A Definition of Abstraction

sophisticatedroboticsaccessible∗

DouglasBlank

HollyYanco

ComputerScienceComputerScienceBrynMawrCollegeUniv.ofMass.LowellBrynMawr,PA19010Lowell,MA01854dblank@cs.brynmawr.edu

holly@cs.uml.edu

Abstract

Aseducators,weareoftenfacedwiththeparadoxofhavingtocreatesimplifiedexamplesinordertodemon-stratecomplicatedideas.Thetrickisinfindingtherightkindsofsimplifications—onesthatwillscaleuptothefullrangeofpossiblecomplexitiesweeventuallywouldlikeourstudentstotackle.Inthispaper,wearguethatlow-costrobotshavebeenausefulfirststep,butarenowbecomingadead-endbecausetheydonotallowourstu-dentstoexploremoresophisticatedroboticsmethods.Wesuggestthatitistimetoshiftourfocusfromlow-costrobotstocreatingsoftwaretoolswiththerightkindsofabstractionsthatwillmakeiteasierforourstudentstolearnthefundamentalissuesrelevanttorobotprogram-ming.WedescribeaprogrammingframeworkcalledPyrowhichprovidesasetofabstractionsthatallowsstu-dentstowriteplatform-independentrobotprograms.

Introduction

TheKarel-the-robotenvironmentwasdesignedtointro-ducestructuredimperativeprogrammingtobeginningprogrammingstudents(RichardE.Pattis1981).Inasimilarway,inexpensiverobotshavemadeintroductoryAItopicsaccessibletoawiderangeofstudents,fromK-12tothecollegelevel.Theavailabilityoflow-costrobotshasledtotheirwidespreaduseintheundergrad-uateartificialintelligencecurriculum(Meeden1996;Turneretal.1996;Kumar&Meeden1998;Beer,Chiel,&Drushel1999;Harlan,Levine,&McClarigan2001;Wolz2001;Gallagher&Perretta2002;Klassner2002).Althoughthistrendhasbeenatremendoushelpinbringingroboticstostudents,webelievetheselow-costrobotplatformsoftenleadtoaroboticsdead-end,muchthesamewaythatoverrelianceontheKarelenviron-mentdidtoadvancedprogrammingparadigms.Whilelow-costrobots,liketheKarelenvironment,provideawonderfulmotivationandagreatstartingpoint,the

∗ThisworkwassupportedinpartbyNSFCCLIgrantDUE0231363.

DeepakKumar

LisaMeeden

ComputerScienceComputerScienceBrynMawrCollegeSwarthmoreCollegeBrynMawr,PA19010Swarthmore,PA19081dkumar@cs.brynmawr.edu

meeden@cs.swarthmore.edu

paradoxisthattheyoftentrapthestudentinasingleparadigm,orworse,asinglehardwareplatform.

Thereareseveralproblemswiththeuseoflow-costrobotsineducation.Thefirstproblemisthateveryrobotplatformcomeswithitsown,oftenproprietary,devel-opmenttoolsthataresubstantiallydifferentfromotherplatforms.Oftentheprimaryprogramminglanguagesusedaredifferentaswell.Moreproblematicmaybeacompletechangeinparadigmfromonerobottoanother.Consequently,evenifoneweretoinvestinlearningtouseonerobotplatform,probablynoneofthecode,andpossiblylittleoftheknowledgewouldtransfertoadif-ferentplatform.Thissituationisperhapssimilartotheoneintheearlydaysofdigitalcomputerswheneverycomputerhadadifferentarchitecture,adifferentassem-blylanguage,andevenadifferentwayofstoringthemostbasickindsofinformation.

Secondly,webelievethatmanyrobotprogrammingparadigmsdonoteasilysupportmoresophisticatedsen-sors.Forexample,low-costrobotsoftenonlycomeequippedwithinfraredrangesensors.Somecanbeexpandedtoincludesonarrangesensorsorevenlaserrangesensors.However,wesuspectthatifevenmoresophisticatedsensorsweretobecomeaffordable,wewouldbeunabletoutilizethembecausethereisnoeasywayofintegratingthemintoexistingrobotsoft-wareparadigms.Thatis,sophisticatedsensorsmaybehardwareaccessible,butnotconceptuallyaccessiblebythestudent.Theproblemisthattheframeworkdoesn’tpedagogicallyscalewell.

WebelievethattheproliferateduseofrobotsinAIeducationwillresultnotfromlow-costhardwareplat-forms,butfromaccessibilityoftheseplatformsviacommonconceptualfoundationsthatwouldmakepro-grammingthemuniformandconsistent.Ourpositionisdefinedmorebystrivingforalowerlearningcostofroboticswithoutsacrificingthesophisticationofad-vancedcontrollers.

