A Cognitive Modeling of Space using Fingerprints of Places for Mobile Robot Navigation

Transcription

1 A Cogniive Modeling of Space using Fingerprins of Places for Mobile Robo Navigaion Adriana Tapus Roland Siegwar Ecole Polyechnique Fédérale de Lausanne (EPFL) Ecole Polyechnique Fédérale de Lausanne (EPFL) Auonomous Sysems Lab Auonomous Sysems Lab 1015, Lausanne, Swizerland 1015, Lausanne, Swizerland Absrac - In his work we address he problem of percepion, spaial cogniion and opological navigaion for a mobile robo. The objecive of his work is o enable he navigaion of an auonomous mobile robo (or vehicle) in an indoor (or oudoor) srucured environmen wihou relying on maps a priori learned and wihou using arificial landmarks. A new mehod for incremenal and auomaic opological mapping and global localizaion using fingerprins of places is presened. The fingerprin-based represenaion permis a reliable, compac and disincive environmen-modeling. Experimenal resuls for mapping indoor and oudoor environmens wih a mobile robo and a SMART vehicle, boh equipped wih a muli-sensor sysem composed of wo 180 laser range finders and an omnidirecional camera are also repored. Index Terms fingerprins of places, opological navigaion, cogniive mapping, muli-modal percepion I. INTRODUCTION A robus navigaion sysem requires a spaial model of physical environmens as a meric [1, 3, 5] or opological map [11, 14]. Approaches using meric maps are suied when i is necessary for he robo o know is locaion accuraely in erms of meric coordinaes. However, he sae of he robo can also be represened in a more qualiaive manner, similar o he way humans do i. The informaion can be sored as cogniive maps erm inroduced for he firs ime in [15] which permi an encoding of he spaial relaions beween relevan locaions in he environmen. This has led o he concep of opological represenaion. The opological map can be viewed as a graph of places, where a each node he informaion concerning he visible landmarks and he way o reach oher places, conneced o i, is sored. The opological represenaion is compac and allows high-level symbolic reasoning for map building and navigaion. In order o have a robus and reliable framework for navigaion (i.e. in order o move wihin an environmen, manipulae objecs in i, avoid undesirable collisions, ec.) space cogniion, percepion, localizaion and mapping are all needed. The objecive of his work is o enable auonomous navigaion wihou relying on maps a priori learned and wihou using arificial landmarks. Therefore, his paper describes a new mehod for incremenal and auomaic opological mapping and global localizaion wih POMDP (Parially Observable Markov Decision Processes) using fingerprins of places. One of he main problems in opological map building is o deec when a new node should be added in he map. Our approach relies on a heurisic ha deecs wheher he curren locaion of he robo is similar o a mapped one or no. The proposed mehod permis a reliable and disincive environmen model ha can be globally handled in an efficien way. Various mehods have been proposed o represen environmens in he framework of auonomous navigaion, from precise geomeric maps based on raw daa or lines o purely opological maps using symbolic descripions. Each of hese mehods is opimal wih respec o some characerisics bu can be very disappoining wih respec o ohers. Alhough lieraure relaed o SLAM (Simulaneous Localizaion and Mapping) is very vas, we only concenrae here on papers ha have direcly influenced our hinking and research work. Topological approaches o SLAM aemp o overcome he drawbacks of geomeric mehods (e.g. problems concerning he global disinciveness and global consisency) by modeling space using graphs. Significan progress has been made since he seminal paper by Kuipers [8], where, an approach based on conceps derived from a heory on human cogniive mapping is described as he body of knowledge represening large scale space. Korenkamp and Weymouh in [7] have also used cogniive maps for opological navigaion. They defined he concep of gaeways which have been used o mark he ransiion beween wo adjacen places in he environmen. The model described in [4] represens he environmen wih he help of a Generalized Voronoi Graph (GVG) and localizes he robo via a graph maching process. This approach has been exended o H-SLAM (i.e. Hierarchical SLAM) in [10], by combining he opological and feaure-based mapping echniques. In [16], Tomais e al. have conceived a hybrid represenaion, similar o he previously menioned work, comprising of a global opological map wih local meric maps associaed o each node for precise navigaion. Topological maps are less complex and permi more efficien planning han meric maps. Moreover, hey are easier o generae. Mainaining global Adriana Tapus was wih Ecole Polyechnique Fédérale de Lausannne, 1015-CH. She is now wih he Roboics Research Lab / Ineracion Lab, Compuer Science Deparmen, Universiy of Souhern California, Los Angeles, USA; adriana.apus@ieee.org.

