■ View full text 

Journal of Manufacturing Science and Engineering (2023): 1-33.

https://url.kr/dogb5u

 

 

■ Researchers

Jinho Park

Indiana Next Generation Manufacturing Competitiveness Center (IN-MaC), Purdue University

Park, J., C. Han, M.B.G. Jun, and H. Yun*

 

 

■ Abstract

Vision-basedrobotshavebeenutilizedforpick-and-placeoperationsbytheirabilitytofind objectposes.Astheyprogressintohandlingavarietyofobjectswithclutteredstate,more flexibleandlightweightoperationshavebeenpresented.Inthispaper,anautonomous roboticbin-pickingplatformisproposed.ItcombineshumandemonstrationwithacollaborativerobotfortheflexibilityoftheobjectsandYOLOv5neuralnetworkmodelforfaster objectlocalizationwithoutpriorcomputer-aideddesignmodelsordatasetinthetraining. Afterasimplehumandemonstrationofwhichtargetobjecttopickandplace,therawcolor anddepthimageswererefined,andtheoneontopofthebinwasutilizedtocreatesynthetic imagesandannotationsfortheYOLOv5model.Topickupthetargetobject,thepointcloud wasliftedusingthedepthdatacorrespondingtotheresultofthetrainedYOLOv5model, andtheobjectposewasestimatedbymatchingthemwithIterativeClosestPoints(ICP) algorithm.Afterpickingupthetargetobject,therobotplaceditwheretheuserdefinedit intheprevioushumandemonstrationstage.Fromtheresultofexperimentswithfour typesofobjectsandfourhumandemonstrations,ittookatotalof0.5storecognizethe targetobjectandestimatetheobjectpose.Thesuccessrateofobjectdetectionwas 95.6%,andthepick-and-placemotionofallthefoundobjectswassuccessful.

 

• computer-integratedmanufacturing
• robotics

• [신진연구자 인터뷰] 윤희택(Huitaek Yun)(KAIST 기계공학과 교수)