Ourgoalistoreducethecostoflearningtoprogramrobotsbycreatinguniformconceptualizationsthatare

independentofspecificrobotplatformsandincorpo-ratetheminanalreadyfamiliarprogrammingparadigm.Conceptualizinguniformrobotcapabilitiespresentsthebiggestchallenge:HowcanthesameconceptualizationapplytodifferentrobotswithdifferentcapabilitiesanddifferentprogrammingAPI’s?Ourapproach,whichhasbeensuccessfultodate,hasbeenshowntoworkonsev-eralrobotplatforms,fromthemost-expensiveresearch-orientedrobot,tothelowest-costLEGO-basedones.Wearestrivingforthe“write-once/run-anywhere”idea:robotprograms,oncewritten,canbeusedtodrivevastlydifferentrobotswithoutmakinganychangesinthecode.Thisapproachleadsthestudentstoconcen-tratemoreonthemodelingofrobot“brains”byallow-ingthemtoignoretheintricaciesofspecificrobothard-ware.Moreimportantly,wehopethatthiswillallowstudentstograduallymovetomoreandmoresophis-ticatedsensorsandcontrollers.Inourexperience,thismoregeneralizedframeworkhasresultedinabetterin-tegrationofrobot-basedlaboratoryexercisesintheAIcurriculum.Itisnotonlyaccessibletobeginners,butisalsousableasaresearchenvironmentforrobot-basedmodeling.

ThePyroFramework

Wehavebeendevelopingarobotprogrammingframe-workthatwecallPyroforPythonRobotics(Blank,Meeden,&Kumar2003).Asthenamesuggests,mostoftheframeworkiswritteninPython.Pythonisaneasy-to-readscriptinglanguagethatlooksverysimilartopseudocode.Pythonfitsverywellwithourgoalsinthatitiseasyforbeginningstudentstolearn,andyetitalsosupportsmanyadvancedprogrammingparadigms,suchasobject-orientedandfunctionalprogrammingstyles.ItalsointegrateseasilywithCandC++codewhichmakesitpossibletoquicklyincorporateexistingcode.TheC/C++interfacealsoletsoneputveryex-pensiveroutines(likevisionprograms)atlowerlevelsforfasterruntimeefficiency.OneinterestingreasonforusingPyroisthattheentiresoftware,fromtheOpenGLinterfacetotheneuralnetworkcode,canbeexploredbythestudent.Inaddition,advancedstudentscancopythecodeintotheirownspaceandchangeanythingthatintereststhem.

WehavealsousedPyrowithbeginningprogrammersandnon-programmers.Forexample,inanintroductiontocognitivesciencecourse,PyrocanbeusedlikeonewoulduseLEGO-basedrobots.However,thestudentsneednotlearnanewinterfaceastheyexploreothercon-trolparadigms,suchasfuzzylogic,neuralnetworks,orgeneticalgorithms.

Inadditiontotheunifiedframework,wehavecreatedsimpleabstractionsthatmakethewritingofbasicrobotbehaviorsindependentofthesize,weight,andshapeof

arobot.Considerwritingarobotcontrollerforobstacleavoidancethatwouldworkona24-inchdiameter,50-poundPioneer2robotaswellasona2.5-inchdiameter,3-ounceKhepera.Thiswasmadefeasiblebymakingthefollowingabstractions:

RangeSensors:Regardlessofthekindofhardwareused(IR,sonar,laser)sensorsarecategorizedasrangesensors.Sensorsthatproviderangeinforma-tioncanthusbeabstractedandusedinacontrolpro-gram.

