人工智能機(jī)器學(xué)習(xí)與監(jiān)管：自動(dòng)駕駛汽車案例-2025-03-自動(dòng)駕駛

International

TransportForum

197

Roundtable

AIMachineLearningandRegulation:TheCaseofAutomatedVehicles

SummaryandConclusions

iiAI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2025

TheInternationalTransportForum

TheInternationalTransportForumisanintergovernmentalorganisationwith69membercountries.ItactsasathinktankfortransportpolicyandorganisestheAnnualSummitoftransportministers.ITFistheonlyglobalbodythatcoversalltransportmodes.TheITFispoliticallyautonomousandadministrativelyintegratedwiththeOECD.

TheITFworksfortransportpoliciesthatimprovepeoples’lives.Ourmissionistofosteradeeperunderstandingoftheroleoftransportineconomicgrowth,environmentalsustainabilityandsocialinclusionandtoraisethepublicprofileoftransportpolicy.

TheITForganisesglobaldialogueforbettertransport.Weactasaplatformfordiscussionandpre-negotiationofpolicyissuesacrossalltransportmodes.Weanalysetrends,shareknowledgeandpromoteexchangeamongtransportdecisionmakersandcivilsociety.TheITF’sAnnualSummitistheworld’slargestgatheringoftransportministersandtheleadingglobalplatformfordialogueontransportpolicy.

TheMembersoftheForumare:Albania,Argentina,Armenia,Australia,Austria,Azerbaijan,Belarus,Belgium,BosniaandHerzegovina,Brazil,Bulgaria,Cambodia,Canada,Chile,China(People’sRepublicof),Colombia,CostaRica,Croatia,CzechRepublic,Denmark,DominicanRepublic,Estonia,Finland,France,Georgia,Germany,Greece,Hungary,Iceland,India,Ireland,Israel,Italy,Japan,Kazakhstan,Korea,Latvia,Liechtenstein,Lithuania,Luxembourg,Malta,Mexico,RepublicofMoldova,Mongolia,Montenegro,Morocco,theNetherlands,NewZealand,NorthMacedonia,Norway,Oman,Poland,Portugal,Romania,RussianFederation,SaudiArabia,Serbia,SlovakRepublic,Slovenia,Spain,Sweden,Switzerland,Tunisia,Türkiye,Ukraine,theUnitedArabEmirates,theUnitedKingdom,theUnitedStatesandUzbekistan.

InternationalTransportForum

2rueAndréPascalF-75775ParisCedex16

contact@

ITFDisclaimer

Citethisworkas:ITF(2025),AI,MachineLearningandRegulation:TheCaseofAutomatedVehicles,OECDPublishing,Paris.

AI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2025iii

Acknowledgements

ThisreportbuildsonexpertdiscussionsduringanITFRoundtable,“ArtificialIntelligence,MachineLearningandRegulation”,heldinParisandvirtuallyon26-27January2023.Dr.MarkusReinhardt(GermanCentreforRailTrafficResearch)chairedtheRoundtable.TheITFwouldliketothankthehigh-levelparticipationoftheGermanFederalMinistryforDigitalandTransport,highlightedbytheopeningremarkofMrStefanSchnorr,theStateSecretaryoftheGermanFederalMinistryforDigitalandTransport,whoremindedparticipantsoftheimportanceofAIinthefutureoftransport.

AttheITF,ChanggiLeecoordinatedtheprojectalongsideCamilleCombe.ChanggiLeeandPhilippeCristauthoredthereportwithsubstantiveinputsfromCamilleCombe.EugineRohelpedwiththeco-ordinationoftheRoundtableeventandtooknotesofdiscussions.PhilippeCristwasresponsibleforoverallqualitycontrolandCamilleLarmanoumanagedtheeditorialprocess.MilaIglesiasandApostolosSkourtashelpedorganisetheRoundtableevent.ThisRoundtableReportispartoftheITF’scoreProgrammeofWorkfor2022-23,co-ordinatedbyJagodaEgelandandOrlaMcCarthy,andhasbeenapprovedbytheITF’sTransportResearchCommittee.

