USF Libraries
USF Digital Collections

Analysis of how mobile robots fail in the field

MISSING IMAGE

Material Information

Title:
Analysis of how mobile robots fail in the field
Physical Description:
Book
Language:
English
Creator:
Carlson, Jennifer
Publisher:
University of South Florida
Place of Publication:
Tampa, Fla.
Publication Date:

Subjects

Subjects / Keywords:
reliability analysis
field work
robotics
meta-study
fault tolerance
Dissertations, Academic -- Computer Science -- Masters -- USF   ( lcsh )
Genre:
government publication (state, provincial, terriorial, dependent)   ( marcgt )
bibliography   ( marcgt )
theses   ( marcgt )
non-fiction   ( marcgt )

Notes

Summary:
ABSTRACT: The considerable risk to human life associated with modern military operations in urban terrain (MOUT) and urban search and rescue (USAR) has led professionals in these domains to explore the use of robots to improve safety. Recent studies on mobile robot use in the field have shown a noticeable lack of reliability in real field conditions. Improving mobile robot reliability for applications such as USAR and MOUT requires an understanding of how mobile robots fail in field environments. This paper provides a detailed investigation of how ground-based mobile robots fail in the field. Forty-four representative examples of failures from 13 studies of mobile robot reliability in the USAR and MOUT domains are gathered, examined, and classified. A novel taxonomy sufficient to cover any failure a ground-based mobile robot may encounter in the field is presented. This classification scheme draws from established standards in the dependability computing 30 and human-computer interaction 40 communities, as well as recent work 6 in the robotics domain. Both physical failures (failures within the robotic system) and human failures are considered. Overall robot reliability in field environments is low with between 6 and 20 hours mean time between failures (MTBF), depending on the criteria used to determine if a failure has occurred. Common issues with existing platforms appear to be the following: unstable control systems, chassis and effectors designed and tested for a narrow range of environmental conditions, limited wireless communication range in urban environments, and insufficient wireless bandwidth. Effectors and the control system are the most common sources of physical failures. Of the human failures examined, slips are more common than mistakes. Two-thirds of the failures examined in 6 and 7 could be repaired in the field. Failures which resulted in the suspension of the robot's task until the repair was completed are also more common with 94% of the failures reported in 13.
Thesis:
Thesis (M.S.C.S.)--University of South Florida, 2004.
Bibliography:
Includes bibliographical references.
System Details:
System requirements: World Wide Web browser and PDF reader.
System Details:
Mode of access: World Wide Web.
Statement of Responsibility:
by Jennifer Carlson.
General Note:
Title from PDF of title page.
General Note:
Document formatted into pages; contains 144 pages.

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
aleph - 001469382
notis - AJR1136
usfldc doi - E14-SFE0000283
usfldc handle - e14.283
System ID:
SFS0024978:00001


This item is only available as the following downloads:


Full Text

PAGE 1

AnalysisofHowMobileRobotsFailintheFieldbyJenniferCarlsonAthesissubmittedinpartialfulllmentoftherequirementsforthedegreeofMasterofScienceinComputerScienceDepartmentofComputerScienceandEngineeringCollegeofEngineeringUniversityofSouthFloridaMajorProfessor:RobinMurphy,Ph.D.KimonValavanis,Ph.D.DeweyRundus,Ph.D.DateofApproval:March3,2004Keywords:faulttolerance,robotics,reliabilityanalysis,meta-study,eldworkcCopyright2004,JenniferCarlson

PAGE 2

DedicationTomyparentsfortheirunwaveringsupport.

PAGE 3

AcknowledgmentsIwouldliketorstthankmyprimaryprofessor,Dr.RobinMurphy.AlsothankstoDr.DeweyRundusandDr.KimonValavanisforbeingonmythesiscommittee.IwouldliketothanktheDepartmentofEnergyandtheOfceofNavalResearchfortheirsupport.Thankstomyco-workerswhohelpedtogathertheinformationwhichmadethisworkpossible,andprovidedfeedbackalongthewayespeciallyMarkMicire,BrianMinten,AaronGage,Dr.AndrewNelson,RichardGarcia,andLauraBarnes.

PAGE 4

TableofContentsListofTablesivListofFiguresvAbstractviiChapterOneIntroduction11.1Motivation21.2ResearchQuestion41.3Contribution41.4OverviewofAnalysisMethod51.5ThesisOrganization7ChapterTwoRelatedLiterature92.1RobotFailureandQualitativeEvaluationStudies102.1.1MultipleDataSourceRoboticManipulatorFailureAnalyses112.1.2SingleDataSourceMobileRobotReliabilityStudies132.1.3QualitativeEvaluationofMobileRobotsforUSAR142.2FailureAnalysisApproaches152.2.1FailureCharacterizationandClassication162.2.2ReliabilityValidationMethods182.3Fault-toleranceSystems212.3.1Model-basedFault-toleranceSystems212.3.2HybridFault-toleranceSystems232.3.3ExpertSystemBasedFault-toleranceSystems242.3.4DataCentricFault-toleranceSystems242.3.5Fault-toleranceinAutonomicComputing252.4Summary28ChapterThreeTaxonomyofFailuresandMetrics303.1Terminology313.2TaxonomyofFailures343.3Calculations383.4Summary39i

PAGE 5

ChapterFourSourceStudies414.1RobotsSurveyed424.2InformationAvailablebyStudy464.3CRASARWTCStudies494.3.1Approach504.3.2SummaryofResults524.4TECO'sStudiesfromFortLeonardWood544.4.1Approach554.4.1.1ARTS.554.4.1.2CBRNLOE.564.4.1.3D-7G.564.4.1.4DEUCE.564.4.1.5PANTHER.574.4.1.6SARGE.574.4.1.7UGVROP.574.4.1.8UrbanRobot.584.4.2SummaryofResults584.5CRASARHCFRDFieldExperiments594.5.1Approach594.5.2SummaryofResults614.6CRASARReliabilityStudies624.6.1Approach634.6.2SummaryofResults644.7Summary67ChapterFiveMeta-StudyResults705.1PhysicalFailures715.1.1RelativeFrequencyofPhysicalClasses715.1.2Effector735.1.3Sensor765.1.4ControlSystem785.1.5Power795.1.6Communications795.1.7Attributes815.1.7.1Impact.815.1.7.2Repairability.825.2HumanFailures835.2.1RelativeFrequencyofHumanClasses835.2.2Interaction845.2.2.1Mistakes.855.2.2.2Slips.875.3Summary88ii

PAGE 6

ChapterSixConclusions916.1Findings926.2Discussion956.3FutureWork97References99Appendices105AppendixA:DenitionsofReliabilityRelatedTerms106AppendixB:ExampleFieldFailureswithClassications109iii

PAGE 7

ListofTablesTable1.ComparisonofRelatedStudiesIncludingSynthesisMethod,andRobotandFailureTypes.13Table2.ComparisonofRelatedStudies:NumberofRobotsandFailureAt-tributesExamined.13Table3.ComparisonofRobotFailureandQualitativeEvaluationStudies.15Table4.ComparisonofFailureandReliabilityAnalysisApproaches.20Table5.StrengthsandWeaknessesofAnalysisApproaches.20Table6.AssumptionsMadebyFault-toleranceApproachesAboutMobileRobotFailures.27Table7.HowFault-toleranceApproachesCanBenetfromthisStudy.27Table8.TheRobotsExaminedinthisMeta-study.43Table9.TheRobots'Characteristics.43Table10.OverviewofDataCollectionMethodInformationAvailablefromEachStudy.48Table11.OverviewofDataandAnalysisInformationAvailablefromEachStudy.49Table12.FailuresEncounteredatWTCfrom[34].53Table13.SummaryofResultsfromtheCRASARReliabilityStudies[6][7].65Table14.ProbabilitybyPhysicalClassfromtheReliabilityStudies[6][7].72Table15.ProbabilitybyPhysicalClassforM1PANTHER[13].73Table16.ComparisonofTerminalVersusNon-terminalFailuresfromtheM1PantherStudy[13].81Table17.ComparisonofField-repairedvs.NotField-repairedFailures.82Table18.HumanFailureAnalysisResults.84iv

PAGE 8

ListofFiguresFigure1.AnInuktunMicroVGTVInsideaConnedSpaceTrainingMaze.3Figure2.AniRobotPackbotClimbingaRubblePileUsedforUSARTraining.3Figure3.TheTaxonomyofMobileRobotFailuresUsedinthisAnalysis.7Figure4.AConnedSpaceTrainingMaze.32Figure5.ARubblePileUsedforUSARTraining.32Figure6.TheTaxonomyofMobileRobotFailuresUsedinthisAnalysis.35Figure7.InuktunMicroVGTVleftandMicroTracsrightRobots.44Figure8.Foster-MillerSolemleftandURBOTBuiltonaSolemBaseright.45Figure9.iRobotATRV-Jrleft.iRobotPackbotrightExploringaRubblePile.45Figure10.Foster-MillerMATILDAleftandTalonright.46Figure11.SARGEisBuiltonaYamahaBreezeATVBase.47Figure12.ARTSBuiltonaAllSeasonsVehiclesMD-70BaseleftandCaterpillarDEUCEright.47Figure13.CaterpillarD-7GandPANTHERBuiltonaUSArmyM1TankBase.47Figure14.ANomadResearchRobotIncludedintheCRASARReliabilityStudies.65Figure15.ProbabilitybyPhysicalClassfromtheFollow-upReliabilityStudy[7].73Figure16.DirtFoundNearSensitiveEquipmentInsideaniRobotUrban.74Figure17.RockStuckinARTSTrackMechanism.75Figure18.FailureEncounteredattheWTC.76Figure19.ComparingWTCWorkingConditionstothatoftheHCFRDStudy.85Figure20.ARTSonitsSideAfteraFall.86v

PAGE 9

Figure21.SummaryofClassicationResultsUsingtheFailureTaxonomy.89vi

PAGE 10

AnalysisofHowMobileRobotsFailintheFieldJenniferCarlsonABSTRACTTheconsiderablerisktohumanlifeassociatedwithmodernmilitaryoperationsinurbanterrainMOUTandurbansearchandrescueUSARhasledprofessionalsinthesedomainstoexploretheuseofrobotstoimprovesafety.Recentstudiesonmobilerobotuseintheeldhaveshownanoticeablelackofreliabilityinrealeldconditions.ImprovingmobilerobotreliabilityforapplicationssuchasUSARandMOUTrequiresanunderstandingofhowmobilerobotsfailineldenvironments.Thispaperprovidesadetailedinvestigationofhowground-basedmobilerobotsfailintheeld.Forty-fourrepresentativeexamplesoffailuresfrom13studiesofmobilerobotreliabilityintheUSARandMOUTdomainsaregathered,examined,andclassied.Anoveltaxonomysufcienttocoveranyfailureaground-basedmobilerobotmayencounterintheeldispresented.Thisclassicationschemedrawsfromestablishedstandardsinthedependabilitycomputing[30]andhuman-computerinteraction[40]communities,aswellasrecentwork[6]intheroboticsdomain.Bothphysicalfailuresfailureswithintheroboticsystemandhumanfailuresareconsidered.Overallrobotreliabilityineldenvironmentsislowwithbetween6and20hoursmeantimebetweenfailuresMTBF,dependingonthecriteriausedtodetermineifafailurehasoccurred.Commonissueswithexistingplatformsappeartobethefollowing:unstablecontrolsystems,chassisandeffectorsdesignedandtestedforanarrowrangeofenvironmentalconditions,limitedwirelesscommunicationrangeinurbanenvironments,vii

PAGE 11

andinsufcientwirelessbandwidth.Effectorsandthecontrolsystemarethemostcommonsourcesofphysicalfailures.Ofthehumanfailuresexamined,slipsaremorecommonthanmistakes.Two-thirdsofthefailuresexaminedin[6]and[7]couldberepairedintheeld.Failureswhichresultedinthesuspensionoftherobot'staskuntiltherepairwascompletedarealsomorecommonwith94%ofthefailuresreportedin[13].viii

PAGE 12

ChapterOneIntroductionThisthesisprovidesadetailedinvestigationofhowground-basedmobilerobotsfailintheeld.Forty-fourrepresentativeexamplesoffailuresfrom13studiesofmobilerobotreliabilityintheUrbanSearchandRescueUSARandMilitaryOperationsinUrbanTerrainMOUTdomainsaregathered,classied,andexamined.Anoveltaxonomyispresentedandusedwhichdrawsfromthedependabilitycomputing[30],human-computerinteraction[40],androbotics[6]communitiesandissufcienttocoveranyfailureaground-basedmobilerobotmayencounterintheeld.Bothphysicalfailuresfailureswithintheroboticsystemandhumanfailuresareconsidered.The13studiescomefromtwoprimarysources.TheCenterforRobot-AssistedSearchandRescueCRASARattheUniversityofSouthFloridaprovidedveofthestudies.CRASARspendsmorethan200hoursperyearusingtherobotsintheeldandcurrentlyhastwenty-onerobotsfromsixmanufacturers.TheothereightcomefromtheTestandEvaluationCoordinationOfceTECO,partoftheManeuverSupportCenteratFortLeonardWood.TECOprovidesoperationaltestandevaluationexpertisetotheChemical,EngineerandMilitaryPoliceSchoolsandassistsinthedevelopmentandexecutionofAdvancedWarghtingExperimentsAWEfortheUSArmy.Overallrobotreliabilityineldenvironmentsislowwithbetween6and20hoursmeantimebetweenfailuresMTBF,dependingonthecriteriausedtodetermineifafailurehasoccurred.Commonissueswithexistingplatformsappeartobethefollowing:unstablecontrolsystems,chassisandeffectorsdesignedandtestedforanarrowrangeof1

PAGE 13

environmentalconditions,limitedwirelesscommunicationrangeinurbanenvironments,andinsufcientwirelessbandwidth.Effectorsandthecontrolsystemarethemostcommonsourcesofphysicalfailures.Ofthehumanfailuresexamined,slipsaremorecommonthanmistakes.Two-thirdsofthefailuresexaminedinCRASAR'sreliabilitystudies[6][7]couldberepairedintheeld.Failureswhichresultedinthesuspensionoftherobot'staskuntiltherepairwascompletedarealsomorecommonwith94%ofthefailuresreportedinTECO'sM1PANTHERIIstudy[13].1.1MotivationTheconsiderablerisktohumanlifeassociatedwithmodernMOUTandUSARhasledprofessionalsineachofthesedomainstoexploretheuseofadvancedtechnology,likerobotics,toimprovesafety.MobilerobotsareappealingforUSARandthemilitarybecausetheycanbesenteitheraheadoforinplaceofhumansinparticularlyhazardoussituations.Theyarealsocapableofdoingthingshumanscannot,likeenduringlowoxygenenvironmentswithoutsupportequipmentandworkingforindenitelylongshiftswithoutbecomingfatiguedifprovidedsufcientpower.Mobilerobotscansendbackawidevarietyofdatafromtheiron-boardsensors.Despitethelureofthesefeatures,onlyafewmobilerobotshavebeenusedinrealUSARe.g.theWorldTradeCenterrescueresponse[34][8]andmilitaryoperationssuchascavereconnaissanceinAfghanistan[2].Thenalacceptanceofrobottechnologyinthesenewapplicationareaswilldependasmuchontheirreliability,asonthecapabilitiesoftherobotplatformsuchastheabilitytodetectchemicalorbiologicalagents.Afragilerobotinconstantneedofmaintenanceandrepairislikelytobeleftbehindtomakeroomformorereliableequipment.Thereareavarietyoffactorswhichmakeeldworkparticularlychallengingforrobotsasopposedtolabconditions.Forexample,conditionswhicharobotmay2

PAGE 14

Figure1.AnInuktunMicroVGTVInsideaConnedSpaceTrainingMaze. Figure2.AniRobotPackbotClimbingaRubblePileUsedforUSARTraining.encounterinaeldenvironmentforUSARand/orMOUTinclude:dirt,standingwater,rain,intenseheat,intensecold,connedspacesseeFigure1,unevensurfacesseeFigure2,thepresenceofobstacleswithunpredictablemovement,andhostileagents.Recentstudiesonmobilerobotuseintheeldhaveshownanoticeablelackofreliabilityinrealeldconditions.In[6]themeantimebetweenfailuresMTBFforeldrobotswasalittleover6hoursandtheavailabilityratewasonly50%.Theanalysisin[34]showedthattetheredrobotsrequiredassistancethroughthetetheranaverageof2.8timesperminute.StudiesperformedbyTECOhavefoundaMTBFlessthan20hours.3

PAGE 15

ImprovingmobilerobotreliabilityforapplicationssuchasUSARandMOUTrequiresanunderstandingofhowmobilerobotsfailineldenvironments.1.2ResearchQuestionHowdoground-basedmobilerobotsfailintheeld?Understandingmobileroboteldfailuresrequiresknowledgeofthecausesoffailuresandtheircharacteristics.Thesecharacteristicsincludethefrequency,symptoms,andimpactofeachtypeoffailure.Itisalsoimportanttoknowhowtherobots'operatingenvironment,anddecisionsmadeduringtheirdesignaffectthesecharacteristics.Awidevarietyofplatformsmustbeexaminedtoisolatetraitsoffailuresthatspanallground-basedmobilerobotswhichcouldbeusedintheeld.Theorganizationswhichoperateinhazardousdomains,likeUSAR,relyheavilyoninteraction[9].Evenafullyautonomousmobilerobotmuststillinteractwithandtakeordersfromhumans[37].Therefore,notonlythefailuresoftherobotplatformbutalsohumanoperatorerrorsmustbeconsideredwhenexaminingmobilerobotreliabilityinthesedomains.1.3ContributionThisworkprovidestwomajorcontributions:1.AtaxonomyofrobotfailuresbuiltfromasynthesisoffailuretaxonomiesinthreeseparatecommunitieswithintheeldofComputerScience.2.Ameta-studyincluding44representativeexamplesofmobileroboteldfailuresdrawnfrom13studiesintheUSARandMOUTdomains,arguablytwoofthemostchallengingelddomainsforground-basedmobilerobots.Theseexamplesdemonstratehowmobilerobotfailurescanbeclassiedusingthenewtaxonomyandthechallengesassociatedwithusingrobotsineldenvironments.4

PAGE 16

Thetaxonomywascreatedfromexperienceintherobotics[6],human-computerinteraction[40]anddependabilitycomputing[30]communities.Failuresarecategorizedbasedonthesourceoffailurephysicalandhuman.Twoattributes,repairabilityandimpact,areusedtocapturetheseverityandrepercussionsofthephysicalfailures.Thoughthetaxonomywasdesignedtocoveronlyeldfailures,itisexpectedtobesufcienttocoveranyapplicationofmobilerobots.Informationonhowandwhenmobilerobotsfailhelpstoidentifytheweaknessesofcurrentmobilerobottechnology,whichinturn,illuminatesthechallengeswhichrobotmanufacturersanddevelopersoffault-tolerantcontrolsystemsmustmeettoimprovetheirreliability.Inaddition,potentialadoptersofmobilerobottechnologycanbenetfromanunbiased,quantitativeassessmentofcurrenttechnology.Thisgivesthemtheabilitytobalancethecapabilitiesamobilerobotwillbringtotheirapplicationdomainagainsttheactualcostofmaintainingtheequipment.Dataonhowmobilerobotsfailcanalsobeusedtoprovidearealisticstartingpointforfaultmodelinginmodel-basedfaulttolerancesystems,suchas[25],[32],[48],[51],[54],and[61].Inaddition,robotfault-toleranceapproacheswhichusecostlydiagnosistechniques,likeactiveprobinggatheringadditionalinformationfromothersensorsorrobotsusedbyLongandMurphy[31],wouldalsobenetfromtheincorporationofprobabilitydatatorankhypotheses,reducingthecostofdiagnosisbyensuringthatthemostlikelyhypothesesarecheckedrst.1.4OverviewofAnalysisMethodThisthesisexaminesmobilerobotfailuredatafrom13studiesfromtheCenterforRobot-AssistedSearchandRescueCRASARattheUniversityofSouthFloridaandtheTestandEvaluationCoordinationOfceatFortLeonardWood[42].TheseincludetheCRASAR'sWorldTradeCenterWTCstudies[34][8],eldexperimentswith5

PAGE 17

HillsboroughCountyFireRescueDepartment[10],andreliabilityanalyses[6][7];aswellasTECO'seightmobilerobotstudiesdescribedin[42].Atotalof28robotswereconsideredinthisthesis,representing15differentmodelsfromsevenmanufacturers.Theyrangefromsmalllessthan10poundstrackedvehiclescapableofchangingtheirgeometry,toamodiedM1tankover60tons.Forthepurposesofthispaper,aeldenvironmentisdenedasanenvironmentwhichhasnotbeenmodiedtoensurethesafetyoftherobotortoenhanceitsperformance,andafailureistheinabilityoftherobotoritssupportequipmenttofunctionnormally.Notethatbothcompletebreakdownsandnoticeabledegradationsareincluded.AsdiscussedinSection1.3anoveltaxonomyofmobilerobotfailureswasdevelopedforthismeta-study.ThetaxonomyshowninFigure3usesclassestocapturethesourceofthefailurewhichcanbeeitherphysicalsystemorrobotorhuman.Fivesubclasses,basedoncommonsubsystemsfoundinallrobotsystems,fallwithinthephysicalbranch.Theseareeffector,sensor,controlsystem,power,andcommunications.Humanfailuresaredividedintodesignandinteractionsubclasses,thelatterofwhichisfurthersubdividedintomistakesandslips.Twoattributesarealsoincludedtodescribetheseverityofthephysicalfailuresintermsofitsrepairabilityandimpactontherobot'smissionatthetimeofthefailure.Thevaluesgiventotheseattributesareeld-repairableandnon-eld-repairable,andterminalandnon-terminalrespectively.Duetodifferencesindatacollectionandreportingmethodsamongthestudies,quantitativeanalysisofthefailuredatawasoftennotpossible.Therefore,themajorityofthendingsarebasedonexaminationoftheexamplesofeldfailuresdescribedineachofthestudies.Wherequantitativeresultscouldbegenerated,standardformulastakenfrom[24]wereusedtoconvertthedataintocommonreliabilitymetrics.OnesuchmetricwasthemeantimebetweenfailuresMTBF,whichprovidesaroughestimateofhowlongonecanexpecttousearobotwithoutencounteringfailures.Projectedavailability,6

PAGE 18

Figure3.TheTaxonomyofMobileRobotFailuresUsedinthisAnalysis.Classesareshownwithsolidlines,andattributeswithdashedlines.alsocalledreliability,wasalsoused.Thismetricisreportedasapercentageandshouldbeinterpretedastheprobabilitythattherobotwillbefreeoffailuresataparticularpointintime.Allotherquantitativeresults,suchastheaveragedowntimeandprobabilitythatafailurewascausedbyacomponentinclassfromthetaxonomyc,werealsocalculatedusingstandardformulas.1.5ThesisOrganizationChapterTwoprovidesanoverviewofrelatedworkinrobotreliability,failureandreliabilityanalysisapproaches,andfaulttolerantsystems,establishingtheuniquenessandpotentialbenetsofadetailedstudyonhowmobilerobotsfailintheeldwithinthelargerroboticscommunity.ChapterThreedescribesthenewmethodforexaminingmobilerobotfailuresusedinthismeta-studyincludingsomebasiccriteria,anoveltaxonomydevelopedformobilerobotfailures,andformulasusedtosummarizethefailuredataintermsofreliabilitymetrics.InChapterFour,the13studiesincludedinthismeta-studyaredescribedindetail,includingtherobotsexamined,theinformationavailablefromeachstudy,theirapproachtofailuredocumentationandanalysis,andasummaryofimportantresults.ChapterFivepresents44representativeexamplesofmobile7

PAGE 19

roboteldfailuresfromthe13studies,andanynumericresultsavailable.Itsorganizationfollowsthenewtaxonomyexactly,providinganexampleofhowthetaxonomycanbeusedtoclassifyfailureeventsand,inturn,usingthoseexamplestohighlightthecharacteristicsoffailureswhichfallintoeachclass.Finally,inChapterSix,asummaryofthendingsofthismeta-studyisgivenfollowedbyadiscussionoftheimplicationsofthesendingsandrecommendedavenuesforimprovement.ChapterSixcloseswithanoverviewoffutureworkonthistopic.8

PAGE 20

ChapterTwoRelatedLiteratureThischapterprovidesanoverviewofrelatedworkinrobotreliability,failureandreliabilityanalysisapproaches,andfaulttolerantsystems.Itestablishesboththeuniquenessandpotentialbenetsofadetailedstudyonhowmobilerobotsfailintheeldwithinthelargerroboticscommunity.Section2.1comparesrelatedworkonrobotfailureanalysisandreliability.Section2.2examinesapproachestofailureandreliabilityanalysisfoundintheliterature,lookingforsuitabletechniquesforameta-studyofmobileroboteldfailures.Section2.3presentsrelevantworkinfault-tolerancesystemsformobilerobots.Finally,Section2.4providesasummaryoftheresultsofthisliteraturereview.Section2.1presentsstudiesfoundintheliteraturewhichattempttocharacterizeandimproverobotreliability,excludingthe13studiesincludedinthismeta-study.Thisincludesstudieswhichseektoimproverobotreliabilitythroughfailureandreliabilityanalyses,aswellasqualitativestudiesofmobilerobotsuitability.Itisshownthatthisthesisisuniqueinthreerespects:bothhumanandrobotfailuresareexamined,itcoversrobotfailuresintheurbansearchandrescueUSARandmilitaryoperationsinurbanterrainMOUTdomains,anditistheonlymeta-studythatcoversmobilerobotfailures.Section2.2coverstwodistinctapproachestofailureandreliabilityanalysisfoundintheliterature:characterizationandformalvalidation.Characterizationapproachesfromthedependabilitycomputing,human-computerinteraction,androboticscommunitiesarepresented.Theseclassifyfailuresbasedonapre-denedsetofcategories.Formalvalidationapproachesfromtheroboticsandcontroltheoreticliteraturearealsoincluded.9

PAGE 21

Thesecreateacompletemodelofthesystemtobestudiedandthenexaminethatmodeltodetermineifitmeetsreliabilitycriteria.Thissectionestablishesthatnoexistingapproachissuitableforthetaskofmobileroboteldfailureanalysis.TheclassicationschemesdevelopedfordependabilitycomputingandHCIarenotsufcient,andformalvalidationmethodscanonlybeappliedtoasinglesysteme.g.robotmodeland/orasingletaskatatime.Thisestablishestheneedforanewmethodofrobotfailureanalysis.ThisthesiscontributessuchamethodthatisdescribedindetailinChapterThree.Failuresappearmostoftenintheroboticsliteratureinstudieswhichendeavortocreatefault-tolerancesystemsthatcandetect,isolate,and/orrecoverfromfailures.Fault-tolerancework,presentedinSection2.3,isofinteresttothismeta-studyfortworeasons.First,afault-tolerancesystemiscreatedforfailureswithcertaincharacteristics.Itisofinteresttoseeifthecommunityasawholeagreesoncommoncharacteristicsofrobotfailures.Second,fault-toleranceresearchersanddevelopersareapotentialconsumeroftheresultsofthisstudy.Thissectionrevealsthat,whilenoneofthestudiesdiscussedpresentedanyevidenceofhavinginvestigatedthekindsoffailuresrobotsencounter,eachapproachmadecertainassumptionsaboutthecharacteristicsofthosefailures.Itisconcludedthatthefault-tolerancecommunitycanbenetfromthismeta-study,whichprovidesinformationoncommonfailuresforelddomainslikeUSARandMOUT.Fault-tolerancesystemswhichuseprobabilisticmodeling,inparticular,canusethefrequencyoffailuresandrelativeprobabilityoffailureforrobotsubsystemspresentedinChapterFiveashigh-delityestimatesforUSARandMOUTenvironments.2.1RobotFailureandQualitativeEvaluationStudiesThissectioncoversstudieswhicharerelatedtothismeta-studybyhavingsimilargoals,namelytocharacterizeandimproverobotreliability.Specically,Sections2.1.1and2.1.210