RobotUnits:Distanceinformationprovidedbyrangesensorsvariesdependingonthekindofsensorsused.Somesensorsprovidespecificrangeinformation,likedistancetoanobstacleinmetersormillimeters.Otherssimplyprovideanumericvaluewherelargervaluescorrespondtoopenspaceandsmallervaluesimplynearbyobstacles.Inourabstractions,inaddi-tiontothedefaultunitsprovidedbythesensors,wehaveintroducedanewmeasure,arobotunit:1robotunitisequivalenttothediameterof1robot,whateveritmaybe.SensorGroups:Robotmorphologies(shapes)varyfromrobottorobot.Thisalsoaffectsthewaysen-sors,especiallyrangesensors,areplacedonarobot’sbody.Additionally,thenumberofsensorspresentalsovariesfromplatformtoplatform.Forexample,aPioneer2has16sonarrangesensorswhileaKheperahas8IRrangesensors.Inordertorelieveaprogram-merfromtheburdenofkeepingtrackofthenum-berofsensors(andauniquenumberingscheme),wehavecreatedsensorgroups:front,left,front-left,etc.Thus,aprogrammercansimplyqueryarobottore-portitsfront-leftsensorsinrobotunits.Thevaluesreportedwillworkeffectivelyonanyrobot,ofanysize,withanykindofrangesensor,yetwillbescaledtothespecificrobotbeingused.MotionControl:Regardlessofthekindofdrivemechanismavailableonarobot,fromaprogram-mer’sperspective,arobotshouldbeabletomovefor-ward,backward,turn,and/orperformacombinationofthesemotions(likemoveforwardwhileturningleft).Wehavecreatedthreemotioncontrolabstrac-tions:translate,rotate,andmove.Thelattersub-sumesbothtranslateandrotateandcanbeusedtospecifyacombinationoftranslationandrotation.Asinthecaseofrangesensorabstractions,thevaluesgiventothesecommandsareindependentofthespe-cificvaluesexpectedbytheactualmotordrivers.Aprogrammeronlyspecifiesvaluesinarange-1.0..1.0(seeexamplesbelow).Services:Theabstractionspresentedaboveprovideabasic,yetimportantfunctionality.Werecognizethattherecanbeseveralotherdevicesthatcanbepresent

frompyro.brainimportBrain

classAvoid(Brain):

defwander(self,minSide):

robot=self.getRobot()

#ifapproachinganobstacleontheleftside,turnright

ifrobot.get(’range’,’value’,’front-left’,’minval’)robot.move(0,-0.3)

#ifapproachinganobstacleontherightside,turnleft

elifrobot.get(’range’,’value’,’front-right’,’minval’)robot.move(0,0.3)#elsegoforwardelse:

robot.move(0.5,0)

defstep(self):

self.wander(1)defINIT(engine):

returnAvoid(’Avoid’,engine)

Figure1:AnobstacleavoidanceprograminPyro

onarobot:agripper,acamera,etc.Wehavede-visedaserviceabstractiontoaccommodateanynewdevicesoradhocprogramsthatmaybeusedinrobotcontrol.Forexample,acameracanbeaccessedbyaservicethatenablesaccesstothefeaturesofthecamera.Further,studentscanexplorevisionprocess-ingbydynamicallyandinteractivelysequencingandcombiningfilters.

Inthefollowingsectionweexploreanexamplethatutilizestheseabstractionsanddemonstratestheeffec-tivenessoftheseabstractionsinwritinggenericrobotcontrollers.

scribedintheprevioussection.Itiswritteninanobject-orientedstyle,andcreatesaclasscalledAvoidwhichinheritsfromaPyroclasscalledBrain.EveryPyrobrainisexpectedtohaveastepmethodwhichisex-ecutedoneverycontrolcycle.Thebrainshownwillcausetherobottocontinuallywanderandavoidobsta-clesuntiltheprogramisterminated.

ItisnotimportanttounderstandallthedetailsofPyroimplementation,butthereadershouldnoticethattheentirecontrolprogramisindependentofthekindofrobotandthekindofrangesensorbeingused.Thepro-gramwillavoidobstacleswhentheyarewithin1robotunitoftherobot’sfrontleftorfrontrightrangesensors,regardlessofthekindofrobot.

Afterlearningaboutdirectcontrol,studentscanmovetoanyoftheothercontrolparadigms.TheparadigmsselectedwoulddependuponthecoursethatPyrowasbeingusedfor.Inacoursethatempha-sizedrobotics,thenextparadigmwouldmostlikelybebehavior-basedcontrol.AnAIormachinelearn-ingcoursewouldlikelyskipbehavior-basedcontrolandmoveimmediatelytoneural-network-basedcontrol.Currently,thefollowingmodulesareimplementedandextensivecourse-stylematerialsareavailable:di-rectcontrol,sequencingcontrol,behavior-basedcon-trol,neuralnetwork-basedlearningandcontrol,self-organizingmapsandothervectorquantizingalgo-rithms,computervision,evolutionaryalgorithms,andmulti-robotcontrol.Otherparadigmsandmodulesareplannedinthefuture.Thesewillincludelogic-basedreasoningandacting,classicalplanning,pathplanningandnavigation.Pyroisanopen-source,freesoftware

AnExample

Inthissection,we’llusetheexampleofavoidingob-staclestodemonstratetheunifiedframeworkthatPyroprovidesforusingthesamecontrolprogramacrossmanydifferentrobotplatforms.

