Decenralized Cooperaive Muli-Robo Localizaion wih EKF Ruihua Han, Shengduo Chen, Yasheng Bu, Zhijun Lyu and Qi Hao* arxiv:1811.76v1 [cs.ro] 19 Nov 218 Absrac Muli-robo localizaion has been a criical problem for robos performing complex asks cooperaively. In his paper, we propose a decenralized approach o localize a group of robos in a large feaureless environmen. The proposed approach only requires ha a leas one robo remains saionary as a emporary landmark during a cerain period of ime. The novely of our approach is hreefold: (1) developing a decenralized scheme ha each robo calculaes heir own sae and only sores he laes one o reduce sorage and compuaional cos; (2) developing an efficien localizaion algorihm hrough he exended Kalman filer (EKF) ha only uses observaions of relaive pose o esimae he robo posiions; () developing a scheme has less requiremens on landmarks and more robusness agains insufficien observaions. Various simulaions and experimens using five robos equipped wih relaive pose-measuremen sensors are performed o validae he superior performance of our approach. I. INTRODUCTION In he sudy of auonomous mobile robos, localizaion is one of he mos fundamenal problems [1]. Mobile robo localizaion, as known as posiion esimaion, is he process of deermining he pose of a robo relaive o a given map of he environmen. Compared o a single robo, muli-robo sysems are more robus and efficien o perform complex asks wihin complex environmens, including rescue and disaser managemen, surveillance and monioring, and so on [2]. For he auonomous muli-robo sysems, he performance of mos funcionaliies heavily relies on he accuracy of robo localizaion resuls, where individual localizaion resuls easily inerfere wih each oher. The main muli-robo localizaion approaches can fall ino hree caegories: (1) exernal, (2) inernal, and () cooperaive. The exernal mehods help each robo o acquire heir posiion informaion independenly hrough hose exernal acive landmarks such as Global Posiion Sysems (GPS) or Ulra-Wideband (UWB) sysems. The localizaion uncerainy of he exernal mehods can be bounded only when hose exernal signals are reliable and available. The inernal mehods enable individual robos o localize heir posiions by uilizing he sensors on each robo such as camera or laser scanner o deec he feaures of he environmen. However, hese mehods can no work well in hose environmens wihou rich feaure informaion. The cooperaive mehods, also known as cooperaive localizaion (CL), allow a group *This work was suppored by he Science and Technology Innovaion Commission of Shenzhen Municipaliy (No. CKFW216117217 and No. GJHZ217111212), and he Naional Naural Science Foundaion of China (No. 6177197). The auhors are wih Deparmen of Compuer Science and Engineering, Souhern Universiy of Science and Technology, Shenzhen, Guangdong, 18, China. {hanrh@mail.susc.edu.cn haoq@susc.edu.cn} *Corresponding auhor (a) (c) (b) Fig. 1. Illusraion of he approach. The example during (a) denoes a he ime, one robo is randomly chose o remain saionary, firsly, he neares moving robo 1 deec he saionary one o esimae is pose. During (b), he nex neares robo o robo 1 esimae is pose by he wo observaion. During (c) and (d) he robo esimae heir pose by he observaion of previous robo. of communicaing robos o esimae heir saes joinly using relaive observaions of each oher. Those mehods are advanageous in heir low cos and high scalabiliy, receiving increasing aenions recenly, especially for hose robo sysems working in feaureless environmens. However, many echnical challenges need o be overcome in using cooperaive localizaion mehods. Generally speaking, he esimaion accuracy in feaureless environmens is inferior o ha in environmens wih rich feaure informaion. Besides, as he group of robos becomes larger, he cos of compuaion and communicaion grows much higher. In pracice, only using relaive observaions of each oher may no be enough o localize muliple robos given various uncerainies and noises. A number of approaches have been proposed o address above challenges. Dead-reckoning() is one common mehod which esimaes he robo pose from is previous pose informaion by using propriocepive sensors [], bu he esimaion errors will be accumulaed over a long disance. To improve he accuracy and consisency of esimaed robo saes, various daa fusion algorihms have been developed, such as Markov localizaion (ML), exended Kalman filer (EKF) localizaion and Mone Carlo localizaion (MCL) []. The main idea of hese schemes is o opimally combine he daa derived from propriocepive and exerocepive sensors and filer ou he moion and measuremen noises o reduce he uncerainy of robo pose esimaion. Usually, he dependency on landmarks limis he applicaions of hose approaches. (d)
