版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
GTI5GandCloudRobotics
WhitePaper
1
PAGE
6
Confidentiality:ThisdocumentmaycontaininformationthatisconfidentialandaccesstothisdocumentisrestrictedtothepersonslistedintheConfidentialLevel.Thisdocumentmaynotbeused,disclosedorreproduced,inwholeorinpart,withoutthepriorwrittenauthorizationofGTI,andthosesoauthorizedmayonlyusethisdocumentforthepurposeconsistentwiththeauthorization.GTIdisclaimsanyliabilityfortheaccuracyorcompletenessortimelinessoftheinformationcontainedinthisdocument.Theinformationcontainedinthisdocumentmaybesubjecttochangewithoutpriornotice.
DocumentHistory
Date
Meeting#
Version#
RevisionContents
DD-MM-YYYY
NA
DD-MM-YYYY
DD-MM-YYYY
DD-MM-YYYY
DD-MM-YYYY
CloudRobotics:Trends,Technologies,
Communications
Abstract
Cloudrobotsarecontrolledfroma“brain”inthecloud.Thebrain,locatedinadatacenter,makesuseofArtificialIntelligenceandotheradvancedsoftwaretechnologiestodealwithtasksthatintraditionalrobotswereundertakenbyalocal,on-boardcontroller.Comparedtolocalrobots,cloudrobotswillgeneratenewvaluechains,newtechnologies,newarchitectures,newexperiencesandnewbusinessmodels,thiswhitepaperwillexploretheseaspects.
Introduction
Cloudroboticsisarelativelyrecentconcept.Earlyworkdatesbackto2010,whentheEuropeanCommission’sRoboEarthiprojectbegan.Thisaimedtoestablisha“WorldWideWebforrobots”.RoboEarthandlaterprojectssuchasRapyutaiiandRobohowiiiformalizedthebasicconceptandtechnologies,andarestillinfluencingcloudroboticresearchtoday.
Therearethreecoreadvantagesofcloudrobotscomparedtostand-alonerobots:
InformationsharingManycloudrobotscanbecontrolledfromonebrain,andthebraincanaccumulatevisual,verbal,andenvironmentaldatafromallconnectedrobots.Intelligencederivedfromthisdatacanbeusedbyalltherobotscontrolledbythebrain.Aswithothercloudservices,informationcollectedandprocessedoneachrobotwillalwaysbeup-to-dateandbacked-upsafely.Developersalsobenefit,astheycanbuildreusablesolutionsforallcloud-connectedrobots.
OffloadedcomputationSomerobottasksrequiremorecomputationalpowerthanalocalcontrollercaneconomicallydeliver.Offloading
totheclouddata-intensivetaskssuchasvoiceandimagerecognition,voicegeneration,environmentalmappingandmotionplanningwilllowerthehardwarerequirementsandpowerconsumptionofrobots,makingthemlighter,smaller,andcheaper.
CollaborationCloudrobotsdonotneedtoworkalone.Usingthecloudasacommonmedium,tworobotscanworktogethertocarryanobjecttooheavyforone,oragroupofsimpleworkerrobotscanworkwithalocalmap,providedbyaleaderrobotwithcostlysensors.
Figure1:Largescaledatacollectionwithanarrayofrobots(14robotsaresharingexperiencesofmachinelearningforgrasping)[Source:
/2016/03/deep-learning-for-robots-learning-from.html]
DistributedversionofAlphaGoexploited40searchthreads,1202CPUsand176GPUsx,noordinaryrobotcaninstallinside.Butcloudrobotcanmakeuseofit.
Applicationsforcloudrobots
Usingcloudresourcesempowersrobotsandgivesthemnewcapabilitiesinmanyareas:
Intelligentvisualprocessing:imageclassification,targetdetection,imagesegmentation,imagedescription,characterrecognition.
Naturallanguageprocessing:semanticunderstandingbasedondepthlearning,accurateidentificationofuserintent,multi-intentionanalysis,emotionalanalysis.Makesuseofapowerfulbackgroundknowledgebase.
Facialrecognition:facedetectionalgorithmbasedondepthlearning;Inthereal-timevideo
streamtoaccuratelydetecttheface;Anyfacemaskandreal-timedetectionundertheviewingangle;Toovercome:thesideface,halfobscured,blurredface;
Extensionfromcurrentrobotapplications:outdoormapnavigation,indoorpositioningandnavigation,typicalproductidentification,universalitemidentification,environmentalunderstanding,textreading,voiceprompts.
