基础模型驱动的工业智能体:技术成熟度、能力变迁与未竟之挑战(英文版)_第1页
基础模型驱动的工业智能体:技术成熟度、能力变迁与未竟之挑战(英文版)_第2页
基础模型驱动的工业智能体:技术成熟度、能力变迁与未竟之挑战(英文版)_第3页
基础模型驱动的工业智能体:技术成熟度、能力变迁与未竟之挑战(英文版)_第4页
基础模型驱动的工业智能体:技术成熟度、能力变迁与未竟之挑战(英文版)_第5页
已阅读5页,还剩38页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

Foundation-Model-BasedAgentsinIndustrial

Automation:Purposes,Capabilities,andOpen

Challenges

VincentHenkel1*†,FelixGehlhoff1†,DavidKube2,

arXiv:2605.02592v1cs.AI42026

[]May

AsaadAlmutareb3,LuisCruz4,BerndHellingrath5,

PhilipKoch6,ChristophLegat7,FlorianMohr8,MichaelOberle9,

FelixOcker10,ThorstenSchoeler11,MarioThron12,

NicoAndreTo…pfer6,LucasVogt13,YuchenXia14

1*InstituteofAutomationTechnology,HelmutSchmidtUniversity/UniversityoftheFederalArmedForcesHamburg,Hamburg,Germany.

2SiemensAG,Nuremberg,Germany;InstituteforTechnologiesand

ManagementofDigitalTransformation,BergischeUniversit“at

Wuppertal,Germany.

3ArtiquareGmbH,Ingolstadt,Germany.

4FacultaddeIngenier´ıaMec´anica,Electr´onicayBiom´edica(FIMEB),UniversidadAntonioNari~no,Bogot´a,Colombia.

5ChairofInformationSystemsandSupplyChainManagement,UniversityofM“unster,M“unster,Germany.

6FraunhoferInstituteforManufacturingTechnologyandAdvancedMaterialsIFAM,Stade,Germany.

7ResearchGrouponCognitiveAutonomy&PredictiveIntelligence,FacultyofElectricalEngineering,TechnicalUniversityofApplied

SciencesAugsburg,Augsburg,Germany.

8BirkenfeldInstitutesofTechnology,TrierUniversityofAppliedSciences,Birkenfeld,Germany.

9FraunhoferInstituteforManufacturingEngineeringandAutomationIPA,Stuttgart,Germany.

10HondaResearchInstituteEurope,OffenbachamMain,Germany.

11FacultyofComputerScience,AugsburgTechnicalUniversityofAppliedSciences,Augsburg,Germany.

1

12InstituteforAutomationandCommunication(ifak),Magdeburg,Germany.

13Process-to-OrderGroup,TUDDresdenUniversityofTechnology,Dresden,Germany.

14InstituteforIndustrialAutomationandSoftwareEngineering,UniversityofStuttgart,Stuttgart,Germany.

*Correspondingauthor(s).E-mail(s):vincent.henkel@hsu.hamburg;Contributingauthors:felix.gehlhoff@hsu.hamburg;

david.kube@;asaad.almutareb@;

luicruz@.co;bernd.hellingrath@ercis.uni-muenster.de;

philip.koch@ifam.fraunhofer.de;christoph.legat@tha.de;

f.mohr@umwelt-campus.de;michael.oberle@ipa.fraunhofer.de;

felix.ocker@honda-ri.de;thorsten.schoeler@tha.de;mario.thron@ifak.eu;

nico.andre.toepfer@ifam.fraunhofer.de;lucas.vogt@tu-dresden.de;

yuchen.xia@ias.uni-stuttgart.de;

tTheseauthorscontributedequallytothiswork.