Generally, he schemes o calculae he daa fusion in CL are eiher cenralized, muli-cenralized or decenralized []. In a cenralized sysem, he compuaions for all robo daa are processed in one cenral robo. Cenral robo provides feedback and required sensory informaion from all oher robos [6]. The major disadvanages of his scheme include he high expense of compuaion and he high vulnerabiliy of he whole sysem. In a muli-cenralized sysem, here are more han one robo performing as he cenral robo o process all he daa of enire group [7]. Consequenly, his sysem needs higher compuaional power. The decenralized sysem means each robo esimaes is own pose respecively hrough processing heir own sensors daa and updaes i locally. Noneheless he problem of double-couning or daa inces may occur when wo uncorrelaed robos share informaion [8]. The decenralized sysem yield he leas compuaional cos and highes sysem scalabiliy. In his paper, we presen a decenralized approach for CL, he approach is illusraed as Fig. 1. In his approach, a each ime insan only one random robo in he group remains saionary, while he oher moving robos esimae heir sae by fusing he daa of odomery and observaions of relaive observaion beween wo robos. The rearward moving robos uilize he observaion wih he forward moving robos o improve he accuracy and sabiliy of he esimaor. And he order depends on he relaive disance. The conribuions of his work include: 1) A decenralized scheme for CL is proposed, where each robo esimaes heir own sae ha reduces he cos of compuaion. 2) An EKF Based algorihm is proposed, where each robo only mainains is own pose and corresponding covariance, and he communicaion beween wo robos only occurs when one robo deec anoher a each ime insan, yielding a compuaional complexiy of O(N). ) This approach only requires he odomery daa and relaive pose measuremens so ha i can perform in GPS-denied feaureless environmens. Even hough here are insufficien relaive pose measuremens, he proposed approach can sill guaranee a moderae performance. The res of his paper is organized as follows. Secion II describes he relaed work. Secion III presens he sysem seup and he localizaion problem. Secion IV gives he deails of he proposed algorihm. Secion V provides simulaion and experimen resuls and discussions. Secion VI concludes he paper and oulines fuure works. II. RELATED WORK In he problem of muli-robo localizaion, he sudy of CL receives considerable aenion for a long period, because of is scalabiliy and superior performance. Many approaches on CL have been proposed based on various schemes and algorihms as menioned above. In his secion, we describe several relevan approaches on CL o highligh he novely of his work. To esimae he sae of robos wih high accuracy and consisency, various filer algorihms have been esed for CL. An EKF-based algorihm for heerogeneous oudoor mulirobo localizaion is described in [9], where individual robo equipped wih encoders and a camera mainains an esimaed pose by using sensor daa fusion hrough EKF. An improved EKF on CL is sudied in [1]. Addiionally, a recen approach ha called recursive decenralized localizaion based on EKF is presened in [8]. The proposed algorihm approximaes he iner-robo correlaions and performs wih asynchronous pairwise communicaion. Oher algorihms such as paricle filer (PF) [11] and maximum a poseriori (MAP) [12] are also exensively sudied for CL. The main limiaion of hese approaches is ha in cerain case, none of he robos in group has an access o he absolue sae informaion ha will reduce he esimaion accuracy and consisency. In our approach, he saionary robo a each ime can improve he accuracy efficienly by an opimized EKF algorihm. To reduce he cos of compuaion, various decenralized schemes are proposed. Luis C e al. presen a Covariance Inersecion (CI)-based algorihm for reducing he complexiy