Theapplicationsthatwillemergeforcloudrobotsareofmanykinds;someareemergingnow–othersareatanearlystageofdevelopment.
Logistics
Amazon,Jingdong,S.F.Expressandothercompanieshavedeployedlogisticsrobotsystems.ThewheeledAGV(AutomatedGuidedVehicle)isthemaintypeoflogisticsrobot(thoughlogisticscompaniesarealsotriallingtheuseofaerialdrones).Byconnectingtothecloud,AGVscanachieveunifiedscheduling(whereallAGVsareworkingasasinglesystemformaximumefficiency).Inaddition,AGVscanbeequippedwithmachinevisionsystems,andvideocanbetransmittedtocloud-basedsystemstohandleavarietyofsituationsontheroad.EventuallythiswillresultinAGVscomingoutofcontrolledareastotakeonmorework,includinginpublicplacesfordeliveryofparcelsorfood.
Securityandsurveillance
Inpublicplaces,cloudrobotscanperform24/7securityinspections,replacingsecuritypersonnel.Thecloudrobotwillcollectvideoandstillimagesandsendthemtothepublicsafetycloudforreal-timeidentificationofsuspiciouspeopleand
Personalassistanceandcare
Providingpersonalassistanceandcarefortheelderlyiswidelyconsideredthe“nextbigthing”inrobotics.Thepowerofthecloudmakescarerobotsbehavemorelikehumans.Theycancarryoutreal-timemonitoringofpersonalhealth,helppeoplemoveabout,andcompletehousework.AnexampleofthistypeofrobotisSoftbank’sRomeo.
activity.SuchrobotsarealreadybeingusedatShenzhenairportinChina.
Guidance
Inpublicplacessuchasenterprises,banksandhospitals,robotslikeSoftbank’sPepperarebeingusedtoguidevisitors.TheyarealsobeingusedtodeliverretailservicesbycompaniesincludingNestle,YamadaElectricandMizuhoBank.Cloudrobotscanmakeuseofavastknowledgedatabaseinthecloud,andcommunicateusingnaturallanguage;theycanevenrecogniseandrespondtopeople’sexpressionsusingcloud-AI-basedimageanalysis,toimprovetheuseexperience.
Education,entertainmentandcompanionship
Inrecentyears,theapplicationofmachinevisionandartificialintelligencehasresultedinthedevelopmentofmanyrobotsforeducationandentertainment.ExamplesincludeJibo,AsusZenbo,andSoftbankNao.Theserobotshaveahumanoidappearanceandtheabilitytousenaturallanguage.Theycandownloadcontentfromthecloudtoprovideeducationandentertainmentservices.
Figure2:Cloud-poweredsmartdevicesandcommunicationrobots[Source:Softbank]
Markettrends
Robotscanbecategorisedasindustrialrobotsorservicerobots,accordingtotheiruse.Service
226.2$bn
Accordingtomarketanalyst
robotscanbefurtherdividedintoprofessional
servicesrobotsandpersonalhomeservicerobots.Professionalservicerobotsareusedinthefieldsofmedicine,construction,underwaterengineering,logistics,defenceandsafety.Personalhomeservicerobotsareusedtoundertakehousework,providecompanionshipandpersonalassistance,andarealsousedinotherfields.
companyTractica,thevalueoftheglobalrobotmarketwillgrowfrom$34.1billionin2016to
$226.2billionin2021,withacompoundannualgrowthrate(CAGR)of46%invalueterms.Mostofthegrowthwillbeinthemarketfornon-industrialrobotsIX.
2bnOneofthemajordriversofthismarketgrowthistheagingpopulation.Therearefewerworking-agepeopletotakecareoftheincreasing
numbersoftheelderly.TheUNhasforecastthatby205021%oftheglobalpopulationwillbeovertheageof60–atotalofover2billionpeople.
Robotshavearoletoplayhere.Inaddition,industrialautomationcontinuestodevelopatarapidpace,withinitiativessuchasIndustry4.0inGermanyandMadeinChina2025.
