pistar06-a VLA That Learns From Experience一个从经验中学习的甚长基线阵列.pdfVIP

π∗:aVLAThatLearnsFromExperience

0.6

PhysicalIntelligence

AliAmin,RaichelleAniceto,AshwinBalakrishna,KevinBlack,KenConley,GraceConnors,JamesDarpinian,KaranDhabalia,JaredDiCarlo,

DannyDriess,MichaelEqui,AdnanEsmail,YunhaoFang,ChelseaFinn,CatherineGlossop,ThomasGodden,IvanGoryachev,LachyGroom,

HunterHancock,KarolHausman,GashonHussein,BrianIchter,SzymonJakubczak,RowanJen,TimJones,BenKatz,LiyimingKe,

ChandraKuchi,MarindaLamb,DevinLeBlanc,SergeyLevine,AdrianLi-Bell,YaoLu,VishnuMano,MohithMothukuri,SurajNair,KarlPertsch,

AllenZ.Ren,CharviSharma,LucyXiaoyangShi,LauraSmith,JostTobiasSpringenberg,KyleStachowicz,WillStoeckle,AlexSwerdlow,

JamesTanner,MarcelTorne,QuanVuong,AnnaWalling,HaohuanWang,BlakeWilliams,SukwonYoo,LiliYu,UryZhilinsky,ZhiyuanZhou

https://pi.website/blog/pistar06

Fig.1:RECAPenablestrainingVLAswithrewardfeedbackandinterventions.Oursystemstartswithapre-trainedVLAthatincorporatesadvantage

conditioning,allowingthemodeltolearneffectivelyfromreal-worldexperience.Foreachtask,wedeploythemodelandcollectbothautonomousrolloutsand

onlinehumancorrections.Wethenfine-tunethevaluefunctiononthisonlinedata,improvingitsestimatesofhowactionsinfluenceperformance.Fine-tuning

andconditioningtheVLAontheseupdatedadvantageestimatesinturnimprovespolicybehavior.

Abstract—Westudyhowvision-language-action(VLA)modelswecanflexiblyspecifytasksforgeneralistrobotsthrough

canimprovethroughreal-worlddeploymentsviareinforcementprompts.Butjustlikepeople,thesemodelswillneedto