基于改进YOLOX-S的机械臂自主识别抓取  

Robot Arm Autonomous Recognition and Grasping Based on Improved YOLOX-S

作　　者：任军胜 晁永生[1] REN Jun-sheng;CHAO Yong-sheng(School of Mechanical Engineering,Xinjiang University,Urumqi 830017,China)

机构地区：[1]新疆大学机械工程学院,乌鲁木齐830017

出　　处：《组合机床与自动化加工技术》2022年第10期59-63,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：自治区自然科学基金(2022D01C37);自治区重点研发计划项目(2020B02013);国家自然科学基金(51565058)。

摘　　要：为实现机械臂对待抓取目标的实时识别与抓取,提出一种基于改进目标检测算法YOLOX-S的机械臂自主识别抓取方法。首先对深度相机进行自标定和手眼标定,尽可能降低定位误差,实验对比了常见的四种手眼标定方法并选择最优解。其次,建立常见机械零件数据集,在使用卷积神经网络目标识别算法YOLOX-S的基础上,引入注意力机制(CBAM),验证网络识别准确率。最后,为验证机械臂抓取效果,在ROS上搭建了眼在手上的目标抓取仿真实验平台,模拟了整个识别和抓取过程;同时在协作机械臂UR5上进行抓取实验。结果表明,机械臂能够自主识别并抓取目标物体,且识别定位误差较小,能够满足机械臂自主抓取的要求。In order to realize the real-time recognition and grasping of the target,a robot arm autonomous recognition and grasping method based on the improved target detection algorithm YOLOX-S is proposed.Firstly,self-calibration and hand-eye calibration for the depth camera are carried out to reduce the positioning error as much as possible.Four common hand-eye calibration methods are compared and the optimal solution is selected.Secondly,the data set of common mechanical parts is established.Based on the convolutional neural network target recognition algorithm YOLOX-S,attention mechanism(CBAM)is introduced to verify the accuracy of network recognition.Finally,in order to verify the grasping effect of the robot arm,an eye-in-hand target grasping simulation experimental platform is built on ROS,and the whole recognition and grasping process is simulated.At the same time,the grabbing experiment is conducted on the cooperative manipulator UR5.The results show that the robot arm can identify and grab the target object autonomously,and the identification and positioning error is small,which can meet the requirements of robot arm's autonomous grabbing.

关 键 词：机械臂 卷积神经网络 注意力机制 实时识别 自主抓取 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]