TheauthorswouldliketothankGianmarcoBaldini(JointResearchCentre,EuropeanCommission),FlorentPerronnin(NaverLabs),MartinRuss(AustriaTech),andWilliamH.Widen(UniversityofMiami)fortheirfeedbackonthereport.TheITFwouldliketothankMarkusReinhardtforchairingtheRoundtable.ThanksalsotoGregorioAmeyugo(CEAList),SiddarthaKhastgir(UniversityofWarwick),LouiseDennis(UniversityofManchester),LatifaOukhellou(UniversityGustaveEiffel),Marry“Missy”Commings(GeorgeMasonUniversity),GianmarcoBaldini(JointResearchCentre,EuropeanCommission),AidaJoaquinAcosta(MinistryofTransport,MobilityandUrbanAgenda,Spain),andMartinRuss(Austriatech)fortheirpresentationsduringtheRoundtable.AnnexAliststhenamesandaffiliationsoftheRoundtableparticipants.

ivAI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2025

Tableofcontents

Executivesummary 1

Enterout themachine:Takingthe(human)driverofthevehicle4

Automatedarenotevenacross vehicledeploymentlevelsmodes4

Safetyfirst,allotherutilitarianconsiderationslater 6

AssessingversustrustAVsystem vehiclesafetyensuringinsafeperformance9

Whatto andwhencertify?11

IncludingAI-specificchallengesinAVassessment safety12

PolicyTakeaways 15

HowAIconfoundshumanoversightandhowtoensureitstrustworthiness 16

Anatomyofdriving:whatconstitutesdrivingtasks,andhowdoesAI performthem?16

EnsuringAI’sKeyAI trustworthiness:elementsandlifecycle20

PolicyTakeaways 25

RegulatoryconsiderationstoensuretrustworthyAIineachdimensionoftheAIlifecycle 26

WhatDataRequiredVehicles? isforAutomated26

DevelopmenttoDeployment:VerifyingAIModels 30

Co-evolvingwithAVs:FromAVtoAVstransport deploymentmakingworkforbetter38

PolicyTakeaways 40

References 41

AnnexA.ListRoundtable ofparticipants48

TABLEOFCONTENTS

AI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2002v

Figures

Figure.MachineLearningConceptsClasses 1and10

Figure 2.Multi-phasetesting,verificationandvalidationofAVs13

Figure.Schematic 3representationofdynamicdrivingtasks16

FigureExamples 4.ofAItechniquesusedforautomateddriving19

FigureSystemstressresponse 5.scenarios22

FigureTheFiveDimensionsAISystemLifecycle 6.ofthe24

Figure.Thetypesamap 7fourofdatainlocaldynamic27

Figure 8Skill-Rule-Knowledge-Expert(SRKE)Taxonomy31

Figure.Machineto 9learningimagerecognitionvulnerabilitiesadversarialattacks33

Figure.Test,Evaluation,VerificationValidation(TEVV)Environments 10and34

Figure.AnAIon 11exampleofmakingdifferentdecisionbasedthesimilardatainterpretations35

Figure.Explainability 12andinterpretabilitybydesignformachinelearningapplications35

Figure.FaultAVs 13tolerancemodesfor36

Tables

Table.SummaryFittsListacrossaspects1ofofstrengthsandweaknessesvariousoffunctionallocation

betweenhumansandhardware/softwaresystems 18

Boxes

Boxreport 1.Terminologyusedinthis5

BoxArtificial 2.Intelligenceinbrief10

Box.Automationassessment3ofdrivingcapabilitiesforroadvehiclesandtheparadigmshiftforvehicle

13

BoxEitheryouorSkipping 4.drive,Idrive:level3inregulation38

viAI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2025

Abbreviationsandacronyms

ADAS

Advanceddriverassistancesystems

ADS

Automateddrivingsystem

ADS-DV

AutomatedDrivingSystems–DedicatedVehicles

AI

Artificialintelligence

ASDE

AuthorisedSelf-DrivingEntity

AV

AutomatedVehicle

*Inthisreport,AVsrefertohighlyorfullyautomatedvehiclescorrespondingtoSAEJ3016Level4andabovethatdoesnotrequirehumaninvolvementinperformingdrivingtaskswithindesignatedODDs(seeBox.1)

DDT

Dynamicdrivingtask

DMV

DepartmentofMotorVehicle

EIBD

ExplainableandInterpretablebyDesign

EM

EmergencyManoeuvre

FHWA

FederalHighwayAdministration

GDPR

GeneralDataProtectionRegulation(EU)

