Paradigms of A.I. Robotics

flybittencobwebΤεχνίτη Νοημοσύνη και Ρομποτική

2 Νοε 2013 (πριν από 3 χρόνια και 7 μήνες)

69 εμφανίσεις

ParadigmsofA.I.Robotics
WeiweiPan
DepartmentofMathematics
WesleyanUniversity
Middletown,CT06457
ParadigmsofA.I.Roboticsp.1/??
IntelligentRobots
WhatisanIntelligentRobot?
Anintelligentrobotisamechanicalagentoperating
autonomouslyandexiblyinitsenvironment.
￿Arobothasitsowngoalsanddecisionmakingprocesses
￿Arobotiscapableofextractinginformationfromits
environment
￿Arobotisabletoachieveitsgoalsusingsensoryinformation
aboutitsenvironment
￿Arobotcanadapttoapartiallyunknownenvironment
ParadigmsofA.I.Roboticsp.2/??
IntelligentRobots
WhydoRobotsNeedA.I.?
￿Sensoryinterpretation
￿Situationawareness
￿PlaningandProblemSolving
￿Interactionwithotheragents
￿Handleunexpectedevents
￿Learning
ParadigmsofA.I.Roboticsp.3/??
Outline
￿RoboticParadigms
￿TheHierarchalParadigm

RobotProle:
SHAKEY
￿TheReactiveParadigm
⋄TwoArchitecturesofRP

RobotProle:
TOTO(etMETATOTO)