AdvancesintechnologiesincludingArtificialIntelligence,theInternetofThingsandwirelesscommunicationsaremakingrobotsmorecapable.Theycannowidentifytheirsurroundings,calibratetheirposition,plantrajectories,andusenaturalinterfacestointeractwithhumans.Therehavebeenincreasesinthecapabilitiesofrobotsusedinindustry,agriculture,logisticsandeducation.Therapidriseintheuseofdronesisalsoevidenceoftheincreasingcapabilitiesofrobots.
CloudrobotswillsoonbecomethenormCloud-basedAIandconnectivitywillshapethedevelopmentoftherobotmarketsignificantlyin
thenextfewyears.Thesetechniqueshavealreadybeguntochangethewaythatpeopleinteract:technologygiantshavedevelopedAI-basedsystemsthatarebecomingwidelyused.ExamplesincludeGoogleCloudSpeechAPI,AmazonAlexa,BaiduDuer,IBMWatson,AppleSiriandMicrosoftCortana.
12%AccordingtoHuaweiGIV,by2025theuseofmobileconnectivityandartificialintelligencewillresultinrobotpenetrationinthefamilyof12%;intelligentrobotswillchangethefaceofallindustriesinthesamewaythattheautomotiveindustrywastransformativeinthe20thcentury.
GTIcloudroboticsworkinggroupresearchforecaststhatby2020connectedrobotswillaccountfor90%ofallrobots,andabout20millionnewconnectionswillbeneededeveryyeartosupporttheirday-to-dayoperations.
GTIcloudroboticsworkinggrouphasexaminedtheroboticsmarketindetail.Itsworksuggeststhatby2020theproportionofconnectedrobotsgloballywillbe90%,andabout20millionnewconnectionswillbeneededeveryyeartosupporttheirday-to-dayoperations.Figures3-7showprojectionsforsalesofconnectedrobots.
Figure3:Connectedrobotsales2016-2020(million)[Source:GTIcloudroboticsworkinggroup]
Figure4:Connectedlogisticssystemrobots(thousand)[Source:GTIcloudroboticsworkinggroup]
Figure5Connecteddomesticrobots(million)
[Source:GTIcloudroboticsworkinggroup] 8
PAGE
10
Figure6:Connectedentertainmentrobots(million)
[Source:GTIcloudroboticsworkinggroup]
Figure7Connecteddisabledcareassistantrobot(thousand)[Source:GTIcloudroboticsworkinggroup]
Inthenextfewyears,domesticrobotsandrecreationalrobotswilloccupymostoftheshipmentsofconnectedrobots.Withtheincrease
inthecapabilityofrobots,theneedsofindividualsandfamiliesforservicerobotswillcontinuetoincrease.
willing unwilling neither
TURKEY
QATARNETHERLANDS
NORWAYGERMANY
UK
60
45
40
35
30
27
Figure8:WillingnesstouseAIandrobotsforhealthcare[Source:PwC]
Figure9:Willingnesstohavesurgeryperformedbyrobot
[Source:PwC]
Thecurrentpublicacceptanceofroboticservices,especiallymedicalservices,isnothigh.Peopleareskepticalaboutwhetherrobotscanreachthelevelsofskillofhumandoctors.However,inthenextfewyears,withrobots’abilitiesgraduallyimproving,people’sacceptanceofroboticmedicalserviceswillincrease.
ResearchpublishedbytheOpenRoboethicsInitiativeshowsthatthemainexpectationofhomeservicerobotsistocompletehouseworktomakelifeeasier.Inaddition,education,inspectionandsecurityneedsarerelativelystrong.
9%
11%
17%
19%
26%
32%
32%
38%
75%
Other
Fancytoy
Petreplacement
Companionforfamlily Forcoolnessfactor Educationtoolforchild
Homesecurity ExtensionofelectronicdevicesHouseholdchores
Figure10:Reasonsforpurchaseahomerobot[Source:OpenRoboethicsInitiative]
Thecloudroboticsvaluechain
ThevaluechainofcloudroboticsisshowninFigure11.Therobotplatformproviderdeliverstherobotwhichrunsapplications;theseapplicationsuseintelligentservicesfromtheAIprovider,makinguseofthemobilenetworktoprovidea“smart”userexperienceforendusers.
Robot
Platform
Application
Provider
Mobile
Network
AIProvider
Endusers
Figure11:Cloudroboticsvaluechain[Source:GTIcloudroboticsworkinggroup]
Robotplatform–thetechnologiesbehindcloudrobots
Thedefinitionofrobotmayvarybycontext,butageneraldefinitionis“Amechanicalsystemwiththreeelements:controller,sensor,and
effector/actuator”.