PAGE 22

coverstudiesthatseektoimproverobotreliabilitythroughfailureandreliabilityanalysesrespectively.InSection2.1.3qualitativestudiesofmobilerobotsuitabilityfortheUSARdomainarepresented.Thestudiescoveredinthissectionarenotincludedinthemeta-studyforoneoftworeasons:theystudyrobotsusedinindoor,controlledenvironments;ortheypresentshortcomingsofmobilerobotsforUSARbasedonthecharacteristicsofthedomain,ratherthantheexaminationofdocumentedfailures.2.1.1MultipleDataSourceRoboticManipulatorFailureAnalysesTwostudieswerefoundintheliteraturethat,likethismeta-study,examinedrobotfailuresfrommultiplesources.Bothcoveredindustrialmanipulatorsusedprimarilyintheautomotiveindustry.BeauchampandStobbe[1]examineddocumentedhuman-robotsystemaccidentsfromninedifferentstudies.Starr,Wynne,andKennedy[52]surveyedfailureandmaintenancereportsfromtwolargeautomotiveplants.In[1]BeauchampandStobbepresentareviewofhumanfactors1experimentalstudiesonhuman-robotsystemaccidents,wheretherobotisanautomatedindustrialmanipulator.Thereviewincludedtwotypesofstudies:studiesofdocumentedhuman-robotsystemaccidentsninesummarized,andrelatedhuman-factorsexperimentsincluded.Thefocusofthepaperwasthelatter.Thereviewofthehuman-robotaccidentstudiesservedmainlyasmotivationforadditionalworkinhuman-factorsonindustrialmanipulatorsystems.Thendingsofeachoftheninestudiesweresummarizedindividually.Noclassicationschemewasdevelopedtodescribethedocumentedaccidents.Neithertherateoffailurenortheimpactofthefailuresonthehumaninvolvedorplantproductivitywasdetermined.Aquicksummaryoftheaccidentstudiesprovidedthefollowingoverallndings:mostaccidentsoccurduringprogrammingtrainingthe 1Humanfactorsisaeldofstudyinvolvingresearchintohumanpsychological,social,physical,andbiologicalcharacteristics,andworkingtoapplythatinformationtothedesign,operation,anduseofproductsorsystemstooptimizehumanperformanceandsafety.11

PAGE 23

manipulatortoperformataskand/ormaintenance,whenapersonislikelytobefoundwithintheoperatingenvelopeareainwhichtherobot'sarmcouldmoveofthemanipulator;andtheaccidentsfrequentlyinvolvedunexpectedrobotmovementwhichwascausedbyeitherequipmentfailuremaintenanceorhumanerrorprogramming.Starr,etal.presentin[52]asurveyofthefailuresofroboticindustrialmanipulators.Thesurveyincludesfailureandmaintenancereportsfromtwoautomotiveplants.Thedatacovered200robotsfromvemanufacturersoveraperiodof21weeksforPlantAand5weeksforPlantB.InPlantA,therobotswerefoundtobedownforrepairormaintenancefor3.95hoursperweekpermanufacturingline,andinPlantB,thedowntimeratewas1.74hours.Robotusagewasnotrecorded,thereforethemean-timebetweenfailurescouldnotbedetermined.Theanalysiscreatedeightcategoriestosummarizethe60and100differentcodesusedtoindicatethetypeoffailureusedinPlantAandPlantB,respectively.Non-robotfailures%ofallfailuresencounteredwereplacedintheirowncategoryandwerenotanalyzedfurther.Oftherobotfailures,positionfailureswerethemostcommonat45%,followedbydrivesystemhardwarefailuresat25%.ThedatafromPlantBwerealsoanalyzedingroupsbymanufacturer.Thedifferencesintheresultsforeachmanufacturerweredeterminedtobeduetodifferencesinthemanipulators'tasksspotweldingversustransportationanddrivesystemelectricalversushydraulic.Forexample,thehydraulicrobotshaddoublethenumberoffailuresrecordedpermanufacturinglinecomparedtotheelectricrobots.Tables1and2provideasummaryofthedifferencesbetweenthetwostudiescoveredinthissectionandthismeta-study.Table1liststhetypeofrobot,typeoffailuresrobotorhuman,andthemethodusedtosynthesizethedataforfurtheranalysis.Table2providesadditionalinformationforcomparisonintermsofthenumberofrobotsandthefailureattributesexamined.Thesetablesshowthatthismeta-studycoversmobilerobotfailuresinamorecompleteincludingbothhumanandrobotsfailuresandorganized12

PAGE 24

Table1.ComparisonofRelatedStudiesIncludingSynthesisMethod,andRobotandFail-ureTypes.Manipulatorsreferstoautomatedindustrialmanipulators. Study RobotType FailureType SynthesisMethod BeauchampandStobbe[1] Manipulators Humanonly None Starretal.[52] Manipulators Robotonly Classiedusingcategoriesdevelopedforthisanalysis Thismeta-study Mobilerobots HumanandRobotFieldFailures Classiedusingacompletefailuretaxonomy Table2.ComparisonofRelatedStudies:NumberofRobotsandFailureAttributesExam-ined. FailureAttributesExamined Study #Robots Frequency Impact RelativeFrequencyofCategories BeauchampandStobbe[1] Unknown No No No Starretal.[52] 200from5manu. No Yes Yes Thismeta-study 28from7manu. Yes Yes Yes fashionthanthetwostudiesofindustrialmanipulatorfailures.Table2alsoshowsthatStarretal.'sstudyislargerinscopeintermsofthenumberofrobotsexamined.Thisisduetothefactthatindustrialmanipulatorshavebeenacceptedasessentialtoolsintheautomotiveindustry.Nosuchindustryexiststodayformobilerobots.2.1.2SingleDataSourceMobileRobotReliabilityStudiesTheWorkshoponRobotsinExhibitionsatIROS2inSeptemberof2002producedtwostudiesonthereliabilityofmobilerobotsactivelyusedforlongperiodsoftime.Nourbakhsh[41]describesasetoffourautonomousrobotsusedforaperiodofveyearsasfull-timemuseumdocents.TheirrobotsreachedameantimebetweenfailuresMTBF 2IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems13

PAGE 25

of72to216hours.Tomatis,Terrien,Piguet,Burnier,Bouabdallah,andSiegwartin[58]describeasimilarproject.Ascomparedtothescopeofthismeta-study,theanalysisdescribedbyTomatiset.al.wasmorenarrowbothintheapplicationsandtherobotsanalyzed.Theoperatingenvironmentwasalsoindoorsandengineeredtoassisttherobot.However,theirMTBFwas7hourssimilartothe8.3hourMTBFfoundintheoriginalreliabilitystudy[6]performedbytheCenterforRobot-AssistedSearchandRescueCRASAR.2.1.3QualitativeEvaluationofMobileRobotsforUSARThreestudieshaveconcentratedonidentifyingtheweaknessesofground-basedmobilerobotsforUSARbasedsolelyonthecharacteristicsoftherobotandtheconstraintsofthedomain.Blitchprovidesasurveyin[2]ofthemobilityproblemskeepingcurrentrobottechnologyfrompopulatingwell-suitednicheswithinUSAR,especiallytheconnedspaceaccessniche.Tumblerecovery,traction,andtheincorrectassumptionofanobstacle-freeworkingenvelopeareidentiedasthekeyproblems.Casper,Micire,andMurphy[9]presentanoverviewoftheUSARdomain,listingtasksthatrobotsarebestsuitedfor,followedbyadiscussionoftheconstraintsthatapplicationdomainplacesonrobotictechnology.Sensorsareidentiedastheareawhichrequiresthemostimprovement,thoughweather-proongandaninvertiblechassiswerealsomentionedasrequiredfeaturesrarelyfoundinrobotsatthattime.In[37]Murphy,Casper,Hyams,Micire,andMintendiscussthesameissuesasCasperetal.[9]butprovidesomeadditionaldiscussionontheneedforadjustableautonomy,ortheabilitytochangetheallocationofcontrolbetweentherobotanditsoperator.Table3providesasummaryofthepapersdiscussedinthissection,includingthismeta-study,forcomparison.Thetypesofrobotsanalyzed,thevarietyofmodels,thetargetenvironmentorapplicationdomain,thekindsoffailureshumanorrobotexamined,and14

PAGE 26

Table3.ComparisonofRobotFailureandQualitativeEvaluationStudies.Forthemeta-studies,thenumberofdatasourcesisindicatedinparenthesesafterthegroupname. Group Robots Models Environment Failures Analysis Thismeta-study Mobilerobots Varied USAR&MOUT HumanandRobot Failure Multiple-sourceStudies[1][52]total IndustrialManipulators Varied Factory HumanorRobot Failure Single-sourceStudies[41][58] Mobilerobots Single Museum Robot Reliability QualitativeEvaluation[2][9][37] Mobilerobots Varied USAR Robot Suitability theanalysistypeareincluded.Table3showsthatthismeta-studyisuniqueinthatitcoversbothhumanandrobotfailures,itisfocusedonroboteldfailuresintheUSARandmilitaryoperationsinurbanterrainMOUTdomains,anditistheonlymeta-studythatcoversmobilerobotfailures.2.2FailureAnalysisApproachesThissectioncoversthetwodistinctapproachestofailureandreliabilityanalysisfoundintheliterature:characterizationandformalvalidation.CharacterizationapproachesSection2.2.1grouporclassifydocumentedfailuresbasedonapre-denedsetofcategories,usuallyexaminingthesetoffailuresasawholeandeachgroupindividually.FormalvalidationapproachesSection2.2.2createamodelofthesystemtobestudiedandthenexaminethatmodeltodetermineifitmeetsreliabilitycriteria.Formalvalidationapproachesareusefulforanalysisofasinglespeciedtaskwhichcanbeexplicitlymodeled.Forcharacterizationapproachesthegeneralityofthecategoriesistheonlylimittothekindsoffailuresthatcanbeexaminedinasinglestudy.Thismeta-studydealswith15

PAGE 27

mobilerobotsusedinawidevarietyofenvironmentsforawidevarietyoftasks,whichiswhythisstudyemploysacharacterizationapproach.2.2.1FailureCharacterizationandClassicationThissectionpresentsexistingclassicationschemesoffailureswhichweredevelopedforthedependabilitycomputing,robotics,andhuman-computerinteractionHCIcommunities.Inmostfailureanalyses,likeStarretal.[52],thecategoriesusedforclassicationweredevelopedthroughexaminationofthecommonattributesofthesetoffailurestobeexamined.Theseclassicationschemestendtobelimitedtothescopeofthesetoffailures,andtheresultsofsuchanalysesaredifculttoapplytootherapplicationsordomains.Thereforethissubsectiononlyincludesrelevantfailureclassicationschemescreatedforanentireclassoffailures,suchashuman-computerinteractionfailures,ratherthanthosecreatedforaspecicfailureanalysis.In1984,Laprie[30]andhiscolleaguesdevelopedasetofconceptsanddenitionsrelatedtothedependabilityofcomputer-basedsystems.AccordingtoLaprieafaultissimplyacause,anerrorisastate,andafailureisanevent.Specically,afailureisdenedasadeviationfromthespeciedserviceasseenbytheclient.Theclientmaybeahumanuseroranothercomponentofthecomputersystemthatistryingtousetheservice.Anerrorisastatewithinthesystemwhichcanleadtoafailure.Afaultisanythingwhichcouldcausethesystemtoenteranerrorstate.Lapriedenestwomajorfaultclasses,namelyphysicalfaultsandhuman-madefaults.Human-madefaultsarefurthersubdividedintodesignfaultsandinteractionfaults.[30]denestwolevelsforseverityforfailures.Theconsequencesofbenignfailuresarecomparabletothebenetsoftheservicetheyarepreventing.Malignorcatastrophicfailureshaveahighercostbyoneormoreordersofmagnitudethantheservice.16

PAGE 28

Laprie'sdependabilitytaxonomyisgeneralenoughthatitcanbeappliedtoalargevarietyofsystemsandapplicationdomains.Unfortunately,itisnotdetailedenoughforanalyzingmobilerobotreliabilityintheeld.Forexample,amobilerobotcansufferfromaninnitevarietyofphysicalfaults.Laprie'slevelsofseverityarealsodifculttoapplysincethebenetofaserviceandthecostofafailuretendtovarywidelybasedonthesituation,thatisamilitarytrainingexerciseversusarealengagement.Nevertheless,itprovidesasolidfoundation,sothetaxonomyusedinthispaperseeSection3.2drawsagreatdealfrom[30].In[28]KokkinakiandValavanispresenttheerrorspecicationusedintheirreliabilityvalidationapproachforcomputer-integratedmanufacturingCIMsystems.TheirspecicationisinterestinginthatitissimilartoLaprie's.Faultsanderrorshaveroughlythesamedenitionsand[28]statesthatfaultsmaybecausedbyaphysicaldefect,theenvironment,oranoperatorthoughnoexplicittaxonomyofcausesiscreated.Faultsareclassiedtemporallyaspermanentwhichexistuntilrepaired,transientwhichdisappearontheirown,andintermittentwhichrepeatedlyappear.Errorsareclassiedbasedontheirscopeofinuence:null-pointerrorsdonotaffectthesystem,single-pointerrorsarelocalizedtoasingleagenttaskwithinthesystem,andmultiple-pointerrorsaffectmorethanoneagent.Afailureisdenedasanerrorthatspansoneormoreagentsinthecurrentplansetoftasksthemachineiscurrentlyexecuting.Afailureisrecoverableifthesystemcancompletetheplaninspiteofthefailure,irrecoverableifitcannot,andcatastrophicifallagentsareeffectedbytheerror.InpreviousworkbyCarlsonandMurphy[6]aclassicationschemewasusedtoexaminefailuresencounteredduringday-to-dayuseofmobilerobotsinlab,ofce,andUSARenvironments.Afailurewasdenedastheinabilityoftherobotortheequipmentusedwiththerobottofunctionnormally.Categoriesweredenedtocapturethesourceofthefailurebasedonthecommonsubsystemsofamobilerobot:effectora.k.a.actuator,17

PAGE 29

sensor,controlsystem,power,andcommunications.Twoattributeswereusedtocapturetheseverityofthefailuresintermsofrepairability,eld-repairableversusnon-eld-repairable,anditsimpactontherobot'smission,terminalversusnon-terminal.Thisschemewasrstveriedthroughinterviewswithexperiencedrobotoperatorsandhardwarespecialists.Thelatterweregraduatestudentswithextensiveexperiencemaintainingandrepairingrobotplatforms.Theschemewasthenvalidatedthroughitsapplicationonawiderangeoffailuresrecordedoveratwoyearperiod.In[40]Normandrawsfromcognitivepsychology,pointingouttheweaknessesandstrengthsofthesubconsciousandconsciousminds.In[40]hediscussesthetypesandsourcesofhumanerror,classifyingthemasslipserrorsinexecutionofaselectedactionandmistakeserrorsinselectionoftheappropriateaction.ThisclassicationschemehasbeenwidelyacceptedbytheHCIcommunity.2.2.2ReliabilityValidationMethodsFormalvalidationmethodshavebeendevelopedforawidevarietyofapplicationsrelevanttomobilerobotics:software,industrialmanipulators,computer-integratedmanufacturingCIMsystems,andautonomoussystemsingeneral.Someexamplesincludetheautomata-basedvalidationapproachforCIMsystemspresentedbyKokkinakiandValavanisin[28];andprobability-basedapproachespresentedbySheldon,Mei,andYangin[49]andTchanganiin[54].Petrinetsappeartobeapreferredtoolforvalidation.In[45]RamaswamyandValavanispresentahierarchicaltime-extendedpetrinetdesignedtobothanalyzeandprovidefaultidenticationandrecoverycapabilitiesfordiscreteeventdynamicDEDsystems.Gonzalez,Mediavilla,Fraile,Gayubo,Turiel,andGarcadenedin[20]apetrinetmodelofamulti-manipulatorsysteminwhichthreemanipulatorsworktogether18

PAGE 30

withinthesameareacelltocooperateonatask.Theeffectofsoftandhardfailures,andperiodicmaintenanceareexaminedintermsofthroughputofthesystem.Simmons,Pecheur,andSrinivasan[50]presenttheuseofSymbolicModelCheckingtoverifyautonomoussystems.Themaingoalof[50]wastocreatetranslatorswhichacceptsourcecodeforanautonomoussystem,andconvertittoSMVaspecicSymbolicModelCheckingsystemspecications.Twotargetlanguages,andsubsequentlyapplications,weretested:theModel-basedProcessingLanguageMPLusedtobuildLivingstone,themodel-basedfaultdiagnosisandrecoverysystemusedonDeepSpaceOneDS1;andtheTaskDescriptionLanguageTDL,anextensionofC++designedtoimplementthemanagementlayertaskdecomposition,synchronization,monitoring,andexceptionhandlingofautonomousmobilerobotsystems.Tables4and5provideasummaryofthefailureandreliabilityanalysisapproachespresentedinthissection.Table4givesthetargetapplicationoftheanalysisapproach,theapplicabilityscopeoftheresultsofananalysiswhichusestheapproach,andthetypesoffailurescovered.Table5presentsthestrengthsandweaknessesofeachapproachinthecontextofstudyinghowmobilerobotsfailintheeld.Thesetablesshowthatnoexistingapproachtofailureanalysisissuitableforthistask.TheclassicationschemesdevelopedfordependabilitycomputingandHCIarecompletefortheirrespectiveareas,butarenotsufcientforcategorizingmobilerobotfailuresintheeld.Formalvalidationmethodsarerestrictedtosystemsandapplicationswhichcanbedescribedusingtheirmodelingtechniquepetrinets,Baysiannetworks,etc.,andcanonlybeappliedtoasinglesysteme.g.robotmodeland/ortaskatatime.AstudyofmobilerobotreliabilityforUSARandMOUTapplicationsingeneralsimplycannotbecompletedusingthesemethods.Therefore,thestrategytakeninthismeta-study,anddescribedindetailinChapterThree,isacharacterizationapproachwhichdenesitsownnoveltaxonomydrawnfromtheclassicationschemespresentedinSection2.2.1.19

PAGE 31

Table4.ComparisonofFailureandReliabilityAnalysisApproaches.HCIreferstohuman-computerinteraction.CIMreferstocomputer-integratedmanufacturingsystems.DEDreferstodiscreteeventdynamicsystems. Approach ApplicationorDomain ScopeofResults FailuresCovered Laprie[30] DependabilityComputing General SystemandHuman Carlson[6] Mobilerobots General Robot Norman[40] HCI General Human Kokkinakiand CIM Specictoasystem System Valavanis[28] Sheldonetal.[49] Software Specictoasystem Software Tchangani[54] General Specictoasystem System Ramaswamyand DED Specictoasystem System Valavanis[45] Gonzalezetal.[20] Manipulators Specictoasystem System Simmonsetal.[50] Autonomoussystems Specictoasystem Software Table5.StrengthsandWeaknessesofAnalysisApproaches. Approach Strengths Weaknesses Laprie[30] Completeness,widelyaccepted Notdetailedenoughformo-bilerobotapplications Norman[40] Completeness,widelyaccepted Doesnotcoversystemfail-ures Kokkinakiand Canvalidateasystem'sreliability System-specic, Valavanis[28] Completemodelrequired Sheldonetal.[49] Canvalidateasystem'sreliability System-specic,Completemodelrequired Tchangani[54] Canvalidateasystem'sreliability System-specic,Completemodelrequired Ramaswamyand Canvalidateasystem'sreliability System-specic, Valavanis[45] Completemodelrequired Gonzalezetal.[20] Canvalidateasystem'sreliability System-specic,Completemodelrequired Simmonsetal.[50] Canvalidateasystem'sreliability System-specic,Fewlanguagessupported 20

PAGE 32

2.3Fault-toleranceSystemsThemajorityoffailure-relatedworkencounteredintheroboticsliteratureisfocusedoncreatingfault-tolerancesystemsthatcandetect,isolate,and/orrecoverfromfailures.Fault-toleranceworkisofinteresttothismeta-studyfortworeasons.First,eachstudythatproducesafault-tolerancesystemforrobotshasassumedcertaincharacteristicsofrobotfailurestobetrue.Forexample,amethodmayrequirethatallfaultstatesbemodeled,assumingthatnovelfailuresarescarceoruninteresting.Itisofinteresttoseeifthecommunityasawholeagreesonthesecharacteristics.Second,fault-toleranceresearchersanddevelopersarepotentialconsumersoftheresultsofthisstudy.Thecommunityasawholecanbenetfromknowledgeofdocumentedfailuresthatwould,forexample,helpthemtodetermineifthemajorityoffailuresareeasyordifculttodetectanddiagnose,orcouldbehandledbyautomaticrecovery.Thissectionprovidesabriefoverviewofthelastsixyearsofworkinthisarea,groupedbyapproach.Purelymodel-basedmethodsarecoveredinSection2.3.1.HybridapproachesarecoveredinSection2.3.2.Fault-tolerancesystemswhichuseexpertsystemstechniquesfortheArticialIntelligencecommunityarepresentedinSection2.3.3.Datacentricprocessingandlteringtechniquesforfault-tolerancearepresentedinSection2.3.4.Finally,relevantworkintheneweldofAutonomicComputing,whichisdevelopingsystemsthatareself-conguring,self-protecting,self-optimizing,andself-healing,issurveyedinSection2.3.5.2.3.1Model-basedFault-toleranceSystemsThemajorityoffault-tolerancesystemsfoundintheliteraturearemodel-based.Thesemethodsusemodelsofthetargetsystemtopredictthecorrectvaluesforinputdata.Thesepredictionsarecomparedtorealdatafromthetargetsystemtodetectandisolatepotentialfaults.Failuresaretypicallyassumedtobecausedbyactuatororsensorfaults,though21

PAGE 33

technicallytherangeoffailuresthatcanbediagnoseddependsonlyontheconstraintsofthemodelingmethod.Catastrophicfailuresaretypicallyassumedtobeeithertrivialtodetectorimpossibletorecoverfrom,sothemajorityofrecenteffortsarefocusedonminorfailureswhicharedifculttodetectanddiagnose,asinslowdegradationsorintermittentfailures.Model-basedmethodsassumethatnovelfailuresarescarceanddifculttoautomaticallyrecoverfromandarethereforeoflessinterestthanknownfailures.Thesemethodsalsoassumethateverythingthatneedstobeknowntoaccuratelydetectanddiagnosefailurescanbemodeled.Someexamplesofpurelymodel-basedfault-tolerancesystemsincludeKawabata,Akamatsu,andAsama's[27]model-baseddiagnosissystemforanautonomousmobilerobotwhichbreaksdownacompleterobotsystemintomodulesthatcanbemodeledexpectedoutputforagiveninputisknown.In[61]WashingtonpresentsapreliminaryattempttocreateafaultdetectionsystemforroversusingacombinationofMarkovmodelsandKalmanlters.Threepurelymodel-basedmethodsweredesignedspecicallyforwheeledmobilerobots.Dixon,Walker,andDawson[17]developedamathematicalmodelofmobilityfaultstodetectchangesinthewheel'sradius,andslippingorskiddingfaults.Goel,Dedeoglu,Roumeliotis,andSukhatmein[19]useaMultipleModelAdaptiveEstimationtechniquebasedonasetofeightKalmanlters,eachwithamodelofexpectedvaluesfornormalandfaultstatessensorandat-tirefailures,whoseoutputislteredthroughaneuralnetworktodecidethestateoftherobot.Meng,Zhen,Biswas,andSarkardescribeafault-tolerancesystemin[26]wherebond-graphsareusedtomodeltherobotsystem,andtemporalcasualgraphsTCG'sareusedtodetectandisolatefaultsinmotors.22

PAGE 34

2.3.2HybridFault-toleranceSystemsPurelymodel-basedmethodsfailwhenanovelfailureisencounteredorwhenthemodelisinaccurate.Hybridfault-tolerancesystemswhichusequalitativereasoningorlearningmethodsalongwithmodelshavebeendevelopedtohelpeliminatethisweakness.Theseapproachestendtoassume,likepurelymodel-basedapproaches,thatinterestingfailuresaredifculttodetectanddiagnose.Unmodeledeffectsliketheenvironment,andstateslikenovelfailuresareassumedtobeimportant.Ontheotherhand,theseapproachestendtoassumethatthoseeffectswillremainstableovertimeorwillchangeslowlyenoughfortrainingtocompensate.Deuker,Perrier,andAmy[16]presentaneuro-symbolichybridsystemfordiagnosisoffaultsinunmannedunderwatervehicles.Combastel,Gentil,andRognonin[12]presentafuzzylogic,model-basedapproachtofaultisolationinelectricalsystems.Wang,Yamasaki,Yumoto,Ohkawa,Komoda,andMyasakapresentin[60]astochasticqualitativereasoningmethodwhichtracksthestateofarealtimesystem.In[29]LamineandKabanzadescribeamonitoringsystembasedontemporalfuzzylogicforusewithbehavior-basedrobots.Someapproachesusetrainingtomitigatetheeffectsofinaccurateandincompletemodels.Mackey,James,Park,andZakpresentin[32]anoverviewofanextensivearchitectureforfailureprediction,detection,andisolationforautonomoussystemswhichusesbothquantitativemodeling,qualitativereasoning,andtrainingmethods.Othertechniquesworkonlyatthesymboliclevel,usingonlycasualorpartialcausalmodelsofthetargetsystem.Amultiplerobotfault-tolerancesystempresentedbyLongandMurphyin[31]usesonlypartialcasualmodelsofthesystemcombinedwithactivediagnosisprobingtoisolatefaults.23

PAGE 35

2.3.3ExpertSystemBasedFault-toleranceSystemsExpertsystembasedfault-tolerancesystemscapturethediagnosisprocessusedbyhumanexperts.Diagnosistechniquesthattakethisapproacharemorecommoninmedicaldiagnosisandhavenot,todate,beendevelopedspecicallyforroboticsystems.Nevertheless,theydohaveanadvantageforchallengingelddomainslikeUSARandMOUTwherefull-automationisnotpossibleintheimmediatefutureinthattheyareeasierforhumanstointeractwith.Threefault-tolerancesystemswhichusethisapproachareofparticularinterestbecausetheyhaveapromisingfeatureformobileroboticsapplications[47]and[46]orhavebeenusedineldenvironments[21].In[47]Rymonpresentsasystemforassistingphysiciansinthetreatmentofpatientswithmultipletraumas.Thisapproachisofinterestbecauseitisrecovery-basedinthatitisnotconcernedwithndingthesourceoftheproblem,butinsteadfocusesonthestepsrequiredtoreturnthesubjectpatienttoanormalstate.Reed[46]presentsadiagnosismethodthatcorrectlyidentiesmultipledefectsusingarecognition-basedreasoningmoduletrainedfromanexistingknowledgebase.Multiple-simultaneousfaultdiagnosisisstillconsideredtobeadifcultprobleminrobotfault-tolerancesystems.Helfman,Baur,Dumer,Hanratty,andIngham[21]describeanexpertdiagnosticsystemwhichhasactuallybeeneldedbytheUSArmy.2.3.4DataCentricFault-toleranceSystemsDatacentricfault-toleranceapproachesarefocusedonusingincomingdatatodetectandisolateanomalies.Thesefault-toleranceapproachesassumethatfailurestatespresentdistinctlydifferentdatasignaturesfromnormalstatesandeachother,andthatallthedataneededtondthosesignaturesarepresentinthesystem.HungandZhao[23]presentadiagnosticapproachforsystemswithlargeamountsofdatacominginfromsimilarand/or24