RobotProle:
ROBOSOCCER
￿TheHybrid-ReactiveParadigm
ParadigmsofA.I.Roboticsp.4/??
PrimitiveFunctionsofA.I.Robots
￿Sense
◦Extractinformationfromtheenvironment
￿Plan
◦Determinedirectivesbasedonsensedorcognitive
information
￿Act
◦Usesensedinformationordirectivetoproduce
actuatorcommands
ParadigmsofA.I.Roboticsp.5/??
TheHierarchicalParadigm
Sense
Plan
Act
Thegeneralmethodcyclesthroughthefollowingthreesteps:
￿Sense:takesensorreadingandupdatetheworldmodel
￿Plan:determinethenextsetofactionsbasedonthecurrentworld
modelandgoals
￿Act:executeactionsspeciedbytheplanner
Remark.Hereaworldmodelconsistsofapriorirepresentation,sensedandcognitive
information.
ParadigmsofA.I.Roboticsp.6/??
CharacteristicsofH.P.
TheHierarchicalParadigmischaracterizedbyanexplicitmonolithic
world-viewandasequentialorderingofthethreeroboticprimitive.
￿Knowledgeabouttheenvironmentmustbeexplicitlyrepresented
￿Theworldmodelmustcontaineverythingthatanagentneedsin
ordertoplanitscourseofaction
￿Theowofcontrolbetweenthethreeprimitiveroboticfunctions
denesalinearordering(thisisalsoknownashorizontal
decomposition).
ParadigmsofA.I.Roboticsp.7/??
ShakeytheRobot
￿Developedin1967-1970attheStanford
ResearchInstitute
￿Majorcontributorsinclude:CharlesRosen,
NilsNilssonet.al.
￿Therstmobilerobottoreasonaboutits
actions
￿Keycomponents:camera,opticalrange
nder,bumpsensors
ParadigmsofA.I.Roboticsp.8/??
ProsandConsofH.P.
Advantages:
￿Hierarchical(top-down)structureallowfortheplanningmoduleto
focusallbehaviourstowardsasinglesetofgoals.
Drawbacks:
￿(Closedworldassumption)Relianceonthestaticworldmodel
duringtheS-P-Acyclecausespoorperformanceindynamic
environments.
￿(Frameproblem)Updatingandmaintainingasufcientlydetailed
worldmodelcanbecomputationallyintractable.
￿Searchingforaplantoachieveacertaingoalisequallycomplex.
ParadigmsofA.I.Roboticsp.9/??
LessonsfromBiology
Ethology:
￿Simpleanimalsexhibitavarietyofintelligentbehaviour
￿(InnateReleaseMechanism)Aspecicstimuliactivateasimple,
involuntary(hardwired)response
Cognitivescience:
￿(Embodiment)Directphysicalexperienceisthekeytocognition
￿(Embeddedness)Theenvironmentshapesthecognitiveprocess.
ParadigmsofA.I.Roboticsp.10/??
TheReactiveParadigm
Sense
Act
Areactiveroboticsystemdecomposesfunctionalityintobehaviors,
whichtightlycoupleperceptiontoactionwithouttheuseofintervening
abstractrepresentations
￿Behaviorsareimplementedascircuits(hardware)oraslow
computationalcomplexityalgorithms(software)
￿Nomemoryrequired.Behaviorsarepurestimulus-response
reexes.
ParadigmsofA.I.Roboticsp.11/??
CharacteristicsofR.P.
1.Robotsaresituatedagentsoperatinginanecologicalniche.
￿Therobotisanintegralpartoftheworld
￿Therobothasitsowngoalsandintentions
￿Therobotiscapableofaffectingtheworld
￿Therobotiscapableofsensingtheworldandsenseinformation
affectsitsgoalsandactions.
ParadigmsofA.I.Roboticsp.12/??
CharacteristicsofR.P.
2.Overallbehaviourisanemergentquality.
￿Behavioursarecomputationallyindependentandoperate
concurrently
￿Nocontrollerthatdeterminesallbehavioursforagiventask
￿Overallbehaviourisaresultofinteractionsofindependent
behaviours
ParadigmsofA.I.Roboticsp.13/??
CharacteristicsofR.P.
3.Sensingislocalandbehaviour-specic.
￿Abstractrepresentationalknowledgeinperceptualprocessingis
avoided
￿Representationisexpressedasrobot-centricinformation(noworld
model).
ParadigmsofA.I.Roboticsp.14/??
ArchitecturesofR.P.
1.Subsumption
￿DevelopedbyRodneyBrooksin1986
￿Behavioursarelayeredwithhigherlevelsabletooverride
(subsume)lowerones
