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

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