Abstract

Foundationmodels,particularlylargelanguagemodels,areincreasinglyinte-gratedintoagentarchitecturesforindustrialtaskssuchasdecisionsupport,processmonitoring,andengineeringautomation.Yetevidenceontheirpur-poses,capabilities,andlimitationsremainsfragmentedacrossdomains.Thisworkexamineshowmaturefoundation-model-basedagentsystemsareinindus-trialcontexts,howtheirfunctionalprofilediffersfromconventionalagentsystems,andwhichlimitationspersist.Asystematicliteraturesurveyfollow-ingthePRISMA2020guidelineispresented,screening2341publicationsandsynthesisingacorpusof88publicationsthroughastructuredcodingscheme.

Theresultsshowthatreportedsystemsarepredominantlyatprototypeandearlyvalidationstages(75.0%atTRL4—6),withdeployment-orientedevidenceremainingrare(9.1%).Operationalgoalsaremostfrequentlypositionedinuserassistance,monitoring,andprocessoptimisation,whileconventionalproduction-controlpurposessuchasplanningandschedulingarelessprominent.Comparedwithanestablishedbaselineforindustrialagentsystems,thecapabilitypro-filerevealssubstantialgainsinhumaninteraction(+37%)anddealingwithuncertainty(+35%),butapronounceddeficitinnegotiation(-39%).Themostwidelyreportedlimitationsconcernlackofgeneralization,hallucinationandoutputinstability,datascarcity,andinferencelatency.Aworkingdefinitionoffoundation-model-basedindustrialagentsisalsoproposed,bridgingconven-tionalagenttheory,automation-engineeringstandards,andthefoundation-modelparadigm.

Keywords:foundationmodels,largelanguagemodels,multi-agentsystems,industrialautomation,systematicliteraturereview

2

Abbreviations

DTDigitalTwin

FMFoundationModel

HMIHuman–MachineInteraction

LLMLargeLanguageModelMASMulti-AgentSystem

MCPModelContextProtocol

MLLMMultimodalLargeLanguageModel

PRISMAPreferredReportingItemsforSystematicReviewsandMeta-AnalysesRAGRetrieval-AugmentedGenerationTRLTechnologyReadinessLevel

1Introduction

SoftwareagentsandMulti-AgentSys-tems(MASs)havebeenstudiedfordecadesasadesignparadigmfordis-tributeddecision-makinginindustrialdomainssuchasproductioncontrol,logistics,andprocessengineering[

1

,

2

].Inconventionalsettings,i.e.agentsystemsthatpredatetheintegrationofFounda-tionModels(FMs),agentsaretypicallyrule-basedoroptimisation-drivenentitiesthatpursuelocallyspecifiedobjectivesunderpredefinedinteractionprotocolssuchastheContractNetProtocol[

3

].WiththeemergenceofFMs,andLargeLanguageModels(LLMs)inparticular,anewclassofagentsystemshasgainedmomentum[

4

].Unliketheirconventionalcounterparts,FM-basedagentscaninter-pretunstructuredandnoisyinformationsuchasnatural-languageinstructions,maintenancelogs,andvisualorsensor-streaminputs,interactwithhumanoper-atorsthroughconversationalinterfaces,andorchestrateheterogeneoustoolchainsbygeneratingexecutablecodeorapplica-tionprogramminginterfacecallsthroughflexiblereasoning[

5

].AccordingtoRenetal.[

6

],FMs-basedindustrialagents

comprisedifferentlevelsoftechnolog-icalcapabilities:LLM-agentsprimarilyextendlanguage-centricreasoningandtooluse,MultimodalLargeLanguageModel(MLLM)-agentsaddmultimodalperceptionacrosstextual,visual,andsen-sordata,whereasAgenticAIreferstoafurthersteptowardself-directed,goal-drivenautonomyindynamicenviron-ments.TheshiftfromconventionalMAStoFM-basedagentsystemsisnotonlyashiftinrealisedcapabilitiesbutalsointheunderlyingcoordinationlogic.Alietal.

[

7

]pointouttheconstrastbetweensym-boliccoordinationthroughexplicitproto-colssuchastheContractNetProtocolorblackboardsystemswithneuralcoordi-nationbasedonstructuredconversation,role-basedworkflows,andprompt-drivenorchestration.