￿Behavioursoperateconcurrentlyandindependently(withsome
mechanismtoresolveconicts)
2.PotentialField
￿DevelopedbyRonaldArkinin(?)
￿Basicbehavioursarerepresentedbyelementarypotentialelds
￿Complexbehaviours(complexelds)arethecombinationof
elementaryones
ParadigmsofA.I.Roboticsp.15/??
TheSubsumptionArchitecture
Denition.Abehaviourinthisarchitectureisanetworkofsensingandacting
moduleswhichaccomplishatask.
Denition.Amoduleisanaugmentednitestatemachine(aFSMwithregisters,
timersandenhancementstopermitinterfacewithothermodules).
1.Modulesaregroupedintolayersofcompetencewithlowerlayers
capturingbasicfunctionsandhigherlayerscreatingmore
goal-directedactions.Eachlayercanbeviewedasabehaviour.
ParadigmsofA.I.Roboticsp.16/??
TheSubsumptionArchitecture
2.Behaviourallayersoperateconcurrently.Conictsareresolvedby
havingmodulesinahigherlayersubsumeoutputfrommodules
onelevellower.Subsumptioncanoccurinoneoftwoways:
￿(Inhibition)Theoutputofthehighermoduleisconnectedtothe
outputofthelowermodule.Iftheoutputofthehigheris"on",
thentheoutputofthelowermoduleisblocked.
￿(Suppression)Theoutputofboththehigherandthelower
moduleisconnectedtotheinputofanothermodule.Ifthe
outputofthehighermoduleisnonzerothentheoutputofthe
highermodulereplacesthatofthelowermodule.
ParadigmsofA.I.Roboticsp.17/??
TheSubsumptionArchitecture
3.Persistentinternalrepresentationoftheworldisavoided.The
robotshoulddependmostlyoninformationcomingdirectlyfromits
environment.Evenwhensomeinternalstateisneededfor
triggeringbehaviourslikescaredorhungrysuchusageshouldbe
minimized.
4.Ataskisaccomplishedbyactivatingtheappropriatelayer(lower
layersarethenactivatedbyreactivitytotheenvironment).
ParadigmsofA.I.Roboticsp.18/??
SubsumptionRobots
RobotsbuiltusingthesubsumptionarchitectureattheMITA.I.RoboticsLaboratory:
ParadigmsofA.I.Roboticsp.19/??
FollowingaCorridor
Supposewehavearobotwithsonarspointingindifferentdirections
andtwoactuators,FORWARDandTURN.
Level0:ObstacleAvoidance
￿SONARDmoduleconvertssonarreadingsintoaplotofpolar
coordinates,(r,θ),centeredaroundtherobot.
￿FEELFORCEconvertseachsonarreadingintoarepulsivevector
andoutputstheirsumtoRUNAWAY
￿RUNAWAYpassesnewheadingtoTURN
￿COLLIDElooksatthesonarplotandoutputs"halt"toFORWARD
whennearnessthresholdismet.
ParadigmsofA.I.Roboticsp.20/??
FollowingaCorridor
Level1:Explore
￿WANDERmodulecomputesarandomheadingeverynseconds
passingittoAVOID
￿AVOIDmoduletakeoutputfromFEELFORCEandcombinesit
withnewheadingandpassesittoTURN
Remark.Herewemeetourrstexampleofsubsumption,i.e.TURNtakesinputfrom
bothRUNAWAYinLevel0andAVOIDinLevel1,possiblyatthesametime,sowho
doesitlistento?Theansweris,TURNgivespreferencetoRUNAWAY.Whenthe
outputofRUNAWAYisnonzeroitreplacestheoutputofAVOID.Thisisanexampleof
subsumptionbysuppression.
ParadigmsofA.I.Roboticsp.21/??
FollowingaCorridor
Level2:FollowaCorridor
￿￿LOOKmoduleexaminesthesonarplotanidentiesacorridorand
passestheheadingtowardsthemiddleofthecorridortoSTAYIN
MIDDLE
￿STAYINMIDDLEpassestheheadingtoAVOID(alsocalculates
coursecorrections)
￿INTEGRATEobservesactualmotionbetweenLOOKupdatesand
computesdeviationfromintendedcourse,passingthisdatato
STAYINMIDDLE
Remark.STAYINMIDDLEsubsumesWANDERbysuppression.
ParadigmsofA.I.Roboticsp.22/??
FollowaCorridor
LOOK
middle
STAYIN
MIDDLE
INTEGRATE
offmiddle
WANDER
S
middle
AVOID
force
SONAR
FEEL
FORCE
force
RUN
AWAY
S
modified
heading
TURN
encoders
polar
plot
COLLIDE
halt
FORWARD
heading
ParadigmsofA.I.Roboticsp.23/??
SummaryofSubsumption