2 consisency is also easier in opological maps compared o meric maps. However, he main problems o deal wih, when working wih opological maps are he percepual aliasing (i.e. observaions a muliple locaions are similar) and he auomaic esablishmen of a minimal opology (nodes). Our mehod uses fingerprins of places o creae a opological model of he environmen. The fingerprin approach, by combining he informaion from all sensors available o he robo, reduces percepual aliasing and improves he disinciveness of places. The main conribuion of his paper is he consrucion of a opological mapping sysem combined wih he localizaion echnique, boh relying on fingerprins of places. This fingerprin-based approach yields a consisen and disincive represenaion of he environmen and is exensible in ha i permis spaial cogniion beyond jus pure navigaion. The res of he paper is srucured as follows. Secion II presens a shor review of he fingerprin concep. Secion III is dedicaed o he new opological navigaion sysem wih fingerprins of places. Experimenal resuls are presened in Secion IV. The sysems (indoor and oudoor) use boh, wo 180 laser range finders and an omni-direcional camera for feaure exracion. Finally, Secion V draws conclusions and discusses furher work. II. FINGERPRINTS OF PLACES Represening and inerpreing a scene from he environmen is a hard ask. Humans use various sensory cues o exrac crucial informaion from he environmen. This is processed in he corex of he brain in order o obain a highlevel represenaion of wha has been perceived. Wih a view of having robos as companion of humans, we are moivaed owards developing a knowledge represenaion sysem along he lines of wha we know abou us. While recen research has shown ineresing resuls, we are sill far from having conceps and algorihms ha represen and inerpre space, coping wih he complexiy of he environmen. The fingerprin of a place concep is used here. Fingerprins of places (i.e. circular lis of significan feaures around he robo) have been proven o be a very promising approach owards effecive place characerizaion and hence environmen modelling [9, 12, 13]. In his work, we choose o use as significan feaure: colour bins and verical edges from he visual informaion and corners from he laser scanner. Therefore, his muli-modal, feaure based represenaion of space reduces he percepual aliasing and improve he disinciveness of space. III. TOPOLOGICAL NAVIGATION Navigaion described by Gallisel in [6], as he capaciy o localize iself wih respec o a map, is an elemenary ask ha a mobile and auonomous robo mus carry ou. To navigae reliably in indoor or oudoor environmens a mobile robo mus know where i is. For his, he robo needs o consruc or mainain a spaial represenaion of he environmen. Here, we approach he SLAM (Simulaneous Localizaion and Mapping) problem ha is of a chicken and egg naure o localize he robo, a map is necessary and o updae a map he posiion of he mobile robo is needed. A. Topological Mapping While navigaing in he environmen, he robo firs creaes and hen updaes he global opological map. One of he main issues in opological map building is o deec when a new node should be added in he map. Mos of he exising approaches o opological mapping place nodes periodically in eiher space (displacemen, d) or ime ( ) or alernaively aemp o deec imporan changes in he environmen srucure. Any of hese mehods canno resul in an opimal opology. In conras, he approach presened in his work is based direcly on he differences in he perceived feaures. Insead of adding a new node in he map by following some fixed rules (e.g. disance, opology) ha limi he approach o indoor or oudoor environmens, he mehod described in his work inroduces a new node ino he map whenever an imporan change in he environmen occurs. This is possible using he fingerprins of places. A heurisic is applied o compare wheher a new locaion is similar o he las one ha has been mapped. Thus, a new node is inroduced in he opological map jus when imporan changes ino he environmen occur. Wih his, a he end, each node will be composed of a se of similar fingerprins of places. In order o compac even more he curren represenaion, a unique idenifier named he mean fingerprin is generaed. This echnique of clusering fingerprins of places ino a single represenaion enables he consrucion of a very disincive and compac represenaion of he environmen. Therefore, a new node conains all poserior knowledge abou he environmen unil he previous node. A more deailed presenaion is given in one of our previous works [12]. The incremenal naure of he approach permis incremenal addiions o he map and yields he mos up-o-dae map a any ime. B. Topological Localizaion wih POMDP