PAGE 36

distinctsetsofsensorswhichusesacombinationofexistingsignalprocessingandreasoningtechniques.MaddenandNolandescribealearningalgorithmcalledIFTin[33]whichcreatesfaulttreesfromclassiedrawsensordata.In[51]Soikapresentsafailuredetectionframeworkbasedonprobabilisticanalysisofcorrelationbetweenredundantsensorreadings.2.3.5Fault-toleranceinAutonomicComputingAnotherareaofresearchconcernedwiththereliabilityofsystemsisAutonomicComputing.Researchersinthisneweldareinterestedincreatingsystemsthatareself-conguring,self-protecting,self-optimizing,andmostimportantlyforthispaperself-healing.Itistheself-reectivenatureofthisapproachthatmakesitseventualapplicationtointelligentroboticslikely.Todate,workintheAutonomicComputingcommunityonself-healingispreliminaryinnature.Ithasnotyetestablishedthecharacteristicsthatwilleventuallysetitapartfromthebroaderfault-tolerancecommunity,thereforeitsapplicabilitytothismeta-studyisthesameasthatofthelargerfault-tolerancecommunity.Thissectionprovidesasurveyofcurrentworkinthisarea,forfuturereference.Candea[5]hasdevelopedanapplication-genericJava-basedfaultdetectionandrecoverysystemdesignedtoenableweb-basedserviceproviderstodevelopsystemsthatarerobusttotransientsoftwarefailures.Incomparisontotraditionalfault-toleranceapproaches,Candea'sisahybridmodel-basedsystem.In[18]Dong'smodel-basedapproachtoself-healingusessoftwareagentsdevelopedtohandleaspecictypeoffailurebydetecting,analyzing,andrecoveringfromthosefailures.Tohma[57]describesadatadrivenmeansofachievingfaulttoleranceindistributedAutonomicComputingenvironmentswheremanyduplicateserviceprovidersareavailable.Thepaperpresentsmechanismswhichcanbeusedtoensurethatthesystem'sregistryonlyprovides25

PAGE 37

healthyalternativestoclients.Asimplevotingschemeamongpeersisusedtoidentifyandsubsequentlyexcludefaultyserviceproviders.OtherresearcheffortsinAutonomicComputinghaveprovidedhigh-levelviewsofhowtoapproachtheproblemofself-healing.Sterritt[53]discusseshowtheneweldofAutonomicComputingcanbeusedtoachievethegoalsofDependableComputing.Thepaperstronglyadvocatesgroundingtheneweldontheconceptsanddenitionsofkeytermslikedependability,failures,errors,faults,andtoleranceseeLaprie'swork[30]longestablishedintheoldereld.Minsky[35]concludesthatself-healingwillbeimpossiblewithoutimposingsomeformofregularityinthesystem.Minsky'ssolutionisasystemwhichprovidesthatregularityvialawswhichareanalogoustothebasiclawsofnature.Thepaperpresentsasystemdesignedtoenforcethoselaws.Tables6and7summarizethegeneralfault-toleranceapproachespresentedinthissectionastheyrelatetothisstudy.Table6listseachapproach'sassumptionsaboutthefailurestheyweredesignedtohandle.Table7providesasynopsisofthebenetsthisstudycanprovidetoresearchersanddesignersofeachapproachtofault-toleranceformobilerobots.AsTable6shows,researchersinthisareadonotagreeonthecharacteristicsoffailuresafault-tolerancesystemshouldbeabletohandle.Noneofthestudiesdiscussedinthissectionpresentedanyevidenceofhavinginvestigatedthekindsoffailuresrobotsencounter.Therefore,itislikelythattheirapproachwasdeterminedbytheirbackgrounde.g.controltheoreticsformodel-basedandhybridsystemsorAIforexpertsystemsinsteadofexperiencewithrobotfailures.Table7showsthatthisisoneofseveralreasonsthatthefault-tolerancecommunitycanbenetfromthismeta-study.Alltheapproachescanbenetfromknowledgeofcommonfailuresforatargetdomain.Model-basedandhybridapproachesinparticularcanusethefrequencyoffailuresandrelativeprobabilityoffailureforrobotsubsystemspresentedinChapterFiveashigh-delityestimatesforUSARandMOUTenvironments.26

PAGE 38

Table6.AssumptionsMadebyFault-toleranceApproachesAboutMobileRobotFailures. Approach Assumptionsaboutfailures Model-based Allfactorsthataffectfailurescanbemodeled. Detectionanddiagnosisofinterestingfailuresishard. Novelfailuresarescarceanddifculttoautomaticallyrecoverfrom. Actuatorandsensorfailuresarecommon. Hybrid Novelfailuresareimportant. Detectionanddiagnosisofinterestingfailuresishard. Somefactorsthataffectfailurescannotbemodeledaccurately. Factorsthataffectfailuresarestableorchangeslowlyenoughfortrain-ingtocompensate. Expertsystem Humaninterventionmaybeneededforsuccessfuldiagnosisandre-covery. Limitedknowledgeofthetargetsystemisrequiredfordiagnosisandrecovery. Datacentric Failurestatespresentdistinctlydifferentdatasignaturesfromnormalstatesandeachother. Allthedataneededtondthosesignaturesarepresentinthesystem. Table7.HowFault-toleranceApproachesCanBenetfromthisStudy. Approach Howapproachcanbenetfromthisstudy Model-basedandHybrid Typicallyusesprobabilisticmethodswhichrequiretheprobabilityofthesystementeringagivenfailurestate. Needstoknowcommonfailurestocheckforwhenmodelingasystem. Expertsystem Iffailurescannotbehandledautomatically,additionalworkinthisareaisneededtocreatetechniciansupportandoperatoralertingsystems. Needstoknowcommonfailurestocheckforwhencreatingaknowl-edgebase. Datacentric Needstoknowcommonfailurestodeterminethesensorsandotherdatarequiredfordetectionanddiagnosis. 27

PAGE 39

2.4SummaryThisChapterhasprovidedasummaryofrelatedworkincludingsevenstudiesonrobotreliability,ninefailureandreliabilityanalysisapproaches,and22faulttolerantsystemsformobilerobots.Ithasestablishedtheuniquenessandpotentialbenetsofthisthesis,adetailedstudyonhowmobilerobotsfailineldenvironments,withinthelargerroboticscommunity.Twostudieswerefoundintheliteraturethat,likethismeta-study,examinedrobotfailuresfrommultiplesources.Bothcoveredindustrialmanipulatorsusedprimarilyintheautomotiveindustry.BeauchampandStobbe[1]examineddocumentedhuman-robotsystemaccidentsfromninedifferentstudies.Starr,Wynne,andKennedy[52]surveyedfailureandmaintenancereportsfromtwolargeautomotiveplants.Thesemeta-studieswerefoundtobelesscomplete,onlyhumanorrobotfailureswereconsidered,thoughStarretal.'sstudyexaminedmorerobotstotalcomparedtothe28consideredhere.OverallSection2.1establishedthatthisthesisisuniqueinthreerespects:itcoversbothhumanandrobotfailures,isfocusedonroboteldfailuresintheUSARandMOUTdomains,andistheonlymeta-studythatcoversmobilerobotfailures.Section2.2examinedexistingfailureclassicationschemesfromthedependabilitycomputing[30],human-computerinteractionHCI[40],androbotics[6]communities,aswellasrelevantreliabilityvalidationmethodsfromtheroboticsandcontroltheoreticliterature.Thissectionestablishedthatnoexistingapproachtofailureanalysisissuitableforthetaskofexaminingmobileroboteldfailures.TheclassicationschemesdevelopedfordependabilitycomputingandHCIarenotdetailedandcompleteenough,respectively.Formalvalidationmethodsarerestrictedtoapplicationswhichcanbespeciedusingtheirrespectivemodelingtechnique,andcanonlybeappliedtoasinglesysteme.g.robotmodeland/orasingletaskatatime.Therefore,thismeta-studyusesa28

PAGE 40

newcharacterizationapproach,anddescribedindetailinChapterThree,whichdenesitsownnoveltaxonomydrawnfromtheclassicationschemespresentedinthissection.Fault-toleranceworkisofinteresttothismeta-studyfortworeasons.Eachstudywhichproducesafault-tolerancesystemforrobotshasassumedcertaincharacteristicsofrobotfailures.Fault-toleranceresearchersanddevelopersarepotentialconsumersoftheresultsofthisstudy.Section2.3providesabriefoverviewofthelastsixyearsofworkinthisarea,groupedbyapproach.AlsoincludedisabriefsynopsisofrecentworkintheneweldofAutonomicComputingonself-healingsystems.Itisconcludedthatresearchersinthisareadonotagreeonthecharacteristicsoffailuresafault-tolerancesystemshouldbeabletohandle,andthatnoneofthestudiespresentedevidenceofhavinginvestigatedthekindsoffailuresrobotsencounter.Therefore,allthefault-tolerantapproachescanbenetfromknowledgeofcommonfailuresforatargetdomain.Inparticular,fault-toleranceapproachesthatuseprobabilitiestodetectand/ordiagnosefailurecanusethefrequencyoffailuresandrelativeprobabilityoffailureforrobotsubsystemspresentedinChapterFiveashigh-delityestimatesforUSARandMOUTapplications.Examplesofthesewouldbe[25],[32],[51],[54],and[61].Inaddition,approacheswhichuseactiveprobing,namelyLongandMurphy[31],wouldalsobenetfromtheincorporationofprobabilitydatatorankhypotheses,reducingthecostofdiagnosisbyensuringthatthemostlikelyhypothesesarecheckedrst.29

PAGE 41

ChapterThreeTaxonomyofFailuresandMetricsThischapterdescribestheapproachtakeninthisthesistoexaminemobilerobotfailuresintheurbansearchandrescueUSARandmilitaryoperationsisurbanterrainMOUTapplicationdomains.AsdiscussedinChapterTwo,existingapproachesforfailureanalysisinthedependabilitycomputing[30],human-computerinteraction[40],androbotics[6]domainsarenotdetailedorcompleteenoughforthepurposesofthismeta-study.Thepastexperienceineachofthesedomainswasappliedtocreatethenewapproach.Thisapproachwasdevelopediterativelyasexamplesandresultsweregatheredfromthe13studiesthatmakeupthismeta-study.Asaresult,itisbothatoolusedtocreatethendingsofthisthesis,andaproductofthosendings.Thefactthatnoapproachformobilerobotfailureanalysisexistedwhenthe13studieswereperformedwasaseriousdrawbacktothiswork.Eachstudyuseditsownapproach,fewofwhichweredescribedinsufcientdetailtoallowconversionoftheresultsintoacommonframework.Thischapterendeavorstocreateapreciselydenedandeasilyappliedapproachtomobilerobotfailureanalysiswhichotherscanuseintheirparticularapplicationdomainwithoutsimilardifculties.Section3.1providesdenitionsofkeytermslikemobilerobots,eldenvironments,andfailureusedthroughoutthepaper.Next,thenoveltaxonomyofmobilerobotfailures,onecontributionofthisthesis,drawnfromtherobotics,human-computerinteractionanddependabilitycomputingcommunitieswillbepresentedinSection3.2.30

PAGE 42

Thetaxonomyusesclassestocapturethesourceofthefailurewhichcanbeeitherphysicalsystemorrobotorhuman.Twoattributesarealsoincludedtodescribetheseverityofthefailureintermsofitsrepairabilityandimpact.Finally,Section3.3describestheformulasusedtoconverttheavailablerawfailureandusagedataintoreliabilitymetrics.3.1TerminologyThissectionwilldeneafewofthekeytermsusedthroughouttherestofthethesis.ForacompletelistofdenitionsseeAppendixA.AlloftheplatformsdescribedinSection4.1canbeconsideredtobemobilerobots.Amobilerobotisdenedasamechanicaldevicethatcansenseandinteractwithitsenvironment.Itmaypossessanylevelofautonomywithrespecttoitshumanoperators,frommanualwherethehumanhascompletecontroltofullyautonomouswheretherobotcancarryoutassignedtasksonitsown.AlloftherobotscanalsobecalledunmannedgroundvehiclesUGV'swhichareground-basedmobilerobots.Themajorityoftherobotsconsideredinall13studiesareteleoperated,ormanuallycontrolledbyanoperatoratadistancethatistoogreatfortheoperatortoseewhattherobotisdoing[36].Thispaperisprimarilyconcernedwithmobilerobotsusedineldenvironments.Aeldenvironmentisdenedasanenvironmentwhichhasnotbeenmodiedtoensurethesafetyoftherobotortoenhanceitsperformance.ConditionsarobotmayencounterinUSARorMOUTeldenvironmentsinclude:dirt,standingwater,rain,intenseheat,intensecold,connedspacesseeFigure4,unevensurfacesseeFigure5,thepresenceofobstacleswithunpredictablemovement,andhostileagents.FieldrobotsusedbythemilitaryandforUSARhavetobepackedandtransportedtoremotelocations.Robotsdesignedforeldenvironmentsareassumedtoworkoutdoors,thoughgenerallynotin31

PAGE 43

Figure4.AConnedSpaceTrainingMaze. Figure5.ARubblePileUsedforUSARTraining.rainorsnow.Theyareexpectedtohandleroughterrain,toleratedirtanddust,andevenmulti-storyfalls.Reliabilitymetricsaresusceptibletodifferencesinthecriteriausedtodetermineifaneventcanbecalledafailure.Wherestandardcriteriadonotexist,reliabilitystudieswillnecessarilyselectcriteriabasedontheneedsoftheindividualorgroupassessingtheusefulnessofthetechnologyforanewtask.Inthesecases,itisdifculttodirectlycompareresultsacrossstudiesortoapplyndingstonewapplications.Thismeta-study32

PAGE 44

usesageneraldenitionoffailuresothatthendingsreportedherecanbereadilyappliedandcomparedtoresultsfromsimilarstudies.Forthepurposesofthispaper,afailureisdenedastheinabilityoftherobotoritssupportequipmenttofunctionnormally.Bothcompletebreakdownsandnoticeabledegradationsareincluded.AnexampleofacompletebreakdownencounteredintheMicroVGTV'sisafailureofthecontrolsystemwheretherobotbecomesunresponsive,orfreezes.Anexampleofanoticeabledegradationisafaultycameracablecausingsignallossfromacameramountedonarobot.Therestoftherobotplatform,includinganyadditionalcameras,wouldnotbeaffectedbythisfailure.Suchdegradationsmayormaynotaffecttherobot'sabilitytocompleteatask.Ataskwhichrequiresstereovision,forexample,couldnotbeperformedwithasinglecamera.Supportequipmentisdenedasequipmentthatisnotphysicallypartoftherobotandisrequiredfortherobottocompleteitsmissionortask.ThisincludestraditionalsupportequipmentliketethersandoperatorcontrolunitsOCU's.Supportequipmentalsoincludesmaintenanceequipmentrequiredtokeeptherobotoperational,suchasbatterychargersandrecordingequipment.Inmostscientic,USAR,andMOUTscenariosthereareoftenmanystakeholdersindividualswhoareinterestedinandareinuencedbytherobot'sactions.Thevariedinformationneededbyeachstakeholdere.g.commandingofcer,medicaldoctor,andhuman-robotinteractionresearcherisrarelydeliverableinrealtime;thereforerecordingequipmentisoftenrequiredfortherobot'smission.Rigorouslyapplyingspeciccriteriatoanewdomainorapplicationareaisachallengingtask.Forexample,thereliabilitystudies[6][7]reliedonthejudgmentofexperiencedrobotoperators.Theoperatorswereaskedtoapplythisdenitionoffailureastheysawt,basedontheirknowledgeoftherobotplatform.Incaseswherethenormalbehavioroftherobotcanbedescribedinmeasurabletermse.g.rateofprogresstoward33

PAGE 45

goalortheamountofinformationprocessedfromincomingsensors,amoreprecisedenitionoffailurecanbeenforced.Fortheeldapplicationscoveredinthispaper,thisinformationisnotavailable.TestandEvaluationCoordinationOfce'sstudiesatFortLeonardWood[43]andtheCenterforRobot-AssistedSearchandRescueCRASAR[4],whichprovidedthe13studiesexaminedinthisthesis,havebeguntheprocessofidentifyingthecharacteristicsofnormaluseofrobotsintheelddomainsofMOUTandUSARrespectively,buttheirworkisstillinthepreliminarystages.3.2TaxonomyofFailuresInordertogaininsightfromrobotfailures,individualfailurescannotbetreatedasuniqueevents.Meaningfulcommonattributesmustbefoundandusedtocategorizefailuresintowelldenedgroups.Toprovideafoundationforsuchinsights,thissectiondenesaclassicationortaxonomywhichcanbemeaningfullyappliedtoanyfailurethatamobilerobotusedintheeldmightencounter.Thetaxonomy,showninFigure6,wascreatednotonlyfromexperiencewithintherobotics[6][7]community,butitalsodrawsfromthehuman-computerinteraction[40]anddependabilitycomputing[30]communitiesaswell.Thoughthetaxonomywasdesignedtocovertherangeofeldfailures,itisexpectedtobesufcientforanyapplicationofmobilerobots.ThetaxonomyusesclassestocapturethesourceoffailureorwhatLaprie[30]wouldcallthefault,whichisrstdividedintophysicalandhumanbranches,followingdependabilitycomputingpractice.Physicalfailuresarefurthersubdividedbasedoncommonsystemsfoundinallmobilerobotplatforms,thesebeingeffector,sensor,controlsystem,power,andcommunications.Thefollowingdenitionswereusedtoclassifyindividualphysicalfailures:34

PAGE 46

Figure6.TheTaxonomyofMobileRobotFailuresUsedinthisAnalysis.Classesareshownwithsolidlines,andattributeswithdashedlines.1.Effector.Anydevicesthatperformactuationandanyconnectionsrelatedtothosedevices.Exampleswouldbethemotors,grippers,andtreadsorwheels.2.Controlsystem.Anydevicesormanufacturer-providedsoftwarethatissuecommandsatthesymbolicorsignalleveltootherdevicesorsoftwarewithintherobotsystemand/orsupportoperatorinteractionwiththesystem.Forexample,anon-boardcomputer,joystick,motorcontroller,ordisplayunit.3.Sensor.Anydevicesthatsensetherobot'sstateorthestateoftheenvironmentandanyconnectionsrelatedtothosedevices.Exampleswouldbecamerasandlaserrangenders.4.Power.Anycomponentthataffectsthepowersystemoftherobot.Exampleswouldbethebatteries,chargers,andvariousconnectionsallowingtherobottobepowered.5.Communications.AnydevicesthatprovidecommunicationbetweentherobotanditsOCU's.Exampleswouldbetethersandwirelessaccesspoints.UsingLaprie'scategorizationfromdependabilitycomputing,humanfailuresaresubdividedintodesignandinteraction.Designfailuresarecausedbyfaultscreatedduring35

PAGE 47

thedesignofarobotsystem,andarenotcoveredinanyofthestudiesthispaperexamines.DrawingfromNorman'staxonomy[40],whichiswidelyacceptedinthehuman-computerinteractioncommunity,interactionissubdividedintomistakesandslips.Mistakesarecausedbyfallaciesinconsciousprocessing,suchasmisunderstandingthesituationanddoingthewrongthing.Slipsarecausedbyfallaciesinunconsciousprocessing,wheretheoperatorattemptedtodotherightthingbutwasunsuccessful.Eachfailure,physicalorhuman,fallsintoexactlyoneoftheseclasses.Physicalfailuresalsohavetwoattributes,repairabilityandimpact.Theimpactofthephysicalfailureisevaluatedbasedonitseffectontherobot'sassignedtaskormission.Aterminalrobotfailureisonethatterminatestherobot'scurrentmission.Anon-terminalfailureintroducessomenoticeabledegradationoftherobot'scapabilitytoperformitsmission.Therepairabilityofthefailureisdescribedaseld-repairableandnon-eld-repairable.Afailureisconsideredtobeeld-repairableifitcanberepairedunderthefollowingconditions:1.Onlytheequipmentthatcommonlyaccompaniestherobotintotheeldisavailable.Forexample,ifasmallrobotwhichistransportedinasinglebackpackencountersafailure,thatfailurecanonlybelabeledaseld-repairableifthetoolsrequiredfortherepairarepartofasmalltoolkitthattsinthebackpackalongwiththerobotanditssupportequipment.2.Favorableenvironmentalconditions.Theenvironmentalconditionsareconsideredtobefavorableifallconditions,suchasdampnesse.g.fromrainandlightlevel,donotinterferewithorpreventtherepair.Veryfewfailurescouldbeclassiedaseld-repairableiftheyhadtoberepairableundertheworstenvironmentalconditionsencounteredinUSARandMOUTenvironments.Therefore,thisconstraintwasincludedtokeeptheclassicationprocesssimple,whileensuring36

PAGE 48

thatafailure'srepairabilitydependsonlyonthedifcultyoftherepairprocess,notonthetimeandlocationoftherepair.3.Theonlypersonnelavailablearetrainedoperators.Thismeansthattheexpertiserequiredtocompletetherepairhastobepartofanoperator'strainingforthattypeofrobot.ItshouldbenotedthatpackingproceduresandoperatortrainingrequirementswerenotstandardizedforrobotsinUSARandMOUTscenariospriortothismeta-study.Therefore,theeld-repairableclassication,asdenedhere,cannotbeapplieddirectlytotheexamplesfoundinthe13studies.Instead,therepairabilityresultspresentedinChapterFiveweregeneratedbasedonwhetherornoteachfailurewasrepairedintheeld,termedeld-repairedversusnon-eld-repaired.Fornow,thisprovidesadeterministicestimatorforeachexamplefailure'srepairability.ThistaxonomyisusedinChapterFivetoclassifyandstudymobilerobotfailuresintheeld.ForCRASAR'sreliabilitystudies[6][7],CRASAR'seldexperimentswiththeHillsboroughCountyFireRescueDepartmentHCFRD[10],andTECO'sstudies[42]itwasusedtoplaceeachfailurereportedintoasingleclassorleaffromFigure6.TheclassicationofthedatafromCRASAR'sWorldTradeCenterWTCstudies[34][8]wasmoredifcult.Asidefromfournotableterminalfailures,thesestudiesrecordedtheoperator'sresponsetominorproblems,ratherthanthefailuresthemselves.Also,thedatapresentedintheWTCstudiesonlyappearsintermsofthecategoriesdenedintheWTCEngineeringstudy[34].Thedetailsneededtoclassifythemorecommonfailuresthemselvesaremissing.Forthisreason,thecategoriesusedintheWTCstudiescouldfallintoseveralclasseswithinthetaxonomy,dependingonthecircumstancessurroundingeachindividualfailure.37

PAGE 49

3.3CalculationsThissectionwillpresentthecalculationsusedtotransformrawfailureandusagedataintoreliabilitymetricsformobilerobotsusedintheeld.Theseequationswereoriginallyandprimarilyusedtoanalyzethedatainthereliabilitystudies[6][7]presentedindetailinSection4.6.TheywerealsousedtosummarizeanydataprovidedbytheHCFRDstudy[10],theWTCstudies[34][8],andTECO'sstudies[42]seeSection4.2.AlltheformulasweretakenfromtheIEEEstandardspresentedin[24].ThemeantimebetweenfailuresorMTBFiscalculatedbyequation.Thismetricprovidesaroughestimateofhowlongonecanexpecttousearobotwithoutencounteringfailures.Thisformulawasslightlymodied,asdenedinequation,forthefollow-upreliabilitystudy[7]inordertoperformstatisticalanalysisontheresults.AllotherMTBFstatisticsreportedinthispaperwerecalculatedusingequation.Anothermetricpresentedinthispaperisthefailurerate,whichissimplytheinverseofMTBF.MTBF=NumberofHoursRobotWasinUse NumberofFailuresMTBF=Pni=2HoursUsageBetweenFiandFi)]TJ/F19 7.97 Tf 6.587 0 Td[(1 NumberofFailures,fF1;F2;:::;FngarefailuresTheprojectedavailabilityoftherobotiscalculatedbyequation,wherethemeantimetorepair,MTTRisdenedbyequation.Availability,alsocalledreliability,shouldbeinterpretedastheprobabilitythattherobotwillbefreeoffailuresataparticularpointintime.Averagedowntimeistheaverageamountoftimebetweentheoccurrenceofthefailureandthecompletionoftherepairthatxeditseeequation.38

PAGE 50

MTTR=NumberofHoursSpentRepairing NumberofRepairsAvailability=MTBF MTBF+MTTR100%AverageDowntime=Pni=1TimeRepairofFiCompleted)]TJ/F18 11.955 Tf 11.955 0 Td[(TimeFiOccurred NumberofRepairsOthervaluesderivedforthismeta-studywerecalculatedusingstandardformulas.Forexample,theprobabilitythatafailurewascausedbyacomponentfromclassfromthetaxonomy,e.g.sensor,effector,slip,etc.cissimply.Pcjfailure=NumberofFailuresCausedbyacomponentfromc TotalNumberofFailures3.4SummaryThischapterdescribestheapproachtakeninthisthesistoexaminemobilerobotfailuresintheeld.Thefactthatnoexistingapproachformobilerobotfailureanalysiswaspresentatthetimeofthe13studieswasaseriousdrawbacktothiswork.Eachstudyuseditsownapproach,fewofwhichweredescribedinsufcientdetailtoallowconversionoftheresultsintoacommonframework.Thereforethisapproachwasdevelopediterativelyasexamplesandresultsweregatheredfromthe13studieswhichmakeupthismeta-study.Asaresultthistaxonomyisbothatoolusedtocreatethendingsofthisthesis,andaproductofthosendings.39

PAGE 51

Section3.1provideddenitionsofkeytermslikemobilerobots,eldenvironments,andfailureusedthroughoutthepaper.Failureisdenedastheinabilityoftherobotoritssupportequipmenttofunctionnormally.Section3.2presentedthenoveltaxonomyofmobilerobotfailures,onecontributionofthisthesis,drawingexistingfailureclassicationschemesfromtherobotics,human-computerinteractionanddependabilitycomputingcommunities.Thetaxonomyusesclassestocapturethesourceofthefailurethatcanbeeitherphysicalsystemorrobotorhuman.Fivesubclasses,basedoncommonsubsystemsfoundinallrobotsystems,fallwithinthephysicalbranch.Theseareeffector,sensor,controlsystem,power,andcommunications.Humanfailuresaredividedintodesignandinteractionsubclasses,thelatterofwhichisfurthersubdividedintomistakesandslips.Twoattributesarealsoincludedtodescribetheseverityofaphysicalfailureintermsofitsrepairabilityandimpactontherobot'smissionatthetimeofthefailure.Thevaluesgiventotheseattributesareeld-repairableandnon-eld-repairable,andterminalandnon-terminalrespectively.Section3.3describestheformulasusedtoconverttheavailablerawfailureandusagedataintoreliabilitymetricswhichweretakenfromtheIEEEstandardspresentedin[24].OnemetricusedwasthemeantimebetweenfailuresMTBF,whichprovidesaroughestimateofhowlongonecanexpecttousearobotwithoutencounteringfailures.Projectedavailability,alsocalledreliability,wasalsoused.Thismetricisreportedasapercentageandshouldbeinterpretedastheprobabilitythattherobotwillbefreeoffailuresataparticularpointintime.Theaveragedowntimeandprobabilitythatafailurewascausedbyacomponentfromclasscwerecalculatedusingstandardformulas.40

PAGE 52

ChapterFourSourceStudiesThischapterprovidesadetaileddescriptionoftheinformationexaminedinthisthesis.Asthisisameta-studyofmobilerobotfailuresintheeld,theinformationwasgatheredfrom13studiesofmobilerobotuseineldenvironments.ThechapterbeginsinSection4.1bydescribingthe28ground-basedmobilerobotscoveredbythismeta-study.Section4.2thenprovidesanoverviewoftheinformationavailablefromeachofthe13studies.Finally,thegoals,experimentalapproach,andrelevantresultsandndingsarepresentedforeachstudyinSections4.3through4.6.Thestudiescomefromtwoprimarysources.TheCenterforRobot-AssistedSearchandRescueCRASARattheUniversityofSouthFlorida,whichstudiestheuseofrobottechnologyinurbansearchandrescueUSARapplicationsandspendsmorethan200hoursayearintheeld,providedveofthe13studies.Theseincludetwostudiesoftherobot-assistedresponsetotheWorldTradeCenterWTCdisastercoveredinSection4.3,tworeliabilitystudiesofday-to-dayuseofmobilerobotsSection4.6,andasetofeldexperimentsconductedwiththeHillsboroughCountyFireRescueDepartmentHCFRDdescribedinSection4.5.TheremainingeightSection4.4wereprovidedbytheTestandEvaluationCoordinationOfceTECOatFortLeonardWood[42],whichperiodicallyconductsexperimentstodeterminethesuitabilityofaroboticplatformforuseinspeciedmilitaryoperationse.g.militaryoperationsinurbanterrainorMOUT.41

