Research On Robot Grasping System Based On Deep Learning

In recent years,with the increasing demand for intelligent grasping robots,artificial intelligence algorithm-based grasping robots have gradually become a popular research direction for scientific institutions and companies.This paper designs a robotic grasping system based on a deep learning detection algorithm that can determine the position of objects according to given images and then implement grasping and placing actions.The main work includes:kinematic modeling of grasping robots,monocular vision-based calculation of target position,target detection based on deep learning,the construction of grasping robot system platform,etc.The kinematic modeling of the grasping robot includes:a DH kinematic model of the grasping robot with five degrees of freedom;an inverse kinematic formula gained by using a geometric solution;validation of the positive and inverse kinematic formulas using Matlab simulation;and an analysis of the effective operating range of the robot on the ground of its structure.The work on target posture calculation based on monocular vision includes:a monocular vision system was established with the eye outside the hand according to the imaging principle of the camera,and 'the conversion relationship between the pixel coordinate system and the world coordinate system was obtained.;solving the internal parameters of the camera by the calibration method of Zhengyou Zhang;calibrating the external parameters by the PNP algorithm to compensate for the effect of changes of scene on external participants;and verification through experiments of the accuracy of posture calculation on the basis of monocular vision.The target detection work based on deep learning includes:training the model of the target object using the YOLO and Mask R-CNN algorithms based on deep learning to categorize the identification and estimate the location of the target object;comparing the two algorithms above,YOLO is suitable for real-time detection of the scene,while Mask R-CNN is appropriate for accurate grasping of the scene;obtaining the smallest external rectangle of the target image by image processing algorithm and calculate the capture position with reference to its center point;planning the motion of the robot based on the grasping position to realize adaptive grasping of the target object.An experimental platform for monocular five-degree-of-freedom grasping robots was established,and the validity of the above algorithm was verified from engineering practice.Compared to traditional methods,the adaptive gripping robot approach based on deep learning reduces labor costs,improves adaptation to environmental changes,improves the success rate of grasping robots,and extends the scope of application of grasping robots.