基于生成式卷积网络的抓取位姿检测算法研究
Grasping Pose Detection Using Generative Convolution Network
本文提出了一种用于生成抓取位姿的轻量化卷积生成式神经网络。首先用数学模型描述了抓取位姿检测问题，确定回归参数，然后设计了以RGB-D图像为输入的生成式神经网络预测回归参数，最后通过回归参数生成抓取框。模型结构由下采样模块，恒等残差块，RFB模块，上采样模块，通道注意力机制组成，网络模型可预测像素级别的抓取检测回归参数，合成抓取检测框，速度可达到16ms/帧。本文使用公开的Cornell抓取数据集进行模型训练，在随意摆放的物体和未知的物体上生成了抓取位姿，取得了98.87%的抓取检测成功率，在该数据集上的达到了一个当前的最佳结果。验证了模型的正确性和有效性。?????
In this paper, we present a lightweight convolutional generative neural network to generate a steady grasp pose for robotic antipodal gripper in the process of robotic arm grasping. The network model presented in this paper takes RGB-D images as input image and consists of downsample module, RFB module, upsampling module, attention module It can generate robust grasp poses at the speed of 16ms per image. we use the Cornell Grasping Dataset to train the model, and has achieved a 98.87% grasping detection success rate, which is the state-of-art success rate in this dataset. We future prove the effectiveness of the model on random set objects and novel objects.
贾庆轩、王童
自动化技术、自动化技术设备计算技术、计算机技术
控制科学与工程机械臂抓取抓取位姿检测生成式神经网络?????
ontrol Science and EngineeringManipulator graspingGrasping Pose DetectionGenerative Neural Network
贾庆轩,王童.基于生成式卷积网络的抓取位姿检测算法研究[EB/OL].(2022-01-07)[2025-08-11].http://www.paper.edu.cn/releasepaper/content/202201-15.点此复制
评论