PAGE 53

4.1RobotsSurveyedThissectiondescribesthetotalof28robotsconsideredinthispaper.Theyrepresent15differentmodelsfromsevenmanufacturersandrangefromsmalllessthan10poundstrackedvehiclescapableofchangingtheirgeometry,toamodiedM1tankover60tonnes.ThelistofrobotmodelsappearsinTable8withthemodelname,manufacturer,totalnumberofrobotsoverallstudies,andthestudieswhichanalyzedtheuseofthatmodel.CRASAR'sreliabilitystudies[6][7]arereferredtoasReliability.CRASAR'sWTCanalyses[34][8]andHCFRDstudy[10]aredenotedbyWTCandHCFRDrespectively.TECO'sstudiesarecollectivelyreferredtoasTECO.Table9presentsbasicinformationoneachmodelincludingthesize,weight,communicationmethod,andtractionmethod.Thesizeofarobotisman-packable,man-portable,ornotman-portable[34].Aman-packablerobotcanbesafelycarriedbyoneortwopeopleinbackpacks.Man-portablerobotsarelargerandcannotbeeasilycarriedintotheeldbyaperson.TheycanbetransportedinaHUMMVorpersonalcarandoneormorepersonscansafelylifttheminandoutofthevehicle.Robotswhicharenotman-portablerequireadditionalusuallyspecializedequipmentfortransport,e.g.aheavytruckortrailer.ThesmallestrobotsexaminedareInuktun'sMicroTracsandMicroVGTVplatformsFigure7whicharenolargerthan15.5by30.5cm.Botharetrackedvehicleswithoutonboardcomputers.Bothhaveamicrophone,speaker,amotor-drivenmanual-focusCCDcamera,andacameratiltunitwithhalogenlighting.MicroVGTVplatformsalsohavetheabilitytoadjusttheshapeoftheirchassistoraiseorlowerthecameratiltunitandchangethetrackprole.MicroVGTV'sarecommerciallyavailableandwereoriginallydesignedforchemicalandnuclearinspection.TheMicroTracsplatformwasanexperimentaldesignwhichattemptedtomeetBlitch'scriteria[2]forUSARandMOUTscenarios.Botharebuiltfromexperience;overtenyearsof42

PAGE 54

Table8.TheRobotsExaminedinthisMeta-study.Thecolumndenotedby#containsthenumberofrobotsofthatmodelincludedinthismeta-study. Model Manufacturer # Studies MicroTracs Inuktun 1 Reliability,HCFRD,WTC MicroVGTV Inuktun 3 Reliability,HCFRD,WTC Urban iRobot 5 Reliability,HCFRD,TECO SOLEM Foster-Miller 2 WTC,TECO URBOT Foster-Miller 2 TECO Packbot iRobot 4 Reliability MATILDA MesaAssoc. 1 TECO Talon Foster-Miller 1 TECO ATRV-Jr iRobot 1 Reliability ATRV iRobot 1 Reliability SARGE Yamaha 1 TECO ARTS AllSeasonsVehicles 1 TECO DEUCE Caterpillar 2 TECO D-7G Caterpillar 1 TECO PANTHER USArmy 2 TECO Summary 28 Table9.TheRobots'Characteristics. Model Size Weightlbs Comms Traction MicroTracs man-packable 8 Tether Track MicroVGTV man-packable 8 Tether Track SOLEM man-packable 33 Both Track URBOT man-packable 33 Both Track Urban man-packable 35 Wireless Track Packbot man-packable 42 Both Track MATILDA man-packable 50 Wireless Track Talon man-portable 85 Both Track ATRV-Jr man-portable 110 Both Wheel ATRV man-portable 260 Both Wheel SARGE man-portable 298 Wireless Wheel ARTS notman-portable 5,800 Wireless Track DEUCE notman-portable 35,000 Wireless Track D-7G notman-portable 59,000 Wireless Track PANTHER notman-portable 60,000 Wireless Track 43

PAGE 55

Figure7.InuktunMicroVGTVleftandMicroTracsrightRobots.MicroVGTVphotocourtesyofInuktunServicesLtd.developmentworkwithsimilarplatformsproceededthedesignoftheserobots.Thoughtheyhavelimitedsensingcapabilities,bothhavebeenshowntobeveryusefulinavarietyofUSARscenarioswithroughly400hoursofeldusageloggedtodatebyCRASAR.Theirsmallsizeenablesthemtoexploreareashumansanddogssimplycannottinside.Theyarealsothemostportable.Therobotandallsupportequipmentthatisneededcanbepackedintoasinglebackpackandcarriedintotheeld.ThenextsizegroupincludestheSOLEMandURBOTFigure8fromFoster-MillerIncorporated,theUrbananditssuccessorthePackbotFigure9fromiRobotCorporation,andMATILDAFigure10fromMesaAssociates.Allaretrackedvehicleswhichcarryoneormorecamerasandlighting.TheUrbansalsohaveasetof13sonarrangesensors.Allweredevelopedbetween1999and2001formilitaryoperations,specicallyMOUT,andcontinuetoundergomodications.Theserobotsrequiretwopeopletocarrytherobotanditssupportequipmentintotheeldandthereforeareatthelimitofwhatcanbeconsideredman-packable.TheytypicallyusewirelesscommunicationstoconnecttotheoperatorcontrolunitOCU.Theycarrytheirownbatteriesandacomputerinsidetheirchassis.Anonboardcomputerandlargersensorpayloadcapabilities,thanInuktun'splatforms,enabletheserobotstofunctionwithvariedlevelsofautonomy.Thetermautonomyappliedtomobilerobotsisthelevelof44

PAGE 56

Figure8.Foster-MillerSolemleftandURBOTBuiltonaSolemBaseright.SolemphotocourtesyofFoster-Miller.URBOTphotocourtesyofUSUnmannedGroundVehi-cles/SystemsJointProjectOfceUGV/SJPO. Figure9.iRobotATRV-Jrleft.iRobotPackbotrightExploringaRubblePile.supervisionrequiredbyhumans.Thoughhighlevelsofautonomye.g.navigationandplanningarestilllargelyexperimentalandthereforehaverarelybeenusedtodateintheeld,theycaneasetheworkloadoftheoperatorandenableoneoperatortocontrolgroupsofrobots.TheTalonFigure10fromFoster-Miller,ATRV-JrFigure9andATRVfromiRobot,andSARGEFigure11areman-portable.Alltherobotsinthisgroupcarryonboardcomputersandoneormorecameras.Theyhaveawiderrangeandincreasedexibilitycomparedtotheman-packablerobotsduetotheirabilitytocarrymorebatteries,sensors,andeffectors.TalonandtheATRV-Jraresmallenoughtobeusedfor45

PAGE 57

Figure10.Foster-MillerMATILDAleftandTalonright.MATILDAphotocourtesyofUGV/SJPO.TalonphotocourtesyofFoster-Miller.bothindoorandoutdoorapplications.TheATRVandSARGEaremuchlargerplatformswhichcanonlyoperateinwideopenareasbutcanbemodiedtocarrysmallerrobots.Theseheterogeneousgroups,usuallyreferredtoasmarsupial,havethelargerrangeofthemotherrobotaswellasthemaneuverabilityofthebabyrobots.Theman-portablegroupvariesinmaturityfromtheTalonwhichwasdevelopedaround1999,totheATRV'swhichhavebeeninproductionforaboutsixyears,toSARGEwhichisbuiltonamaturerecreationalAllTerrainVehicleplatform.ThelargestgroupofrobotsismadeupoftheARTS,DEUCEFigure12,D7-G,andPANTHERFigure13.Alloftheseplatformshavebeenadaptedfromcommerciallyavailableheavyconstructionequipmentormilitaryplatforms.Theyrequirespecialequipmenttotransportandweighinexcessof5tons.TheseplatformsaretoolargefortypicalUSARorsquad-levelMOUTscenarios.TECOhasequippedeachofthemwithastandardteleoperationinterfaceandtestedthemfordemininganddebrisclearingtasks.4.2InformationAvailablebyStudyThissectionpresentsanoverviewoftheinformationavailablefromthe13studieswhichhaveexaminedmobilerobotuseintheeld.Table10outlinesthedatacollectionand46

PAGE 58

Figure11.SARGEisBuiltonaYamahaBreezeATVBase.SARGEPhotocourtesyofUGV/SJPO. Figure12.ARTSBuiltonaAllSeasonsVehiclesMD-70BaseleftandCaterpillarDEUCEright.ARTSphototakenfromTECO'sARTSstudy[59].DEUCEphotocour-tesyofCaterpillar. Figure13.CaterpillarD-7GandPANTHERBuiltonaUSArmyM1TankBase.PhotoscourtesyofUGV/SJPO.47

PAGE 59

Table10.OverviewofDataCollectionMethodInformationAvailablefromEachStudy.Rel.referstoCRASAR'sreliabilitystudies. Study FailureDened FailuresClassied Perspective GranularityofDataCollection Rel.[6][7] Yes Yes Engineering Hours WTC[34][8] Yes Yes Both Seconds HCFRD[10] Yes Yes HRI Hours PANTHER[13] No Yes Both Unknown DEUCE[14] No No Both Unknown ARTS[59] No No Both Unknown CBRNLOE[55] No No Both Unknown D7[3] No No Both Unknown SARGE[44] No No Both Unknown UGVROP[62] No No Both Unknown URBOT[43] No No Both Unknown analysismethodsusedineachstudy.Foreachstudy,whetherornotadenitionoffailureswasprovided;whetherornotthefailureswereclassied;theperspectiveofthestudyasinengineering,human-robotinteractionHRI,orboth;andthegranularityofthedatacollectionprocessareincluded.Table11providesanoverviewoftheinformationthatwasavailableandexaminedforthismeta-study.Theavailabilityofusageinformation,countanddescriptionsoffailures,andtheformatoftheresultsareincluded.ThesetablesshowthatkeyinformationismissingfrommostofTECO'sstudies[42]including:thedenitionoffailureused,datacollectionmethods,andnumericresults.Theyalsoshowthatthereliabilitystudies[6][7]andtheWTCstudies[34][8]differedinbothdatacollectionandresultgenerationmethods.Forthisreason,atruecomparativestudycannotbeperformeddirectlyonthissetofstudies.Agoalforthismeta-studywastogatherasmuchinformationaspossiblefromthematerialavailable,andusethatinformationtobegindevelopingstandardmethodsforcollectingandanalyzingmobilerobotfailureinformation.ThetaxonomypresentedinSection3.2istheresultofthateffort.48

PAGE 60

Table11.OverviewofDataandAnalysisInformationAvailablefromEachStudy.Rel.referstoCRASAR'sreliabilitystudies. SourceDataAvailable Study Usage #Failures Descriptions ResultsPresentedAs Rel.Original[6] Yes Yes Yes Standardreliabilitymetrics Rel.Follow-up[7] Yes Yes Yes Aboveandstatisticalanalysis WTC[34][8] No No No Frequencyandpercentages HCFRD[10] Yes Yes Some List PANTHER[13] Yes Yes Yes Summary DEUCE[14] Partial Yes No Summary ARTS[59] No No No Summary CBRNLOE[55] No No No Summary D7[3] No No No Summary SARGE[44] No No No Summary UGVROP[62] No No No Summary URBOT[43] No No No Summary 4.3CRASARWTCStudiesTwostudiesweregeneratedfrompost-hocanalysisofdatacollectedduringCRASAR'srobot-assistedsearchandrescueresponseattheWTC.AdetailedanalysisonthefailuresencounteredbyCRASARwhileusingrobotswasreportedintheWTCEngineeringstudy[34].TheWTCHuman-RobotInteractionHRIstudy[8]examinedthehuman-robotinteractionandhuman-humaninteractionbetweentherobots,theiroperators,andtheotherUSARprofessionalsthatworkedwithCRASARduringtherescuephaseoftheWTCresponse.Thestudiescoverfourrobotsfromtwomanufacturersrepresentingthreedifferentmodels.ThedataspanaperiodofninedaysbetweenSeptember12thand21stof2001.TheEngineeringstudycontributedsevenoverallndingsandadetailedtaxonomyoftheenvironmentalandrobotrelationships.TheHRIstudyprovidesseventeenndingsandelevenrecommendationsforfuturework.49

PAGE 61

4.3.1ApproachTheWTCstudies[34][8]bothdrewinvaryingamountsfromthreesources:videooftherobots'cameraview,eldnotes,andinterviewsoftherobots'operators.Videooftherobots'cameraviewwasrecordedeachtimearobotwasdeployedusedtosearchaholeintherubble.Thisresultedin5.5hoursofvideoavailableforanalysis.TwosetsofeldnoteswereopportunisticallymaintainedbyCRASARmembers.Asthisdatawereanalyzed,theoperatorswereinterviewedforadditionaldetailsortoclarifyambiguouscuesintherecordedvideoandeldnotes.TheWTCHRIstudy[8]examinedallthedataintermsoftheoperatingenvironmentandconditions,agentsandskillshumanandrobot,whichofthetraditionalUSARtaskstherobotswereusedtoperform,overallworkow,andcommunications.TheWTCEngineeringstudyfocusedprimarilyonthevideodata,attemptingtoidentifycuesforminortoseriousfailuresonthepartoftherobotortherobot'soperator.FortheEngineeringanalysisafailurewasdenedasanyeventwhichhinderedtheprogressoftherobotinitssearchtask.Twolevelsofseverityweredened:catastrophicandsignicant.Acatastrophicfailurerequiredthattherobotberemovedfromthevoiditwassearchinginordertoberepaired[34].Asignicantfailureintroducedsub-optimalperformance[34]butdidnotkeeptherobotfromcontinuingitsmission.Duringtheresponse,newtechniquesweredevelopedforoperatingtetheredrobots.Theseallowedthepersonfeedingthetethertoassisttherobotoperatorinovercomingproblems.Thisindividualwasreferredtoasthetethermanager.FivecategoriesweredenedforthesignicantcommonfailuresencounteredwhileusingthesmallInuktunrobots.Twocategorieswereprovidedforfailureswherethetethermanagerwasrequiredtoassisttherobotthroughthetether:gravityassistandstuck50

PAGE 62

assist.Theotherthreecategoriesweretrackslippage,occludedcamera,andincorrectlighting.Theyweredenedin[34]asfollows:1.Gravityassist.Eventsthatoccurwhenthetethermanagerisrequiredtoprovidesupportfortherobotthroughthetether.Reportedasthenumberofinstancesoftheseevents.2.Stuckassist.Eventswhichoccurwhenthetethermanagermustusethetethertofreetherobotfromsomeobstacleorterrainthatdoesnotpermittherobottomove.Reportedasthenumberofinstancesoftheseevents.3.Trackslippage.Theamountoftimethatthetrackmechanismsdonothavesufcientcontactwiththegroundsurface.Reportedasasumofthedurationofeachinstanceinminutes.4.Occludedcamera.Amountoftimethatthecameraviewiscompletelyoccludedbyobjectsanddebris.Reportedasasumofthedurationofeachinstanceinminutes.5.Incorrectlighting.Amountoftimethatthelightswerecompletelyofforinatransitionbetweenintensities.Reportedasasumofthedurationofeachinstanceinminutes.ThistaxonomywasdenedfortheMicroTracsandMicroVGTVaswellasthene-graineddatacollectionmethodsusedintheWTCvideoanalysis.Itthereforecannotbeappliedtotheother11studiescoveredinthispaper.ItshouldbenotedthatInuktun'srobotswerenotdesignedtoperformUSARsearchtasks.Thus,thedenitionoffailureusedintheWTCEngineeringstudy[34]istechnicallyabroaderdenitionoffailurethantheoneusedinthispaperseeSection3.1.Forthepurposesofthismeta-study,theWTCEngineeringstudy'sdenitionoffailureis51

PAGE 63

accepted.ThisisdoneintheinterestsofeventuallyproducingrobotswhichcanbeexpectedtoperformbetterundertheextremeconditionsencounteredduringtheWTCrescueresponse.4.3.2SummaryofResultsThissectionwillreporttheoverallresultsoftheWTCEngineeringstudy.InstudyingtheseresultsitisimportanttokeepinmindthattherobotanditsoperatorweretreatedasonesystemintheWTCEngineeringstudy.NotalloftheminorproblemsreferredtoinTable12werethefaultoftherobotorthephysicalcomponentsthatmakeuptherobot.Somefailureswerethefaultofthehumansthatinteractedwithit,andotherscouldnotbehelpedduetotheextremeworkingconditions.IntheWTCEngineeringstudy[34]thefrequencyandimpactoffailureswasanalyzedonamuchsmallerscalethantheother11studiescoveredinthisthesis.Thisanalysisrevealedthattherobotsoftenencounteredminorproblemssuchastrackslippage,1.4timesperminuteonaverage.Table12presentsoverallstatisticsfromthisanalysis.Thedataisbrokendownbyattempt.Anattemptisaneventinwhicharobotisinsertedintoavoidinordertosearchthatvoid.Onlyattemptsforwhichtherearerecordeddataareincluded.Theyarenumberedsequentiallyforsimplicity.See[34]formoredetailsontheconditionsunderwhicheachattemptoccurred.TheWTCEngineeringstudyreportedthedurationinsecondsofthetrackslippage,occludedcamera,andlightingincorrectfailuremodes.Thenumberofoccurrenceswasnotprovidedfortheseclasses,thoughitwasforgravityandstuckassists.Withoutaccesstotherawdata,itisimpossibletocombinethesetwodistinctmeasurementsintoasinglestatistic.Table12presentsthebestavailablesummaryofunclassiedroboteldfailuresfromtherobot-assistedresponseattheWTC.Itincludesthefrequencyperminuteofgravityand52

PAGE 64

Table12.FailuresEncounteredatWTCfrom[34]. Model Attempt #/Min. %ofTime MicroTrac 1 1.8 0.7 MicroTrac 2 1.3 16.7 MicroTrac 3 1.3 4.0 MicroTrac 4 0.3 1.7 MicroTrac 5 4.7 21.3 MicroTrac 6 1.3 32.3 MicroTrac 7 0.0 13.3 MicroVGTV 8 0.1 0.7 OverallAverage 1.4 11.7 stuckassists;andthepercentageoftimespentinthetrackslippage,occludedcamera,orlightingincorrectfailuremodes.Accordingto[34],assistancefromthetethermanagerwasneededanaverageof2.8timesperminute.Onaverage11%ofthesearchtimewaslostperattemptduetotractionslippage,18%duetocameraocclusion,and6%duetolightingadjustments.AninterestingattributeoftheWTCEngineeringstudyisthattherewereonlyfourfailuresthatwouldhavebeenrecordedwithoutthedetailedanalysisthatwasperformedonthevideooftherobots'cameraview.Thedetailedvideoanalysisuncovered136casesinwhichthetethermanagerhadtoassisttherobot,andthat33%ofthetimewasspentinfailuremodestracksslipping,cameraoccluded,etc..Theseminorfailures,notrecordedbytheotherstudies,hadasignicantimpactontherobots'performance.SeveraloftheWTCHRIstudy[8]ndingsregardingtheenvironmentandconditions,robotskills,andcommunicationareofinterestinstudyinghumanfailuresinHRI.TheenvironmentwasfoundtobebenignascomparedtomostUSARenvironments.TheonlyrealenvironmentalhazardsencounteredbyCRASARpersonnelwerethedust,whichcouldbemanagedwithstandardissuerespirators,andtherubbleitself.Asforworkingconditions,CRASARpersonnelwerefoundtobecognitivelyfatiguedmainly53

PAGE 65

duetolackofsleep.ItwasalsodeterminedthatnoneoftherobotsusedattheWTCweredesignedorratedforuseinUSAR,andthatkeysensorsthatmighthavebeenneededcouldnotbeportedtothemostsuitableplatformsforconnedspaceoperation.Consideringthesetaskanalysisresults,theperformanceoftherobotsasreportedintheEngineeringstudyaresurprisinglygood.Finally,thecommunicationanalysisrevealedthattwotypesofvitalinformationwerenotgettingfromtherobottotheoperator:thestateoftherobot,andthestateoftherobot'senvironment.ThesourceforthedatafromtheWTCstudies[34][8]wasnotavailable.Therefore,statisticspresentedinChapterFivearetakendirectlyorderivedfromtheinformationprovidedinthestudies.Derivedstatisticsweregeneratedinthesamemannerastheircounterpartsinthereliabilitystudies[6][7].ChapterThreeoutlinedthedenitions,taxonomy,andcalculationsappliedtoderivethestatistics.4.4TECO'sStudiesfromFortLeonardWoodTheresultsfromeightstudiesconductedbytheTestandEvaluationCoordinationOfceTECO,partoftheManeuverSupportCenteratFortLeonardWood,havebeenpostedtotheDepartmentofDefenseJointRoboticsProgramJRPlibrary[42].TECOprovidesoperationaltestandevaluationexpertisetotheChemical,EngineerandMilitaryPoliceSchoolsandassistsinthedevelopmentandexecutionofAdvancedWarghtingExperimentsAWE.TheoverallgoaloftheirstudieswastoevaluatethefeasibilityofusingtheroboticplatformsfortheirrespectiveassignedtasksintheFutureCombatSystemFCS.Theexperimentsfocusedonsafety,maintenance,andpossibletactics,techniques,andproceduresTTP'sforeachplatform.Thesestudieswereperformedonthefollowing:anAll-PurposeRemoteTransportSystemARTSforclearinganddemining;integrationofChemical,Biological,Radiological,andNuclearCBRNsensormodulesonexistingrobotplatformsURBOT54

PAGE 66

andMATILDA;thedeliveryofnon-lethalmunitionsfromanexistingrobotplatformSARGE;aUGV-basedrapidobscurationsystem;aD-7Gbulldozer,DeployableUniversalCombatEarthmoverDEUCE,andanM1tankeachequippedwithastandardizedteleoperationsystem;aswellasavarietyofsmallerplatformsURBOT,Urban,SOLEM,andTalon.Allofthesestudieswerecarriedoutinmockmilitaryoperationswhichcanbeconsideredtobehigh-delitycloseenoughtoarealscenariotoproducesimilarresultseldenvironments.4.4.1ApproachForallbuttheDEUCEstudy,experimentalscenariosweredevelopedandmanagedbyTECOpersonnelandsubjectmatterexpertswhereneeded.Questionnairesandreportsweredevelopedandusedtorecord:operatorperformance,operatorfeedback,platformandpayloadperformance,anddocumentationofunanticipatedeventse.g.equipmentfailure.Thesewereusedtodeveloptheassessmentsandrecommendationswhichappearedineachofthestudies'ExecutiveSummaries.Unfortunately,therawdatafromthesequestionnairesandreportswerenotavailable.ThedocumentsstoredintheJRPlibrary[42]foreachstudyconsistedmainlyoftheExecutiveSummary,PatternofAnalysisadetailedlistofissuesexploredbythestudy,blankversionsofanyquestionnairesorreportsdeveloped,andoftenpicturesoftheplatforms.OnlytherepositoriesfortheCBRNLOE,DEUCE,andM1PANTHERIIstudiesincludedacompleteTestReport.Thefollowingparagraphsprovideabriefsummaryofeachstudy.4.4.1.1ARTS.TheAll-PurposeRemoteTransportSystemorARTSplatformwasdesignedtoconductunmannedmineclearingandproongverifyingthatagivenpathisfreeoflivemines,boobytrapproong,unmannedbreachingofwireobstacles,andmaterialshandling.ThepurposeofTECO'sstudy[59]wastosubjecttheunmannedplatformtovariousterraintypesandrealisticmissionsoverlargedistances.Thestudywas55

PAGE 67

alsodesignedtodeterminewhichsoilconditionsandslopestheARTSplatformcouldhandle.4.4.1.2CBRNLOE.In[55]TECOdescribedalimited-objectiveexperimentLOEontheintegrationofChemical,Biological,Radiological,andNuclearCBRNsensorsonaground-basedmobilerobot.Thesensorswerepackagedasmoduleswhichcouldbeeasilyaddedorremovedfromtherobots'payload.Twosmallrobotplatformswereused:MATILDAandURBOT.ThestudywasperformedoveraperiodoffourdaysinadecommissionedcoastaldefensebunkerinCalifornia.Twooperatorscontrolledtherobotsduringtheexperiments.TheyweretrainedbeforehandontheCBRNpayloadandrobotteleoperation.4.4.1.3D-7G.TheD-7Gmodelbulldozerwasmodiedforthepurposesofremotemineclearing,rubbleremoval,andhazardousmaterialshandling.TECO'sstudydescribedin[3]wasafollow-upononethatwascompletedafewmonthsbefore.Thepurposeofthefollow-upstudywastoevaluatethechangesmadetotheopticalsystemrecommendedintherststudyandtoexplorenightscenarioswhichwerenotconductedintherststudy.Ineveryotheraspectthetwostudieswerethesameincludingtherobot,operators,andscenariosdesignedtotesttheD-7'scapabilities.Forward-lookinginfraredFLIRcamerasweredeterminedtobeessentialfornightoperations.4.4.1.4DEUCE.TECO'sstudyoftheDeployableUniversalCombatEarthmoverDEUCEplatform[14]wasexploratoryinnature.ItwasdesignedtodetermineifDEUCEcouldsupportheavyforcesinacombatenvironmentwhilestillmeetingtheweightrequirementsfortransportandretainitstopspeedof30mphkph.Unliketheotherstudiestheoperator'staskswerenotdeterminedbyTECOstaff.AbattalionfromFortLeonardWoodwasusingtheplatformduringtheirNationalTestingCenterNTCrotation,whichlasts28days,atFortIrwin.Asidefromaoneweekperiodinwhich56

PAGE 68

TECO'sstaffwasallowedtoconductformalexperimentswiththesoldiers,theymainlyservedasobservers.Thesoldiersweretrainedontheplatformpriortothistestperiod.4.4.1.5PANTHER.TECO'sM1PANTHERIIstudy[13]examinedthemodiedtank'sutilityformineproonginapotentiallyhazardousarea.Theexperimentswereconductedoveraperiodof32daysatTECO'stestsiteatFortLeonardWood.Nineoperatorsweretrainedandevaluatedduringtheexperiments.First,thesoldiers'abilitytocontrolthevehicleandtohandlepayloadoperationswasassessedinabenignenvironment.ThenthePANTHERwasoperatedinmineproongscenariosinvariousterrainandweatherconditions.Thestudyfocusedonhowwelltheintegratedteleoperationsystemperformed,theplatform'slimitations,howquicklyitcouldbeconvertedfrommanualtoremoteoperation,andthecapabilityoftheinstalledcamerastoassistinremotesteering.4.4.1.6SARGE.InTECO'sSARGEstudy[44]theobjectivewastoevaluatetheeffectivenessoftwonon-lethalcrowdcontrolmunitionsdeliveredfromamobilerobotplatform.Themunitionswerethe40mmCrowdDispersalCartridgeandthe37mmGrabNet.TheywerepropelledfromtheremotevehicleofaSARGEsystemconsistsofatwo-mancontrolvehicle,aHUMMV,andasmallerremotevehicle.Themajorityofproblemsreportedinthisstudydealtwiththerecentlydevelopednon-lethalmunitionsratherthantherobotplatformitself.4.4.1.7UGVROP.Theuseofaroboticplatformwithaddedrapidobscuration,andobscurantgenerationcapabilitieswasevaluatedinTECO'sUnmannedGroundVehicleUGVRapidObscurationPlatformROPstudy[62].Themobilerobotplatformwasnotspecied,buttheimagesincludedintherepositoryshowalargerHUMMVvehicledeliveringtheobscurantandthecontrolvehiclefromtheSARGEsystem.ThestudywasconductedoveratwoweekperiodatFortLeonardWood.AswiththeSARGEstudy,the57

