Research On Part Recognition And Robot Grasping Based On Deep Learning

With the continuous development of technology,the application range of robots has become more and more extensive,and has entered many different fields such as home entertainment and industrial production.At the same time,it also brings some new challenges,such as requiring robots to adapt to environmental changes,the ability to work independently and higher performance.Intelligent robots should not only be able to perceive the environment,but also need to be able to interact with the environment.Among all these capabilities,the recognition and grasping of goals is one of the most basic and important capabilities,because it will bring enormous productivity to society.Deep learning has achieved significant success in computer vision applications.It can learn from the success of deep learning in the field of computer vision,using deep learning instead of traditional control methods in the field of part recognition and grasping of industrial robots,and then get the feature representation with better generalization ability,thus improving the part recognition and grasping algorithm.The performance compensates for the shortcomings of traditional methods.This paper first studies the conversion relationship between each coordinate system,the principle of camera calibration and the principle of robot hand eye calibration.Based on V-REP,a virtual environment for robotic grasping was built,which included the parts to be grabbed,the robotic arm and the camera.Two common manipulator motion planning algorithms,PRM and RRT,are studied to solve the motion planning problem in the process of robot grasping.A part identification network based on SSD target detection is established.By generating a series of fixed-size bounding boxes and classifying the categories of the parts contained in the bounding box,the final prediction result is generated by the non-maximum suppression.The data enhancement method solves the problem of too little collected data,and effectively improves the recognition rate of parts.A robot grasping network model based on deep reinforcement learning is established,and the task of grasping parts as a Markov decision process is established.The fully convolutional network is used to approximate the action-value function in Q learning,which selects the best grasping action by maximizing the Q function.Then,the performance test simulation experiment of the grasping part was carried out in V-REP.Finally,the experimental platform was built in the real environment,and the camera calibration,robot handeye calibration and other preliminary preparations were completed,and the robot's grasping parts experiment was carried out.The experiment verified the feasibility of the established model.