GPS

Globalpositioningsystem

ITF

InternationalTransportForum

LDM

Localdynamicmap

LIDAR

Lightdetectionandranging

MRM

MinimumRiskManoeuvre

NHTSA

NationalHighwayTrafficSafetyAdministration

NUiC

No-user-in-charge

OBU

On-boardunit

ODD

Operationaldesigndomain

OEM

Originalequipmentmanufacturer

PTO

Publictransportoperator

R&D

Researchanddevelopment

RSU

Roadsideunit

SAE

SocietyofAutomotiveEngineers

TEVV

Test,Evaluation,VerificationandValidation

UiC

User-in-charge

UNECE

UnitedNationsEconomicCommissionforEurope

V2I

Vehicle-to-infrastructure

V2V

Vehicle-to-vehicle

V2X

Vehicle-to-everything

AI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2025vii

Glossary

Automateddrivingsystem

ThehardwareandsoftwarethatarecollectivelycapableofperformingtheentireDDTonasustainedbasis,regardlessofwhetheritislimitedtoaspecificoperationaldesigndomain(ODD);thistermisusedspecificallytodescribeaLevel3,4,or5drivingautomationsystem.

(SAEInternational,2021a)

AuthorisedSelf-DrivingEntity

TheentitythatputsanAVforwardforauthorisationashavingself-drivingfeatures.Itmaybethevehiclemanufacturer,orasoftwaredesigner,orajointventurebetweenthetwo.

(LawCommissionofEnglandandWales&ScottishLawCommission.,2022)

AutomatedVehicle

AmotorvehicleequippedwithADSandthuscapableofperformingdynamicdrivingtasks.(seeBox.1forfurtherdetailsontheusageofthewordinthisreport)

Explainability

ThepropertyofanAIsystemtoexpressimportantfactorsinfluencingtheAIsystemresultsinawaythathumanscanunderstand.

ISO/IEC22989:2022(en),3.5.7

Interpretability

ThepropertyofanAIsystemthatelementsorfeaturescanbeassignedmeanings.DINSPEC92001-3:2023-04

PositiveRiskBalance(PRB)

Thepropositionthatacomputerdrivershouldbenolesssafe(andideallysaferthan)ahumandriver.

Koopman&Widen,2023Koopman&Widen,2024

Robustness

ThedegreetowhichanAIsystemcanmaintainitsleveloffunctionalcorrectnessunderanycircumstances.

ISO/IEC25059:2023(en)

Safety-criticalsystem

Asafety-criticalsystemdescribesasystemthatdirectlyaffectsthesafety,healthandwelfareofthepublicandwhosefailurecouldresultincriticalsafetyissuessuchasinfringementsofprivacy,financialloss,environmentalharm,seriousinjuries,orlossoflife.

(Laplanteetal.,2020;Moteff&Parfomak,2004;SrinivasAcharyulu&Seetharamaiah,2015)

Trustworthiness

Abilitytomeetstakeholderexpectationsinademonstrable,verifiableandmeasurableway

ISO/IEC20924:2024(en),3.1.33

1AI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2025

Executivesummary

Whatwedid

ThisreportexaminesthemainchallengesthatArtificialIntelligence(AI)posesinautomatedtransportsystemsandtheregulatoryapproachestoaddressthem.Thesediversechallengesbroadlyrelatetotechnical,regulatory,economic,societalandenvironmentalissues,includingissuesrelatingtotrainingdataqualityandrepresentation,developmentandverificationofAImodels,increasedvehicletravelandland-useimpacts,deskillingvehicleoperatorsandwiderlabourimpacts.ThereportprovidesacommonunderstandingofAI-basedautomatedtransportsystemsandtheprinciplesthatcanformthebasisofinstitutionalandregulatoryactionstoincreasethesafetyandsocialacceptabilityofusingAI-basedtransportsystems.ThereportisbasedondiscussionsheldatanITFRoundtableinJanuary2023andmaterialspreparedforit.Whilerecognisingtheuniquespecificitiesofeachtransportmode,thisreportmainlyfocusesontheautomationofroadvehicles.Nonetheless,somelessonsfromroadautomatedvehicles(AVs)willbeapplicabletoregulationsonAVsinotherdomains.

