2023年第5期仪表技术与传感器InstrumentTechniqueandSensor2023No.5基金项目:国家重点研发计划“智能机器人”重点专项(2018YFB1307100);安徽省教育厅科学研究重点项目(KJ2020A0364)收稿日期:2022-11-01基于PointNet++的机器人抓取姿态估计阮国强1,曹雏清1,2(1.安徽工程大学计算机与信息学院,安徽芜湖241000;2.哈尔滨工业大学芜湖机器人产业技术研究院,安徽芜湖241000)摘要:为解决在无约束、部分遮挡的场景下对部分遮挡的物体生成可靠抓取姿态的问题,基于PointNet++网络改进了一种抓取姿态估计算法,该算法可直接从目标点云中生成二指夹具的抓取姿态。由于该算法降低了抓取姿态的维度,将抓取的7自由度问题转变成4自由度问题处理,从而简化学习的过程加快了学习速度。实验结果表明:该算法在无约束、部分遮挡的场景中,能够生成有效的抓取姿态,且较Contact-GraspNet算法成功抓取率提升了约12%,能够应用于家用机器人的抓取任务。关键词:点云;位姿估计;抓取估计;深度学习;损失函数中图分类号:TP391文献标识码:A文章编号:1002-1841(2023)05-0044-05RobotGraspingAttitudeEstimationBasedonPointNet++RUANGuo-qiang1,CAOChu-qing1,2(1.SchoolofComputerandInformation,AnhuiPolytechnicUniversity,Wuhu241000,China;2.HarbinInstituteofTechnologyWuhuRobotTechnologyResearchInstitute,Wuhu241000,China)Abstract:Inordertosolvetheproblemofgeneratingreliablegraspattitudeforpartiallyoccludedobjectsinunconstrainedandpartiallyoccludedscenes,thispaperproposedagraspingattitudeestimationalgorithmbasedonPointNet++network.Theal-gorithmcandirectlygeneratetwofingergrippergraspattitudefromthetargetpointcloud.Becausethealgorithmreducedthedi-mensionofgraspingattitude,ittransformedtheproblemof7degreesoffreedomintotheproblemof4degreesoffreedom,acceler-atingtherateoflearningbysimplifyingtheprocess.Experimentalresultsshowthatthealgorithmcangenerateeffectivegrabatti-tudeinunconstrainedandpartiallyoccludedscenes,andimprovesthegrabbingratepredictionbyabout12%comparedwithCon-tact-GraspNet,whichcanbeusedforgrabbingtaskswithhomerobots.Keywords:pointcloud;poseestimation;graspestimation;depthlearning;lossfunction0引言智能机器人在执行抓取任务时需要获取目标的位...