推断紧凑表示以实现对机器人指令高效自然语言理解.pdfVIP

2019InternationalConferenceonRoboticsandAutomation(ICRA)

PalaisdescongresdeMontreal,Montreal,Cay20-24,2019

InferringCompactRepresentationsfor

—Thespeedandaccuracywithwhichrobotsare

abletointerpretnaturallanguageisfundamentaltorealizing

effectivehuman-robotinteraction.Agreatdealofattention

hasbeenpaidtodevelopingmodelsandapproximateinference

algorithmsthatimprovetheefficiencyoflanguageunderstand-

ing.However,existingmethodsstillattempttoreasonovera

representationoftheenvironmentthatisflatandunnecessarily

detailed,whichlimitsscalability.Anopenproblemisthen

todevelopmethodscapableofproducingthemostcompact

environmentmodelsufficientforaccurateandefficientnatural

languageunderstanding.Weproposeamodelthatleverages

environment-relatedinformationencodedwithininstructionsto

identifythesubsetofobservationsandperceptualclassifiers

necessarytoperceiveasuccinct,instruction-specificenviron-

mentrepresentation.Theframeworkusesthreeprobabilistic

graphicalmodelstrainedfromacorpusofannotatedinstruc-

(a)amobilerobotreceivinganaturallanguageinstruction

tionstoinfersalientscenesemantics,perceptualclassifiers,and

groundedsymbolperimentalresultsontworobotsoperat-

ingindifferentenvironmentsdemonstratethatbyexploiting

thecontentandthestructureoftheinstructions,ourmethod

learnscompactenvironmentrepresentationsthatsignificantly

improvetheefficiencyofnaturallanguagesymbolgrounding.

I.INTRODUCTION

Theabilityforrobotstoperformcomplextasksi