Whatwefound

Theautomatedoperationofvehicles–whetherbasedonorsupportedbyAIapplications–holdsgreatpotentialtomeetfuturemobilityneedsinanefficientandsafemanner.Torealisethefullpotentialofautomatedtransport,twoessentialconditionsmustbemettoensureitssafeandsecuredelivery:trustworthinessanddependability.Overcomingtechnicalandregulatorychallengesandminimisingriskswillhelpenhancesocialacceptanceanduptake.

TheSafeSystemapproachprovidesarobust,safety-firstframeworkfordevelopingAVregulations.TheSafeSystemapproachassumesthatmistakesandunexpecteddrivingandoperatingbehavioursareunavoidableandensuresthatthesedonotcontributetoseriousinjuriesordeaths.ThetenetsoftheSafeSystemapproachapplyequallytohuman-basedandAI-enabledvehicleoperation.PublicdialogueonidentifyingacceptablelevelsofriskinlinewiththeSafeSystemapproachisfundamental.TheuseofsimplecomparativeriskmetricsbetweenAVsandhuman-operatedvehiclesraisesrealpracticalchallengesinthecontextofAVcertification.Suchmetricsshouldbesupplementedbyamorefine-grainedapproachthatcompareslike-for-likesafetyperformanceandaddresseschangesinthedistributionofrisksamongthepopulation.

AVsperformoperatingtasks–previouslyperformedbyhumans–usinganAI-basedautomateddrivingsystemknownasADS–anAI-basedoperatingsystemfortrains,vesselsandaircraft.TheseAIsystemshavefundamentallydifferentdecision-makingprocessesfromhumans.Therefore,thetwoseparateregulatorysystemsthathavebeendevelopedforhumanvehicleoperatorsandhuman-operatedvehiclesarenotadaptedtoAVs.Consequently,automatedvehiclesrequirenewinstitutionalandregulatoryarrangementscoveringtheentiretyoftheAI-automatedoperatingsystem.

ThewholeAIlifecycle–includingthecontextinwhichtheAVsareoperated,thedatausedforAVs,theAImodels,outputsandtheirimpactonsociety–shouldbetakenintoaccountwhendevelopingregulationsforAVstoassurethattheAVsaresafe,secureandbeneficialenoughtobecomepartofourtransportsystems.Thoseregulationsmustincludebothtechnicalandnon-technicalmeasures.

EXECUTIVESUMMARY

AI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF20252

ThekeyfactorthatwillunderpinthesustaineddeploymentandbroaderuseofAVsistrustinAVs’abilitytobesafe,sociallyacceptable,andbeneficial.Inadditiontotechnicalrobustnessandsafety,privacyprotection,unbiasedandethicalhandlingofdata,fairness,explainability,andtransparencyarerequiredatallstagesoftheAIlifecycletomakeAVssociallybeneficialandacceptable.

Toaddressdata-relatedissuessuchaspotentialbiasandprivacyinfringement,regulationsareneededtoverifythelawfulandethicalcollectionanduseofdata.Alsoimportantaretheregulatoryarrangementswhichallowpublicauthoritiesorvettedthirdpartiestoverifythatethicalrequirementsaresatisfiedintheacquisitionandprocessingofdata.SyntheticdatacanbebeneficialtotrainAIinrarecaseswhererealdatainputisscarce,butitcouldcreatenewbiasesifnotadequatelymanaged.PrinciplesandguidanceontheuseofsyntheticdataforthetrainingofAIsystemsare,therefore,crucial.

ThetrustworthinessofdataanditsvalidationdependonitsfairandaccurateselectionforthespecificAVusecase.Toprovideassuranceoftheirsafeandpredictableperformanceandrobustness,AIsystemsusedinAVoperationswillneedtobeverifiedandvalidated.ThefunctionsperformedbyAImodelsusedforvehicleoperationincludelocalisation,dynamicsceneunderstanding,pathplanning,control,andmanaginguserinteraction.Eachofthesefunctions–andoverallbehaviour–needstobeevaluatedusingsimulation,testsincontrolledenvironments,andtestsonrealtrafficsituations.Whilescenario-basedtestscanprovideassuranceforcommonscenarios,itmayprovemoredifficultforrareandedgecasesbecauseofthescarcityofavailabledata.Also,evaluationbasedonpredefined,knownscenarioscanleadtooverfittingbymanufacturers.Continuousscenarioupdatesanddiversification–includingbyusingrandomisedscenarios

–areessential.