Controller
Astherobotgainscomplexityanddemandsbecomemoreadvanced,thecontrollerparthasalsodevelopedandtoday’srobotsareoftencontrolledbyOSorrichmiddleware,suchasROS(
/
),OpenRTM-aist,middlewarecompliantwithObjectManagementGroup(OMG)RoboticTechnologyComponent(RTC)Specificationiv,andNAOqi(OSusedinSoftbank’sPepper).
Incloudrobots,thecontrollerpartisachievedbycoordinationofcloudandlocalsystems.
Sensors
Robotsusemanydifferenttypesofsensorsrelevanttotheirfunction.Themostimportanttypesare:
CamerasandmicrophonesSophisticatedcamerasandmicrophonesarerequiredtosensetheenvironment.Forinstance,Softbank’shuman-sizedcommunicationrobot
Peppervusesa3DcameraandtwoHDcameras(seeFigure12),andfourdirectionalmicrophonestodetectwheresoundsarecomingfromandlocateuser’sposition.
Figure12:MicrophonearrayandtopcamerainPepperrobot[Source:
http://techon.nikkeibp.co.jp/article/COLUMN/201506
23/424503/?P=2]
3Dcamerasareusedtoprovidepositiondetectionandmapping(oftenreferredtoasSLAM(simultaneouslocationandmapping)).Other3Dpositioningsensorsandtechnologiesarealsoused,frominexpensiveproximitysensing,sonarandphotoelectricsensingtomoreaccurateandcostlytechniquessuchasLiDARthatcanbeusedtobuilduphighresolution3Dpicturesacrossawidecoveragearea.
Wirelessnetworksshouldprovidesufficientbandwidthandlatencyperformancetosendsensordatatothecontroller.Astheaccuracyof
thesensorincreases,sodoesthebandwidthrequired.
Approach
Accuracy
Range
DataRate
3Dcamera
Stereotriangulation/structuredlight
Accurate
Middle
2.8Mbit/s(1280*960@16fps
binocular)
Sonar
Sonicwavemeasurement
Proximity
Short
<1kbit/s
Photoelectricsensor
Photoelectricsignal
measurement
Proximity
Short
<1kbit/s
LiDAR
Timeofflight
Accurate
Wide
0.1Mbit/s(4000
samples@10Hz)
Figure13:Imagesensordescriptionandrequirements[Source:GTIcloudroboticsworkinggroup]
Figure14:SLAMprocessvisualizedonRVIZ,visualizationtoolforROS,andasampleofobtainedmapdata[Source:SoftBank]
Gyroscopes,accelerometers,magnetometersandothersensorsThesesensorsenablearobottoknowitsownorientation,rotationandlocation
Sensor/technologyforlocation
Function
InertiaMeasurementUnit(IMU)
Orientationandrotation
Opticalandquantum-basedsensors
Orientationandrotation
Touchsensor
Contactdetection
GPS
Outdoorlocation
Cellularnetworkdata
Indoor/Outdoorlocation
Bluetoothbeacon
Indoorlocation
Ultrasoundsystem
Objectdetection
Effectors/actuators
Mostactuatorsusedforrobotsareelectric,thoughhydraulicandpneumaticactuatorsarealsoused.Eachtypehasadvantagesanddisadvantages(seeFigure15).
Electrical
Hydraulic
Pneumatic
Operatingprinciple
Electricity,electromagneticforce
Pressurechangeinliquids(oil,water)
Compressedgasisusedtopowerthesystem
Formfactor
Motors(DC,AC,geared,directdriveetc.)andcontrolcircuits
Cylinder,fluidmotor
Cylinder,pneumaticartificialmuscles(PAM)
Advantages
Easytostoreanddistributeelectricenergy,highcontrolflexibility,lowcost
Quickmovementsandgreatforce
Cleanerthanhydraulic,easyinstallation,lightweight
Disadvantages
Producedtorquesaresmallerthanhydraulicorpneumatic
Requirepump,liquidcancausecontamination,difficulttocontrolprecisely
Requirecompressor,lessforceandslowerspeedthanhydraulicduetocompressibility
Figure15:Comparisonofrobotactuatorprinciples[Source:Softbank]
Newdevelopmentsinmobility
Robotplatformneedstoevolve.Robotsneedtohavelongeruptime,highermobilityandrange,thecapabilitytounderstandtheirsurroundings,andtocarryoutsimultaneouslocalizationandmapping(SLAM).
