AoE：用于具身人工智能的始终在线第一人称视频采集AoE Always-on Egocentric Human Video Collection for Embodied AI.pdfVIP

DIGITAL

TECHNOLOGIES

AoE:Always-onEgocentricHumanVideoCollectionfor

EmbodiedAI

∗1∗1∗1∗†121

BowenYang,ZishuoLi,YangSun,ChangtaoMiao,YifanYang,ManLuo,Xiaotong

1134455

Yan,FengJiang,JinchuanShi,YankaiFu,NingChen,JunkaiZhao,PengweiWang,

54,531‡1

GuocaiYao,ShanghangZhang,HaoChen,ZheLi,KaiZhu

1AntDigitalTechnologies,AntGroup2InstituteofAutomation,ChineseAcademyofSciences

3ZhejiangUniversity4PekingUniversity5BeijingAcademyofArtificialIntelligence

6

2‡kaile.zk@

0

2

r

a

MAbstract

2

Embodiedfoundationmodelsrequirelarge-scale,high-qualityreal-worldinteraction

]dataforpre-trainingandscaling.However,existingdatacollectionmethodssuffer

Vfromhighinfrastructurecosts,complexhardwaredependencies,andlimitedinter-

C

.actionscope,makingscalableexpansionchallenging.Infact,humansthemselves

s

careidealphysicallyembodiedagents.Therefore,obtainingegocentricreal-world

[interactiondatafromgloballydistributed“humanagents”offersadvantagesoflow

2costandsustainability.Tothisend,weproposetheAlways-onEgocentric(AoE)

vdat