Thus,eventhoughtherehavebeenattemptstoaddresschallengessuchasinteractionwithhumanoperators,toolorchestration,andgeneraloptimisationapplicationsinthepre-FMera,thesenewtechnologiesmakesuchsystemsmuchmoreaccessible,easiertodevelopandmaintain,aswellasconsiderablymorecapableandadaptive.

Thesedevelopmentshaveledtorapidadoptionacrossabroadrangeofindustrialapplications,fromengi-neeringdesignautomationandshop-floorcontroltoenergy-systemoperationandinformation-technologyinfrastruc-turemanagement[

8

10

].Atthearchi-tecturallevel,LLM-basedagentsarecommonlyintegratedascentralcogni-tivecomponentsthatinterpretcontext,generateplansorrecommendations,andinvokeexternaltools,oftenaugmentedbyretrievalmechanismstogrounddeci-sionsindomain-specificknowledge[

5

,

10

].RecentsurveysonLLM-basedagentsingeneral-purposesettingshavemapped

3

thesearchitecturalpatterns,reasoningstrategies,andtool-usemechanisms[

4

,

5

,

11

].However,thedegreetowhichthesegeneralfindingstransfertoindustrialcontexts,wheresafety,determinism,andintegrationwithlegacysystemsimposeadditionalconstraints,remainsanopenquestion[

12

].

Despitethisgrowingadoption,thereisnoestablishedworkingdefinitionthatbridgesconventionalagentconceptsandstandards,suchasautonomousactioninanenvironment[

1

]orencapsulatedenti-tieswithcontrolobjectives[

13

],withtheFMparadigm,resultinginpossibleconfusionwithapproachesthatmerelyuseLLMsfortextgenerationorcon-versationalinteraction.Recentsurveysacknowledgethisgap:Jinetal.[

14

]notethatamongresearchersthereisno“cleardistinctionbetweenLLMsandLLM-basedagents”andthat“unifiedstandardandbenchmarking”remaininanearlystage,whileZhouetal.[

15

]callfora“unifiedtaxonomy”toaddressthecurrently“fragmentedapproach”toclassifyingFM-basedagentarchi-tectures.Arelatedproblemrefersto“conceptualretrofitting”,i.e.,theten-dencytodescribemodernLLM-basedsystemsusingclassicalagentconceptssuchasBelief-Desire-Intention(BDI)orperceive-plan-act-reflectloops,despitesubstantialdifferencesintheiroper-ationalmechanisms[

7

].Thislackofconceptualconsolidationisalsoevidentinrecentmanufacturing-focusedlitera-ture,whichexplicitlynotesthatthedefinitions,capabilityboundaries,andinterconnectionsofLLM-agents,MLLM-agents,andAgenticAIremaininsuf-ficientlyclarified[

6

].Compactopera-tionalcharacterisationshavebeenpro-posed,e.g.,definingagenticLLMsas

systemsthat“reason,act,andinter-act”[

16

],yetnoconsolidateddefinitionexiststhatintegratestheseperspectiveswithautomation-engineeringstandards.Withoutsuchadefinition,asystem-aticliteraturereviewlacksareproducibleinclusioncriterion.

Atthesametime,thelevelofmatu-rityofFM-basedagentsisunclear,sinceevidenceonsuchsystemsisfrag-mentedacrossapplicationdomains,levelsoftechnologicalmaturity,andevalua-tionpractices.Domain-specificsurveysconfirmthattheintegrationofLLMsinmanufacturingis“stillinitsini-tialstages”[

17

],andarecentmeta-surveyonagentevaluationdescribesthefieldasa“complexandunderdevelopedarea”andaimstobring“claritytothefragmentedlandscapeofagentevalua-tion”[

18

].Individualpublicationsreportprototypesinmanufacturing[

8

,

19

],logis-tics[

20

],energysystems[

10

],orengi-neeringdesign[

9

],butnoconsolidatedoverviewmapsthecapabilitiesandmatu-rityoftheseapproachestoacommonframework.AnotherstudypointsoutthatexistingindustrialFM-basedagentapproachesappeartoclusteraroundassistiveandtask-orientedroles,whereasstrongerformsofself-directedgoalfor-mulationandsystem-levelorchestrationarestilllargelydiscussedasanemergingAgenticAIvisionratherthanasestab-lishedindustrialpractice[