Finding an efficien soluion o he robo localizaion problem is necessary for he robos o be inegraed ino our daily lives. Mos asks for which robos are well suied demand a high degree of robusness in heir localizing capabiliies. A series of localizaion echniques based on he fingerprin concep have been already presened in [13]. These approaches perform a fingerprin-maching operaion so as o localize he robo. The maching mehods compare he observed feaures encoded in he fingerprins of places wih he map fingerprins. Only he exereocepive sensory informaion conained in fingerprins of places is used for maching, wihou aking ino accoun he moion of he robo and he previous esimaion. Hence, for opological navigaion, a Parially Observable Markov Decision Process (POMDP) model [2] is used here. The POMDPs inegrae boh he moion and sensor repors daa o deermine he pose disribuion. Thus, by adding he

3 moion informaion o he sysem, new knowledge abou he robo s posiion is acquired. The probabiliy of being in a place is calculaed in funcion of he las probabiliy disribuion, and he curren acion and observaion. A POMDP is defined as <S,A,T,O>, where: S is a finie se of environmen saes; A is a finie se of acions; T(s,a,s ) is a ransiion funcion beween he environmen saes based on he acion performed; O is a finie se of possible observaions; OS is an observaion funcion. Wih his informaion, he probabiliy of being in a sae s (belief sae of s ) afer having made observaion o, while performing acion a, is given by: OS ( o, s' ) T ( s, a, s' ) SE + 1 S s S SE S ' = P( o a, SE ) where SE S is he belief of sae S a he las sep, SE is he belief sae vecor a he las sep, and P ( o a, SE ) is he normalizing facor. The key idea is o compue a discree approximaion of a probabiliy disribuion over all possible poses in he environmen. An imporan feaure of his localizaion echnique is he abiliy o globally localize he robo wihin he environmen. More deails abou his approach are given in [2]. In our approach, he se of observaions O is composed of he fingerprins of places generaed by he robo in he environmen. These observaions are very disincive since disinciveness is one of he main characerisics of he fingerprins of places. An observaion conains informaion given by he exereocepive sensors and designaes a subse of he world sae. The informaion for he observaion funcion OS wihin he opological framework is given by he fingerprin maching algorihm: Global Alignmen wih Uncerainy [13]. C. Map Updae While navigaing in he environmen, he robo firs creaes and hen updaes he global opological map. By using a POMDP (Parially Observable Markov Decision Process), a discree approximaion of a probabiliy disribuion over all possible poses in he environmen is compued. The enropy of a probabiliy disribuion is used here. The lower he value is, he more cerain he disribuion is. When he robo is "confused", he enropy is high. So he POMDP is confiden abou is sae if he enropy is smaller han a fixed hreshold. Therefore, he sraegy of updaing he map will be he following: a) When he enropy of he belief sae is low enough, he map will be updaed and so he fingerprin and he uncerainy of he feaures will also be updaed. b) If he enropy is above he hreshold, hen he updaing will no be allowed, and he robo will ry o reduce he enropy by coninuing he navigaion wih localizaion. When he robo feels confiden concerning is sae, i can decide if an exraced feaure is new by comparing he observed fingerprin o he fingerprin from he map, corresponding o he mos confiden sae. This can happen eiher in an unexplored porion of he environmen, or in a known porion where new feaures appear due o he environmenal dynamics. When a feaure is re-observed, he uncerainy of he feaure from he map fingerprin is weigh averaged wih he uncerainy of he exraced one. The weigh depends on he ype of feaure. Since he exracion of feaures wih he laser scanner is more robus han he ones exraced wih he camera, a higher weigh is given o hem. Oherwise, if he robo does no see an expeced feaure he uncerainy is decreased. When he uncerainy of a feaure from a map fingerprin is below a minimum hreshold, han he feaure is deleed, allowing in his way for dynamics in he environmen. D. Closing he Loop One fundamenal problem in SLAM is he idenificaion of a place previously visied, if he robo reurned o i. This is known as he closing he loop problem since he robo s rajecory loops back on iself. Thus, for opological maps, his means ha if a place (i.e. a node) has been visied before, and he robo reurns o i, he robo should deec i (see Figure 1). Figure 1: Loop Closing Problem. The robo sars in place A and afer moving hrough he environmen arrives in place Q. The