2019ieee rsj智能机器人与系统国际iros 20.pdfVIP

2019IEEE/RSJInternationalConferenceonInligentRobotsandSystems(IROS)

Macau,,November4-8,2019

RecallingCandidatesofGrasMethodfromanObjectImage

usingNeurawork

MakotoSANADA,TadashiMATSUO,NobutakaSHIMADA,andYoshiakiSHIRAI,Member,IEEE



—Robotsarerequiredtosupportpeople’swork.Ingrastypessuitabletheobjectshape.Ekvalletal.prepared

ordertoalleviatetheburdenonpeople,itisdesirablethatrobotprimitiveobjectshapeinformationandgrastypesfrom

canautomaticallygenerateandexecutecomplicatedmotionsdemonstrationdatathatahumangraspsanobject.Therobot

accordingtosimpledirectionsfrompeople.However,therearesearchesandexecuteagrastypesuitabletheobjectshape

multiplegrasmethodsforoneobject.Inordertoselectafromthegraspexperiencedatabase[1].Nagataetal.prepared

motionsuitableforthedirection,itisimportanttoestimateprimitiveobjectshapeinformationandgrastypessuitable

candidatesofgrasmethodfortheobject.Inthisresearch,foreachobjectshape[2].Intheseresearches,itiseffectiveonly

wepurposetorecallcandidatesofgraspositionandhandforobjectinformationpreparedinadvance.Alotofadvance

shapefromanobjectimage.Inlearning,workthatoutputsdataisrequiredtograspwithmanyobjectshape.Inadditionto

apluralityofgrasmethodcandidatesforoneobjectimage