PAGE 69

majorityofproblemsreporteddealtwiththeobscurantringandgenerationsystemsratherthantherobotplatformitself.4.4.1.8UrbanRobot.Severalsmaller,man-packableplatforms[34]wereexploredinTECO'sUrbanRobotstudy[43].ThisstudyfocusedontheoperationaleffectivenessofsmallrobotusageforreconnaissanceofbunkersandMOUTsubterraneanoperations.EighttrainedoperatorswereaskedtooperatetherobotsoveracoursedesignedbyTECOstaffwhichincludeddebris,inclines,andstairs.Therobotsusedwere:twoUrbans,aSOLEM,anURBOT,andaTalon.Thestudylastedforfourweeks.TECO'sstaffdeterminedthattheSOLEMandUrbanswerenotsuitedforsuchoperationsastheydidnothavethemaneuverabilityneededtotraversethecourseinthetargetperiodoftime.TheUrbanswerealsodeemeduntduetotheirinabilitytokeepmud,sand,andothersmalldebrisencounteredinthetestcoursefromdamagingthedriveandmanipulatorsystems.4.4.2SummaryofResultsTECOhasreportedameantimebetweenfailuresMTBFoflessthan20hours,thoughtheirstudiesdidnotprovideenoughinformationseeSection4.2tovalidatethatgure.OnlytheM1PANTHERIIstudy[13]provideddetailsoneachfailureencountered.NoneprovidedsufcientlydetailedusageinformationtocalculateMTBF.Duringthe32dayperiodofthePANTHERexperiments,atotalof35failureswerereported.Most%ofthefailureswereterminal.Experimentswhichreliedonthatrobotstoppeduntiltheplatformwasrepaired.Severalsensorfailureswerenon-terminal,butreducedthequalityoffeedbackprovidedtotheoperatorsfromtherobotplatform.Theaveragedowntimewas7.31hoursoverall,or7.75hoursexcludingnon-terminalfailures.AccordingtoTECO7.2dayswerelostduetounscheduledvehiclemaintenanceoccurringonbothrobotssimultaneously.58

PAGE 70

ThesourceforthedatafromTECO'sM1PANTHERIIstudy[13]wasnotavailable.ThereforethestatisticspresentedinChapterFivearetakendirectlyorderivedfromthedocumentsprovidedin[42]forthestudy.Derivedstatisticsweregeneratedinthesamemannerastheircounterpartsinthereliabilitystudies[6][7].ThemethodwasdescribedindetailinChapterThree.4.5CRASARHCFRDFieldExperimentsInJulyof2001CRASARperformedapreliminaryeldstudywithHillsboroughCountyFireRescueDepartmentHCFRD.Thepurposeofthestudywastosimulatetheuseofrobotswitharealrescueteamrespondingtoanincident,andtocollecthuman-robotinteractionHRIdataontheevent.ThestudywasconductedinabuildingscheduledfordemolitionindowntownTampa.Duringtheexperimentalscenarios,CRASARmembersservedasrobotoperatorsandcollecteddatawhileHCFRDmembersdirectedtheoperatorsandassistedincompletingtheobjectivesofeachscenario.Thedatacollectedduringtheeldexcursionincludedapproximately8hoursofvideo,twosetsofeldnotes,andsummariesofinformalinterviewswiththereprofessionalsandpost-excursionmeetings.Fromthisdata,thestudyproducedtwonewscriptsasetofactionswhichcanbeperformedbyarobotsequentiallyand/orinparallelforUSAR,determinedasuitablehuman-to-robotratioforteleoperatedrobots,andidentiedandcategorizedthefailuresencountered.Foracompletediscussionofthiseventsee[10].4.5.1ApproachThetaskswereselectedbyClintRoberts,IncidentCommandertheofcerinchargeatasearchandrescueeventandHCFRDmember,andChiefofSpecialOperationsRogers.Fourtasksweredevised:climbstairsandinvestigatetheupperoors,searchadarkandclutteredareaforanunconsciousremansimulatedusingadummy,searchthesame59

PAGE 71

areaforanunconsciousvictimusingaFLIR,andexploreaoorbyenteringfromaholeintheceilingverticalentry.Threetrackedrobotswereusedfordatacollection:anUrban,aMicroVGTV,andaMicroTracs.TheUrbanwasrunintwodifferentcongurations,withandwithoutaIndigoAlphaforward-lookinginfraredcameraFLIR.TheIncidentCommanderworkedcloselywiththeprimaryrobotoperatorsduringthefourtasks,providingthemwithasecondaryviewpointoftheinformationcomingbackfromtherobot.ThestairclimbingtaskutilizedtheUrbantonavigateupthestairswhileassessingtheenvironmentforstructuralintegrityandenvironmentalindicatorsofhazardse.g.smoke,vaporclouds.Thistasklasted24minutesduringwhich3.5ightsofstairsweretraversed.ThedownedremantaskwastoteleoperatetheUrbaninadarkenedandclutteredoorsearchingforasimulateddownedremaninasmokybuilding.Therobotenteredtheoorfromthestairwaylanding,foundthereman,andreturnedtothelanding.Thetotalexecutiontimewas18minutes.TheoperatorandIncidentCommanderwerelocatedinaroomonthefarendoftheoor,wheretheycouldnotseeexceptthroughtherobot'scameraviewtheareacoveredbytherobot.Thethirdtaskwasidenticalwithtwoexceptions:alivevictimincivilianclothingwasusedinplaceofthedummy,andtheFLIRwasaddedtotheUrban'ssensorsuite.Thethirdtasktook10minutesand40secondstocomplete.TheverticalentrytaskwastoteleoperatethetwoInuktunrobotsdownahallway,throughaholeinawall,intoaroomwithaholeintheoor,entertheloweroorthroughthehole,andsearchit.TheMicroVGTVleadrobotwasusedtoexplorewhiletheMicroTracsrobotprovidedanexternalviewofotherrobot'sprogress.Thetasktookapproximately30minutes.60

PAGE 72

Theexperimentalmethod1includesaworkowanalysisofthereprofessionalsutilizingrobotsinthespeciedUSARtasks,andinformalinterviews.Theworkowwasrecordedusingfoursynchronizedcamerastorecordfourdifferentaspectsofthetasksperformed.Thefourviewpointsvideotapedwere:theoperatorandcontrolunit,therobot,thererescueprofessionalswithinthevicinityoftheoperator,andanadditionalviewpointwhichvariedbytask.Fieldnoteswereusedtorecordthestartandendtimes,andanyinterestingeventse.g.failureswhichoccurredduringthetasks.Theestimatedlocation,testenvironmentstatus,andweatherconditionswerealsorecorded.Informaldiscussionstookplacewiththeparticipatingreprofessionalsandoperatorsaftereachtask,andpost-excursionmeetingincludedtheindividualswhovideotapedandobservedtheevents.Relevantcommentsandsuggestionsfrombothwererecorded.4.5.2SummaryofResultsTheHCFRDstudy[10]foundthatthehumantorobotratioforteleoperatedrobotswasoften2:1twohumanstoonerobotand3:2atbest.Itwastherstofseveralstudiessee[4]tondevidenceofaheavycognitiveloadassociatedwithteleoperatingasmallrobot,withlimitedsensingcapabilities,inaUSARenvironment.Inthisstudytherobotoperatordrovebyalivevictiminplainview,whileconcentratingonnavigatingtherobotthroughtheenvironment.Thererescuepersonnelservedasasecondpairofeyes,focusedonthetraditionalsearchtaskincludinglookingforstructuralhazardsandvictims,aswellaskeepingtrackoftheareacovered.Themostcommonerrorswerecollisions,inwhichtherobotcollidedwithsomeobstacle,followedbycommunicationfailureswiththewirelessUrbanplatform.One 1ApprovedbytheInternalResearchBoardIRBforhumansubjects.61

PAGE 73

navigationerrorduringthestairclimbingtaskwasmentioned.Inthiscase,theoperatorlettherobotslipdownonestairwhenhemisinterpretedacueinthecameraview.AswiththeWTCstudiesseeSection4.3,thesourceforthedatafromtheHCFRDstudy[10]wasnotavailable.ThereforethestatisticspresentedinChapterFivearetakendirectlyorderivedfromthepublishedstudy.DerivedstatisticsweregeneratedusingthemethodsdescribedindetailinChapterThree.4.6CRASARReliabilityStudiesCRASARcurrentlyhastwenty-onerobotsfromsixmanufacturers,andspendsmorethan200hoursperyearintheeld.InadditiontothepublishedstudiesalreadymentionedinSections4.3and4.5,CRASARhasalsodocumenteditsexperienceusingmanuallogging.Overthepastthreeyearsthishasproducedareasonabledatabaseofmobilerobotphysicalfailuresandtheircharacteristics.Twostudieshavebeenproducedfromthisdatabase.CRASAR'soriginalreliabilitystudy[6]analyzedfailureandusagedatacollectedduringthersttwoyears,includinginformationon13robotsrepresentingthreemanufacturersandsevenmodels.Thefollow-upreliabilitystudy[7]expandedontherstwithanadditionalyear'sworthofdata,twomorerobots,andastatisticalanalysisoftheresults.Thestudieswerenotlimitedtoeldenvironments,butincludedusageandfailureswhichoccurredinthelabaswell.Thefailuredatawereanalyzedusingstandardmanufacturingmeasuresforthereliabilityofaproduct,likemeantimebetweenfailuresandavailability.Therelativefrequencyofthephysicalclasseswerealsodetermined.Theoriginalstudy'sresultsshowedanaveragemeantimebetweenfailuresMTBFof8hoursforeldrobotsandavailabilityoflessthan50%%foreldrobots.Thefollow-upstudyrevealedthatMTBFhadimprovedbyafactorofthreehoursbutthatavailabilityremainedlowandthegulfbetweenthereliabilityofindoorresearchusedonlyinthelabandsimilarofcestyleenvironmentsandeldrobotshadwidened.Inthe62

PAGE 74

originalstudy,theeffectorswerethemostcommonsourcesoffailures%foreldrobots.Thecontrolsystemwasthemostfrequentsourceoffailuresat32%inthefollow-upstudy.Forthepurposesofthismeta-studyonlyeldincludingbothofceandUSARdomainsuseandfailureswereconsidered.4.6.1ApproachUserandfailurelogsservedasthesourcesofdataforthisanalysis.Atotalof171failureswererecordedoveraperiodofthreeyears,specicallyJune21,2000throughJanuary10,2003.PriortoFebruary2002informalrecordswerekept,includingchangestotherobotsandinformationaboutongoingrepairs.StartinginFebruary2002formalfailureanduserlogswerekept.Theuserlogswereenteredbyrobotoperatorsandthefailurelogswererecordedbytheindividualwhoperformedtherepair.Sincethenover2100hoursofusagehavebeenlogged,including500hoursofeldwork.Thefollowinginformationwasgatheredforquantitativeanalysis:whichrobotwasinvolved,whorepairedit,thedatethefailurewasdiscovered,thedatethefailurewasxed,thetotalrepairtime,whichcomponentfailed,wherethefailureoccurred,andwheretherepairwasperformed.ChapterThreedescribesindetailhowthisinformationwasanalyzed.Abriefsynopsisisprovidedhere.Thesourceofthefailurewascategorizedaseffector,sensor,controlsystem,power,orcommunications.Humanfailuresinthefollow-upstudyaredividedintomistakesandslips.Afailure'srepairabilitywasconsideredtobeeld-repairableornon-eld-repairable.Sincetheusageandfailurelogscoveredbothlabandeldevents,adistinctionwasmadebetweendatawhichcamefromthelabversusfromtheeld.Labusageandfailuresoccurredinthelab.Fieldusageandfailuresoccurredoutsideofthelab,usuallyduringdemos,outdoortesting,ortrainingsessions.63

PAGE 75

Standardformulastakenfrom[24]wereusedtoconverttherawdataintocommonreliabilitymetrics.Thestatisticalanalysisperformedinthefollow-upstudyconsistedofcalculatingthecondenceintervalsforthemean-basedandtheprobability-basedresults.Themean-basedresultswereanalyzedusingthestandardequationforthe95%condenceintervalwheremrepresentsthesamplemean.Condenceintervalsfortheprobability-basedresultsweresimilarlycalculatedusingequationwheresrepresentsthesampleprobability.m)]TJ/F15 11.955 Tf 11.955 0 Td[(1:96s Px)]TJ/F21 11.955 Tf 11.955 0 Td[(m nm+1:96s Px)]TJ/F21 11.955 Tf 11.955 0 Td[(m ns)]TJ/F15 11.955 Tf 11.956 0 Td[(1:96s s)]TJ/F21 11.955 Tf 11.955 0 Td[(s ns+1:96s s)]TJ/F21 11.955 Tf 11.955 0 Td[(s n4.6.2SummaryofResultsTable13summarizesthegeneralndingsfromthedatacollectedforthereliabilitystudies[6][7].Here,weconsideronlytheportionofdatausageandfailurerecordsgeneratedineachplatform'stargetenvironment,groupedbymanufacturer.FortheNomadindoorresearchrobotsseeFigure14,onlyin-labdatawasincluded.FortheInuktunsandiRobotmodelsonlyeldusageandfailureswereused.Theplatformtypeeldversusindoorresearch,percentageofusageinthetargetenvironmentoveralltherecordedusage,andthemeantimebetweenfailuresMTBFareincludedtodescribetheoverallfrequencyoffailures.Availabilityandaveragedowntimeareincludedtoshowtheimpactoffailures.NotethattheMTBFdoesnotincludeidletimeseeSection3.3.Table13showsthatMTBFbyitselfdoesnotpaintacompletepicture.Forexample,theoverallMTBFforInuktunandiRobotintheoriginalreliabilitystudy[6]werethesame,butavailabilityisquitedifferent.Theprimaryreasonforthisisthatthe64

PAGE 76

Table13.SummaryofResultsfromtheCRASARReliabilityStudies[6][7].Researchreferstoplatformssuitedforindoorresearchonly. Manu. Type %ofUsage MTBFhrs Availability Ave.Downtimehrs Inuktun Field 94% 6.14 90% 177 iRobot Field 28% 6.27 36% 207 Nomad Research 100% 19.50 94% 61 Inuktun Field 80% 10.27 27% 39.65 iRobot Field 24% 4.57 88% 0.66 Nomad Research 94% 149.08 99% 0.3 Figure14.ANomadResearchRobotIncludedintheCRASARReliabilityStudies.Plat-formslikethesewhicharedesignedforindoorresearchonlycannothandleeldconditionsandarethereforeexcludedfromtheresultspresentedinChapterFive.65

PAGE 77

Inuktunrobotstendedtosufferfromminorfailuresinthersttwoyears,whichoftentooklessthanaminutetox.TheiRobotplatformsweremorelikelytosufferfromseriousfailuresthattookhourstorepair.Inthefollow-upreliabilitystudy[7]theoppositewastrue.TheiRobotmodelsencounteredfailuresmorefrequentlyintheeldbuthadamuchhigheravailabilityrate.Thisisduetothefactthatthefailureswereeasiertorepair,basedonanaveragedowntimeof40minutescomparedtoInuktun'snearly40houraveragedowntime.Incomparisontotheoriginalstudy,thegulfbetweeneldandindoorrobotsincreaseddramaticallyinthefollow-upstudy.Thisappearstobeduetotheinnovativecapabilitiesoftheserobots,andtheinherentdifcultyinconstructingrobotswhichcanoperateinunstructured,outdoorenvironments.RobotsmanufacturedbyiRobotinparticularhadamuchlowerMTBFintheeldcomparedtotheircombinedeldandlabMTBFalmost16hours.Thelikelyreasonforthisisthateldenvironmentsaremorechallenging.AnotherreasonisthatthelessreliableiRobotplatformswereusedmoreoftenintheeldthenthelessfragilebutlargerplatformsandthatonly28%ofiRobotusagewasintheeld.Inthefollow-upstudy,therefore,themorereliableplatformshadabetterchanceofinuencingtheoverallresults.TheoverallMTBFfromthefollow-upstudyimprovedbyalmostafactorofthreefromtheoriginalanalysisresults.Sinceonlyayearhadpassed,andthemajorityofrobotswereexaminedbybothstudies,itisunlikelythatthisresultedfromanactualimprovementinthereliabilityoftherobotsthemselves.Instead,anadditionalyear'sworthofusagelogsprovidedbetterrecordsandsubsequentlybetterestimatesoftherobots'behavior.Thestatisticalanalysisshowedthatthetimebetweenfailures,thetimetorepair,andthedowntimevarywidelythereforenoneofthedifferencesbetweenrelatedmeanscanbeconsideredtobereliablepredictorsforfuturefailures.Regardless,theydoprovideagoodsummaryoftheinformationfoundinthelogsandageneralassessmentofrobotreliability.66

PAGE 78

4.7SummaryThisChapterhasprovidedadetaileddescriptionoftheinformationexaminedinthisthesis.Thirteenstudiesfromtwoprimarysources,vefromCRASARandeightfromTECO[42],weredescribedindetail.TheseincludetheCRASAR'sWTC[34][8],HCFRD[10],andreliabilitystudies[6][7],aswellasTECO'sstudies[59][55][3][14][13][44][62][43].Thechapterbeganwithadescriptionofthetotalof28robotsconsideredinthisthesis.Theyrepresent15differentmodelsfromsevenmanufacturersandrangefromsmalllessthan10poundstrackedvehiclescapableofchangingtheirgeometry,toamodiedM1tankover60tonnes.TheninSection4.2,anoverviewoftheinformationavailablefromeachstudywasprovided.Thissectionrevealedthatkeyinformationthedenitionoffailureused,datacollectionmethods,andnumericresultsismissingfrommostofTECO'sstudies[42].Italsoshowedthatthereliabilitystudies[6][7]andtheWTCstudies[34][8]differedinbothdatacollectionandresultgenerationmethods.Therefore,atruecomparativestudycouldnotbeperformeddirectlyonthissetofstudies.Thegoalforthismeta-studywastogatherasmuchinformationasitcouldfromthematerialavailable,andusethatinformationtodevelopastandardmethodforcollectingandanalyzingmobilerobotfailureinformation.ThetaxonomypresentedinSection3.2istheresultofthateffort.Section4.3coveredthetwostudiesgeneratedfromapost-hocanalysisofdatacollectedduringtherobot-assistedsearchandrescueresponseattheWTC.AdetailedanalysisonthefailuresencounteredwhileusingrobotswasreportedintheWTCEngineeringstudy[34].TheWTCHuman-RobotInteractionHRIstudy[8]examinedthehuman-robotinteractionandhuman-humaninteractionbetweentherobots,theiroperators,andtheotherUSARprofessionalsthatworkedwithCRASARattheWTC67

PAGE 79

response.ThedataexaminedspanaperiodofninedaysbetweenSeptember12thand21stof2001,andfourrobotsfromtwomanufacturersrepresentingthreedifferentmodels.Accordingto[34]minorfailuresoccurred2.8timesperminute.Onaverage11%ofthesearchtimewaslostduetotractionslippage,18%duetocameraocclusion,and6%duetolightingadjustments.TheWTCHRIstudy[8]foundthatCRASARpersonnelwerecognitivelyfatiguedmainlyduetolackofsleep.ItwasalsodeterminedthatnoneoftherobotsusedattheWTCweredesignedorratedforuseinUSAR,andthattwotypesofvitalinformationwerenotgettingfromtherobottotheoperator:thestateoftherobot,andthestateoftherobot'senvironment.InSec4.4,TECO'seightstudiesweredescribed.Thesestudieswereperformedonawidevarietyofplatforms:smallmobileplatforms,severalbulldozers,andamodiedM1tank.Experimentswerecarriedoutinmockmilitaryoperationswhichcanbeconsideredtobehigh-delityeldenvironments.TECOhasreportedameantimebetweenfailuresMTBFoflessthan20hours,thoughtheirstudiesdidnotprovideenoughinformationtovalidatethatgure.Duringthe32dayperiodofthePANTHER[13]experiments,atotalof35failureswerereported.Most%ofthefailureswereterminal.Theaveragedowntimewas7.31hoursoverall,or7.75hoursexcludingnon-terminalfailures.Section4.5coversCRASAR'seldstudywiththeHillsboroughCountyFireRescueDepartmentHCFRD.Thepurposeofthestudywastosimulatetheuseofrobotswitharealrescueteamrespondingtoanincident,andtocollecthuman-robotinteractionHRIdata.Thedatacollectedduringtheeldexcursionincludedapproximatelyeighthoursofvideo,twosetsofeldnotes,andsummariesofinformalinterviewswiththereprofessionalsandpost-excursionmeetings.Thisstudyfoundthatthehumantorobotratioforteleoperatedrobotswasoften2:1twohumanstoonerobotand3:2atbest.Themost68

PAGE 80

commonerrorswerecollisions,inwhichtherobotcollidedwithsomeobstacle,followedbycommunicationfailureswiththewirelessUrbanplatform.Thereliabilitystudies[6][7]weredescribedinSection4.6.ThesestudiesexaminedCRASAR'sdatabaseofmanuallyloggedusageandfailurerecordsoverthepastthreeyears.Theoriginalreliabilitystudy[6]analyzedfailureandusagedatacollectedduringthersttwoyears,includinginformationon13robotsrepresentingthreemanufacturersandsevenmodels.Thefollow-upreliabilitystudy[7]expandedontherstwithanadditionalyear'sworthofdata,twomorerobots,andastatisticalanalysisoftheresults.Theoriginalstudy'sndingsshowedanaverageMTBFof8hoursforeldrobotsandavailabilityoflessthan50%%foreldrobots.Inthefollow-upstudy,theMTBFhadimprovedto24hoursbutthatavailabilityremainedlowandthegulfbetweenthereliabilityofindoorresearchusedonlyinthelabandsimilarofcestyleenvironmentsandeldrobotshadwidened.Intheoriginalstudy,theeffectorswerethemostcommonsourcesoffailures%foreldrobots.Thecontrolsystemwasthemostfrequentsourceoffailuresat32%inthefollow-upstudy.Forthepurposesofthismeta-studyonlyeldincludingbothofceandUSARdomainsuseandfailureswereconsidered.Thoughmoresourcedataanddetailswereavailableforthereliabilitystudies[6][7]thantheothers,andonlythreeofthestudieswerefocusedonmobilerobotfailures;allhavecontributedvaluableexamplesandinsightsintohowmobilerobotsfailintheeld.Theircontributionshaveproduced44representativeexampleslistedinAppendixBofmobileroboteldfailures,andenabledthecreationoftheframeworkforfuturerobotfailureanalysespresentedinChapterThree.69

PAGE 81

ChapterFiveMeta-StudyResultsThischapterexaminesthemobilerobotfailuresreportedin13studiesofeldworkintheapplicationareasofurbansearchandrescueUSARandmilitaryoperationsinurbanterrainMOUT.ResultsandndingsfromthesestudieswerenotgatheredunderthesameconditionsseeSection4.2andforthisreasoncannotbesynthesizedintoasinglesetofmetricswhichdescribemobilerobotfailuresintheeld.ThischapterthereforeexploresrepresentativeexamplestotalwhichdemonstratehowmobilerobotfailurescanbeclassiedusingthetaxonomypresentedinChapterThree,andthechallengesassociatedwithusingrobotsineldenvironments,comparingnumericresultswheneverpossible.ForacompletedescriptionofeveryexampleincludedseeAppendixB.Thischapter'sorganizationfollowsthemobilerobotfailuretaxonomypresentedinChapterThreeexactly.ThistaxonomyusesclassestocapturethesourceofthefailurewhichcanbeeitherphysicalsystemorrobotcoveredinSection5.1,orhumanexaminedinSection5.2.Foreachbranch,therelativefrequencyoftheclassesunderthebranchandtheexampleswhichfallintoeachofthoseclassesareincluded.Thisistrueofalloftheleafclassesinthetaxonomywiththeexceptionofthedesignfailureclass,whichnoneofthe13studiesexplored.Thetaxonomyalsoincludestwoattributes,examinedinSection5.1.7,todescribetheseverityofthefailureintermsofitsrepairabilityandimpactontherobot'smissionatthetimeofthefailure.Anyinformationgatheredfromthesourcestudieswhichrelatestotheseattributesisincludedinthissection.70

PAGE 82

5.1PhysicalFailuresThissectioncoversindetailthefailuresreportedinthe13studiesthismeta-studyexamineswhichdonotdirectlyinvolveahumanthatis,physicalfailures.Takenasawhole,thevastmajorityoffailuresfoundinthestudiesfallunderthephysicalbranch.SincedetailsoneachindividualfailureencounteredattheWorldTradeCenterWTCwasnotavailableseeSection4.2,thefailuresarepresentedbycategoryasdenedintheWTCEngineeringstudy[34]performedbytheCenterforRobot-AssistedSearchandRescueCRASAR.Thesewere:stuckassist,gravityassist,trackslippage,occludedcamera,andlightingincorrect.ChapterFourprovidesdetaileddenitionsofthesecategories.EachreportedfailurefromCRASAR'sreliabilitystudies[6][7]andthestudiesperformedbytheTestandEvaluationCoordinationOfceTECOatFortLeonardWood[42]wereindividuallyclassied.BoththeCRASARWTCHRIstudy[8]andCRASAR'seldexperimentswiththeHillsboroughCountyFireRescueDepartmentHCFRD[10]werefocusedonhuman-robotinteractionHRIissues,thereforetheirdataarenotcoveredinthissection.FirstthefrequencyoffailureswithintheclassesthatfallunderthephysicalfailurebranchofthetaxonomyarepresentedinSection5.1.1.InSections5.1.2through5.1.6,examplesoffailureswhichfallintoeachcategoryareprovided.Eachofthesesectionswillalsoincludesomediscussionofgeneraltraitsofthatphysicalfailureclass.5.1.1RelativeFrequencyofPhysicalClassesTable14presentsdatafromthereliabilityanalysesdescribedinSection4.6.Itshowstherelativefrequencyofeachofthephysicalclassesintheformofprobabilitiesthatafailurewascausedbycomponentsfromeachclass.Thereliabilitystudies[6][7]coveredday-to-dayusageandfailuresinbothlabandeldenvironments.Therefore,itis71

PAGE 83

Table14.ProbabilitybyPhysicalClassfromtheReliabilityStudies[6][7].Resultsfromtheoriginalstudyappearabovewiththefollow-upstudy'sresultsbelow. Manufacturer Effector ControlSystem Power Comms Sensing Inuktun 0.50 0.34 0.03 0.00 0.13 iRobot 0.58 0.17 0.25 0.00 0.00 Overall 0.50 0.33 0.09 0.00 0.09 Inuktun 0.45 0.32 0.00 0.02 0.12 iRobot 0.22 0.30 0.15 0.22 0.11 Overall 0.36 0.31 0.06 0.10 0.12 importanttonotethatthisinformationwastakenfromthesourcedata1forboththestudies,ratherthanthestudiesthemselves.Allusageandfailureeventsrecordedinthelabwerefactoredoutofthestatisticsreportedhere.Thefailuresaregroupedbymanufacturerwiththeoverallprobabilitiesforeachclassprovidedatthebottomofthetable.SincesufcientdescriptionsofallthefailuresencounteredinTECO'sM1PANTHERIIstudy[13]wasprovided,asimilartable,Table15,wascreatedforthatstudyaswell.Figure15wasgeneratedfromresultsfromthefollow-upreliabilitystudy[7]includingthestatisticalanalysis.Asintheprevioustablethefailuresaregroupedbymanufacturerwithoverallprobabilitiesforeachcategoryshownintheright-mostset.Theprobabilitiesareshownasbarswiththe95%condenceintervalsfromthestatisticalanalysisindicated.Thedifferencebetweeneffectorandcontrolsystemrelativefrequenciesissignicantonlyifa50%orlesscondenceintervalisused.Botharesignicantlymorecommonthantheothercategories.Table14showsthateffectorsarethemostcommonsourceofphysicalfailures,followedbythecontrolsystem,sensors,communications,andpower.Table15forthePANTHERshowsverydifferentresults.InTECO'sstudytheprimaryproblemwasanewteleoperationsysteminstalledonthemodiedtank.Thisresultedinahighpercentageofcontrolsystemfailures.Thesensorsthatcamewiththecontrolsystemweresecond,with 1Databaseofusageandfailurelogsenteredbytheoperatorandrepairerresp.72