6

].Asaconse-quence,itremainsdifficulttoassesstheoveralltechnologicalmaturityofthefield,toidentifyrecurringlimitationsacrossdomains,andtoestablishacomparableevaluationbasis,leavingpractitionersindoubtaboutthecapabilitiesandmeritsofadoptionofFM-basedagentsystemsinindustrialcontexts.

4

Furthermore,itisunclearhowFMintegrationchangesthefunctionalpro-fileofindustrialagentscomparedtoclassicalMAS.Earliersystematicworkonsoftwareagentsinindustrialproduc-tion[

2

]providesabaselineofpurposes(e.g.,planning,scheduling,control)andcapabilities(e.g.,negotiation,coordina-tion,reactivity).Initialevidencesug-gestsashift:Greisetal.[

21

]explic-itly“contrastthecapabilitiesofclassicautonomoussoftwareagentsandLLMsoftwareagents”inadigital-twin-enabledmanufacturingcontext,andZhaoetal.

[

22

]observethatconventionalagentnego-tiationrelieson“pre-definedandfixedheuristicrules”thatareill-suitedtodynamicdisturbances,motivatingamul-timodalLLM-basedalternative.Thissug-geststhatthetwoparadigmsmaybecomplementaryratherthanFM-basedagentsbeingthesuccessorofconventionalapproaches.However,thishasnotbeensystematicallyexaminedacrossdomains.

Againstthisbackground,theguidingresearchquestions(RQs)ofthisworkare:

•RQ1:HowcanFM-basedindustrialagentsbedefined,andwhatisthecur-rentmaturityofsuchsystemsinindus-trialandindustriallyrelevantresearch,includingtheirtechnologyreadiness,applicationdomains,andusecases?

•RQ2:WhichsystempurposesandcapabilitiesdoFM-basedagentsexhibit,andhowdoestheirfunc-tionalprofiledifferfromconventionalindustrialagentsystems?

•RQ3:Whichlimitations,challenges,andfutureworkdirectionsaremostfre-quentlyreportedforFM-basedagentsystemsinindustrialcontexts?

Toaddressthesequestions,thisworkfollowsaPreferredReportingItemsforSystematicReviewsandMeta-Analyses

(PRISMA)-stylesystematicreview(Section

3

).Foreachincludedpublica-tion,technologyreadiness,applicationdomain,systempurposes,capabilities,reportedlimitations,andfutureworkdirectionsareassessedandconsolidated.

Themaincontributionsofthisworkarethreefold.First,itproposesaworkingdefinitionofFM-basedtechnicalagentsthatbridgesconventionalagentthe-ory,automation-engineeringstandards,andtheFMparadigm,andevaluateswhetherthisdefinitionadequatelycap-turesthesystemsreportedinthecorpus(Section

5.1

).Second,buildingonthisdefinition,itprovidesacross-domainsyn-thesisoftechnologicalmaturity,systempurposes,andcapabilityprofilesofFM-basedagents,includingadescriptivecom-parisonwithanestablishedbaselineforindustrialagentsystems.Third,itcon-solidatesrecurringlimitationsandfutureworkdirectionsintostructuredthemesthatcaninformthedesign,evaluation,anddeploymentofFM-basedagentsys-temsinindustrialcontexts.

2Relatedwork

Thissectionrelatesthepresentworktothreestreamsofpriorresearch:industrialsoftwareagentsandMASs(Section

2.1

),FM-basedagentsinindustrialcontexts(Section

2.2

),andexistingtaxonomiesforpurposesandcapabilities(Section

2.3

).Section

2.4

introducesthebaselineusedforcomparativeanalysis.

