cvpr18-neural kinematic networks for unsupervised motion用于无监督运动重定向神网络.pdf

NeuralKinematicNetworksforUnsupervisedMotionRetargetting

RubenVillegas1,*JimeiYang2DuyguCeylan2HonglakLee1,3

1UniversityofMichigan,AnnArbor

2AdobeResearch

3Brain

Time

n

o

i

t

o

M

t

u

p

n

I

1

r

e

t

c

a

r

a

h

C

t

e

g

r

a

T

2

r

e

t

c

a

r

a

h

C

t

e

g

r

a

T

Figure1:Ourend-to-endmethodretargetsagiveninputmotion(toprow),tonewcharacterswithdifferentbonelengthsand

proportions,(middleandbottomrow).Thetargetcharactersareneverseenperformingtheinputmotionduringtraining.

1.Introduction

Imitationisanimportantlearningschemeforagentsto

Weproposearecurrentneuralnetworkarchitecturewithacquiremotorcontrolskills[32].Itisoftenformulatedas

aForwardKinematicslayerandcycleconsistencybasedlearningfromexpertdemonstrationswithaccesstosample

adversarialtrainingobjectiveforunsupervisedmotionre-trajectoriesofstate-actionpairs[3,15].However,thisfirst-

targetting.Ournetworkcapturesthehigh-levelpropertiesimitationassumptionmaynotalwaysholdsince1)

ofaninputmotionbytheforwardkinematicslayer,and