Oneapproachtoachievehighermobility,especiallyinroughterrain,ortodealwithstairsanddoors,istheuseofbipedalorquadrupedsystem.Butcontinuousbalancingisrequiredinthesesystemsandthisrequiresgreaterpower,andtherearesomesafetyconcerns.Safetyrulesforrobotsmayvarybycountryandlocalarea:onepossiblearrangementmaybetotreatrobotsaspedestrians,ormobilityscooters.Speedlimitsand
Mobilenetworksupport
5GOverview
5Gisthenextgenerationofmobilecommunicationtechnology.Itisexpectedtobedefinedbytheendofthisdecadeandtobewidelydeployedintheearlyyearsofthenextdecade.
Thekeycapabilityof5Gisthepeakrateofmorethan10Gbit/s,1millionconnectionspersquarekilometer,andlessthan1msend-to-enddelay.Threeapplicationscenariosfor5Ghavebeendefined:eMBB(EnhancedMobileBroadband),mMTC(MassiveMachineTypeCommunications),andURLLC(Ultra-reliableandLow-latencyCommunications).
remotemonitoringmayberequired(perhapsnotasstrictaswithautonomousvehicles).SafetystandardsthatalreadyapplytorobotsincludeISO13482;otherrelevantstandardsarethosecoveringhomeelectricalappliancesandradiowavetransmitters.
Amorepracticalapproachthanarobotwithlegsisawheeledrobotequippedwith3Dcamerasandrangesystems,asdescribedabove.Anotherapproachisthe“wearable”robot–suchas
CloudMinds’Metaheadset,whichprovidessophisticatedvisualrecognition,SLAM,anddirectionindicationusingvibration.
Todeliverservicesforthesethreescenarios,theconceptofnetworkslicinghasbeendeveloped.Itisexpectedtoimprovetheoperationofcommunicationnetworks.Thisconceptessentiallyconsistsincreatingdifferentinstancesofnetworktechnologiessuitablefordifferentapplicationswithdifferentrequirements.Suchadynamicandflexiblecommunicationnetworkparadigmwillbeenabledbyanewcloud-basednetworkarchitecture,encompassingSoftwareDefinedNetworking(SDN)andNetworkFunctionVirtualization(NFV).
Figure16:5Gcloudarchitecturetosupportmultipleapplications[Source:HuaweiXLabs]
5Gwillmeetthenetworkrequirementforcloudrobotics
Incloudrobotics,fourtypesofbasicconnectionareneeded:
Monitoringandstatusreporting–therobotuploadingdataaboutitsstatustothecloudbrain
Real-timecontrol–mission-criticalcontrolsignalstotelltherobotwhattodo
Videoandvoiceprocessing–tousepowerfulcloudresourcestohelptherobotunderstanditsenvironment,andtointeractwithusers
Softwareandservicesdownload–for
updatingtherobot’ssoftware,ordownloadingusercontentsuchasmapsoreducationalmaterial.Figure17showstherequirementsofthoseconnectiontypes.
Bandwidth
Latency
Reliability(%uptime)
Summary
Monitoringandstatusreporting
Uplink:1kbit/s
1s
99.9%
Highconnection
density
Real-timecontrol
Downlink:10kbit/s
20ms
99.999%
Lowlatency
Videoandvoiceprocessing
Uplink:3.3Mbit/s(1080p/H.264/30fps)
20ms
99.9%
Highuplinkbandwidthand
lowlatency
Softwareandservicesdownload
Downlink:10Mbit/s
100ms
99.9%
Highdownlink
bandwidth
Figure17:Robotnetworkrequirementanalysis[Source:HuaweiXLabs]
Figure18characterisesthenetworkrequirementsoffullycloudifiedversionsofcurrentrobottypes.Existingnetworkswillfinditdifficulttosupport
newrobotapplications,but5G’shighbandwidth,lowlatencyandhighreliabilitycanproviderobustsupportforfuturerobotapplications.