2.1Industrialsoftwareagentsandmulti-agentsystems

IndustrialsoftwareagentsandMASs

havelongbeenstudiedasadesignparadigmfordistributeddecision-makingandcontrolinproductionandrelatedindustrialdomains[

23

].Inconventional

5

settings,agent-basedapproachesaretypi-callymotivatedbytheneedtodecomposecomplexobjectivesintolocallymanage-ablesub-problems(e.g.,planningandscheduling,dispatchingdecisions,shop-floorcontrol,anddiagnosisandfaultmanagement)andtocoordinatethesedecisionsacrossheterogeneousresourcesunderoperationalconstraints[

24

,

25

].Atthesametime,industrialapplicationsimposestrongnon-functionalrequire-ments,includingdeterminism,safety,andreal-timesuitability,whichshapehowautonomyandinteractionmechanismscanberealizedinpractice[

26

,

27

].

Fromaconceptualperspective,Rus-sellandNorvig[

28

]provideabroadcharacterizationofanagentas“any-thingthatcanbeviewedasperceiv-ingitsenvironmentthroughsensorsandactinguponthatenvironmentthroughactuators”,adefinitionthatisdelib-eratelyparadigm-agnosticandaccom-modatesbothsymbolicandsubsym-bolicrealizations.Withinthemorespe-cificMAStradition,Wooldridge[

1

]emphasizesautonomousaction,reactiv-ity,andcoordinationamongmultipleentities,includinginteractionpatternssuchasnegotiationwhereapplicable.Inindustrialsurveywork,thesecon-ventionalexpectationsareoperational-izedasconcretepurposesandcapabilitiesthatcanbeevidencedfromapplicationreportsinproductioncontexts[

2

].Tra-ditionalapplicationsinclude,forexam-ple,heuristics-baseddecentralisedpro-cessplanningandschedulingaswellascontrollingresourcesanddiagnosistasksontheshopfloor[

25

].

2.2Foundation-model-based

agentsinindustrial

contexts

RecentFM-basedindustrialagentsys-temsbuildonthesefoundationsbutoftenrepurposethemtowardsassistiveanddecision-support-centricroles(e.g.,maintenancedecisionsupportormanu-facturingassistance)ratherthantowardsfullydecentralisednegotiation-heavycon-trol[

8

,

9

,

19

].FMs,andinparticularLLMs,providegenericlanguageunder-standingandgenerationcapabilitiesthatcanbeadaptedtovariousdownstreamtasks[

29

].Inindustrialandindustri-allyrelevantagentsystems,LLMsarecommonlyintegratedascentralcogni-tivecomponentsthat(i)interprethumaninstructionsandcontextualinformation,(ii)generateplansorrecommenda-tions,and(iii)orchestratetool-mediatedactionsbyinteractingwithexternalsoft-warecomponents(e.g.,databases,simu-lationmodels,codegenerators,ordomainservices)[

30

].ArecurringarchitecturalpatternistocomplementLLMswithexplicitgroundingmechanismssuchasRetrieval-AugmentedGeneration(RAG)toimprovefactualityandtocon-nectagentdecisionstodomain-specificknowledgesources[

10

].Renetal.

[

6

]alsoemphasizethatindustrialtasksalsodependonthejointinterpretationof,e.g.,enterprisedata,maintenancerecords,sensorstreams,andmachine-visioninputs,whichiswhymultimodalLLM-basedagentsareincreasinglydis-cussedasarelevantarchitecturalexten-sionthatsupportscontext-awarepercep-tion,diagnosis,anddecisionsupport.

Priorworkhasnotedlimitationsrelatedtorobustness,reliability,andexternalvalidity,particularlywhenLLMsareusedfordecision-makingunder

6

incompleteinformationorwhengener-atedoutputsmustbeexecutableintech-nicalsystems.Practicalconstraintssuchaslatency,cost,andintegrationeffortarealsodiscussed[

8

,

9

,

19

,

31

].

