Abstract. This paper describes the AGILO RoboCuppers 1 the RoboCup team of the image understanding group (FG BV) at the Technische Universit?t München. With a team of four Pioneer I robots, all equipped with CCD camera and a single board computer, we’ve
tofasterrecovertheirpositionsaftertheyhavelosttrackofthem.Adetaileddescrip-tionoftheselflocalizationalgorithmcanbefoundin[8]andthealgorithmsusedforcooperativemulti-objecttrackingareexplainedin[13,12].
Ourvisionalgorithmscanprocessupto25framespersecond(fps)ona200MHzPentiumPC.Theaveragenumberofimagesprocessedduringamatchisbetween12and17fps.Thisisduetocomputationalresourcesbeingsharedwiththepathplanningandactionselectionmodules.
3.2ExperienceBasedLearningforSituatedActionSelection,PathPlanning
andMovementControl
Anothermajor eldofourresearchactivitiesisautomaticrobotlearningbasedonexperiencesgainedfromexploration.Experiencebasedlearningprovidesapowerfultoolfortheautomaticconstructionofhigh-performanceactionselectionandlow-levelrobotcontrol.Inthisrespectexperiencebasedlearningcaneffectivelycomplementothermethodsfordevelopingsuchcontrollers,inparticularthehandcodingofcon-trollers.Weuselearningfromexperienceinseveralpartsofoursystemsuchaslowlevelrobotcontrol,pathplanningandactionselection.
InlowlevelrobotcontrolwerepresentthestateofaPioneerIrobotasaquintuple
,whereandarecoordinatesinaglobalsystem,istheorienta-tionoftherobotandandarethetranslationalandrotationalvelocities,respectively.
Thelow-levelrobotcontrolleracceptscommandsoftheform.Aneuralnet-workmapsthedesiredstatechangestolowlevelrobotcommands:
Totrainthisnetworkwemeasureahugenumberofstatechangesaccordingtodifferentexecutedlowlevelcommands[6].Doingsoourneuralcontrollerisbasedonnothingbutexperiencenotmakinganyassumptions.
Inorderto ndtheoptimalpathplanningalgorithmforourRoboCuprobotswesta-tisticallyevaluateddifferentmethodsandfoundoutthatthereisnooptimalalgorithmbutanumberofnavigationproblemclasseseachperformedbestwithacertainalgo-rithm/parameterization[6].Theseclassesarede nedwiththehelpofafeaturelan-guage.Inordertoselectthebestmethodforthegivensituationwe’velearnedadecisiontree[11].Thetrainingdataisobtainedfromaccuraterobotsimulationswhereahugenumberofpathplanningproblemswereperformedwithdifferentalgorithmseach.
Theselectionofanappropriateactionisperformedonthebasisofafusedenvi-ronmentalmodel.Asetofpossibleactionssuchasgo2ball,shoot2goal,dribble,block...isde ned.Forallrobotsandeachofthoseactionssuccessrates
[5].Fromallpromisingactions,whichexceedapre-andgainsareestimatedde nedthresholdtheonewiththehighestgainischosentobecarriedout.
3.3Plan-basedActionControl
Whileoursituatedactionselectionaimsatchoosingactionsthathavethehighestex-pectedutilityintherespectivesituationitdoesnottakeintoaccountastrategicassess-mentofthealternativeactionsandtherespectiveintentionsoftheteammates.Thisisthetaskoftheplan-basedactioncontrol.
Inordertorealizeanactionassessmentbasedonstrategicconsiderationandonaconsiderationsoftheintentionsoftheteammates,wedeveloparobotsoccerplaybook,alibraryofplanschematathatspecifyhowtoperformindividualteamplays.Theplans,orbetterplays,aretriggeredbyopportunities,forexample,theopponentteamleaving.