Figure18:Networkrequirementsforcloudrobotapplications[Source:HuaweiXLabs]
5Gnetworkslicingandmobileedgecomputingarewellsuitedforcloudroboticsapplications
Networkslicesthathavedifferentspecificperformancecharacteristicscanmatchtherequirementsofcloudrobotics,matchtheneedsforpowerconsumptionattherobotterminal,andprovideappropriateroaming.Usingtheseapproaches,5Gnetworkswillalsobeabletomeetthemostdemandingrequirementsintermsofbandwidth,latencyandsecurity.
Mobileedgecomputing(MEC)providesappropriatenetworkandothercomputingandstorageresourceslocatedatthemostappropriatepointtomeetthecloudroboticsapplicationrequirements.Byplacingresourcesclosertothe
user,networklatencycanbereduced.MECsolutionsmaybedeployedwiththeMECserverdeployedatagatewayorinthebasestation,providinglocalcontentcache,wirelessawareness-basedbusinessoptimization,localcontentforwarding,andnetworkcapability.Securityisalsoenhancedasmoredataisretainedclosertotheuseranddoesnottraversethecorenetwork.Forcloudrobotics,thecloserAIresourcescanbedeployedtotheenduserthelowerthelatency.
Softwarecontrolofvirtualizedresourcesthroughoutthenetworkwillensurethattheoptimumbalanceisachievedbetweenuseofcentralizedcloudresourcesanduseofmorelocaledge-basedresources,dependingonthelatencyrequirements.
Figure19:Cloudrobotfunctiondeploymentaccordingtolatency
AIprovider–deliveringcloudAIandMachineLearning
AI,MLandDL
AddingthepowerofcloudcomputingtoroboticswillenableArtificialIntelligence(AI),MachineLearning(ML)andDeepLearning(DL)tobeappliedtoabroadsetofnewapplicationswhererobotswillbeverymuchmorecapable,powerfulandintelligentthanbefore.Thiswillinturnaffectindustriesrangingfromsecuritytomanufacturing.DefinitionsofAI,MLandDLarenotuniversallyagreed,butinthispaper:
ArtificialIntelligenceisacomputersystemabletoperformtasksnormallyrequiringhumanintelligence(includingvisualperception,speechrecognition,decision-makingandtranslation)
MachineLearningistheuseofalgorithmsandmethodssuchasdecisiontrees,neural
networksandcase-basedreasoningtoimproveperformancethroughtraining
DeepLearningreferstotheuseofmulti-layeredartificialneuralnetworksthatenablethetrainingtobecarriedoutonahugescale,withtheresultthatdecisionsareverymuchbetter.
Theseconceptsenablerobotstobetaughttodoatask–andtolearnhowtoimprove–ratherthansimplyrespondingtoaprograminacontrolsystem.Machinelearningalgorithmsofvariouskindshelpcomputerstointerpretdataandmakedecisionsbasedonthedata.Theycanbetrainedtounderstandwhentheirdecisionsarerightorwrongsothattheirdecisionsgetbetterovertime.
Usingmachineordeeplearning,arobotcanbecomebetterabletocompleteatask,ortoundertakeanewone,throughanimprovedawarenessofitsenvironmentandthecontextofthetask.Theseapproacheswillalsoreducetheneed–andcost–toprogramrobotsforeachnewtask.Thisinturnopenstheprospectsformoreflexibleindustrialrobotsthatcancopewithchangesinfactoryconfigurationsandshorterproductionruns,andcapableofoptimizingtheprocessesthattheyarerequiredtoperform.Innon-industrialsettings,AIandmachinelearning
enableimagesandspokenwordstobeinterpretedandforrobotstorespondappropriately.Accesstothecomputingpowerrequiredformachineanddeeplearningisgreatlyenhancedthroughhigh-speednetworksandtheuseofcloudresources.
Thekeyareasinwhichthesetechnologieswillbeappliedincloudroboticsareinintelligentvisualprocessingforarealearningandautonavigation,facerecognition,andnatural-language(speech)processing.Theserequiredataprocessingpowerbeyondthatwhichissensiblybuiltintoarobotlocally.