AVsarenotimmunetoprogrammingerrorsorunexpectedbehaviour.Therefore,processesthataddressthisuncertaintyarenecessary,alongwithpoliciestobothmitigateAV’simpactsandimprovetheirsafetyperformanceex-post.Inlinewithexistingapproachesintheaviationsector,AVroll-outshouldincludeformalprotocolstoensurelessonsarelearnedandintegratedbyallactorsfollowingsafetyincidents.Such“antifragile”approacheswillhelptomaximiseAVsystemsafetydespiteuncertaintyaboutspecificfailuremodes.

Publicauthorities’institutionalcapacityandregulatorymeasuresshouldguaranteethetransparencyofAIdevelopmentanddeploymentprocess.TheyshouldalsoguaranteeasufficientlevelofexplainabilityofAIsystems–evenmoresowhenself-learningAItoolsareused.Thisrequirespublicauthorityinstitutionsandstafftocontinuouslybuildknowledgeandacquireskills.Achievingthisinthefaceofthecurrentconcentrationofskillsintheprivatesectorisachallenge.

TheoperationalenvironmentsofAVsplayanimportantroleintheirsafeuse.ExistingoperatingenvironmentsshouldbeimprovedtomakeiteasierforAVstofunctionsafely,withalowerchanceofencounteringrarebutriskycases.BetterinformationexchangebetweenAVsandotherroaduserscanhelpavoidpotentiallydangeroussituations.Machine-readablelawsandenhancingtheabilityforvehiclesandinfrastructuretocommunicatewillbebeneficialinreducingsuchrisks.

EXECUTIVESUMMARY

3AI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF2025

Whatwerecommend

BaseAIregulatoryandinstitutionalmeasuresonsharedfundamentalprinciples

Fundamentalhumanrightsandthevaluesderivedfromthemlikesafety,fairness,explainability,andhumanoversightshouldformthecornerstoneofAVregulation.ToimprovethesafetyofAVsandtheentiretransportsystem,authoritiesshouldadopttheSafeSystemapproach.Publicauthoritiesmustensureallstakeholdersunderstandwhatconstitutesasafeandacceptablelevelofuncertainty.Moreover,authoritiesmustdesignandimplementappropriateregulatoryinterventionsmeanttodeliversafeoutcomes.

EnsurethatAIremainsexplainable,andthattrainingdataiscollectedandhandledinatransparentandverifiableway

AIsystemsshouldbedesignedinawaythatexplainshowspecificdecisionsaremadebasedonspecificinputs.Thisensuresthatidentifiedrisksdonotresurface.DataisacoreelementofAIsystems.DatahandlingproceduresandsystemsshouldensurethatAIsystemstrainingdatalendsitselftoidentifyingbiases,qualityissues,privacyissues,contaminationfromadversarialattacksorencoding/humanerrors

Mandatereportingofsafety-relevantdatafromautomatedvehicles

Incidentdataispartoftheessential“soft”infrastructurethatensuressafety.Publicauthoritiesshouldmandatethereportingofincidentdatathatissafety-relevant,includingwhenautomatedvehicleoperatingsystemsdisengageduringtestoperating.Dataregardingnearmissesmayalsoproverelevant.Publicauthoritiesshouldestablishmonitoring,reportingandevaluationprocessesthatimproveoverallsafetyperformanceaftereachincident.Metadataonsystemcapabilitiesshouldaccompanythesereports.Allthesedatashouldbeaccompaniedbyaggregateexposuredataondistancescoveredandtheenvironmentsinwhichthevehiclesoperated.Transportauthoritiesmustalsobuildinstitutionalcapacityandtechnicalproficiencytoenforceregulatorymeasures.

DevelopandupdateAVtestscenariosandprocedures

Scenario-basedtestswillplayacentralroleinassessingAVs'abilitiesinaholisticandsaferway.PublicauthoritiesshouldestablishinstitutionalandregulatorymechanismstoensurethattestscenariosarecontinuouslyupdatedandrandomisedtopreventmanufacturersfromdesigningAVperformancethatonlymeetsalimitedsetofpotentialscenarios.