of CL [1], where each robo locally mainains is sae and covariance. Amanda e al. provide a PF-based algorihm ha is compleely decenralized and a low-cos paricle clusering mehod o reduce he compuaional complexiy [1]. In a decenralized sysem, i is imporan o avoid he problem of double-couning or daa inces. Masinjila e al. presens a heurisically uned EKF o address his problem [], which proposes an empirical mehodology for improving he consisency of EKF esimaes by means of arificial inflaion of he landmark covariance. This problem is also discussed in [1] and [8]. Our approach also employs he similar decenralized scheme o reduce he sysem complexiy. The approach of CL wih saionary robos was firs sudied by Kurazume e al. [16] [17]. The auhor proposes a mehod called Cooperaive Posiioning Sysem(CPS). Wihin he CPS, a eam of robos are divided ino a saionary group, which acs as landmarks, and a mobile group. The roles of groups would reverse a he nex ime insan unil he eam reaches he desinaion. More relaed issues are discussed in [18] and [19]. A leader-assisive approach is presened in [2] and [21], where a eam of robos are divided ino wo groups, he performance-plus robos ac as leader and he res as child. The leaders have high-resoluion esimaors and help o localize he child robos. Such an approach is furher exended in [22]. The main limiaion of hese approaches is ha oo many consrains(such as he order of movemen) are required on he robo moion. In our approach, each robo moves wih is own pah. III. SYSTEM SETUP AND PROBLEM STATEMENT Consider a group of N robos navigaing in 2D environmen, each robo is equipped wih propriocepive and exerocepive sensors. The sae of individual robo a ime is denoed as X = [x, y, θ ] T, which conains he informaion of posiion and orienaion. In his approach, we adop he velociy moion model wih he assumpion ha a mobile robo can be conrolled hrough a roaional and ranslaional velociy denoed by v and ω respecively.
Thus, he conrol vecor is denoed as u = [ v ω ] T. I derives from he propriocepive sensors such as encoder. On he basis of moion model, a ime, he sae of robo can be presened as following equaion: X = g (u, X 1) + N (, R ) v ω sin θ 1 + v ω sin(θ 1 + ω ) g (u, X 1) = X 1 + v ω cos θ 1 v ω cos(θ 1 + ω ) ω (1) where is he duraion of a ime sep, N (, R ) is he moion noise in an addiive Gaussian wih zero mean and covariance marix R. Addiionally, when one robo deecs anoher by exerocepive sensors a ime, he relaive measuremen z = T r φ, r and φ are relaive disance and bearing respecively beween wo robos. The measuremen model for robo i measuring robo j a ime is z i,j h X i, X j = ( = h X i, X j ) + N (, Q ) ( (x j x i ) 2 +(y j y i ) 2 a an 2(y j y i, x j x i ) θ i ) (2) Here, X i and X j are he acual sae of robo i and robo j. N (, Q ) is he addiive Gaussian noise wih zero mean and covariance marix Q. A each ime sep, each robo is provided wih an esimae X of is own sae and he corresponding covariance marix Σ which represens he uncerainy. The available daa is odomery of each robo and all relaive measuremens beween each wo robos. For cenralized scheme, cener robo sores and processes all daa including he sae and measuremen of each robo. I can be denoed as: X = [ˆx 1, ˆx 2, ˆx, ˆx...ˆx i ] Z = [z 1,2, z,...z i,j ] [ x 1, x 2, x, x... x i ] = filer(x, Z) For our decenralized scheme, each robo i only sores is las sae, odomery and measuremen o esimae curren sae. Depending on he odomery and moion model, each robo i has a prediced sae: X i = g ( u i, X i ) 1 () When he robo i have several relaive measuremens wih n robos. The objecive of our approach is o find a proper esimaor which mees he following equaion: X i = arg min = arg min (z i,j z i,j ) ( ) z i,j h X i, X j IV. PROPOSED APPROACH () () The purpose of our approach is o localize a group of mobile robos navigaing in feaureless environmen depending on he relaive measuremen beween wo robos. We have he following assumpions: i) Each robo is equipped wih propriocepive and exerocepive sensors o obain daa of odomery and relaive measuremen. ii) Each robo is marked wih ags in order o be deeced. iii) Robo is able o communicae wih each oher. In our localizaion algorihm, a any given ime, firsly, one robo is chose randomly o remain saionary. Secondly, he moving robo which is closes