摘要
为了解决机器人抓取领域传统方法中目标建模和锚框回归导致的计算时间长、检测精度低的问题,提出了一种基于anchor-free的机器人抓取检测方法。所提出的方法可以直接在图像的每个像素处生成抓取姿势和抓取质量。该方法使用双模态特征编码分别从RGB图像和深度图像中提取特征,并使用混合注意力机制来减少特征冗余。在实验中,该方法在Cornell抓取数据集和Jacquard数据集上的准确率分别达到了98.8%和96.4%。单张图像检测速度可达23 ms/张,可满足机器人抓取未知物体的闭环需求。在Franka 7自由度机器人上进行验证,实现了97.3%的抓取成功率。
In order to solve the problems of long calculation time and low detection accuracy caused by target modeling and anchor box regression in traditional methods in the field of robot grasping,an anchor-free robotic grasping method is proposed.The proposed method can generate the grasping pose and grasping quality directly at each pixel of the image.This method uses bimodal approach to extract features from RGB images and depth images respectively,and uses a hybrid attention mechanism to reduce feature redundancy.In the experimental evaluation,the proposed method achieved the accuracy of 98.8%and 96.4%on Cornell grasping datasets and Jacquard datasets,respectively.The single image detection speed can reach 23 ms/image,which can meet the demand of closing the loop for robotic grasping of unknown objects.Finally,it also demonstrates a grasping success rate of 97.3%using a Franka 7 DOF robot.
作者
吕泽
蔡乐才
成奎
高祥
罗春兰
LYU Ze;CAI Lecai;CHENG Kui;GAO Xiang;LUO Chunlan(Artificial Intelligence Key Laboratory of Sichuan Province,Sichuan University of Science&Engineering,Zigong 643002,China;Sanjiang Artificial Intelligence and Robot Research Institute,Yibin University,Yibin 644007,China)
出处
《无线电工程》
北大核心
2023年第4期936-945,共10页
Radio Engineering
基金
四川省科技厅面上项目(19ZDYF2284)。
关键词
机器人抓取
视觉检测
双模态特征编码
混合注意力机制
深度学习
robotic grasping
visual detection
bimodal feature encoding
hybrid attention mechanism
deep learning