quesion o answer is: Has he robo reurned o an already visied place or no? (i.e. Is place A equivalen o place Q?) Conrary o oher mehods used for solving his problem, based usually on he percepion, loops are idenified and closed wih he help of he localizaion echniques. In order o accomplish consisency of he opological map, a mehod similar o he one described in [16] is used. In his work he mehod employed is a non-explici loop closing algorihm. Our loop closing mehod is based on he localizer (i.e. he POMDP). The robo is moving hrough he environmen and incremenally builds he opological map. As soon as he robo reurns in an already visied place (i.e. node) he probabiliy disribuion poenially should spli up. Two candidaes hypoheses should appear: one for he new place (i.e. node) currenly creaed by he robo (e.g. in Figure 1, node Q) and anoher one for he previously creaed node already presen in he map (e.g. in Figure 1, node A). As soon as he POMDP is unconfiden, he algorihm racks he wo highes probabiliy disribuions showing ha he disribuion diverged in wo peaks. A loop is hus idenified if he probabiliy disribuion

4 given by he localizer converges in wo peaks ha move in he same direcion. In order o deec where he loop was closed, he wo hypoheses are backracked wih localizaion unil a single one remains. IV. EXPERIMENTAL RESULTS Our approach for opological SLAM using he fingerprin of places echnique has been implemened and evaluaed in various real world indoor and oudoor environmens. In his secion we presen some of he indoor experimens carried ou wih our indoor robo, a fully auonomous mobile robo and he firs aemps for oudoor opological mapping using he SMART vehicle (Daimler-Chrysler). Boh mobile plaforms (indoor and oudoor) are equipped wih wo 180 laser range finders and an omni-direcional camera. The omni-direcional camera sysem uses a mirror-camera sysem o image 360 in azimuh and up o 110 in elevaion. The firs se of experimens demonsraes he robusness of he mapping module in wo indoor real world scenarios and he firs aemps o map urban oudoor environmens. In paricular, i illusraes he consrucion of disincive and compac maps (composed only of local feaures, which is an advanage of his fingerprin-based mapping echnique). A. Indoor Topological Mapping The firs indoor experimen was conduced in a porion of our insiue building shown in Figure 2 and he second experimen was performed in anoher building from our campus (see Figure 3(a)). The firs es seup was he following: he robo sared a he poin S and ended a he poin E, as illusraed in he Figure 2, he disance raveled being of 75m. For he second es he robo raveled a disance of 67m. While he robo explored he environmen, i recorded, a every d (disance) (e.g. in our case d =15cm), daa readings from sensors (i.e. an image from he omnidirecional camera and a scan from he laser scanner) in order o exrac he fingerprins. The robo has a mid-line following behavior in he hallways and cener of he free space behavior in he open spaces. We assume ha he posiion in he room wih he maximum free space around i, is he one wih he highes probabiliy of exracing numerous and characerisic feaures. This ensures high disinciveness of he observaion. The map building process was performed off-line. The hreshold θ, defined as he maximum allowable dissimilariy and used for auomaic mapping is calculaed experimenally. I is calculaed for a small porion of he environmen (i.e. 5 m), so ha he map obained maches he real srucure of he environmen. Once his hreshold is deermined, i is fixed for he res of he indoor experimens. Figure 2(b) shows he opological map obained by he sysem in he firs es environmen (i.e. in our laboraory), superimposed on an archiecural skech of he environmen. The resuling map is composed of 20 nodes as shown in he Figure 2. Each node is represened by a mean fingerprin which is an aggregaion of all he fingerprins composing he respecive node. Typically, he nodes are posiioned in he rooms and in he hallway. Raw Scan Map (for reference only) (a) Coffee Room (b) Corridor Bill s Room Figure 2: (a) Floor plan of he firs environmen where he experimens have been conduced. The robo sars a he poin S and ends a he poin E. The rajecory lengh is 75 m. During his experimen, he robo colleced 500 daa ses (i.e. images and scans) from he environmen. The exraced opological map is superimposed on an archiecural skech of he environmen. (b) The exraced opological map given by our mehod, superimposed on he raw scan map. (a) Raw Scan Map (for reference only) (b) Figure 3: (a) The second es environmen wih he rajecory raveled by he robo. (b) The map of he second es environmen wih he graph represening he opological map.

