Whendevelopingahumanoidmyo-controlhand,notonlythemechanicalstructureshouldbeconsideredtoaffordahighdexterity,butalsothemyoelectric(electromyography,EMG)controlcapabilityshouldbetakenintoaccounttofullyaccomplishtheactuationtasks.Thispaperpresentsanovelhumanoidroboticmyocontrolhand(ARhandⅢ)whichadoptedanunderac-tuatedmechanismandaforearmmyocontrolEMGmethod.TheARhandⅢhasfivefingersand15joints,andactuatedbythreeembeddedmotors.Underactuationcanbefoundwithineachfingerandbetweentherestthreefingers(themiddlefinger,theringfingerandthelittlefinger)whenthehandisgraspingobjects.FortheEMGcontrol,twospecificmethodsareproposed:thethree-fingeredhandgestureconfigurationoftheARhandⅢandapatternclassificationmethodofEMGsignalsbasedonastatisticallearningalgorithm-SupportVectorMachine(SVM).Eighteenactivehandgesturesofatesteearerecognizedef-fectively,whichcanbedirectlymappedintothemotionsofARhandⅢ.Anon-lineEMGcontrolschemeisestablishedbasedontwodifferentdecisionfunctions:oneisforthediscriminationbetweentheidleandactivemodes,theotherisfortherecog-nitionoftheactivemodes.Asaresult,theARhandⅢcanswiftlyfollowthegestureinstructionsofthetesteewithatimedelaylessthan100ms.