￿Behavioursaredenedasanetworktightlycouplingsensingand
acting
￿Behavioursareorganizedintolayers(lowerlayersservebasic
functionsandhigherlayersdirectmoregoal-orientedactions)
￿Higherlayersmayoverridelowerones
￿Perceptionisego-centricanddistributed
￿Nopersistentinternalrepresentationoftheworld.
ParadigmsofA.I.Roboticsp.24/??
TototheRobot
ParadigmsofA.I.Roboticsp.25/??
ThePotentialFieldArchitecture
Treattherobotasapointparticleinaforceeld,motionsoftherobot
aretherebyrepresentedbypathsofaparticlethroughthiseldwith
obstaclesandgoalsactingappropriatelyontheparticleviaattractive
orrepulsiveforces.
￿Foreachbehaviourtherobot"feels"aforce.Obstaclesactas
repulsiveforcesandgoalsasattractiveforces.
￿Complexbehavioursresultfromapplyingthesumofallforcesto
therobot.
￿Theoverallbehaviourofarobotinanenvironmentcanbe
representedbyapotentialeld,wherethevectorassociatedto
eachpointrepresentstheforcetherobotwouldfeelatthatpoint.
ParadigmsofA.I.Roboticsp.26/??
PrimitivePotentialFields
(a)Uniformeld,(b)Perpendiculareld
(c)Attractioneld,(d)Repulsioneld,(e)Tangentialeld
ParadigmsofA.I.Roboticsp.27/??
ObstaclesandGoals
ParadigmsofA.I.Roboticsp.28/??
MovementviaPotentialField
Therobotsmovementsaredeterminedbythefollowingthreecases:
1.Onlythegoalisperceived:thesonarreturnsadistancetothegoal
andavectorofattractioniscomputedbasedonthatdata.
2.Onlytheobstacleisperceived:thesonarreturnsadistancetothe
obstacleandavectorofrepulsioniscomputed.
3.Boththegoalandtheobstacleareperceived:thevectorof
repulsionandattractionissummedforanewforcingpushingthe
robotbothawayfromtheobstacleandtowardsthegoal.
Remark.Eventhoughtherobot'smovementsthroughagivenenvironmentcanbe
visualizedatonceasanarticialpotentialeld,thereisnointernalrepresentation!!!
ParadigmsofA.I.Roboticsp.29/??
MovementviaPotentialField
ParadigmsofA.I.Roboticsp.30/??
MagnitudeProles
Question:Howshouldtherobotmodeltheforceitfeelsfromanobjectinits
environmentasafunctionofdistance?
Answer:Therobotcanchoosefromavarietyofmagnitudeprolesforforcesbeing
exertedbyobstaclesandgoals.
Denition.Thewaythemagnitudeofvectorschangeinthepotentialeld(aroundanobstacleoragoal)is
calledthemagnitudeprole.
Threebasictypesofmagnitudeproles:
￿Constant-theforceexertedbyanobjectisofconstantmagnitudewithinaxed
radiusandiszeroelsewhere.
￿Linear-themagnitudeofforceisinverselyandlinearlyproportionaltothedistance
totheobject.
￿Exponential-themagnitudeofforceisinverselyandexponentiallyproportionalto
thedistancetotheobject.
ParadigmsofA.I.Roboticsp.31/??
MagnitudeProles
(a)LinearDropOff(b)ExponentialDropOff
ParadigmsofA.I.Roboticsp.32/??
GeneratingPotentialFields
Denition.GivenS⊂R
2
,avectoreldonSisafunctionF:S→R
2
.Thatis,F
associatesavectortoeverypointinS.
FACT:Everypotentialeldisavectoreld,andmostpotentialeldsare
conservativevectorelds,i.e.theyare(moreorless)thegradientofpotential
energyfunctions.
Remark.Thinkofpotentialenergyfunctionsasalandscapewithhillsandvalleysandthe
potentialeldisthevectoreldgeneratedbymappingtheinstantaneouschangeineach
directionateverypointinthelandscape.
FACT:Magnitudeprolesdenepotentialenergyfunctions!I.e.Givena
potentialenergyfunction,U:R
2
→R
+
,thepotentialeldassociatedtoUis
F(~r)=−▽U(~r),~r∈R
2
.
ParadigmsofA.I.Roboticsp.33/??
GeneratingPotentialFields
Let~r∈R
2
andp(~r)=k~r−~r
obj
k.Chooseamagnitudeprole,U(~r),foranobject.
￿Anattractiveeld(standard):
U
att
(~r)=
1
2
k
att
p
2
(~r)
then,
F
att
(~r)=−k
att
(~r−~r
goal
)
￿Arepulsiveeld(standard):
U
rep
(~r)=
8
<
:
1
2
k
rep