PAGE 84

Figure15.ProbabilitybyPhysicalClassfromtheFollow-upReliabilityStudy[7].Table15.ProbabilitybyPhysicalClassforM1PANTHER[13]. Model Effector ControlSystem Power Comms Sensing PANTHER 0.11 0.54 0.09 0.00 0.26 26%ofthefailures,followedbytheeffectors.Onlypowerhasasimilarprobabilityofcausingfailuresacrossthethreestudies.ThedifferenceisprobablyduetothematurityoftheM1Tankplatform,whichhasbeenusedbytheUSArmyfor20years.Incomparison,theplatformsexaminedinthereliabilitystudieswerelessthan10yearsold.5.1.2EffectorThemostcommontypeoffailureacrossthe11studieswhichexploredphysicalfailureswasfailureofcomponentsthatperformactuationandtheirconnections,oreffectorfailures.Commonfailuresourcesintheoriginalreliabilitystudy[6]weretheshearpinandpiniongearinthegeometryshiftingmechanismontheMicroVGTV.Iftherobotencountersresistancewhileshiftinglowclearanceinaconnedspaceforexamplethe73

PAGE 85

Figure16.DirtFoundNearSensitiveEquipmentInsideaniRobotUrban.shearpinwillbreak.Thepiniongearisaproblembecausetheareathathousesitisopentotheenvironment.Dirtandotherdebrisgetinsidethatareaandcauseprematurewear.AccordingtoTECO,theUrbanplatformsuffersfromsimilarproblems[43].Opengearingforthearticulatingarmsanddrivemotorcollectdebriswhichcausesthemtostopworking.Theseexamplesillustrateacommonlyfoundsourceofproblemsforeldrobots,namelydirtandothersmalldebris.TheseparticleshavebeenshowntoappearineveryunsealedareainsideamobilerobotseeFigure16.Resultsfromthereliabilitystudies[6][7]showthattrackedvehiclesweremorepronetoeffectorfailurethanwheeledplatforms.Intheoriginalreliabilitystudy[6]96%oftheeffectorfailuresoccurredontrackedratherthanwheeledvehicles.57%oftheeffectorfailureswerethetracksworkingofftheirwheelsknownasde-tracking,usuallyduetoexcessivefrictionwiththegroundsurface.TheUrbanRobotstudy[43]describedthesameproblemwiththeUrbansusedintheireldexperiments.ThestudyoftheARTSvehicleperformedbyTECO[59]mentionedtwoinstancesinwhichrocksbecamestuckinthetrackguidesandsprocketsseeFigure17.ThePANTHERalsothrewatrack,afailurewhichtookdaystorepair.Severalofthestudies,theWTCstudies[34][8],74

PAGE 86

Figure17.RockStuckinARTSTrackMechanism.PhototakenfromARTSstudy.[59]TECO'sD-7Gstudy[3],andTECO'sDEUCEstudy[14],mentionedrepeatedproblemswithtrackslippage.ThetrackslippagecategorydenedintheWTCEngineeringstudy[34]wasusedtocatalogtheamountoftimethetrackswerespinningwhiletherobotremainedinplace.Thetrackswereoftenslippingonpilesofloosesandorpaper,aproblemwhichtheeffectorsonasmalltrackedvehiclethatweighslessthan10poundscannoteasilyovercome.Thereforeonlysituationsinwhichthetracksshouldhavehadenoughfrictionwiththegroundsurfaceunderthoseconditions,wouldbeconsideredtobeaneffectorfailure.TheWTCEngineeringstudy[34]donotprovidesufcientlydetaileddescriptionsofthefailuresthemselvestodeterminewhatpercentageofthetrackslippagetimewouldmeetthiscriteria.TwofailuresfromtheWTCdoclearlyfallintotheeffectorcategory.Inonedocumentedcase,thevoidarobotwasexploringexceeded122degreesFahrenheitandsoftenedthetrackenoughthatitbecamelooseandfelloff.Inanothercase,analuminumrodbecamelodgedintothetrackmechanismonaMicroTracsrobotwherethespacetolerancebetweenthetrackandtheplatformwaslessthanone-eighthofaninch75

PAGE 87

Figure18.FailureEncounteredattheWTC.seeFigure18.Theseareclassiedasterminalfailures,astherobotshadtobepulledoutofthevoidandrepairedbeforebeingusedagain.Effectorfailuresincludemorethanjustmobilityfailures.Forexample,TECO'sDEUCEstudy[14]mentionedproblemswiththeripperappendagemountedtotherobot.Thetoolgetsboggeddownwhenusedinparticularlyhardrock.TECO'sM1PANTHERIIstudy[13]describedproblemswithitshydraulicsystemwhichmanifestedassmokeissuingfromthetank'sturret.5.1.3SensorThesensorcategorycoversfailedsensorsandproblemswiththeirconnections.Thesefailurestendtobelesscommonthaneffectorandcontrolsystemfailures,withonly9%ofthefailuresanalyzedintheoriginalreliabilitystudy[6]and11%offailuresinTECO'sM1PANTHERIIstudy[13].AttheWTCsensorsweremoreofaproblem.Duetoincorrectlightingandoccludedcameraviews,anaverageof24%ofsearchtimewaslosteachtimetherobotwasusedtosearchavoid[34].Byfarthemostcommonfailedsensor76

PAGE 88

inallofthestudieswasthecamera.Itwasalsotheonlysensorcommontoalloftherobots'sensorsuites.Sensorswerethemostraresourceoffailuresintheoriginalreliabilitystudy[6],andtiedwiththePowerclassforleastcommoninthefollow-upreliabilitystudy[7].Themostcommonsourceofsensorfailuresinthesestudieswasthesensor'sconnectiontothecontrolsystem.Notethatthisdoesnotincludeproblemsinthesmallerrobots'tetherorthelargerrobots'wirelessconnection,asthosefailureswouldfallunderthecommunicationsclass.Examplesincludefaultycablingandbrokenorlooseconnectionsateitherendofthecabling.Statisticalanalysisinthefollow-upreliabilitystudy[7]showsthatthesefailuresareequallyuncommonacrossalltherobotmodelsexaminedinthatstudy.TheWTCEngineeringstudy[34]identiedtwocategoriesofintermittentsensorfailures:occludedcameraandincorrectlighting.Occludedcamerawasdenedasastateinwhichtheentirecameraviewisblockedbyobstacles.Thisfailurewasfoundtooccurduring18%,onaverage,ofthetotaltimetherobotsspentsearchingavoid.Notethatthispercentageishighdespitethefactthat100%obstructionwasrequired.Thelightingincorrectcategoryincludedstatesinwhichthelightswerecompletelyoffinwhichcasetheoperatorcouldnotseeorwereintransitionbetweenintensities.Thisfailurewaslesscommon,occurring6%,onaverage,ofthetotaltimetherobotsspentsearching.LightingproblemswerealsomentionedinTECO'sARTSstudy[59].Thecamera'sautomaticirisdidnotadjustenoughfortheoperatortoseetomaneuvertherobot.TECO'sM1PANTHERIIstudy[13]citedsensorproblemswhichdonottendtooccurunderlabconditions.Bumpyterrain,suddenchangesinlighting,andrainyweathercausedproblemsfortheon-boardcameras.Sincehumanoperatorsrelyheavilyoncameraviewswhileteleoperatingarobot[10]theseminorfailuresmadeitdifculttocontrolthe77

PAGE 89

robotfromtheremoteoperatorcontrolunitOCU.ThePANTHERstudyalsomentionscaseswherecameralenseswerecoveredinmoisture,dirt,ormud.5.1.4ControlSystemThecontrolsystemfailureclassincludesanyproblemscausedbytheon-boardcomputer,manufacturerprovidedsoftware,andOCUs.TheywerethemostcommonfailuresinTECO'sM1PANTHERIIstudy[13]%,andthesecondmostcommoneldfailuresinboththereliabilitystudies[6][7].NonewerereportedintheWTCEngineeringstudy[34].InTECO'sM1PANTHERIIstudy[13]morethan6daysofthe32availablefortestingwerelostduetodiagnosisandmaintenanceoftheteleoperationsystem.Themostfrequentcontrolsystemfailuresintheoriginalreliabilitystudy[6]werecasesinwhichtherobotwassimplyunresponsive%ofeldcontrolsystemfailuresandthesolutionwastocyclethepower.Sincerebootingtherobotand/orOCUsolvedtheproblem,itwasassumedthattheproblemwasduetothecontrolsystem.Bythetimeofthefollow-upreliabilitystudy[7],thecontrolsystemfailureshadbecomemoreserious.ThemostcommonproblembecameanoverloadoftheelectricalsystemeitherontherobotorwithintheOCUwhichcyclingpowercouldnotx.IneachcasetherobotorOCUhadtobedismantledtoreplaceeitherafuseoraburntcomponent.Theseproblemsoccurasfrequentlyinthelabastheydointheeld.TECO'sM1PANTHERIIstudy[13]reportedawidervarietyofcontrolsystemfailures.Thesteeringsystemwasthemostfrequentsourceofproblems.Symptomsrangedfromsluggishnesstoacompletelossofsteering,sometimesmanifestinginonlyonedirectionatatime.Theemergencystopswitchfailedmultipletimes.ThePANTHER'scontrolsystembehaviorwaserraticandunstableinsomecases.Reportedfailuresinclude:uncontrolledacceleration,theRPMsshootinguptoacriticallevelfornoapparentreason,andasystemshutdownwhentheoperatortriedtoswitchtoteleoperation78

PAGE 90

mode.TECO'sUGVROPstudy[62]describedsimilarproblemswiththesameteleoperationsystem,buttheyappearedtobelessfrequent.5.1.5PowerBasedontheresultsfromthereliabilitystudies[6][7]andTECO'sM1PANTHERIIstudy[13],powerfailuresdonotproducemanyofthefailuresthatoccurintheeld.TheWTCEngineeringstudy[34]revealednofailuresduetobatteriesandtheirrelatedconnectionsduringthetwoweekrescueresponse.Thiswasprobablyduetothefactthattherobotswerenotusedforanextendedperiodoftime.Thelongestperiodoftimearobotwascontinuouslyusedwasalittleover24minutes,thereforethebatterieswerenotheavilytaxed.Powermaybemorereliablethantheothersystemssinceitistheleastaffectedbyenvironmentalhazards.Inthereliabilitystudies[6][7]halfofthepowerfailuresontherobotsareduetothebatteryanditsconnections.ThePANTHERplatformsufferedrepeatedlyfromlowbatteriesandlowfuel.TECOhadrecurringfailuresduringtheDEUCEstudy[14].OneofthetwoDEUCEplatformssufferedfromcloggedfuellters,requiringareplacementroughlyeverysixhours.5.1.6CommunicationsThemajorityofcommunicationsfailuresintheeldwerefoundanddescribedbyTECO,withtheWTCEngineeringstudy[34]providingoneexamplearobotwaslostduetocompletecommunicationsdropout.Thisisdueinparttothefactthatwirelesscommunicationsisaknownproblemineldenvironments[34][10].Wiredrobotsaremorecommonlyused%ofInuktunusagewasintheeld[6]thenthelargerwirelessrobots8%ofusageintheeld.CRASARthereforeencounteredfewercommunicationsfailuresintheeldduetotheuseofrelativelyreliablewiredcommunication.TECOdid79

PAGE 91

nothavethisluxury.Mostoftherobotstheyexaminedcouldonlyusewirelesscommunicationsandnonehadadurabletether,likethoseusedwiththeInuktuns.ExperiencewiththewirelessSOLEMplatformattheWTC[34]providesagoodexampleofwhytheserobotsaredifculttouseremotelyineldenvironments.ThestructuralsteeloftheWorldTradeCenterhadasignicantimpactontherangeofthe2watttransmittertheplatformwascarrying.Insteadoftheusualmileormore,therobotlostcommunicationwiththeOCUinunder20feet.Evenuptothatpointthesignalwasnotverystable,23.8%ofthe7minutestherobotspentsearchingtherubblepileresultedincompletelyuselessvideoduetowirelessdropout.TherobotnallycompletelylostcontactwiththeOCU.Itwasneverrecovered.SimilarproblemsmayhavebeenfoundwiththePANTHERplatformtestedbyTECO[13].14%ofthefailuresencounteredduringthatstudywereduetovideodropout,agoodindicationofcommunicationsfailures.TECO'sD-7Gstudy[3]concludedthatthenon-line-of-sightcontrolrequirementfortheplatformcouldnotbemet.Thiswasbecausetheteleoperationequipmentcouldnottransmitvideothroughinterposedmaterialsorfoliage.Videobandwidthandreliabilitylimitationsevenimpactedtheperformanceofoperatorsinline-of-sightscenarios.OccasionalstaticwasaproblemforARTSoperatorsaswell[59].InthestudyconductedbyTECOonthesmallermobilerobotplatformstheUrbanrobotstudy,additionalantennaswereusedinanattempttoimprovethequalityofthevideosignal.Theanalogsignalwasdescribedasbeingadequatebutwasbreaking-upoftenenoughtodistracttheoperatorsfromthetestscenario.Ifcommunicationtechnologyimproves,newproblemsarelikelytoemerge.Forexample,theproblemofsharinglimitedbandwidthamongmanywirelessrobotsinthesamearea.Limitedbandwidthisalreadyaproblemforthemilitary,whererulesonallowedfrequencieswerefoundbyTECO[59]tobeahindrancetoimprovedsignalstrengthandreliability.Itisalsoachallengetoensurethatwirelesstransmittersadhereto80

PAGE 92

Table16.ComparisonofTerminalVersusNon-terminalFailuresfromtheM1PantherStudy[13]. Impact # AverageDiagnosisTimehrs AverageDowntimehrs Terminal 33 0.33 7.75 Non-terminal 2 12.50 0 Overall 35 1.02 7.31 theestablishedrules.TheARTSforexamplewouldbleedovertoradiofrequenciesitwasnotallowedtouse.5.1.7AttributesInthetaxonomydenedinChapterThreetwoadditionalcharacteristicsofphysicalfailuresweredenedinadditiontothecauseorfaultofthefailure:therepairabilityofthefailureandtheimpactofthefailure.FirsttheimpactattributeisexploredinSection5.1.7.1,followedbytherepairabilityattributeinSec5.1.7.2.Eachsectionwilldiscusskeytraitsoftheattributesbasedonanydataavailablefromthe13studies.5.1.7.1Impact.Theimpactofafailureisspeciedbythetermsterminalandnon-terminal.Itisdeterminedbasedontheeffectthefailurehadontherobot'sassignedtaskormissionatthetimeofthefailure.Sinceinformationontherobot'smissionwasnotconsistentlyrecordedforthereliabilityanalyses[6][7],thelargestsourceofdocumentedfailuresinthismeta-studycouldnotbeusedtodescribethetraitsofthisattribute.Instead,TECO'sM1PANTHERIIstudy[13]providedsufcientinformationtoreliablydistinguishbetweenterminalandnon-terminalfailures.Table16showsthenumberofterminalversusnon-terminalfailures,andtheaveragediagnosistimeanddowntimeforeachtype.Overallresultsareprovidedatthebottomofthetable.Table16showsthatterminalfailureswerefarmorecommon.Theaveragediagnosistimeshowsthatintermittentnon-terminalfailurestendtorequiremore81

PAGE 93

Table17.ComparisonofField-repairedvs.NotField-repairedFailures.Resultsfromtheoriginalstudyappearabovewiththefollow-upstudy'sresultsbelow. Repairability % MTBFhrs AverageDowntimehrs Field-repaired 65% 9.5 0.14 Non-eld-repaired 35% 17.7 553 Overall 6.17 185 Field-repaired 60% 12.90 0.16 Non-eld-repaired 40% 32.06 50.91 Overall 12.23 22.49 diagnosistimeandadditionaltechnicalknowledgeoftheroboticsystem.Forexample,anunresponsivecontrolsystemmaytakelessthanaminutetodiagnoseandxpowercycle.Thisisstillconsideredtobeaterminalfailureastherobotcannotcontinueitsmissionuntilitisrepaired.Anon-terminal,butstillsignicantsteeringproblemmaydegradetheoperators'performanceandreducethesafetylevelofoperationespeciallyforavehicleaslargeasthePANTHERforanextendedperiodoftime.5.1.7.2Repairability.Sincethismeta-studyisconcernedwitheldmobilerobotfailures,repairabilityisdenedintermsofeld-repairabilityseeChapterThree.Table17comparestheratesofphysicalfailuresthatwereeld-repairedandthosethatwerenotusingonlyonlydatacollectedintheeldfromthereliabilitystudies[6][7].Thetablepresentsthepercentageoffailures,MTBF's,andaveragedowntimeforeld-repairedandnon-eld-repairedfailures.Averagedowntimeistheaverageamountoftimebetweentheoccurrenceofthefailureandthecompletionoftherepairthatxedit.BasedonTable17,eldfailuresaremorelikelytoberepairedintheeldthaninthelab.ThisresulthasremainedstabledespitetheincreaseintheMTBFandthedropinaveragedowntimeoverallbetweenthetwostudies.Theaveragedowntimeforeldrepairedfailureshasalsoremainedlowcomparedtothosethatwerenoteldrepaired.82

PAGE 94

5.2HumanFailuresThissectioncoversthehumanbranchofthetaxonomypresentedinChapterThreeusinghumanfailuresthatemergedfromtheWTCstudies[34][8],theHCFRDstudy[10],thereliabilitystudies[6][7],andTECO'sstudies[42].Designfailureswerenotexploredbyanyofthe13studies,thereforethissectionislimitedtointeractionfailures.AsmentionedbyLapriein[30],allfailurescaneventuallybetracedtoahumanerroratsomelevel.Thissectioncoversonlydirectorifnotdirectatleastimportantconnectionsbetweenhumanerrorandthefailuresthatresultedfromthem.Section5.2.1explorestherelativefrequencyofclasseswithinandinteractionbranchandSection5.2.2providesexamplesofhuman-robotinteractionfailures.ThissectionwillpresentresultsfromasmallersetofrecordedhumanfailuresascomparedtothephysicalfailurescoveredinSection5.1.Thereliabilitystudies[6][7]andtheWTCEngineeringstudy[34]werefocusedonphysicalratherthenhuman-robotinteractionHRIfailures.WhileanobjectiveofTECO'sstudieswasanassessmentoftheroboticsystem'susability,noneoftheirstudiesincludedformalHRIexperiments.BoththeWTCHRIstudy[8]andtheHCFRDstudy[10]werefocusedonHRIissuesinUSAR,notfailuresandthereforeprovideamorelimitedsetofrecordedexamples.5.2.1RelativeFrequencyofHumanClassesTable18isolatesthehumanfailuresdescribedintheHCFRDstudy[10]andtheWTCstudies[34][8]andcategorizesthembasedonthetaxonomypresentedinChapterThree.Theeldeventandthetasktheoperatorwasaskedtoperformwiththerobotareincluded.Table18alsoincludesthetotaldurationofthattask,totalnumberoffailures,meantimebetweenfailuresMTBF,seeChapterThreeinhours,percentageofmistakes,and83

PAGE 95

Table18.HumanFailureAnalysisResults. FieldEvent Task Timemin # MTBFhrs Mistakes Slips HCFRD[10] ClimbStairs 24 3 0.13 33% 67% WTC[8][34] SearchSmallVoids 55 16+ 0.06 38% 63% Overall 79 19 0.28 37% 63% percentageofslips.Theresultsarebrokendownbyeventwithoverallvaluesprovidedatthebottom.Itisalsoimportanttonotethat,for13oftheWTCfailuresreportedhere,thedurationwasrecordedratherthenthenumberofindividualfailures.Morespecically,thesixmistakeswerefromfailurescategorizedasincorrectlightingseeSection4.3,andsevenoftheslipswerefromthetrackslippagecategory.Forthepurposesofthismeta-study,eachdurationvaluerecordedwasconsideredtobeasinglefailure.Therefore,thenumberoffailuresreportedhererepresentstheminimumthatactuallyoccurred.Table18showsthathumanfailuresoccurredmoreoftenduringtheactualUSARresponseattheWTCascomparedtotheeldexperiments.ThisresultisexpectedconsideringthedifcultyofnavigatingacollapsesiteasextensiveandcompactastheWTCdisasterseeFigure19,compoundedbyfatigueandrisktopersonalsafety.Theratioofmistakestoslipsissimilardespitethesedifferences.MoredataisneededtodetermineifthisisauniversalattributeofHRI.5.2.2InteractionTheinteractioncategorycapturesallofthefailurescausedatthehumanendofHRI.Formobilerobotsthismeansfailurescausedbytherobot'soperators,aswellasanysecondaryandtertiarystakeholderswhomaybedirectingtheoperator'sactions[10].Bothmistakes,orcognitiveerrors,coveredinSection5.2.2.1andslips,orunconsciouserrors,describedinSection5.2.2.2fallunderthiscategory.84

PAGE 96

Figure19.ComparingWTCWorkingConditionstothatoftheHCFRDStudy.ImagetakenattheWTCdisasterleft.AMicrotracsandMicroVGTVcooperativelynavigatingasmallpileofdebrisattheHCFRDstudyright.5.2.2.1Mistakes.IntheWTCEngineeringstudy[34]manyofthefailurescouldbecategorizedasmistakes,wheretheoperatorwasplanninghisactionsbasedonincompleteknowledge.Itisdifcultforanyhumantomaintainacognitivemodeloftheenvironmenttherobotisinwithonlyasinglecolorcamerawithalimitedeldofviewand2-wayaudio.Anoperatorwhoisphysicallyandcognitivelyfatiguedforinstanceduetoperceivedthreatstopersonalsafetywillhavemoreproblems.Gravityassistsforexample,mayhaveoccurredwhentheoperatordrovetherobotintoanareawheretheinclinewastoosteepforit,becausehecouldnotjudgetheverticalorientationofthevoid.Otherexampleswouldbestuckassistsneededwhentheoperatorlosttrackoftheobstaclesaroundtherobot,andoccludedcamerawhenhedidnotknowwhichwaytogotogetaroundtheobstacleblockingitsview.Someofthetimespentintheincorrectlightingfailuremodewasdenitelyduetomistakesontheoperator'spart.Hewastryingtoimprovethecameraviewofthevoidbyadjustingtherobot'shalogenlights,butatthesametimethecamcorderwasautomaticallyadjustingtothenewconditions.Theoperatorknewaboutthisfeatureofthecamcorder,butdidnotimmediatelyrealizewhatwashappening.85

PAGE 97

Figure20.ARTSonitsSideAfteraFall.PhototakenfromARTSstudy.[59]Asensorimpoverishedrobotcanleadtodecitsinanawarenessofthestateoftherobot,aswellastheenvironment.InoneexampledescribedbyTECOtheoperatorsattemptedtodrivetheARTSupa30%slope[59].TheplatformbecameunstableandtheoperatorscouldnotrecoverintimetopreventtherobotfromrollingonitssideseeFigure20.TECO'sSARGEstudy[44]notednavigationerrorspositioninganddrivingmadebytheoperatorsduetothedifcultyofjudgingdistanceandposition.Thestudyexplicitlyblamedtheseproblemsonthefactthattheinstalledcameradidnotprovideadequateperipheralvisionordepthperception.ItisimportanttonoteherethattheWTCstudies[34][8]andTECOinthreeseparatestudiesspecicallycitedthelackofdepthperceptionasaproblemencounteredintheeld.Thisappearstobeanimportantfeaturelackingincurrentsensorsuites.Itoccursregardlessofthepayloadcapabilityorthelocationofthecamera,astheARTSvehicleisconsiderablylargerthantherobotsusedattheWTC.ThedifferenceinpayloadcapabilitiesledTECOtorecommendusingadditionalcameras,whereastheWTCstudyrecommendedsoftwareand/orcognitivescience-basedsolutions.86

PAGE 98

Undetectedphysicalfailurescanalsoleadtohumanmistakes.TECOalsohadproblemswiththeUrbanplatform[43]duetothefactthattheoperatorsdidnotrecognizethatthedrivesystemwasfrozen,andwouldovertaxthepowersysteminavainefforttogettherobottomove.Evenexperiencedoperatorswhoarenotintimatelyfamiliarwiththecapabilitiesofaroboticsplatformmayoverestimatethem,causingadditionalfailures.Anexampleofsuchamistakefromthereliabilitystudies[6][7]waswhenanoperatorattemptedtouseanATRV-Jrtopushaheavyloadupahill.Thisresultedinablownmotoramplierwhichimmobilizedtherobot.AsimilarcasewithinexperiencedoperatorswasmentionedinTECO'sDEUCEstudy[14].Theoperatortriedtosteertherobotwiththerippersembeddedinrock,notrealizingthatthisactionwouldbendthetool.5.2.2.2Slips.Thehalogenlightsfailure[34]attheWTCwasagoodexampleofaslip.Inthiscasethehalogenlightsfailedduetoanenergyspikewhentherobot'stetherwasremovedandthenreconnectedtotheOCU.Theoperatorwhocausedthisfailureknewthepotentialproblemsofpluggingelectricalcomponentsbacktogetherwhileoneisstillpowered.HeprobablyintendedtoturnofftheOCU,justincase,butduetodistractionsand/orfatiguedidnotdoso.Minorslipscanoccurfrequentlywhenoperatorsdonothaveenoughexperiencewithaparticularrobotsystem.Twoexamplesfromthereliabilitystudies[6][7]wouldbeajoystickthatwasdroppedinsand;andatetherthatwasnotproperlyconnectedtotheOCU.Inbothcasestheoperatorshadlearnedenoughabouttheplatformstounderstandhowtousethemproperly,butdidnothavealotofpractice.Alloftheslipsmentionedsofararenotuniquetomobilerobotsusedineldenvironments.Thesetypesofhumanerrorarecommonfeaturesofhuman-computerinteractionHCIorevenjustman-machineinteraction.CertainaspectsofHRI,especiallyineldenvironments,leadtoadditionalmistakesandslipsnotcommonlyfoundinotherelds.Forexample,adistinctattributeofteleoperationcontrollingtherobotfroma87

PAGE 99

remotelocationisthatitremovestheoperatorfromtheenvironmentoftherobot.InTECO'sUrbanRobotstudy[43]theoperatorwouldoftendrivetherobotdangerouslyclosetosuspiciousobjectsinordertodetermineiftheyweremines,grenades,orsomeotherharmfuldevice.Ifthesoldiershadbeenthereinplaceoftherobottheyprobablywouldhaveactedmorecautiously.SimilarteleoperationslipsfromtheHCFRDstudy[10]includedtwocollisionswithwallsastheoperatorwastryingtonavigatearobotupaightofstairs.5.3SummaryThischapterhasreportedexampleeldfailuresandrelevantndingsfrom13studiesofeldworkwithmobilerobotsintheapplicationareasofUSARandMOUT.TheexamplesdemonstratehowmobilerobotfailurescanbeclassiedusingthenovelmobilerobotfailuretaxonomypresentedinChapterThree,andthechallengesassociatedwithusingrobotsineldenvironments.Numericresultswerecomparedwheneverpossible.Thechapter'sorganizationfollowedthetaxonomy.Section5.1coveredthephysicalfailures,Section5.2exploredthehuman-robotinteractionHRIfailures,andSection5.1.7presenteddataontherepairabilityandimpactattributes.Figure21providesasummaryofthendingsintermsofthetaxonomypresentedinChapterThree.Theprobabilitythatagivenfailurebelongstoagivenclassisdisplayedbeneaththatclassleafnodeinthetaxonomytree.Ifmultiplesourcestudiesprovideddata,theprobabilityisshownasarange.Designfailuresunderhumanfailureswerenotexploredbythe13studiesthismeta-studydrewfrom,andthereforenoprobabilityisincludedforthisclass.Ideallytheprobabilitiesofsiblingsshouldsumto1.0butsincetheyareextractedfrommultiplesources,thatisnotalwaysthecase.AsseeninFigure21,theeffectorsandthecontrolsystemarethemostcommonsourcesoffailures,withatmosthalfofthefailuresfallingintooneoftheseclasses.88