2.3Taxonomiesforpurposes,

properties,and

capabilities

Systematiccomparisonacrossagentsys-temsrequiresarepresentationthatseparateswhatasystemisintendedtoachievefromhowitachievesit.M“ulleretal.[

32

]characteriseindustrialautonomoussystemsthroughfourhigher-leveldimensions—systematicprocessexe-cution,adaptability,self-governance,andself-containedness—andrelatethesetolower-levelabilitiessuchaslearningabil-ity,decision-makingcapacity,cooperabil-ity,reactivity,andself-explanation.Sincetheserelationsaremany-to-manyratherthanone-to-one,theirframeworksug-geststhathigher-levelsystemqualitiesshouldbeanalysedseparatelyfromtheconcretefunctionalmeansbywhichtheyarerealised.Kaber[

33

]providesanadditionalconceptualbasisforseparat-inghigher-levelsystemqualitiesfromconcretefunctionalitybydistinguishingautomationfromautonomy.Ratherthantreatingautonomyasahigherlevelofautomation,hedefinesitthroughthreeconditions—viability,independence,andself-governance—andfurtheropera-tionalisesthedistinctionintermsofthedemandsasystemplacesonitsenviron-ment,humancollaborators,andtaskpro-tocols.Inthiswork,theanalysisfollowsthepurpose-property-capabilityframingemployedbyReinpoldetal.[

2

],whichbuildsontheworkbyM“ulleretal.[

32

].Inthisframing,systempurposesdescribeoperationalgoalcategories(e.g.,plan-ning,scheduling,control,monitoring,

userassistance),capabilitiesdescribecon-cretefunctionalskillsevidencedbythesystem(e.g.,interaction,communication,coordination,reasoning),andpropertiesaggregatecapabilityevidenceintohigher-leveldimensionsthatsupportcorpus-levelcomparison(seeSection

3.1

fortheworkingdefinitionusedinthisreview).ThisdistinctionisparticularlyusefulforFM-basedagentsystemsbecauseLLMscansimultaneouslyshifttheoperationalemphasistowardsassistivepurposesandaffecttheevidencingofcapabilitiessuchashumaninteractionanduncertaintyhandling.

2.4Baselineforcomparativeanalysis

Reinpoldetal.[

2

]systematicallycom-pareindustrialsoftwareagentsandDig-italTwins(DTs)inproductioncontextsthroughaPRISMA-alignedliteraturereviewcovering145publications.Theirworkprovidesthepurposeandcapa-bilitytaxonomyadoptedinthisreview(seeSection

2.3

)togetherwithcompositescoresthatsummarizecapabilityprofilesatpropertylevel,therebyestablishingaquantitativereferencepointforcorpus-levelcomparison.Giventhedifferenttem-poralscopesofthetworeviews,withReinpoldetal.[

2

]largelypredatingthebroadadoptionofFMsinindustrialagentresearch,substantialoverlapbetweenthetwocorporaisunlikely,andthetwostudiescanbeviewedaslargelycom-plementarysamplesofconventionalandFM-basedagentresearch,respectively.

Inthiswork,Reinpoldetal.[

2

]isusedasabaselinefordescriptivecompar-ison,enablingthecomputationofcom-parablecoverageprofilesanddifferencesinpercentagepoints.Differencesincor-puscompositionandinclusioncriteria

7

betweenthetwostudiesshouldbecon-sideredwheninterpretingthecomparison(seeSection

5.3

).

3Method

ThisworkfollowsthePRISMA2020guidelineforsystematicliteraturereviews[

34

]andadaptsittotheanalysisofFM-basedagentsystemsinindustrialandindustriallyorientedresearch.Thepro-cedurecomprisesthestandardPRISMAstagesof(1)identificationofpotentiallyrelevantrecordsviaastructuredmulti-databasequeryand(2)relevancefilteringaccordingtopredefinedeligibilitycrite-ria,followedby(3)aconsolidationstepinwhichtheresultingcorpusistransformedintoaconsistent,machine-readablerep-resentationforquantitativeanalysis.

