Intelligent Target Grabbing With Low-cost Eye-to-hand Robotic Arm

How to make robots work in industrial efficiently and cost-effectively has been the highlight of researchers all over the world.High-end robots with high precision and speed cost a lot,meanwhile middle-end robots are cheap and can't meet the needs because of their poor accuracy,speed and stability.This project starts research on low-cost robotic gripper system,making use of deep learning algorithm and reinforcement learning algorithm,to achieve the goal of higher speed,accuracy and intelligent crawling.This article takes the low-cost robotic arm grasping system as the subject,and focuses on three modules: the pose estimation module,the controller algorithm module and the target detection module of the robotic arm grasping.This article expands on the basis of the predecessors,first designed the overall framework and process of the system,and then designed the core program of the system.This article focuses on improving the grasping speed,accuracy and intelligence of the robot arm,and has designed four aspects for the above three modules.For the pose estimation module,the end-to-end robotic arm training requires a large data set.In this paper,the method of virtual data set and domain adaptation is adopted to reduce the cost of manual labeling.According to whether the pose of the camera relative to the robot arm changes during pose estimation,a pose estimation method based on semiautonomous calibration is designed in this paper to reduce the dimension of pose estimation parameters and improve the accuracy of pose estimation slightly.Next,greatly speed up the system's crawling speed.For the controller algorithm module,reinforcement learning and traditional control have their own characteristics.For the control of the manipulator,this paper analyzes and compares the two controller algorithms.Aiming at the problem of reinforcement learning in the control of the robotic arm,as the number of steps increases,the difference between the coordinates of the end of the robotic arm and the target coordinate does not continue to decline,this paper studies the algorithm of the reinforcement learning controller based on the error Acting on the reinforcement learning controller,the algorithm accelerates the network convergence to some extent.For the target classification and detection module,in order to facilitate testing,this article first made its own data set.For the difference of training samples,this article introduces the definition of sample priority,using the classification loss and regression loss of the picture as the sample priority,and increasing the number of trainings for the sample with larger loss,thereby improving the accuracy of the target classification and detection module.This system is an intelligent grasping system.In order to realize the grasping of the robot arm,an efficient coordinate conversion method is designed to realize the conversion of the coordinates of the target to be grasped from the pixel coordinate system to the robot arm coordinate system.In order to test the performance of the system,this paper conducts experiments on each module of the system,and finally conducts experiments on the whole system to prove the effectiveness of the design.