PAGE 100

Figure21.SummaryofClassicationResultsUsingtheFailureTaxonomy.Includesprob-abilitiesforeachleafclassandattributevalue.Forexample,theprobabilitythatafailurewasaneffectorfailureisatleast.11andatmost.5dependingonthesourceandtheprobabilitythatthefailurewillbeeldrepairableisbetween0.60and0.65.Sensorfailuresarelessfrequentwithatmost26%ofthefailures,followedbypowerat9%,andcommunicationsfailuresat1%.Thoughcommunicationsfailuresappeartobetheleastfrequent,itshouldbenotedthatseveralofTECO'sstudies,whichdidnotprovideenoughdetailstoderiverelativefrequencies,reportedchroniccommunicationsproblems.Withinthehumanfailuresbranch,slipsaremorecommonwithatmost67%ofthefailureswhilemistakescontributednomorethan38%ofthefailuresexamined.Sincethephysicalandhumanfailureresultscamefromdifferentsources,therelativefrequencyofphysicalversushumanfailurescannotbedeterminedfromthesestudies.ThereliabilityanalysesshowthattheMTBFformobilerobotsintheeldisbetween6and12hours.TECOreportedthattherobotstheyexaminedfailedwithin20hoursofuse[56],thoughtheirstudiesdidnotprovideenoughinformationseeChapterFourtovalidatethatgure.Theattributesaresimilarlymarkedwiththeprobabilitythatagivenfailurewillhavethatattributevalue.Field-repairableandterminalfailuresaremorecommonthantheiropposites,withupto65%and94%respectivelyofthefailurescoveredinthismeta-study.Thestudiesdidnotprovidesufcientdatatocalculateconditional89

PAGE 101

probabilitiesbetweentheclassesandattributes,forexample,theprobabilitythatagiveneffectorfailurewillbeterminaloreld-repairable.Adetailedexaminationoftheeldfailuresdescribedinthe13studiesuncoveredthefollowingcommonissues:1.unstablecontrolsystems2.chassisandeffectorsdesignedandtestedforanarrowrangeofenvironmentalconditions3.limitedwirelesscommunicationrangeinurbanenvironments4.insufcientwirelessbandwidthforvideo-basedfeedback90

PAGE 102

ChapterSixConclusionsThisthesishasexploredthequestionofhowground-basedmobilerobotsfailintheeld.Twocontributionswereproducedfromthiseffort:1.Anoveltaxonomyofmobilerobotfailuresthatreconcilesfailureclassicationschemesfromthedependabilitycomputing[30],human-computerinteractionHCI[40],androbotics[6]communities.2.Ameta-studyincluding44representativeexamplesofmobileroboteldfailuresdrawnfrom13studiesintheurbansearchandrescueUSARandmilitaryoperationsinurbanterrainMOUTdomains.TheCenterforRobot-AssistedSearchandRescueCRASARattheUniversityofSouthFloridaprovidedveofthestudies,andtheremaindercomefromtheUSArmy'sTestandEvaluationCoordinationOfceTECOatFortLeonardWood.Thefailureexamplesdrawnfromthesestudieswereusedtodemonstratehowmobilerobotfailurescanbeclassiedusingthenewtaxonomyandthechallengesassociatedwithusingrobotsineldenvironments.AreviewofrelatedworkpresentedinChapterTwoestablishedthatthisworkwasuniqueinthreerespects:itcoversbothhumanandrobotfailures,itisfocusedonroboteldfailuresintheUSARandMOUTdomains,anditistheonlymeta-studythatcoversmobilerobotfailures.Itwasalsofoundthatnoexistingapproachtofailureanalysisissuitableforthetaskofmobileroboteldfailureanalysis.91

PAGE 103

ChapterThreepresentedanewapproachtomobilerobotfailureanalysisusedinthisthesis,includinganoveltaxonomybuiltonexistingwork.FollowingLaprie[30],failuresweredividedintophysicalandhumanclasses,withhumansubdividedintodesignandinteraction.Thehuman-computerinteractionclassesofmistakesandslips[40]fellunderinteractionfailures.PhysicalfailureclassesweretakenfromCarlsonandMurphy[6].Theseareeffector,sensor,controlsystem,power,andcommunications.Thenewtaxonomywasthenusedtoexplore,inChapterFive,howmobilerobotsfailintheeldbasedon13studies[3][6][7][8][10][13][14][34][43][44][55][59][62]describedindetailinChapterFourwhichexaminedmobilerobotperformanceintheeld.Foreachclassoffailure,exampleswereprovidedwhichillustratethenatureofmobileroboteldfailureswhichfallunderthatclass.EachexamplefailureisdescribedindetailinAppendixB.Thischapterprovidesasummaryofthendingsofthismeta-studyinSection6.1.ThisisfollowedbyadiscussioninSection6.2oftheimplicationsofthosendings,aswellaspossiblesolutionstotheproblemsencountered.Section6.3closeswithabriefoverviewoffuturework.6.1FindingsHowdoground-basedmobilerobotsfailintheeld?Thenineoverallndingslistedbelowanswerthisquestionbasedontheinformationfoundinthe13studiescoveredinthisthesis.Thesendingscoverstudiesfromtwoeldapplicationdomains:veinUSAR,andeightinMOUT.28robotswereexaminedwhichrepresent15differentmodelsfromsevenmanufacturers,andrangefromsmalllessthan10poundstrackedvehicles,toamodiedM1tankover60tons.Duetodifferencesindatacollectionandreportingmethodsamongthe13studies,aprecisequantitativesummaryoftheirexperienceintermsofreliabilitymetricswasnotpossible.Therefore,themeantimebetweenfailures92

PAGE 104

andrelativefrequencyoftheclassiedfailureswerepresentedasestimatesonly.Sincetheseweretheonlystudiesfoundintheliteraturethatdescribemobileroboteldfailures,thesenumbersrepresentthebestanswerthatcanbemade,atthistime,tothequestionofhowground-basedmobilerobotsfailintheeld.1.Robotreliabilityineldenvironmentsislow,meantimebetweenfailuresMTBFisbetween6and24hours.2.Commonissuesacrossthe13studiesarethefollowing:aunstablecontrolsystemsbchassisandeffectorsdesignedandtestedforanarrowrangeofenvironmentalconditionsclimitedwirelesscommunicationrangeinurbanenvironmentsdinsufcientwirelessbandwidthforvideo-basedfeedback3.Trackedvehiclesaremorepronetoeffectorfailuresthantheirwheeledcounterparts.4.Effectors%andthecontrolsystem%arethemostcommonsourcesofphysicalfailures.5.Sensorfailuresmakeupbetween9%and26%ofthefailures.6.Power%andcommunications%faultsweretheleastfrequentsourcesoffailure.7.Slips%aremorecommonthanmistakes%.8.Field-repairablefailuresmakeuptwo-thirds%ofthefailuresexamined.9.Terminalfailuresaremorecommonthannon-terminalwith94%ofthefailuresreportedinTECO'sPANTHERstudy[13].93

PAGE 105

Asthendingsstate,overallrobotreliabilityineldenvironmentsislow,between6and24hoursMTBF,dependingonthecriteriausedtodetermineifafailurehasoccurred.Effectorsandthecontrolsystemarethemostcommonsourcesofphysicalfailuresacrossthe11studieswhichreportedphysicalfailures.ThisresultwasveriedthroughstatisticalanalysisinCRASAR'sfollow-upreliabilitystudy[7].Thoughcommunicationsfailuresappeartobetheleastfrequent,itshouldbenotedthatseveralofTECO'sstudiesreportedchroniccommunicationsproblemsbutdidnotstatethenumberofcommunicationsfailures.Therefore,thequantitativeresultsprobablyunderestimatethefrequencyofcommunicationsfailuresformobilerobotsusedintheeld.Withinthehumanfailuresbranch,slipsaremorecommonthanmistakes.Humandesignfailureswerenotcoveredinanyofthe13studies.SincethemajorityofphysicalandhumanfailureresultscamefromdifferentsourcesonlyCRASAR'sWorldTradeCenterWTCEngineeringstudy[34]contributedbothphysicalandhumanfailures,therelativefrequencyofphysicalversushumanfailurescannotbedetermined.Field-repairablefailuresmakeuptwo-thirdsofthefailuresexamined.Terminalfailuresarealsomorecommonwith94%ofthefailuresreportedintheM1PANTHERIIstudy[13]performedbytheTECO.Theresultsofthismeta-studysupporttheconcept,theorizedintheoriginalreliabilitystudy[6],thatmaturityhasalargeimpactonthereliabilityofaplatform.Theresultsfromtheoriginalreliabilitystudy[6]indicatedthatmatureplatformsaremorereliable,eveninapplicationsandenvironmentsthattheywerenotdesignedfor,thanlessmatureplatformsdraftedforthatapplicationalone.Inaddition,itappearsthatthematurityofeachsubsystemwithinarobotplatformalsoinuencestheoverallsystem'sreliability.Forexample,thenewteleoperationsysteminstalledonthePANTHERfailedfarmorefrequentlythantheplatform'seffectors,whichhavebeenactivelyusedforover20yearsaspartoftheUSArmy'sM1tankplatform.AnotherexampleistheInuktun94

PAGE 106

platformswhichfellfrom90%to27%availabilitybetweentheoriginalandfollow-upreliabilitystudies[6][7]duelargelytoupgradesinthecontrolsystem,whichdidaddvaluablefeaturestotheplatform,butalsoaddedtothecomplexityandreducedthereliabilityofthatsubsystem.6.2DiscussionDeterminingtheunderlyingcausesforthendingspresentedinSection6.1isbeyondthescopeandexpertiseofthisthesis.Theimplicationsofthesendingsforpotentialusersineldapplications,mobilerobotdesigners,andfault-toleranceresearchersanddevelopersareasfollows:1.Astateoftheartmobilerobotcannotbeexpectedtocompleteanentireshiftwithoutincident.2.A50%availabilityrate[6][7]impliesthatadditionalbackupequipmentmustbeavailableforeldapplications.3.Themostcommonsourcesoffailureinmodernroboticsystemsarecustom-builtbyhandandincreasinglycomplexcontrolandeffectorsystems.4.Themostreliablecomponentsinmodernrobotsystemsaresimplepowerand/ormass-producedsensors.5.Mostfailureswillinterrupttherobot'smissionbutrequirerelativelyminorrepairs.PotentialusersinelddomainssuchasUSARandMOUTshouldbeawarethatastateoftheartmobilerobotcannotbeexpectedtocompleteanentireshifthoursforUSARor20hoursfortheDepartmentofDefensewithoutincidentandrobotsareexpectedtohavea50%availabilityrate[6][7].Thisimpliesthatadditionalbackup95

PAGE 107

equipmentmustbeavailableforeldapplications,doublingtheresourcesandlogisticsneededtogetasinglerobotintheeld.Ineffectthisraisesthebarformobileroboteldapplications,makingitmoredifcultforaroboticsystemtoprovethatitisinfactusefulenoughtooffsetthesecosts.RobotmanufacturersshouldbecognizantofthefactthatmobilerobottechnologyissufferingfromthesamecreepingincreaseincomplexityidentiedbyNormaninthecomputerindustry[39].Toexacerbatethisproblem,thecontrolanddrivesystemsareusuallycustomdesignedforaspecicrobotmodel,andarebuiltbyhandthisisthecaseforthe15modelscoveredinthisthesis.Qualitycontrolforacomplex,custom-builtsystemisdifculttomanageevenwithsufcientresources.Ultimately,manufacturersmustacceptthattheirrobotswillsufferfromfailuresintheeld,andmustdesignformaintainability.Regularmaintenancetasksshouldbemadeaspainlessaspossiblefortheend-userandanycustompartsshouldbereadilyavailableintheeventthatseriousfailuresoccur.Basedonthecommonissuesforalltherobotplatforms,limitedmobility[2][9]andunreliablewirelesscommunicationsareproblemswhichneedtobeaddressed.Thiswillrequiretheattentionofrobotdesignersinbothindustryandresearchdomains.Someresearchers[22]havespentyearsdevelopingrobotplatformswithadvancedmobilitycapabilitieslikewallclimbing.Thesesolutionsneedtobeexploredandhardenedaspartofacompleterobotsystem,withsufcientpayloadcapabilitiestocarrythematerialstherobotsneedtocompletetheirassignedtask.Alternativesolutionstothewirelesscommunicationproblemforexample,theuseofacombinationofwiredandwirelesscommunicationorrepeaterstoboostsignalstrengthhavealsobeenexplored[38][11]andneedtobedevelopedfurtherforeldapplications.Fault-tolerancesystemssimilarto[17]and[19],whichhandleeffectorfailuresinwheeledrobots,needtobedevelopedfortrackedvehicles,whicharemorepronetothis96

PAGE 108

classoffailure.Researchersinfault-toleranceshouldalsonotethatcomplexcomponentsaremorelikelytofailthansimpleones.Thisfavorsmodel-basedmethodswhicharebetterathandlingthislevelofcomplexity.Ontheotherhand,thesamecomponentsareusuallycustombuiltbyhand,whichmeansthatasufcientlyprecisemodelislikelytoapplytoexactlyonerobot.Basedonthendingsofthisthesis,hybridfault-tolerancesystemswhichuselearning[32],qualitativereasoning[60],oractiveprobinggatheringadditionalinformationfromothersensorsorrobots[31]toaugmentlessprecisemodelsappeartoholdthemostpromiseforeldroboticsapplications.6.3FutureWorkThisthesiswastherstofitskind,andhaslaidthegroundworkforfuturestudiescharacterizinghowmobilerobotsfailineldenvironments.Thissectiondiscussesthreekeyavenuesforsuchwork:improveddatacollectionmethods,human-robotinteractionfailures,andnewmethodsforapplyingresultsofstudieslikethisonetoimproveexistingfault-toleranceapproaches.Automatedblackboxdatacollectionmethods,likethoseusedonmodernairplanes,areneededformobilerobotstoautomaticallyrecordbothusageandfailuredataforfuturestudy.Duetolimitedcommunicationsbandwidth,operatorcontrolunitsoftencannotreceivealloftheinformationgeneratedbytherobot'ssensorsandcontrolsystematagiventime.Thereforeloggersareneededthatresideinarobot'sonboardcomputerandhaveaccesstovitalinformationwhicharobotoperatorcannotreadilyaccess.Ascomputingresourcesonmobilerobotsareoftenlimited,loggersneedtohaveaminimalcomputationaloverhead.Further,constraintsoncommunicationsbandwidthandstoragerequirecompactrepresentationsofinformation,andthatthisinformationbelteredforrelevancetotheoperatorandotherconsumersofusageandfailuredatai.e.onlinefaultdetectionandofinefailureanalysismodules.97

PAGE 109

On-goingresearchinhuman-robotinteraction[4][63]islikelytodemonstratethattherearesignicantdifferencesbetweenitandthehuman-computerinteractioneld.Morestudiesareneededtoexplorehumanfailureswithmobilerobotsaspartofthiseffort.Reliabledatacollectionmethodswhichcanbeusedforextendedperiodsoftime,likethoseinplaceforphysicalfailures,needtobedevelopedandimplemented.Thehumanbranchofthefailuretaxonomyusedinthisthesiswassufcientforthelimitednumberlessthan20ofhumanfailuresexaminedhere.Studiesofmoreextensiverecordsofhuman-robotinteractionfailuresarelikelytoleadtorenementofthehumanfailurebranch,justassimilarstudiesofphysicalfailures[6]leadtoamoredetailedclassicationschemeforthatclassoffailures.Theincorporationofprobabilityestimates,likethosegiveninSection6.1,isoftenstraightforwardinfault-toleranceapproachesthatuseprobabilitymodelstodetectand/ordiagnosefailuressuchas[25],[32],[51],and[61].Forotherapproachestofault-tolerancewhichdonotmodelthestateofthesystemexplicitlyanduseactivediagnosisprobingforadditionalinformationfromsensorsand/orotherrobots,like[31]and[15],theincorporationofprobabilitydataispossiblebutmoreproblematic.Systemsareneededwhichcanreliablyrankfailurehypothesisbasedonanyinformationaboutthecurrentstateofthesystemgatheredduringthedetectionphaseandthelikelihoodofthatfailurebasedontheestimatedprobability.Suchsystemswouldreducethecostofthediagnosisprocesswhichcanbequitehighifotherrobotsarerecruitedasin[31]toassistinhypothesistestingbyensuringthatthemostlikelyhypothesesarecheckedrst.98

PAGE 110

References[1]Y.BeauchampandT.J.Stobbe.Areviewofexperimentalstudiesonhuman-robotsys-temsituationsandtheirdesignimplications.InternationalJournalofHumanFactorsinManufacturing,5:283,1995.[2]JohnG.Blitch.Adaptivemobilityforrescuerobots.InProceedingsofSPIE.Sensors,andCommand,Control,Communications,andIntelligenceC3ITechnologiesforHomelandDefenseandLawEnforcementII,volume5071,pages315,2003.[3]GlennBoxley.Teleoperationalddozerconceptevaluationprogramcepexecutivesummary.Conceptexperimentationprogramcepreportandexecutivesummary,USArmy,TestandEvaluationCoordinationOfceTECO,Ft.LeonardWood,1999.Fromarchivesatwww.jointrobotics.com.[4]JenniferL.Burke,RobinR.Murphy,andMichaelD.Coovert.Moonlightinmiami:Anethnographicstudyofhuman-robotinteractioninthecontextofanurbansearchandrescuedisasterresponsetrainingexercise.ToappearinHuman-ComputerInter-action.[5]G.Candea,E.Kiciman,S.Zhang,P.Keyani,andA.Fox.Jagr:Anautonomousself-recoveringapplicationserver.InFifthAnnualInternationalWorkshoponActiveMiddlewareServices.ProceedingsoftheAutonomicComputingWorkshop.,pages169,2003.[6]J.CarlsonandR.Murphy.Reliabilityanalysisofmobilerobots.InProceedingsoftheIEEEInternationalConferenceonRoboticsandAutomation,2003.[7]JenniferCarlson,RobinR.Murphy,andAndrewNelson.Followupanalysisofmo-bilerobotfailures.ToappearatIEEEInternationalConferenceonRoboticsandAutomation,2004.[8]J.CasperandR.Murphy.Human-robotinteractionsduringtherobot-assistedurbansearchandrescueresponseattheworldtradecenter.IEEETransactionsonSystems,Man,andCyberneticsPartB,33:367,June2003.[9]JenniferCasper,MarkMicire,andRobinMurphy.Issuesinintelligentrobotsforsearchandrescue.InSPIEGroundVehicleTechnologyII,volume4,pages41,2000.99

PAGE 111

[10]JenniferCasperandRobinMurphy.Workowstudyonhuman-robotinteractioninusar.InProceedingsoftheIEEEInternationalConferenceonRoboticsandAutoma-tionICRA,2002.[11]A.CerpaandD.Estrin.Ascent:Adaptiveself-conguringsensornetworktopologies.InProceedingsIEEETwenty-FirstAnnualJointConferenceoftheIEEEComputerandCommunicationSocietiesINFOCOM2002,volume3,pages1278,2002.[12]C.Combastel,S.Gentil,andJ.-P.Rognon.Asymbolicreasoningapproachtofaultdetectionandisolationappliedtoelectricalmachines.InProceedingsoftheIEEEIn-ternationalConferenceonControlApplications,volume1,pages475,Septem-ber1998.[13]F.W.Cook.Testreportforthepantherii.Testreport,USArmy,TestandEval-uationCoordinationOfceTECO,Ft.LeonardWood,1997.Fromarchivesatwww.jointrobotics.com.[14]F.W.Cook.Limitedobjectiveexperimentationreportforthedeployableuniver-salcombatearthmoverdeuce.Executivesummary,USArmy,TestandEval-uationCoordinationOfceTECO,Ft.LeonardWood,2000.Fromarchivesatwww.jointrobotics.com.[15]J.deKleerandO.Raiman.Tradingoffthecostsofinferencevs.probingindiag-nosis.InProceedingsoftheFourteenthInternationalJointConferenceonArticialIntelligence,pages1736,1995.[16]B.Deuker,M.Perrier,andB.Amy.Fault-diagnosisofsubsearobotsusingneuro-symbolichybridsystems.InIEEEOceanicEngineeringSociety.OCEANS'98.Con-ferenceProceedings,pages830,1998.[17]W.E.Dixon,I.D.Walker,andD.M.Dawson.Faultdetectionforwheeledmobilerobotswithparametricuncertainty.InProceedingsofIEEE/ASMEInternationalCon-ferenceonAdvancedIntelligentMechatronics,volume2,pages1245,2001.[18]X.Dong,S.Hariri,L.Xue,H.Chen,M.Zhang,S.Pavuluri,andS.Rao.Autonomia:Anautonomiccomputingenvironment.InProceedingsoftheIEEEInternationalPer-formance,Computing,andCommunicationsConference,pages61.IEEE,2003.[19]P.Goel,G.Dedeoglu,S.I.Roumeliotis,andG.S.Sukhatme.Faultdetectionandidenticationinamobilerobotusingmultiplemodelestimationandneuralnetwork.InProceedingsofIEEEInternationalConferenceonRoboticsandAutomation,vol-ume3,pages2302,2000.[20]J.L.Gonzalez,M.Mediavilla,J.C.Fraile,F.Gayubo,J.P.Turiel,andF.J.Garcia.In-uencesofrobotmaintenanceandfailuresintheperformanceofamultirobotsystem.InProceedings2002IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems,volume3,pages2801,2002.100

PAGE 112

[21]R.Helfman,E.Baur,J.Dumer,T.Hanratty,andH.Ingham.Turbineenginediagnos-ticsted:Anexpertdiagnosticsystemforthem1abramsturbineengine.InProceed-ingsoftheFifteenthNationalConferenceonArticialIntelligence,pages1032,1998.[22]S.HiroseandH.Arikawa.Coupledanddecoupledactuationofroboticmechanisms.InProceedings2000ICRA.MillenniumConference.IEEEInternationalConferenceonRoboticsandAutomation.,volume1,pages33,2000.[23]E.HungandF.Zhao.Diagnosticinformationprocessingforsensor-richdistributedsystems.InProceedingsofthe2ndInternationalConferenceonInformationFusionFusion99,1999.[24]InstituteofElectricalandElectronicsEngineers,Inc.Ieeerecommendedpracticeforthedesignofreliableindustrialandcommercialpowersystems.IEEEStd493-1997,1998.[25]E.Jackson.Real-timemodel-basedfaultdetectionanddiagnosisforautomatedsys-tems.InDynamicModelingControlApplicationsforIndustryWorkshop,IEEEIn-dustryApplicationsSociety,pages26,1997.[26]MengJi,ZhenZhang,G.Biswas,andN.Sarkar.Hybridfaultadaptivecontrolofawheeledmobilerobot.IEEE/ASMETransactionsonMechatronics,8:226,2003.[27]K.Kawabata,T.Akamatsu,andH.Asama.Astudyofself-diagnosissystemofanautonomousmobilerobot:expansionofstatesensorysystems.InProceed-ingsIEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsCat.No.02CH37332C,volume2,pages1802,2002.[28]A.I.KokkinakiandKimonValavanis.Errorspecication,monitoring,andrecoveryincomputer-integratedmanufacturing:Ananalyticapproach.InIEEEControlTheoryandApplications,volume143,pages499,November1996.[29]K.B.LamineandF.Kabanza.Historycheckingoftemporalfuzzylogicformulasformonitoringbehavior-basedmobilerobots.InProceedings12thIEEEInternationalsConferenceonToolswithArticialIntelligence,pages312,2000.[30]J-C.Laprie.Dependablecomputingandfault-tolerance:Conceptsandterminology.InTwenty-FifthInternationalSymposiumonFault-TolerantComputing,pages27,1995.Reprint,originallypublishedin1985.[31]MattLongandRobinR.Murphy.Distributedmulti-agentdiagnosisandrecoveryfromsensorfailures.InIEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsIROS,volume3,pages2506,October2003.101

PAGE 113

[32]R.Mackey,M.James,HanPark,andM.Zak.Beam:Technologyforautonomousself-analysis.In2001IEEEAerospaceConferenceProceedings,pages2989,2001.[33]M.G.M.MaddenandP.J.Nolan.Monitoringanddiagnosisofmultipleincipientfaultsusingfaulttreeinduction.InIEEProceedings-ControlTheoryandApplications,vol-ume146,pages204,1999.[34]MarkMicire.Analysisoftherobotic-assistedsearchandrescueresponsetotheworldtradecenterdisaster.Master'sthesis,UniversityofSouthFlorida,July2002.[35]N.H.Minsky.Onconditionsforseal-healingindistributedsoftwaresystems.InFifthAnnualInternationalWorkshoponActiveMiddlewareServices.ProceedingsoftheAutonomicComputingWorkshop,pages91,2003.[36]RobinR.Murphy.IntroductiontoAIRobotics.TheMITPress,2000.[37]RobinR.Murphy,JenniferCasper,JefferyHyams,MarkMicire,andBrianMinten.Mobilityandsensingdemandsinusar.InIEEEInternationalConferenceonIndus-trialElectronics,ControlandInstrumentation,October2000.[38]H.G.Nguyen,H.R.Everett,N.Manouk,andA.Verma.Autonomousmobilecom-municationrelays.InProceedingsoftheSPIE-TheInternationalSocietyforOpticalEngineering,volume4715,pages50,April2002.[39]DonaldA.Norman.TheInvisibleComputer.TheMITPress,1999.[40]DonaldA.Norman.TheDesignofEverydayThings,chapterToErrIsHuman,pages105.TheMITPress,2002.[41]IllahR.Nourbakhsh.Themobotmuseumrobotinstallations.InProceedingsoftheIEEE/RSJIROS2002WorkshoponRobotsinExhibitions,pages14,2002.[42]DepartmentofDefense.Departmentofdefensejointroboticsprogram.http://www.jointrobotics.com.[43]G.Piskulic.Militarypolice/engineerurbanroboturbotconceptexperimentationprogramcepreport,engineermodule.Conceptexperimentationprogramcepre-port,USArmy,TestandEvaluationCoordinationOfceTECO,Ft.LeonardWood,2000.Fromarchivesatwww.jointrobotics.com.[44]G.Piskulic.Militarypolicecepforsmallrobotsinsupportofdismountedtroopsutilizingnon-lethalweapons.Conceptexperimentationprogramcepreport,USArmy,TestandEvaluationCoordinationOfceTECO,Ft.LeonardWood,2001.Fromarchivesatwww.jointrobotics.com.102

PAGE 114

[45]S.RamaswamyandKimonValavanis.Hierarchicaltime-extendedpetrinetsh-epnsbasederroridenticationandrecoveryformultilevelsystems.InIEEETransactionsonSystems,ManandCybernetics,PartB,volume26,pages164,February1996.[46]N.E.Reed.Constructingthecorrectdiagnosiswhensymptomsdisappear.InProceed-ingsFifteenthNationalConferenceonArticialIntelligenceAAAI-98,pages151156,1998.[47]R.Rymon.Goal-directeddiagnosis-diagnosticreasoninginexploratory-correctivedomains.InProceedingsoftheThirteenthInternationalJointConferenceonArticialIntelligence,pages1488,1993.[48]C.Sary,C.Peterson,J.Rowe,T.Ames,K.Mueller,W.Truszkowski,andN.Ziyad.Trendanalysisforspacecraftsystemsusingmultimodalreasoning.InMultimodalReasoning.Papersfromthe1998AAAISymposium,pages157,1998.[49]F.T.Sheldon,H.Mei,andS.-M.Yang.Reliabilitypredictionofdistributedembeddedfault-tolerantsystems.InFourthInternationalSymposiumonSoftwareReliabilityEngineeringProceedings,pages92,1993.[50]R.Simmons,C.Pecheur,andG.Srinivasan.Towardsautomaticvericationofau-tonomoussystems.InProceedingsofIEEEInternationalConferenceonIntelligentRobotsandSystems,volume2,pages1410,2000.[51]M.Soika.Asensorfailuredetectionframeworkforautonomousmobilerobots.InProceedingsofthe1997IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems.InnovativeRoboticsforReal-WorldApplications,pages1735,1997.[52]A.G.Starr,R.J.Wynne,andI.Kennedy.Failureanalysisofmaturerobotsinauto-matedproduction.InProceedingsoftheInstitutionofMechanicalEngineers,PartBJournalofEngineeringManufacture,volume213,pages813,1999.[53]R.SterrittandD.Bustard.Autonomiccomputingameansofachievingdepend-ability.InProceedingsIEEEInternationalConferenceandWorkshopontheEngi-neeringofComputer-BasedSystems,pages247,2003.[54]A.P.Tchangani.Reliabilityanalysisusingbayesiannetworks.StudiesinInformaticsandControl,10,2001.[55]TestandEvaluationCoordinationOfceTECO.Finalreportforthechemicalbi-ologicalradiologicalnuclearcbrnsensormodulewithroboticplatformlimitedob-jectiveexperimentloe.Finalreport,USArmy,TestandEvaluationCoordinationOfceTECO,Ft.LeonardWood,2003.Fromarchivesatwww.jointrobotics.com.[56]RoxanaTiron.Lackofautonomyhamperingprogressofbattleeldrobots.NationalDefense,May2003.103

