cscmuedu thrun In Proceedings of Uncertainty in AI U AI 2002 Abstract In recent years particle 64257lters ha solv ed se eral hard perceptual problems in robotics Early successes of particle 64257lters were limited to lo wdimensional esti mation probl ID: 21996
Download Pdf The PPT/PDF document "article Filters in Robotics Sebastian Th..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
ParticleFiltersinRoboticsSebastianThrunComputerScienceDepartmentCarnegieMellonUniversityPittsburgh,PA15213http://www.cs.cmu.edu/ thrunInProceedingsofUncertaintyinAI(UAI)2002AbstractInrecentyears,particleltershavesolvedseveralhardperceptualproblemsinrobotics.Earlysuccessesofparticlelterswerelimitedtolow-dimensionalesti-mationproblems,suchastheproblemofrobotlo-calizationinenvironmentswithknownmaps.Morerecently,researchershavebegunexploitingstructuralpropertiesofroboticdomainsthathaveledtosuccess-fulparticlelterapplicationsinspaceswithasmanyas100,000dimensions.Thefactthateverymodelnomaterhowdetailedfailstocapturethefullcomplex-ityofeventhemostsimpleroboticenvironmentshasleadtospecictricksandtechniquesessentialforthesuccessofparticleltersinroboticdomains.Thisarti-clesurveyssomeoftheserecentinnovations,andpro-videspointerstoin-deptharticlesontheuseofparticleltersinrobotics.1INTRODUCTIONOneofthekeydevelopmentsinroboticshasbeentheadop-tionofprobabilistictechniques.Inthe1970s,thepredomi-nantparadigminroboticswasmodel-based.Mostresearchatthattimefocusedonplanningandcontrolproblemsun-dertheassumptionoffullymodeled,deterministicrobotandrobotenvironments.Thischangedradicallyinthemid-1980s,whentheparadigmshiftedtowardsreactivetech-niques.ApproachessuchasBrooks'sbehavior-basedar-chitecturegeneratedcontroldirectlyinresponsetosensormeasurements[4].Rejectionsofmodelsquicklybecametypicalforthisapproach.Reactivetechniqueswerear-guableaslimitedasmodel-basedones,inthattheyreplacedtheunrealisticassumptionofperfectmodelsbyanequallyunrealisticoneofperfectperception.Sincethemid-1990s,roboticshasshifteditsfocusedtowardstechniquesthatuti-lizeimperfectmodelsandthatincorporateimperfectsensordata.Animportantparadigmsincethemid-1990swhoseorigincaneasilybetracedbacktothe1960sisthatofprobabilisticrobotics.Probabilisticroboticsintegratesim-perfectmodelsandandimperfectsensorsthroughproba-bilisticlaws,suchasBayesrule.Manyrecentlyeldedstate-of-the-artroboticsystemsemployprobabilistictech-niquesforperception[12,46,52];somegoevenasfarasusingprobabilistictechniquesatalllevelsofperceptionanddecisionmaking[39].Thisarticlefocusesonparticleltersandtheirroleinrobotics.Particlelters[9,30,40]compriseabroadfam-ilyofsequentialMonteCarloalgorithmsforapproximateinferenceinpartiallyobservableMarkovchains(see[9]foranexcellentoverviewonparticleltersandapplica-tions).Inrobotics,earlysuccessesofparticlelterimple-mentationscanbefoundintheareaofrobotlocalization,inwhicharobot'sposehastoberecoveredfromsensordata[51].Particlelterswereabletosolvetwoimportant,previouslyunsolvedproblemsknownasthegloballocal-ization[2]andthekidnappedrobot[14]problems,inwhicharobothastorecoveritsposeunderglobaluncertainty.Theseadvanceshaveledtoacriticalincreaseintherobust-nessofmobilerobots,andthelocalizationproblemwithagivenmapisnowwidelyconsideredtobesolved.Morere-cently,particleltershavebeenatthecoreofsolutionstomuchhigherdimensionalrobotproblems.Prominentex-amplesincludethesimultaneouslocalizationandmappingproblem[8,27,36,45],whichphrasedasastateestima-tionprobleminvolvesavariablenumberofdimensions.Arecentparticle-lteralgorithmknownasFastSLAM[34]hasbeendemonstratedtosolveproblemswithmorethan100,000dimensionsinreal-time.Similartechniqueshavebeendevelopedforrobustlytrackingothermovingentities,suchaspeopleintheproximityofarobot[35,44].However,theapplicationofparticlelterstoroboticsprob-lemsisnotwithoutcaveats.Arangeofproblemsarisefromthefactthatnomatterhowdetailedtheprobabilisticmodelitwillstillbewrong,andinparticularmakefalseindependenceassumptions.Inrobotics,allmodelslackim-portantstatevariablesthatsystematicallyaffectsensorandactuatornoise.Probabilisticinferenceiffurthercompli-catedbythefactthatrobotsystemsmustmakedecisionsinreal-time.Thisprohibits,forexample,theuseofvanilla(exponential-time)particleltersinmanyperceptualprob-