Directcontrolisnormallythefirstcontrolmethodin-troducedtostudents.Itisthesimplestapproachbe-causesensorvaluesareusedtodirectlyaffectmotoroutputs.Forexample,thefollowingpseudocoderep-resentsaverysimplealgorithmforavoidingobstacles.

ifapproachinganobstacle

ontheleftside,turnrightifapproachinganobstacle

ontherightside,turnleftelsegoforward

TheprogramshowninFigure1implementsthepseudocodealgorithmaboveusingtheabstractionsde-

project,andwehopetogetcontributionsfromotherin-terestedusers.

Conclusions

Wehavearguedthatitismoreimportanttostriveforeasilylearnablerobotprogramminginterfacesthanforlow-costrobotplatforms.WehavetriedtoavoidtheKarel-the-robotparadoxbycarefullydesigningusefulanduniversalconceptualizations.Theseconceptualiza-tionsnotonlymaketherobotprogramsmoreversatile,theyalsohelpinroboticsresearch.Specifically,themodelingofrobotbehaviorscannowbetestedonsev-eralrobotplatformswithouthavingtochangethepro-grams.Thisaddsmuchcredibilitytothetestedmodelsastheresultswillhavebeenconfirmedonseveralrobotplatforms.

Webelievethatthecurrentstate-of-the-artinrobotprogrammingisanalogoustotheeraofearlydigitalcomputerswheneachmanufacturersupporteddifferentarchitecturesandprogramminglanguages.Regardlessofwhetheracomputerisconnectedtoanink-jetprinteroralaserprinter,acomputertodayiscapableofprintingonanyprinterdevicebecausedevicedriversareinte-gratedintothesystem.Similarly,weoughttostriveforintegrateddevicesonrobots.Obviouslywe’renotthereyet.Ourattemptsatdiscoveringusefulabstractionsareafirstandpromisingstepinthisdirection.Webelievethatdiscoveriesofgenericrobotabstractionswill,inthelongrun,leadtoamuchmorewidespreaduseofrobotsineducationandwillprovideaccesstorobotstoanevenwiderrangeofstudents.

Acknowledgments

Pyrosourcecode,documentationandtutorialsareavail-ableatwww.PyroRobotics.org.ThisworkisfundedinpartbyNSFCCLIGrantDUE0231363.

References

Beer,R.D.;Chiel,H.J.;andDrushel,R.F.1999.Us-ingAutonomousRoboticstoTeachScienceandEngi-neering.CommunicationsoftheACM.

Blank,D.;Meeden,L.;andKumar,D.2003.Pythonrobotics:anenvironmentforexploringroboticsbe-yondlegos.InProceedingsofthe34thSIGCSEtechni-calsymposiumonComputerscienceeducation,317–321.ACMPress.

Gallagher,J.C.,andPerretta,S.2002.WWWAu-tonomousRobotics:EnablingWideAreaAccesstoaComputerEngineeringPracticum.ProceedingsoftheThirty-thirdSIGCSETechnicalSymposiumonCom-puterScienceEducation34(1):13–17.

Harlan,R.M.;Levine,D.B.;andMcClarigan,S.2001.TheKheperaRobotandthekRobotClass:APlatformforIntroducingRoboticsintheUndergrad-uateCurriculum.ProceedingsoftheThirty-secondSIGCSETechnicalSymposiumonComputerScienceEducation33(1):105–109.

Klassner,F.2002.ACaseStudyofLEGOMindstormsSuitabilityforArtificialIntelligenceandRoboticsCoursesattheCollegeLevel.ProceedingsoftheThirty-thirdSIGCSETechnicalSymposiumonCom-puterScienceEducation34(1):8–12.

Kumar,D.,andMeeden,L.1998.ARobotLabora-toryforTeachingArtificialIntelligence.ProceedingsoftheTwenty-ninthSIGCSETechnicalSymposiumonComputerScienceEducation30(1).

Meeden,L.1996.UsingRobotsAsIntroductiontoComputerScience.InStewman,J.H.,ed.,Proceed-ingsoftheNinthFloridaArtificialIntelligenceRe-searchSymposium(FLAIRS),473–477.FloridaAIRe-searchSociety.

RichardE.Pattis.1981.KareltheRobot.JohnWileyandSons,Inc.

Turner,C.;Ford,K.;Dobbs,S.;andSuri,N.1996.Robotsintheclassroom.InStewman,J.H.,ed.,Pro-ceedingsoftheNinthFloridaArtificialIntelligenceResearchSymposium(FLAIRS),497–500.FloridaAIResearchSociety.

Wolz,U.2001.TeachingDesignandProjectMan-agementwithLEGORCXRobots.ProceedingsoftheThirty-secondSIGCSETechnicalSymposiumonCom-puterScienceEducation33(1):95–99.