PAGE 115

[57]Y.Tohma.Faulttoleranceinautonomiccomputingenvironment.InProceedingsPa-cicRimInternationalSymposiumonDependableComputing,pages3,December2002.[58]N.Tomatis,G.Terrien,R.Piguet,D.Burnier,S.Bouabdallah,andR.Siegwart.De-signandsystemintegrationfortheexpo.02robot.InProceedingsoftheIEEE/RSJIROS2002WorkshoponRobotsinExhibitions,pages67,2002.[59]R.R.WalkerandM.R.Scarlett.Testreportfortheall-purposeremotetransportsystemartsexecutivesummary.Executivesummary,USArmy,TestandEval-uationCoordinationOfceTECO,Ft.LeonardWood,1999.Fromarchivesatwww.jointrobotics.com.[60]P.Wang,T.Yamasaki,M.Yumoto,T.Ohkawa,N.Komoda,andF.Miyasaka.Anefcientstochasticqualitativereasoningmethodforrealtimefaultdiagnosisofbuild-ingairconditioningsystem.InProceedingsofIEEEInternationalConferenceonSystems,Man,andCybernetics,volume5,pages854,1999.[61]R.Washington.On-boardreal-timestateandfaultidenticationforrovers.InProceedings2000ICRA.MillenniumConference.IEEEInternationalConferenceonRoboticsandAutomation.SymposiaProceedings,pages1175,2000.[62]R.Weiss.Finaltestreportfortheunmannedgroundvehiclerapidobscurantplatformugvropchemicalcorpsconceptexplorationprogramugvrop-cep.Conceptexper-imentationprogramcepreport,USArmy,TestandEvaluationCoordinationOfceTECO,Ft.LeonardWood,2001.Fromarchivesatwww.jointrobotics.com.[63]HollyA.YancoandJillL.Drury.Ataxonomyforhuman-robotinteraction.InAAAIFallSymposiumofHuman-RobotInteraction,pages111,November2002.Uni-versityofMassachusettsLowell.104

PAGE 116

Appendices105

PAGE 117

AppendixADenitionsofReliabilityRelatedTermsThefollowinglistofdenitionscoverstheterminologyusedinthispaperandisprovidedasareference.1.availability.Probabilitythatasystemwillbeerrorfreeatsomegivenpointintime.See.Availability=MTBF MTBF+MTTR100%2.autonomy.Levelofsupervisionrequiredbyhumans.Acontinuumfromnoautonomyteleoperatedtofull-autonomyinwhichnosupervisionisneeded.3.controlsystem.Arobotsubsystemthatincludestheonboardcomputer,manufacturerprovidedsoftware,andanyremoteoperatorcontrolunitsOCU.4.effector.Anydevicethatperformsactuationandanyconnectionsrelatedtothosecomponents.5.error.Astatewithinthesystemwhichcanleadtoafailure.6.failure.Theinabilityoftherobotoritssupportequipmenttofunctionnormally.7.fault.Anythingwhichcouldcausethesystemtoenteranerrorstate.8.favorableenvironmentalconditions.Environmentalconditionsareconsideredtobefavorableifallconditions,suchasdampnessandlightlevel,donotinterferewithorpreventagivenrepairprocedure.9.eldenvironment.Anenvironmentwhichhasnotbeenmodiedtoensurethesafetyoftherobotortoenhanceitsperformance.106

PAGE 118

AppendixAContinued10.eld-repairablefailure.Afailureisconsideredtobeeld-repairableifitcanberepairedunderthefollowingconditions:onlytheequipmentthatcommonlyaccompaniestherobotintotheeldisavailable,environmentalconditionsarefavorable,andtheonlypersonnelavailablearetrainedoperators.11.mistakes.Humanfailurescausedbyfallaciesinconsciousprocessing.12.MTBF.MeanTimeBetweenFailures.See.MTBF=NumberofHoursRobotWasinUse NumberofFailures13.MTTR.MeanTimetoRepair.See.MTTR=NumberofHoursSpentRepairing NumberofRepairs14.mobilerobot.Amechanicaldevicethatcansenseandinteractwithitsenvironment.15.non-eld-repairablefailure.Afailurethatisnoteld-repairable.16.non-terminalfailure.Afailurethatintroducessomenoticeabledegradationoftherobot'scapabilitytoperformitsmission.17.slips.Humanfailurescausedbyfallaciesinunconsciousprocessing.18.supportequipment.Equipmentthatisnotphysicallypartoftherobotandisrequiredfortherobottocompleteitsmissionortask.19.teleoperated.Manuallycontrolledbyanoperatoratadistancethatistoogreatfortheoperatortoseewhattherobotisdoing.[36].107

PAGE 119

AppendixAContinued20.terminalfailure.Afailurethatterminatestherobot'scurrentmission.21.UGV.UnmannedGroundVehicle.Aground-basedmechanicaldevicethatcansenseandinteractwithitsenvironment.108

PAGE 120

AppendixBExampleFieldFailureswithClassicationsThisappendixprovidesdescriptionsandclassicationsofeveryexamplemobileroboteldfailureusedinthispaper.Notethatthelevelofdetailprovidedandtheaccuracyofthisdataislimitedtothegranularityandaccuracyofthesourcematerial.ForCRASAR'sreliabilitystudies[6][7]thiswasdeterminedbythedatacollectionmethoddescribedinSection4.6.Fortherestthiswaslimitedtotheinformationprovidedinthestudiesthemselves.Foreachexamplefailurethefollowingisprovided:robotmodelnames,thestudyorstudiesthatreportedthefailure,classication,repairability,impact,cause,symptom,andabriefdescriptionwithcommentswhereneeded.SomeoftheWTCEngineeringfailurecategorieswereplacedinmultipleclasses.Eachadditionalappearanceispresentedbelowtherstandcontainsonlytheclassication,cause,anddescription.Articiallightingfailure RobotsMicroTracsStudyWTCEngineeringstudy[34]ClassslipRepairabilitynon-eld-repairedImpactterminalwhenadditionallightisrequiredCauseenergyspikeinelectricalsystemSymptomsvideobecamedarkDescriptionTheoperatorreconnectedtherobottotheOCUwhilethelatterwasstillpowered.Thisisconsideredtobeaslipbecausetheoperatorwasahardwarespecialistwithenoughtechnicalknowledgetoknowthatthisactionwasriskyatbest,butwasheavilyfatiguedatthetime.109

PAGE 121

AppendixBContinuedBentripperappendage RobotsDEUCEStudyTECO'sDEUCEstudy[14]ClassmistakeRepairabilitynon-eld-repairedImpactnon-terminalunlessripperappendageisrequiredCauseripperisembeddedinrockwhilerobotisturningSymptomsunresponsiveripperappendageDescriptionTheoperatortriedtosteertherobotwiththerippersembeddedinrock,notrealizingthatthisactionwouldbendthetool.Blownmotoramplier RobotsATRV-JrStudyOriginalCRASARReliabilitystudy[6]ClassmistakeRepairabilitynon-eld-repairedImpactterminalCauseovertaxedthemotorSymptomsmotorunresponsiveDescriptionTheoperatorattemptedtousetherobotpushaheavyloadupahill.110

PAGE 122

AppendixBContinuedBoggeddownripperappendage RobotsDEUCEStudyTECO'sDEUCEstudy[14]ClasseffectorRepairabilitynon-eld-repairedImpactnon-terminalunlessripperappendageisrequiredCauseusedinparticularlyhardrockSymptomsunresponsiveripperappendageDescriptionTheripperappendagebecomesboggeddownandcannotmovewithintherock.Cannotcontrolrobotremotely RobotsD-7GStudyTECO'sD-7Gstudy[3]ClasscommunicationsRepairabilityeld-repairedifatallImpactterminalCausecannottransmitthroughinterposedmaterialsSymptomsfrequentlossofvideo,nosignalfromrobotDescriptionThestudyconcludedthatthenon-line-of-sightrequirementcouldbemetbecausetheequipmentcouldnottransmitthroughinterposedmaterials.111

PAGE 123

AppendixBContinuedCloggedfuellter RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasspowerRepairabilityunspeciedImpactterminalCauseunspeciedSymptomsunspeciedDescriptionThefuellterononeofthetwotanksrepeatedlyclogged,requiringareplacementroughlyeverysixhours.Collision RobotsUrbanStudyCRASARHCFRDstudy[10]ClassslipRepairabilitydependsondamagefromthecollisionImpactusuallynon-terminal,dependsondamagefromthecollisionCauseoperatorsremovedfromrobot'soperatingenvironmentSymptomsusuallyvisiblethroughvideoDescriptionOperatorcausedtherobottocollidewiththewallswhiletryingtonavi-gateupaightofstairs.112

PAGE 124

AppendixBContinuedCommunicationsdropout RobotsAllwirelessplatformsStudyWTCEngineeringstudy[34]ClasscommunicationsRepairabilityeld-repairedifatallImpactterminalCausestructuralsteeloftheWTCinterferedwiththewirelesssignalSymptomsfrequentlossofvideo,nosignalfromrobotDescriptionInsteadoftheusualmileormorerangeofthe2watttransmittertherobotwascarrying,communicationwaslostwithin20feet.Therobotwasneverrecovered.De-tracking RobotsMicroVGTV,UrbanStudyOriginalCRASARReliabilitystudy[6],TECO'sUrbanRobotstudy[43]ClasseffectorRepairabilityeld-repairedImpactterminalCausevariesSymptomssignicantdecreaseinmobilityDescriptionAtrackworksitswayoffitswheelsand/orguides.Commonfailurefortrackedplatforms.113

PAGE 125

AppendixBContinuedEmergencystopswitchfailure RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasscontrolsystemRepairabilitynon-eld-repairedImpactterminalCauseunspeciedSymptomsswitchdidnotstoptherobotDescriptionOperatorsweretaughttotesttheemergencyswitchandotherkeysafetyfeaturesbeforeusingtherobotsineldtests.Itwasnotedeachtimethesefeaturesdidnotwork.Failureofarticulatingarmsordrivemotor RobotsUrbanStudyTECO'sUrbanRobotstudy[43]ClasseffectorRepairabilitynon-eld-repairedImpactusuallyterminalCausedirtandothersmalldebrisgetintoeffectors'housingSymptomseffectornolongerrespondsDescriptionOpengearingforthearticulatedarmsandthedrivemotorcollectdebrisuntilthosecomponentsstopworking.114

PAGE 126

AppendixBContinuedGravityassistWTCCategory RobotsMicroVGTV,MicroTracsStudyWTCEngineeringstudy[34]ClassmistakeRepairabilityeld-repairedImpactnon-terminalCauseinclineofgroundsurfacewastoosteepforrobottokeepitselffromslip-pingSymptomstensionontetherrequiredtokeeprobotfromslippingDescriptionThisfailureisamistakewhentheoperatormis-judgedtheverticalorien-tationofthevoid.Heatinducedtrackfailure RobotsMicroVGTVStudyWTCEngineeringstudy[34]ClasseffectorRepairabilityeld-repairedImpactterminalCausetemperatureexceeding122degreesFahrenheitSymptomssignicantdecreaseinmobilityDescriptionThetrackbecamehotenoughtoexpandandsoftenuntilitfelloffofitswheels.Thisfailurecouldbeeld-repairedwithabackuptrack.115

PAGE 127

AppendixBContinuedInsufcientautomaticadjustmenttolightingconditions RobotsARTSStudyTECO'sARTS[59]studyClasssensorRepairabilitynon-eld-repairedImpactdependsonlightingconditionsCausecamera'sautomaticirisdidnotadjustenoughSymptomsdark,unclearcameraviewDescriptionTheirisdidnotadjustenoughfortheoperatortoseetomaneuvertherobot.Jammedjoystick RobotsPackbotStudyOriginalCRASARReliabilitystudy[6]ClassslipRepairabilitynon-eld-repairedImpactterminalifjoystickisrequiredCauseinexperiencedoperatorSymptomsjoystickunresponsiveDescriptionThejoystickwasdroppedinthesandbyarst-timeoperator.116

PAGE 128

AppendixBContinuedLightingincorrectWTCCategory RobotsMicroVGTV,MicroTracsStudyWTCEngineeringstudy[34]ClasssensorRepairabilityeld-repairedImpactterminalwhenaclearcameraviewisrequiredCausefailureofarticiallightingSymptomsdark,unclearcameraviewDescriptionStateinwhichsufcientlightisnotpresent. ClassmistakeCauseautomaticwhite-balancecompensationforarticiallightingDescriptionTheoperatorattemptedtoimprovethecameraviewbyadjustingthear-ticiallighting,butthecamcorderwasautomaticallyadjustingtocom-pensate.117

PAGE 129

AppendixBContinuedLoworun-powered RobotsAllplatformsStudyOriginalCRASARReliabilitystudy[6],TECO'sM1PANTHERIIstudy[13]ClasspowerRepairabilityusuallyeld-repaired,dependsoncauseandplatformImpactterminalifun-powered,impactiflowdependsontheplatformCauselowbattery,fuelorloosebatteryconnectionsSymptomsnopower,looseconnectionscanbedetectedifpowerlosscorrespondswithcertainmotionsDescriptionAloworfailingbatterycancauseawidevarietyofproblems,dependingonthecharacteristicsoftheelectricalsystem.Deadbatteries,orlooseconnectionswillleavetherobotdead.Mostcommonpowerfailure.Malfunctionoftrackmechanism RobotsARTSStudyTECO'sARTSstudy[43]ClasseffectorRepairabilityeld-repairedImpactterminalCauserockskepttrackmechanismfromfunctioningproperlySymptomssignicantdecreaseinmobilityDescriptionRocksbecamestuckintrackguidesandsprockets.118

PAGE 130

AppendixBContinuedMalfunctioninghydraulicsystem RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasseffectorRepairabilitynon-eld-repairedImpactterminalCauseunspeciedSymptomssmokeissuingfromthetank'smodiedturretDescriptionTECOreportedthatthehydraulicsystemfailed.Whenthisoccurredsmokewasseenissuingfromthetank.Navigationerrors RobotsSARGEStudyTECO'sSARGEstudy[44]ClassmistakeRepairabilityeld-repairedImpactnon-terminalCauseinsufcientinformationfromremotesensorsSymptomsoperatoruncertainofthecurrentpositionoftherobotDescriptionPositionanddrivingerrorsduetodifcultyjudgingdistanceandpositionthroughtheteleoperationinterface.119

PAGE 131

AppendixBContinuedObscuredcameraview RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasssensorRepairabilityeld-repairedImpactnon-terminalwherecameraisnotrequiredCauselensescoveredinmoisture,dirt,ormudSymptomsunclearorobscuredcameraimageDescriptionCameralensbecomescoveredinmoisture,dirt,ormud.Afrequentlyencounteredeldfailure,especiallyinwetorrainyenvironments.Occasionalstatic RobotsARTS,URBOT,TalonStudyTECO'sARTS[59],andUrbanRobotstudies[43]ClasscommunicationsRepairabilityeld-repairedImpactnon-terminalCauseunstablewirelessconnectionSymptomsstaticinvideoDescriptionOccasionalstaticdisruptedtheARTSoperators.IntheUrbanRobotstudyadditionalantennaswereusedtoimprovethequalityofthesignal,butthisfailurestilloccurred.120

PAGE 132

AppendixBContinuedOccludedcameraWTCCategory RobotsMicroVGTV,MicroTracsStudyWTCEngineeringstudy[34]ClasssensorRepairabilityeld-repairedImpactterminalCauseobstacleordebrisinfrontofcameraSymptomsoccludedcameraviewDescriptionStateinwhichthecameraviewiscompletelyoccludedbyobstacles ClassmistakeCauserobotplacedorleftinapositionwherethecameraviewisofnouseDescriptionTheoperatordoesnotknowhowtonavigatearoundorawayfromtheobstacleordebrisblockingtherobot'scameraview.Thisclassicationwouldbeusedtodescribeadditionaltimespentinthisstateaftertheoperatorhasdetectedtheproblem.121

PAGE 133

AppendixBContinuedOvertaxedpowersystem RobotsUrbanStudyTECO'sUrbanRobotstudy[43]ClassmistakeRepairabilitynon-eld-repairedImpactterminalCausetryingtouseafrozendrivesystemSymptomsdependsontheplatformDescriptionTheoperatorsdidnotrealizethatthedrivesystemwasfrozenandwouldovertaxthepowersystemattemptingtogettherobottomove.Piniongearstripped RobotsMicroVGTVStudyOriginalCRASARReliabilitystudy[6]ClasseffectorRepairabilitynon-eld-repairedImpactnon-terminalwhereshapeshiftingisnotrequiredCausedirtandothersmalldebrisgetintogear'shousingandcauseprematurewearSymptomsfailuretoshiftthoughshiftmotorcanbeheardrunningDescriptionPiniongearinsidethegeometryshiftingmechanismbecomesstripped.Acommonfailureforthismodel.122

PAGE 134

AppendixBContinuedPuncturedtrackmechanism RobotsMicroTracsStudyWTCEngineeringstudy[34]ClasseffectorRepairabilitynon-eld-repairedImpactterminalCausealuminumrodlodgedintothetrackmechanismSymptomssignicantdecreaseinmobilityDescriptionTherobotwasrestingontheroduntiltheoperatortriedtodrivetherobot.Thetrackmechanismpulledtherodintothethinspacebetweenthetrackandtheplatform.RPMsatcriticallevel RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasscontrolsystemRepairabilitynon-eld-repairedImpactterminalCauseunspeciedSymptomsRPMshighwhilerobotisstationaryDescriptionTheRPMsshotupfornoapparentreason.Oneofseveralproblemsencounteredwiththeteleoperationsysteminstalledonthemodiedtank.123

PAGE 135

AppendixBContinuedRobotplacedinunsafeposition RobotsURBOT,TalonStudyTECO'sUrbanRobotstudy[43]ClassslipRepairabilityeld-repairedImpactnon-terminalintraining,potentiallyterminalinarealscenarioCauseoperatorsremovedfromrobot'soperatingenvironmentSymptomsrobotcasuallyplacedinpositionswhichtheoperatorwouldnotplacethemselvesDescriptionTheoperatorwouldoftendrivetherobotdangerouslyclosetosuspiciousobjectstodetermineiftheyweremines,grenades,oranotherharmfuldevice.Rolledonitsside RobotsARTSStudyTECO'sARTSstudy[59]ClassmistakeRepairabilitynon-eld-repairedImpactterminalCauseplatformbecameunstableon30%slopeSymptomsvisiblethroughvideoDescriptionTheoperatordidnothaveasenseoftheinclineangleofthegroundsurface,andtherobot'sunstablestate.124

PAGE 136

AppendixBContinuedShearpinbroken RobotsMicroVGTVStudyOriginalCRASARReliabilitystudy[6]ClasseffectorRepairabilitynon-eld-repairedImpactnon-terminalwhereshapeshiftingisnotrequiredCauserobotencountersresistancewhileshiftingSymptomsfailuretoshiftthoughshiftmotorcanbeheardrunningDescriptionShearpininsidethegeometryshiftingmechanismbreaks.Acommonfailurewhichoftenoccursinconnedspaces.Spontaneousshutdown RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasscontrolsystemRepairabilityeld-repairedImpactterminalCauseunspeciedSymptomsrobotisshutdownDescriptionShutdownwhentheoperatortriedtoswitchfromautomatictoteleoper-ationmodeonthecontroller.Oneofseveralproblemsencounteredwiththeteleoperationsysteminstalledonthemodiedtank.125

PAGE 137

AppendixBContinuedSteeringlost RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasscontrolsystemRepairabilitynon-eld-repairedImpactterminalCauseunspeciedSymptomsnoresponsetosteeringcommandDescriptionOneofseveralproblemsencounteredwiththeteleoperationsystemin-stalledonthemodiedtank.Sometimesmanifestedinonlyonedirectionatatime.Steeringslowtorespond RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasscontrolsystemRepairabilitynon-eld-repairedImpactnon-terminalCauseunspeciedSymptomsdelayinrobot'sresponsetosteeringcommandDescriptionOneofseveralproblemsencounteredwiththeteleoperationsystemin-stalledonthemodiedtank.Sometimesmanifestedinonlyonedirectionatatime.126

PAGE 138

AppendixBContinuedStuckassistWTCCategory RobotsMicroVGTV,MicroTracsStudyWTCEngineeringstudy[34]ClassmistakeRepairabilityeld-repairedImpactnon-terminalCauseobstaclesorpilesofdebriskeeptherobotfrommovingSymptomstracksturningbuttherobotisnotmovingDescriptionThefailureisamistakewhentheoperatorlosttrackofordidnotseetheobstacles.Throwntrack RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasseffectorRepairabilitynon-eld-repairedImpactterminalCauseunspeciedSymptomssignicantdecreaseinmobilityDescriptionTECOreportedonecaseinwhichthemodiedtankthrewatrack.127

PAGE 139

AppendixBContinuedTrackslippage RobotsD-7G,DEUCEStudyTECO'sD-7[3]andDEUCE[14]studiesClasseffectorRepairabilityeld-repairedImpactnon-terminalCausevariesSymptomstreadsrotatingwithoutacorrespondingchangeinpositionDescriptionTrackscannotmaintainsufcientfrictionwiththegroundsurface.128

PAGE 140

AppendixBContinuedTrackslippageWTCcategory RobotsMicroVGTV,MicroTracsStudyWTCEngineeringstudy[34]ClasseffectorRepairabilityeld-repairedImpactnon-terminalCausevariesSymptomstreadsrotatingwithoutacorrespondingchangeinpositionDescriptionTrackslippageattheWTCwouldbeaneffectorfailurewhenthereshouldhavehadsufcienttractionintheenvironmentalconditionstherobotwasexperiencingatthetimeofthefailure.Thestudydidnotpro-videenoughinformationtodetermineiftheseconditionswereevermetornot. ClassslipCauserobothighcenteredorinalessthanoptimalcongurationDescriptionTheoperatorisnavigatingthroughanareawheretherobotcanbecomehigh-centeredtreadscannottouchthegroundduetouneventerrainorsmallobstaclesorhasdifcultymaintainingsufcientfrictionwiththegroundsurface.Duringthistasktheoperatorerrorsandleavestherobothigh-centeredorusesthewrongcongurationsee[34]formoredetailsontheimpactofaMicroVGTV'scongurationonmobility.129

PAGE 141

AppendixBContinuedTransmittingoverunauthorizedfrequencies RobotsARTSStudyTECO'sARTSstudy[59]ClasscommunicationsRepairabilitynon-eld-repairedImpactnon-terminalCauseimprecisetransmittingequipmentSymptomssignaldetectedinunauthorizedfrequencyDescriptionThetransmitterwouldbleedoverfromauthorizedfrequenciesintounau-thorizedones.Uncontrolledacceleration RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasscontrolsystemRepairabilitynon-eld-repairedImpactterminalCauseunspeciedSymptomsstillacceleratingwhenoperatorisnolongertryingtosteertherobotDescriptionOneofseveralproblemsencounteredwiththeteleoperationsystemin-stalledonthemodiedtank.130

PAGE 142

AppendixBContinuedUnresponsiveminor RobotsMicroVGTV,UrbanStudyOriginalCRASARReliabilitystudy[6]ClasscontrolsystemRepairabilityeld-repairedImpactterminalCauseunknownSymptomscyclingpowerxestheproblemDescriptionTherobotisunresponsiveorfrozenandthereisnoobvioussourceoftheproblem,likeadeadbatteryorblowncomponent.Therobotbeginsrespondingagainafteroneortwopowercycles.Unresponsiveserious RobotsMicroVGTVStudyFollow-upCRASARReliabilitystudy[7]ClasscontrolsystemRepairabilitynon-eld-repairedImpactterminalCauseoverloadintheelectricalsystemontherobotorintheOCUSymptomscyclingpowerdoesnotxtheproblemand/orsmallamountsofsmokeDescriptionTherobotconsistentlyrunsforashortperiodoftimeandthenfails.131

PAGE 143

AppendixBContinuedUnstablecamerasignal RobotsPANTHERStudyTECO'sM1PANTHERIIstudy[13]ClasssensorRepairabilitydependsoncauseImpactnon-terminalwherecameraisnotrequiredCausebumpyterrain,suddenchangesinlighting,rainSymptomsunclear,intermittent,orlostsignalDescriptionIntermittent,unclear,orlostsignalfromacamera.Unstablesensorsignal RobotsAllplatformsStudyOriginalCRASARReliabilitystudy[6]ClasssensorRepairabilitydependsoncauseandplatformImpactnon-terminalwherethesensorisnotrequiredCausefaultycabling,orbrokenorlooseconnectionsSymptomsunclearorintermittentsignal,looseconnectionscanbedetectedifsignaldropoutcorrespondswithcertainmotionsDescriptionIntermittent,unclear,orlostsignalfromasensorwherethesensoritselfisfunctioningproperly.Mostcommonsensorfailure.132

PAGE 144

AppendixBContinuedUnstableteleoperationcontrol RobotsMicroVGTVStudyOriginalCRASARReliabilitystudy[6]ClassslipRepairabilityeld-repairedImpactterminalCausetetherwasnotproperlyconnectedtotheOCUSymptomslaginleftturnsandintermittentcommunicationsdropoutDescriptionTheoperatorthatsetupthesystemdidnotproperlyconnectthetethertotheOCU.133


xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 001469382
003 fts
006 m||||e|||d||||||||
007 cr mnu|||uuuuu
008 040524s2004 flua sbm s000|0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0000283
9 035
AJR1136
b SE
SFE0000283
040
FHM
c FHM
090
QA76
1 100
Carlson, Jennifer.
0 245
Analysis of how mobile robots fail in the field
h [electronic resource] /
by Jennifer Carlson.
260
[Tampa, Fla.] :
University of South Florida,
2004.
502
Thesis (M.S.C.S.)--University of South Florida, 2004.
504
Includes bibliographical references.
516
Text (Electronic thesis) in PDF format.
538
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
500
Title from PDF of title page.
Document formatted into pages; contains 144 pages.
520
ABSTRACT: The considerable risk to human life associated with modern military operations in urban terrain (MOUT) and urban search and rescue (USAR) has led professionals in these domains to explore the use of robots to improve safety. Recent studies on mobile robot use in the field have shown a noticeable lack of reliability in real field conditions. Improving mobile robot reliability for applications such as USAR and MOUT requires an understanding of how mobile robots fail in field environments. This paper provides a detailed investigation of how ground-based mobile robots fail in the field. Forty-four representative examples of failures from 13 studies of mobile robot reliability in the USAR and MOUT domains are gathered, examined, and classified. A novel taxonomy sufficient to cover any failure a ground-based mobile robot may encounter in the field is presented. This classification scheme draws from established standards in the dependability computing [30] and human-computer interaction [40] communities, as well as recent work [6] in the robotics domain. Both physical failures (failures within the robotic system) and human failures are considered. Overall robot reliability in field environments is low with between 6 and 20 hours mean time between failures (MTBF), depending on the criteria used to determine if a failure has occurred. Common issues with existing platforms appear to be the following: unstable control systems, chassis and effectors designed and tested for a narrow range of environmental conditions, limited wireless communication range in urban environments, and insufficient wireless bandwidth. Effectors and the control system are the most common sources of physical failures. Of the human failures examined, slips are more common than mistakes. Two-thirds of the failures examined in [6] and [7] could be repaired in the field. Failures which resulted in the suspension of the robot's task until the repair was completed are also more common with 94% of the failures reported in [13].
590
Adviser: Robin R. Murphy
653
reliability analysis.
field work.
robotics.
meta-study.
fault tolerance.
690
Dissertations, Academic
z USF
x Computer Science
Masters.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.283