Perceptual Homing by an Autonomous Mobile Robot using Sparse Self-Organizing Sensory-Motor Maps

摘要: PerceptualHomingbyanAutonomousMobileRob otusingSparseSelf-OrganizingSensory-MotorMapsRa jesh P.N. RaoandOlac FuentesDepartment of Computer ScienceUniversity Ro chesterRo chester, NY 14627AbstractWe present a biological ly -motivated approach to the problem p erception-based homing by an autonomousmobile rob ot.A three-layeredself-organizingnetwork is used autonomouslylearn desired mappingfromp erceptions actions.The network, which b ears some similarities structure mammalian cereb ellum,is initially trained teleop erating ot on small numb er paths within given domain interest.During training, connectionsb etween input sensory layer and hidden laer as well those eteen thehidden motor output are mo di ed according well-known comp etitive Hebbian learningrule. By employing opulation averaging scheme for computing vectors, can subsequentlyhome from arbitrary lo cationswithin based solely current erceptions.We describ e preliminaryresults simulation actual mobile ot, equipp with simple photoreceptors infrared receivers,navigating enclosed obstacle-ridden arena.1Intro ductionA central in otics that autonomous goal-directed navigation unstructured environments.Traditional metho ds designing navigational controllers involve prewiring xed set strategies heuristicsand knowledge.Such systems however su inherent inexibility utilizing prede ned ehaviorsand such unable cop variations characteristic environments.Recentwork ehavioral has shown many instances,the uncertainties osed unstructuredenvironments circumvted large extentby endowing ability toautonomouslylearnnavigational ehaviors (for example, [3] [6]).In spirit this recent trend, we biologically-motivatedframework acquisition erception-basedhoming ots.Homing ede agent navigate particular \home" cation cationswithinasp eci cenvironment.Homing ehaviorsarealmostuniversalinanimals[10].Assumingthatcomplexanimal ehaviorsemergedfrompre-existingsimplerones,itisreasonabletoassumethattheacquisitionofhoming ehavior representsa signi cant steptowards acquiring more complex ehaviors.Indeed, thegeneral learning cations decomp into thesimpler onents navigating one-one, many-one, one-many Figure 1 (a).In pap er, prop ose three-layer network architecture allows autonomously learn homebased only its erceptions.The ellum,employscompetitiveHebbianlearningtomo dify theconnectionsb etweentheinputsensorylayerandhiddenlayeraswelltheconnectionsb eteenhiddenandmotoroutputduringaninitialtrainingp erio dwhichinvolves erationof therob ot(Figure1(b))in anenclosedarena1 (c)).Theisequipp four orthogonally-placed detectors measuring strength dulated lightb eing emitted eacon placed at near arena), six horizontally-p ositioned photoreceptors(for theamount light sourcelo catednearthearena),and tiltedphotoreceptors(formeasuring intensity value due color o or surrounding obstacles).This work supp orted NSF research grant no. CDA-8822724, ARPA MDA972-92-J-1012 ONR no.N00014-93-1-0221.