Research On Intelligent Grasping Technology Of Robotic Arm Based On Vision

With the development of artificial intelligence technology,robots become more and more intelligent,and the research of intelligent grasping technology of robotic arm,as the actuator for robot control and operation,is an important way to improve robot intelligence.The traditional robotic arm grasping is usually predefined for the grasping position,and the grasping scene is single and the grasping position is fixed,so when the environment or the position of the object changes,it is difficult for the robotic arm to achieve effective grasping of the object.Therefore,in this paper,we focus on the grasping technology of robotic arm,which is equipped with vision sensing device and combined with corresponding deep learning algorithm to effectively improve its autonomous grasping ability in complex environment.(1)A kinematic model based on the D-H parameter method is established for the Dofbot robot arm,and the conversion relationship between the coordinate systems of each joint is obtained through the forward and reverse kinematic analysis of the robot arm;the coordinate origin of the end-effector is randomly sampled by the Monte Carlo method,and the working space of the robot arm is simulated graphically by using the Robotics Toolbox.(2)By studying and analyzing the types of depth cameras and their principles,the Real Sence D455 depth camera was selected as the vision sensing device,and then the camera calibration experiment of the Real Sence D455 camera was conducted by using Matlab to obtain the camera internal reference;the online hand-eye calibration method based on Ch Aruco code was adopted by The Ch Aruco code-based online hand-eye calibration method is used to complete the hand-eye calibration experiments of camera and robotic arm by combining the camera internal reference with the "eye outside the hand" hand-eye model,and the conversion relationship between the base coordinate system of robotic arm and the camera coordinate system is obtained.(3)In order to solve the problem that the GR-CNN grasping detection method has the problem that the grasping pose is influenced by the image background area,we propose the RCAN network structure that fuses the CA coordinate attention mechanism with the residual network;in order to solve the problem that the gradient of the network disappears due to the non-negative value of the output of the Re LU activation function,we adopt the smoother Swish function as the activation function of the network;before each output The dropout layer is added before each output to reduce the risk of overfitting of the network and enhance the model expression capability.In the experimental stage,the improved network effectively avoids the problems of false detection,missed detection and low-quality grasping regions in the GR-CNN method for single-target object and multi-target object grasping pose prediction,and the improved network model achieves 98.0% detection accuracy and 25 ms fast inference speed in the evaluation of the network.(4)Simulation of grasping and real grasping experiments based on ROS robot operating system,building Gazebo simulation grasping platform,real-time prediction of grasping pose by grasping detection network,and completing simulation grasping experiments;using computer and Raspberry Pi 4B as the upper and lower computer of robotic arm intelligent grasping system,respectively,and completing the grasping of multi-target object scenes by using distributed control Experiment.The experimental results of simulated and real grasping show that after accurate kinematic analysis of the robotic arm and calibration of the vision system,combined with the improved GR-CNN grasping detection method,the intelligent grasping with autonomous decision making of the robotic arm is finally realized.