5 The doors of some rooms remained closed a he ime of experimenaion; his explains why no node is presen in fron of he respecive rooms (see Figure 2). Figures 3(a) - (b) show he second es environmen wih he corresponding opological map, formed using he approach oulined in his work. The mapping sysem added a new node auomaically each ime a very disincive measure (i.e. disincive fingerprin) was encounered. The graph-like map hus obained conains 8 nodes, as shown in Figure 3(b). The same hreshold used for he firs es (hreshold calculaed experimenally) was employed here also, indicaing he robusness of he overall mehod. The represenaions hus obained (see Figure 2 and 3(b)) reproduce correcly he srucure of he physical space, in a manner ha is compaible wih he opology of he environmen. They also permi a disincive modeling of i. I is imporan o menion ha he maps are obained by using only locally disincive feaures composing he fingerprins of places. B. Firs Aemps o Oudoor Topological Mapping Compared wih indoor environmens, urban oudoor environmens presen many challenges for an auonomous vehicle. Coarse localizaion is ofen available from GPS. Mos of he ime, i is more useful o know he posiion of he robo wih respec o buildings, rees, inersecions, ec., han he exac laiude and longiude. In order o validae and o show he robusness of our approach, we also esed i in an oudoor environmen. The approach has been esed in a par of our campus (highly srucured environmen), shown in Figure 4, on a 1.65 km of rajecory. The sysem mouned on he SMART vehicle acquired daa, boh from he lasers and omni-direcional camera every 110 ms. A new hreshold for oudoor environmens was calculaed experimenally in a small porion of he campus. Differen hresholds can be used in funcion of he granulariy of he environmen ha i is desired. High granulariy maps, wih numerous nodes, may be obained by seing small hresholds. Alernaively, seing high values for he hreshold yields maps wih fewer nodes (low granulariy). The oudoor hreshold for obaining high granulariy maps is he same as he one used for indoor environmens. For geing maps wih fewer nodes, he oudoor hreshold is se hree imes bigger han he indoor hreshold. We have obained a map composed of 209 nodes for a high granulariy and a map of 64 nodes conaining only he big changes in he environmen (i.e. inersecions, new buildings, ec). A small example is depiced in Figure 6(b), which represens a low granulariy opological map obained for a 200 m secion of he environmen (i.e. he zoomed view of he magnifying glass shown in Figure 4(a)). The map conains 7 nodes. I can be noiced ha he nodes are usually placed in fron of buildings, a he crossings and when "big" changes occur (e.g. a building disappears from he field of view of he vehicle and driving signs, lamp-spos and rees appear). The map hus obained for he enire rajecory shown in Figure 4 is compaible wih he srucure of he oudoor environmen, aking ino accoun he rees, he buildings and he lamp-poss. C. Indoor Localizaion wih POMDP The qualiy of he opological maps obained wih our fingerprin-based echnique, can be evaluaed by esing he localizaion on i. Localizaion experimens were conduced so as o show his. To es he localizaion, more han 1000 new fingerprin samples, acquired while he robo was ravelling new pahs of 250 m, were used o globally localize he robo (a) (b) Figure 4: (a) The oudoor es environmen (a par of he EPFL campus) wih he rajecory of 1.65 km long raveled by he Smar vehicle. The magnifying glass represens he par of he environmen used for he oudoor opological map exemplificaion; (b) The low granulariy oudoor opological map superimposed on an archiecural skech of a par of he EPFL campus. wih he POMDP. A