202205 谷歌 通用智能体Gato.pdf

PublishedinTransactionsonMachineLearningResearch(11/2022)

AGeneralistAgent

*,†**†

ScottReed,KonradŻołna,EmilioParisotto,SergioGómezColmenarejo,AlexanderNovikov,

GabrielBarth-Maron,MaiGiménez,YurySulsky,JackieKay,JostTobiasSpringenberg,TomEccles,

JakeBruce,AliRazavi,AshleyEdwards,NicolasHeess,YutianChen,RaiaHadsell,OriolVinyals,

MahyarBordbarandNandodeFreitas†

*Equalcontributions,†Equalseniorcontributions,AllauthorsareaffiliatedwithDeepMindreedscot@

ReviewedonOpenReview:/forum?id=1ikK0kHjvj

2

2Abstract

0

2

vInspiredbyprogressinlarge-scalelanguagemodeling,weapplyasimilarapproachtowards

obuildingasinglegeneralistagentbeyondtherealmoftextoutputs.Theagent,whichwe

NrefertoasGato,worksasamulti-modal,multi-task,multi-embodimentgeneralistpolicy.

ThesamenetworkwiththesameweightscanplayAtari,captionimages,chat,stackblocks

1

1witharealrobotarmandmuchmore,decidingbasedonitscontextwhethertooutputtext,

jointtorques,buttonpresses,orothertokens.Inthisreportwedescribethemodelandthe

]data,anddocumentthecurrentcapabilitiesofGato.

I

A

.

s

c

[

3

v

5

7

1

6

0

.

5

0

2

2

:Whatisthecapitalof

vFrance?

i