EnsurethatphysicalanddigitalinfrastructuressupportsafeAVs

Enablingmachine-perceivablesignage,markings,andotherimportantvisualcuesinAVoperatingenvironmentsenhancessafeperformance.TofurtherincreasethesafetyofAVoperationandAVinteroperabilityacrossmultipleregionsandcontexts,establishmachine-readablerulesandregulations,andacommonframeworkforvehicle-to-vehicleandvehicle-to-infrastructurecommunications.ThebenefitsfromthesemeasuresextendbeyondtherealmofAVstoallinfrastructureusers.

AI,MACHINELEARNINGANDREGULATION:THECASEOFAUTOMATEDVEHICLES?OECD/ITF20254

Enterthemachine:Takingthe(human)

driveroutofthevehicle

Technologicaldevelopmentsencompassingbothvehiclesandsoftwarehaveenabledhighlevelsofvehicularautomation.Acrossallmodes,includingroad,railandshipping,highly-andfully-automatedvehicles(hereafterAVs)thatarecapableofoperatingthemselveswithouthumaninterventionwithinadesignatedoperatingenvironmentareexpectedtobroadlyimpactsocieties(Bahamonde-Birkeetal.,2018).

AVsareexpectedtohavemultiplefirst-orderimpacts(i.e.directeffectsontransport).AVsareexpectedbymanytoincreasetransportsafetyandimproveaccessibility(EuropeanCommission,2018;ITF,2023b).AVsarealsoexpectedtoreducegeneralisedtransportcostsduetothereplacementofqualifieddrivers,conductors,pilotsorcaptainsbyAI-enabledtechnologies.Second-orderimpacts(i.e.indirecteffectsontransport)includeimpactsontraveldemand,publicrevenue,andlabour,amongothers(ITF,2023a,2023b).However,alltheseexpectedimpactshaveyettomaterialiseasAVdeploymentcurrentlyremainsquitelow.Nonetheless,vigilanceiswarrantedtoensurethatAVdeploymentdoesnotsimplyreplacehumanerrorwithtechnologicalfailuresorflaws.

ForAVstobewidelydeployedandused,specifictechnicalandsocietalchallengesmustbeaddressedandovercome.AkeychallengeistodeveloptherightregulationstoensurethetrustworthinessofAVsinthesensethattheyarebothsafeenoughtobeoperatedalongsidehuman-operatedvehiclesandthattheiruseshouldworkforachievingvaluedsocietalgoalssuchasimprovedaccessibility,enhancedequity,reducedenvironmentalimpact,andeconomicdevelopment(ITF,2023b)

Automatedvehicledeploymentlevelsarenotevenacrossmodes

AVdeploymenthasprogressedunevenlyacrossroads,railways,andwaterways(Fiedleretal.,2019;ITF,2023b).Dependingonthemodeconsidered,differentlevelsofautonomy-correspondingtodifferentcapabilities-havebeendeveloped(e.g.SAElevels,MASSlevels,GradesofAutomationforrailways)(IEC,2014;IMO,2021;SAEInternational,2021a).

Theextenttowhichavehiclecanbeautomatedwilldependondifferentfactors,namely:

?thetypeofinfrastructure(e.g.road,rail,waterways),

?thedegreeofcontrolovertheoperationenvironment(i.e.openenvironment,closedenvironment),

?andthetypeofserviceconsidered(e.g.passengersorgoods).

Forexample,technologyforautomatedsubwayoperationiswidelyavailable,andautomatedsubwayprojectdeploymentstartedinthe1960s(ITF,2023b).However,unlikesubways,whicharetypicallyseparatedfromthesurroundingenvironmentbytunnelsorbarriers,automatedroadvehiclesincitiesinteractwithvariouselements,includingpedestriansandgenerallyhaveacomplexanddynamicoperationaldomain.Foralongtime,automatedvehiclesincitieswillhavetointeractwithabroadmixofvehiclesofvariousagesrunningontechnologieswithdifferingmaturitylevelsanddifferentlevelsofsophistication.Theirdeploymentisthusmuchmorecomplicated.Deploymentofhighlyautomatedvehiclesinprotectedcontexts(likeports,airports,andrail)posesfewerc