mission is considered successful if he place found, which corresponds o he world sae wih he highes probabiliy, is he same wih he correc node in he real world. TABLE 1: Summary of he indoor localizaion experimens. Fingerprins 1024 samples Disance Travelled 250 m Scenarios 10/10 Kidnapping 7/7 Fingerprin Maching (GA) 81% POMDP localizaion 100% The resuls are summarized in Table 1. I can be noiced ha he resuls wih he POMDP localizaion have given for he se of scenarios esed in his work a percenage of successful maches of 100%. The kidnapping problem (i.e. recovering from a los posiion he robo hinks ha i is in a posiion where i is no) has also been esed. This was performed seven imes and he robo succeeded o recover all he seven imes, afer one or wo seps because of he very disincive observaions ha corresponds o he fingerprins. D. Closing he Loop The localizaion wih POMDP is also used for idenificaion of loops. As explained earlier, he robo moves hrough he environmen and incremenally builds he opological map. The loop closing problem was esed 5 imes in differen siuaions wihin he environmen. The robo succeeded o close he loop in all he siuaions. Figure 5 shows only a simple example ha is explained below. In Figure 5, i can be noiced ha he robo sared in he corridor, in poin S. I raveled in he corridor ill he door ha separaes he wo hallways was deeced (i.e. imporan change ino he environmen - node N1), coninued in he corridor (i.e. node N2), hen enered and wen ou he Room 3 (i.e. node N3). Once i reurned in he corridor, he robo urned lef and

6 enered in an already visied place, corresponding o node N2. The robo emporarily creaes a new node N4. As soon as he robo reurned in an already visied place, he POMDP became unconfiden and he probabiliy disribuion divided in wo possible candidae saes. Two hypoheses appeared: one for he new place (i.e. node N4 circled in red on Figure 5) currenly creaed by he robo and anoher one for he previously creaed node already presen in he map (i.e. node N2). The auomaic mapper is urned off. The robo moved oward node N1 and labels i a node N5. Node N5 was very similar o node N1, and he correc mach is made. A loop is hus idenified if he probabiliy disribuion given by he localizer converges in wo peaks. In order o deec where he loop is closed, he auomaic mapping sysem is urned off and he wo hypoheses are backracked wih localizaion unil a single one remains. In he presen case, his occurred when node N5 was deeced. A ha poin he robo realized ha node N4 is node N2 and ha node N5 is node N1. Thus, he loop was closed correcly. In order o make use of he informaion obained when a place is revisied, he map is updaed. The nodes N1 and N2 are updaed wih he daa brough by he revisied nodes N5 and N4, respecively. Figure 5: Loop Closing: Shows he direced pah ha he robo raveled. The robo sars in poin S. I can be noiced ha he robo arrives in a visied place (i.e. node N4) once i goes ou he office (ROOM 3) and goes o he lef, revisiing again node N4. As explained earlier, due o he fac ha he offices are quie small, he fingerprins of places are very similar, and hus a single node per room is enough. Since a node conains a poserior knowledge abou is environmen and is he aggregaion of all he fingerprins of places beween he las node and he curren place where an imporan change ino he environmen occurred, closing he loop problem does no appear in hese cases (i.e. when one node per office is sufficien). V. CONCLUSIONS AND FUTURE WORKS This paper presened a new echnique for opological SLAM using fingerprins of places. A fingerprin of a place provides a compac and disincive mehodology for space represenaion and place recogniion i permis encoding of a huge amoun of