception and motion control.During these years,the most promising learning algorithm,deep learning has achieved a great success in the field of pattern recognition.However,it is still in the initial stage to apply deep learning to robotic control.This thesis studies and discusses the important technology of intelligent robotic arms and deep learning,focuses on the trajectory planning and grasping detection and put forward original proposals.In the aspect of trajectory planning,a planning method based on binocular vision is proposed in the thesis.The robot arm theory and binocular stereo vision theory are used in revising the trajectory according to the state of the joint of intelligent industrial robot and the visual feedback.This method improves the absolute positioning accuracy of the intelligent industrial robot,and it will be the basis for our further study about the deep learning based control of the intelligent robot.In the aspect of grasping discriminating,this thesis proposes a deep learning based method.This thesis uses the ResNet network model to automatically get the image features and effectively improve the accuracy of discrimination.In the real experiment of capturing electronic components,the thesis presents a simple transformation matrix method of converting the camera coordinates frame to the world coordinate frame,and uses the deep neural network to identify the parts to be grasped.In the theoretical aspects of deep learning,the thesis summarizes the relevant technologies for the development of deep learning,and designed an image generation network to analyze the impact of network structure on the quality of image reconstruction.These relevant robotic grasp methods achieve good results in the experiments.The research results have high theoretical significance and application value.