1
p(~r)

1
r
0

2
ifp(~r)≤r
0
,
0ifp(~r)>r
0
,
then,
F
rep
(~r)=
8
<
:
k
rep

1
p(~r)

1
r
0

1
p
2
(~r)
~r−~r
goal
p(~r)
ifp(~r)≤r
0
,
0ifp(~r)>r
0
,
wherek
att
,k
rep
areconstants,andr
0
istheradiusofinuencefortheobstacle.
ParadigmsofA.I.Roboticsp.34/??
FollowingaCorridor
ParadigmsofA.I.Roboticsp.35/??
FollowingaCorridor
￿RUNAWAYpf:usesrepulsiveeldtogeneratenewheading.Each
sonarreadingactivatesaninstanceofRUNAWAYpf,theresulting
vectorsaresummedandthenpassedtoTURN.
￿COLLIDE:cannotbemappedasabehaviourwiththePFmethod.
Instead,collisionstriggeremergencyresponsesoutsidethe
potentialeldframework.
￿WANDERpf:generatesarandomheadingandauniformeldin
thedirectionofthatheading(theoutputofWANDERpfisnaturally
summedwiththeoutputofRUNAWAYpfviacombinationof
potentialelds).
￿FOLLOWCORRIDOR:isthepotentialeldpicturedinthe
previousslide.
ParadigmsofA.I.Roboticsp.36/??
Robosoccer
ParadigmsofA.I.Roboticsp.37/??
PFRobots
RonArkinandhisrobotsatGeorgiaTech:
ParadigmsofA.I.Roboticsp.38/??
DifcultieswithPotentialFields
Programmingbehavioursviapotentialeldsencountertwocommon
problems:
￿(Localminimaproblem)Ideally,whentherobotfeelsnoforce(i.e.
▽U(~r)=0)itisatthegoalposition(i.e.globalmininmum).
Unfortunately,givenanarbitrarypotentialenergyfunctionUthere
maybeplacesasidefromthegoalpositionwhere▽U(~r)=0
(saddlepoints).Thiscanberesolvedbyaddingasmalldegreeof
random"noise"toheadingvectors.
￿Dependingontheterrainoftheenvironment,badchoicesof
magnitudeprolescanleadtodisastrousbehavioursalongcertain
paths.Thiscanberesolvedbychoosinga"smarter"modelfor
attractiveandrepulsiveforces.
ParadigmsofA.I.Roboticsp.39/??
ComparisonsBetweenArchitectures
￿Subsumption
+Modulardesign(eachbehaviourisindependentlytestable)
-Nottaskable
￿PotentialFields
+Hasrepresentationthatiseasytovisualizeoveralargeregion
+Anyreal-valuedcontinuousfunctiononR
2
canbeusedasa
potentialenergyfunction.
+Isportable
-Localminimaproblem
ParadigmsofA.I.Roboticsp.40/??
SummaryoftheReactiveParadigm
PositiveAttributes:
￿Reactiverobotswithsimplebehaviourscansolverelativelycomplex
problems
￿Thelackofinternalstatessimpliescomputation(reactivesystemsare
fastandinexpensivetobuild)
￿Performwellindynamicenvironments
Drawbacks:
￿Breakingdowndesiredbehaviourintosimplebehavioursisanart
￿Hardtopredictemergentbehaviours
￿Reactiveagentscanonlyaccomplishtasksthatinvolvereexive
behaviours(cannotperformplanningorreasoning).
ParadigmsofA.I.Roboticsp.41/??
TheHybrid-ReactiveParadigm
Plan
Sense
Act
ParadigmsofA.I.Roboticsp.42/??
References
[1]R.R.Murphy,IntroductiontoAIRobotics,Cambridge,
MA:MITPress,2000.
ParadigmsofA.I.Roboticsp.43/??