place-relaed informaion in a single circular sequence of feaures. This represenaion is suiable for boh indoor and oudoor environmens. The experimens verify he efficacy and reliabiliy of our approach. The POMDP localizaion shown here improves he previously resuls obained. Adding he moion of he robo enables o decrease furher he pose uncerainy o a level ha could never be reached by fingerprin maching alone. A success rae of 100% was obained for he ess performed in his work. However, he approach has o be more exensively esed in differen ypes of environmen in order o make a real esimaion of he qualiy of he mehod. In his work, lowlevel feaures (such as verical edges) have been used. An ineresing exension of he model is he addiion of oher modaliies and feaures o he fingerprin framework (e.g. audiory, smell, or higher level feaures such as doors, able, fridge, ec). This will help o improve he reliabiliy and accuracy of he mehod and o add semanics o i. ACKNOWLEDGMENTS This work was suppored by he European projec BIBA IST EU projec. REFERENCES [1] Arras, K.O. and Siegwar, R, Feaure Exracion and Scene Inerpreaion for Map-Based Navigaion and Map Building, In Proceedings of he Symposium on Inelligen Sysems and Advanced Manufacuring, Pisburgh, USA, Ocober 13-17,1997. [2] Cassandra, A., Kaelbling, L., e al(1996). Acing under Uncerainy: Discree Bayesian Models for Mobile-Robo Navigaion. IEEE Inernaional Conference on Roboics and Auomaion (ICRA), Osaka, Japan. [3] Casellanos J.A., and Tardos J.D. (1999), Mobile Robo Localizaion and Map Building: Mulisensor Fusion Approach, Kluwer. [4] Chose, H., and Nagaani, K.(2001), Topological Simulaneous Localizaion and Mapping (SLAM): Toward Exac Localizaion Wihou Explici Localizaion, IEEE Trans. On Roboics and Auomaion, Vol 17, No.2, April. [5] Dissanayake, Newman, Clark, Durran-Whye and Csorba (2001), A Soluion o he Simulaneous Localizaion and Map Building (SLAM) problem, IEEE Trans. On Roboics and Auomaion, Vol 17, No.3, June. [6] Gallisel, R. (1990). The Organizaion of Learning. M. Press. Cambridge, MA [7] Korenkamp, D. and Weymouh, T. (1994), Topological mapping for mobile robos using a combinaion of sonar and vision sensing, In Proceedings of AAAI-94, Seale, WA. [8] Kuipers, B. J. (1978), Modeling Spaial Knowledge, Cogniive Science, 2: , [9] Lamon, P., Tapus A., Glauser E., Tomais N., Siegwar R. (2003), Environmenal Modeling wih Fingerprin Sequences for Topological Global Localizaion, Proceedings of he IEEE/RSJ Inernaional Conference on Inelligen Robo and Sysems(IROS), Las Vegas, USA [10] Lisien, B., e al. (2003), Hierarchical Simulaneous Localizaion and Mapping, In Proceedings of he IEEE/RSJ Inernaional Conference on Inelligen Robo and Sysems, Las Vegas, USA, Ocober [11] Owen, C. and Nehmzow, U.(1998), Landmark-based navigaion for a mobile robo, in : Meyer, Berhoz, Floreano, Roibla and Wilson (Eds.), From Animals o Animae 5, Proceedings of SAB 98, MIT Press, Cambridge, MA, pp [12] Tapus, A., and Siegwar, R., (2005) Incremenal Topological Mapping wih Fingerprins of Places, In Proceedings of he IEEE/RSJ Inernaional Conference on Inelligen Robo and Sysems (IROS), Edmonon, Canada. [13] Tapus, A., Tomais, N. and Siegwar, R., (2004) Topological Global Localizaion and Mapping wih Fingerprin and Uncerainy. In Proceedings of he ISER, Singapore, June [14] Thrun, S. (1998), Learning meric-opological maps for indoor mobile robo navigaion. In Arificial Inelligence 99(1): [15] Tolman, E. C. (1948), Cogniive maps in ras and men, Psychological Review, 55: [16] Tomais, N., I. Nourbakhsh, and R. Siegwar (2003). Hybrid simulaneous localizaion and map building: a naural inegraion of opological and meric. Roboics and Auonomous Sysems, 44:3-14.