Accesstothecomputingpowerrequiredformachineanddeeplearningisgreatlyenhancedthroughhigh-speednetworksandtheuseofcloudresources
28%
26%
16%
12%
7%
3.60%
3%
2010201120122013201420152016
Figure20:ImageNetLargeScaleVisualRecognitionChallenge(ILSVRC)errorrateofclassification(%)[Source:ImageNet]
Thankstotheimprovementincomputingpowerandthecontinuousimprovementofalgorithms,AItechnologyisprogressingrapidly.Asanexample,theerrorrateforobjectclassificationrecognitionintheannualImageNetcontesthasbeenreducedtolessthan3%.Itisworthnotingthatthecurrentvisualrecognitionhasnotyetreachedthecorrectrecognitionrateof100%,whichforhighsecurity
Figure21:objectclassification
applications,suchasautonomousvehicles,isstillachallenge
IntheImageNet2015competition,NVDIAandIBMprovidedtheparticipantswithacloudGPU(NVDIAK80s),demonstratingthefeasibilityofcloudAI.
FrombigdatatechnologystacktoAIstackDeeplearninghasemergedasthebestwaytoperformimageanalysis,inapplicationssuchasmedicalradiography,aswellasinlow-latency
applicationssuchasremovalofstreamingvideocontentthatisinbreachofpolicies.ThebiggestITcompaniessuchasBaidu,GoogleandFacebookhavecreatedspecializedAIinfrastructuretohandleAIusecases,butmanycompaniesdonothavethein-houseexpertiseorresourcestoexploit
thenewtechnologies.Anewbackendinfrastructureisrequired,andthiswillbeachievedwiththeuseofnewacceleratorchipssuchasGPUs(graphicsprocessorunits).Butasthesetechnologiesrequiremoreprocessingpower,puttinginfrastructureintopracticeishardandCIOswillneedtobecomemorefamiliarwiththesenewtrends.
Currently,companies’ITarchitecturesaredesignedtomakeuseoffaulttolerant,lowcoststoragethatallowsforeasyextensionofresourceclustersandcanmitigateequipmentfailure.ButAIrequiresthatbigdataanalyticssoftwareunderstandsbetterhowtoruncompute
workloadsbytakingfulladvantageofthesenewaccelerators.BigdatatechnologystackswillshifttoAIstacksthatwillallowenterprisestocapturemorevaluefromdatathatiscapturedbysensorsonrobotsandelsewhere.
AIPlatform
BigDataCloud
DataCollection&Connectivity
ImplementingcloudAIrequires:verylargestoragecapacityandcomputingpower,forward-lookinginvestmenttoattractsoftwaredevelopment,APIsandopensourcelibraries.HardwareneedstomovefromCPUstoGPUsorevenAIdedicatedprocessors
Robotsrequireslargedatacloudstostorethedata.Therearegreaterdemandsonsecurity,anonymityanddistributioncapacitythaninthepast.
HundredsofmillionsofconnectedsensorsarerequiredtocollecttrainingAIdatafromhumans,assetsandtheenvironment.
Figure22:AIstack[Source:GTIcloudroboticsworkinggroup]
Lowlatencyiscriticalforrobotexperience100msThedelayofthehumanneuralnetworkis100ms,andiftherobotcanrespondwithinthisdelay,itcanbeconsidered"seamless".Toachieve"seamless"robotresponsecapabilitiesneedsvideocapture,videocoding,networ
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 资金管理与优化实践总结
- 广西河池市环江县2022-2023学年六年级上学期英语期末试卷
- 《演讲中的自我介绍》课件
- 2025年山西省、陕西省、宁夏、青海省八省联考高考地理模拟试卷
- 2023年广西壮族自治区柳州市公开招聘警务辅助人员辅警笔试自考题1卷含答案
- 2024年山西省朔州市公开招聘警务辅助人员辅警笔试自考题2卷含答案
- 《全身麻醉》课件
- 机电部的口号和目标
- 辽宁省本溪市(2024年-2025年小学六年级语文)统编版综合练习((上下)学期)试卷及答案
- 《慢阻肺健康大课堂》课件
- 子长市长征文化运动公园项目社会稳定风险评估报告
- 浙教版七年级科学上册期末综合素质检测含答案
- 2024年北京市离婚协议书样本
- 2019年海南省公务员考试申论真题(乙类)
- 北京邮电大学《操作系统》2022-2023学年期末试卷
- 2024-2025学年人教版高二上学期期末英语试题及解答参考
- 2023年税收基础知识考试试题库和答案解析
- 热气球项目可行性实施报告
- 双向进入交叉任职制度
- 合成纤维的熔融纺丝工艺研究考核试卷
- 管道改造施工方案
评论
0/150
提交评论