| Mechanical arms are widely used in scenes such as workpiece sorting,handling,stacking,and metal processing.Traditional mechanical arms are mostly programmed to run repeatedly,which leads to poor flexibility.Based on visual servo control,the perception ability of mechanical arms is greatly improved,providing the basis for intelligent control of mechanical arms.However,the efficiency and accuracy of visual detection have become one of the difficult research points.In addition,with the continuous development of deep reinforcement learning theory,it has become possible to control the mechanical arm to interact with the environment,learn the optimal action,and then control the mechanical arm to complete the work task.Therefore,based on visual object detection and deep reinforcement learning control methods,this dissertation aims to establish an intelligent perception and control system for a robotic arm,in order to improve its level of intelligence.First,in order to solve the problem of high-dimensional input dimension of deep reinforcement learning,the dissertation designs an algorithm based on plane 3D pose estimation to explicitly express the pose of the object to be grabbed.Based on the selfmade VOC format data set,the real-time detection of the target to be grabbed is realized based on the YOLOv5 network through transfer training.Then,based on the color and shape features of the target to be grabbed,the background is segmented based on the average background method.Finally,the rotation angle of the object to be grabbed is calculated and the 3D pose information of the target object on the plane is extracted.Secondly,in order to solve the problem of low effective sampling rate in the initial training phase of the DDPG algorithm in the mechanical arm grasping control,a DMP teaching data generation module is designed to improve the sampling efficiency of the effective data of the network training in the initial phase.After analyzing the classical mechanical arm grasping control process,the visual detection information and the state information of the mechanical arm are integrated,and the input state of reinforcement learning is improved.Finally,an experimental platform is built to verify the algorithm in this dissertation.Firstly,the 3D pose extraction module and the control strategy module of deep reinforcement learning are trained separately,and the physical platform of the UR3 e manipulator is built.Based on the Moveit motion control tool library in ROS,the effectiveness of the visual detection module is experimentally verified.Secondly,the simulation environment of UR3 e is built in Gazebo,and the improved algorithm integrating DMP teaching data generation module is trained in this environment.By comparing with the original algorithm,the effectiveness of the improved algorithm is verified. |