3.1Workingdefinition

Agentshavealong-standinghistorywithinacademia[

1

].Whiletheseclassi-calfoundationsprovidetheessentialbasisforthisreview,themodernliteratureoftenusestheterm“agent”interchange-ablywithrelatedconceptssuchasagenticworkflowsorsimply“LLMs”.Tomain-tainaclearscope,thisworkdistinguishesitsfocusfromtheseadjacenttermsandprovidesaworkingdefinitionthatis(i)groundedinestablishedagentcon-cepts,(ii)compatiblewithautomation-engineeringnotionsoftechnicalagents,and(iii)explicitaboutwhatitmeansforasystemtobe“FM-based”.

Intheagentliterature,Wooldridge[

1

]definesanagentas“acomputersys-

temthatissituatedinsomeenviron-ment,andthatiscapableofautonomousactioninthisenvironmentinorder

tomeetitsdesignobjectives”.Inthecontextofautomationengineering,the

VDI/VDE2653guidelinedefinesatech-nicalagentas“anencapsulatedhardware

orsoftwareentitywithspecfiedobjec-tivesregardingthecontrolofatechnical

systemorapartthereof”[

13

].Founda-tionmodels,inturn,arecharacterizedasmodels“trainedonbroaddata(gener-

allyusingself-supervisionatscale)thatcanbeadapted(e.g.,fine-tuned)toa

widerangeofdownstreamtasks”[

29

].AccordingtoRenetal.[

6

],systemswithhighdegreesofautonomythatusetheseFMsexhibitcapabilitiessuchasseman-

ticretrievalandcontext-awareness,adap-tivereasoning,andautonomousdecision-making,whichincludestherealsation

ofselectedactions,i.e.influencingtheactualprocess.

Combiningthesestreamsandcon-cepts,thefollowingdefinitionisproposedthatguidesthefocusofthiswork:

Throughoutthereview,eachcategoryisassessedbasedonpaper-levelevidence:ifthepublicationdoesnotprovidesuffi-cientdetail,thecorrespondingcategoryistreatedasnotevidenced.

Workingdefinitionusedinthisreview

Afoundation-model-basedindustrialagentisanencapsulated

hardwareorsoftwareentityactinginthecontextofanindustrialsystemthatiscapableofautonomousactioninordertomeetitsspecfieddesignobjectives,andthatusesafounda-tionmodelasacentralcomponentforcontextinterpretation,decision-making,aswellasactionselection

andexecution.

8

3.2Datasourcesandsearchstrategy

Recordswereretrievedfromfourbibli-ographicandpreprintsources:Scopus,SemanticScholar,arXiv,andOpe-nAlex.Thesearchcoveredpublicationsfrom2020onwardsandtargetedLLMsandrelatedFMsinagentormulti-agentsettingswithinindustrialandindustriallyrelevantdomains,includingmanufactur-ing,logistics,energysystems,andtheengineeringlifecycle(productdevelop-mentandprocessengineering),aswellascross-domainfunctionalitiessuchasmaintenance,qualitymanagement,andHuman-MachineInteraction(HMI).Thecorequerycombined(i)termsforLLMsandFMs,(ii)termsforagentsandMASs,and(iii)termsforindustrialcontexts.Acrossallsources,3025validresultsand2341uniquerecordswereobtained.

ThecompositesearchqueryisshowninTable

1

.ThesearchandexportwereexecutedonSeptember8,2025.

3.3Screeningandeligibility

Eligibilityisdefinedatthelevelofindi-vidualpublications.Arecordisincludedifitmeetsallofthefollowingcriteria:

1.Thepublicationdescribesaconcreteapplication,implementation,orcasestudy(notapurelyconceptualcontri-bution,high-levelvisionpaper,survey,orreviewwithoutpracticalrealisationorempiricalresults).

2.ThesystemusesoneormoreLLMs,MLLMs,orrelatedFMsforcontrolordecision-makingwithinanagentoragent-basedarchitecture.

3.Theapplicationcontextisindustrialorindustriallyrelevant,coveringtheabove-mentioneddomains,engineer-inglifecyclephases,orcross-domainfunctionalities.

Giventhesizeoftheinitia

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论