o he saionary robo esimaes is sae by localizaion algorihm wih available relaive measuremen. Sequenially, he res of moving robos can uilize he measuremen wih former esimaed robo. This process is usually broken ino wo seps, namely predicion and correcion. In he following subsecion, we would derive he equaion of his process. A. Predicion Suppose ha here are one saionary robo j and several moving robos i a ime. Robo predics is sae mainly by he odomery noed as u. Thus, he prediced sae of saionary robo j is same as ha a previous ime sep 1: X j = X j 1 Σ j = Σ j 1 Moving robos i change is sae during each ime sep. Therefore, depending on sandard EKF, he prediced sae and he corresponding covariance marix can be derived by he following equaion: X i = g ( u, X i ) 1 Σ i = G Σ i 1G T + V M V T where, X i 1 and Σ i 1 is he esimaed sae and corresponding covariance marix a ime 1. G Σ i 1G T represens he uncerainy of robo i and V M V T provides an approximae mapping beween he moion noise in conrol space o he moion noise in sae space. M is he covariance marix of he noise in conrol space. And Jacobian G is he derivaive of he funcion g wih respec o he sae vecor, V is he derivaive of he funcion g wih respec o conrol vecor. The esimae is calculaed based on he odomery in he predicion sep. However, his sep is similar o ha would have a cumulaive error over a long disance. Hence, we need fuse he relaive measuremen o correc he esimae and reduce he uncerainy. B. Correcion The order of he robos o be correced wih he prediced sae depends on he relaive disance o he saionary robo j. We assume he closes one is robo k, is prediced sae and corresponding variance marix a ime are X k and Σ k. Hence, we have he predicion of relaive measuremen beween k and j: z k,j = h X k, X j (8) (6) (7)
We uilize he difference beween he measuremen z k,j colleced from exerocepive sensors and he prediced measuremen z k,j calculaed in (8) o correc he esimaed sae. There are hree facors ha would influence he accuracy of measuremen, including he uncerainy of saionary robos, he uncerainy of moving robo and he Gaussian noise in exerocepive sensors. Hence, we have: S k,j = H k Σ k [ ] H k T + H j Σ j [ H j ] T + Q (9) Here, H k is he Jacobian of h wih respec o he sae vecor of moving robo k. H j is he Jacobian of h wih respec o he sae vecor of saionary robo. Q is he covariance marix of measuremen noise. This equaion represens he noise mapping ino he measuremen space. Then, depending on sandard EKF, we can calculae he Kalman gain as following: K k,j = Σ k [ H k ] T ( S k ) 1 (1) Fusing he previous formulas, he esimae X k and corresponding variance marix Σ k of moving robo k can be updaed by he following equaion: X k = X k + K k,j z k,j z k,j ( ) (11) Σ k = I K k,j H k Σ k Here I is he ideniy marix of he same dimensions as K k H k. So far, we obain he esimae sae and corresponding covariance marix of moving robo k by fusing he odomery and relaive measuremen wih saionary robo j a ime. Essenially, we have he assumpion ha all probabiliies of measuremen beween robos are independen ha enable us o incremenally add he informaion from muliple robos ino our filer [2]. Therefore, he res of he robos can uilize he measuremen of esimaed moving robo like robo k o esimae is sae efficienly. For he remainder moving robo m and n esimaed robo: X m = Σ m = ( ( K m,j z m,j n ( I K m,j H m )) z m,j ) (12) Accordingly, he belief derive from he saionary robo propagaes o all he robos ha reduces he uncerainy of moving robo and improves he sabiliy. C. Accuracy and Consisency The esimaed sae of robo is represened by he approximaion and corresponding covariance marix. To evaluae he accuracy of his esimaor, we calculae he roo mean square error(rmse) beween wo sae vecors. The definiion is as following: posiion y (uni:m) 1-1 -2 - - - -6 groundruh -7 6 posiion x (uni:m) Fig. 2. The rajecory of one robo calculaed from and De-EKF and he groundruh RMSE := 1 n X X 2 (1) i=1 X and X is he acual and esimaed sae. RMSE indicaes how similar he acual and esimaed sae are over n ime seps. A. Simulaion V. SIMULATION AND EXPERIMENT The performance of our approach was esed in simulaion wih N = robos navigaing in an area of approximaely 1 m 1 m for 1 ime seps, each ime sep has a duraion of. sec. In order o validae he performance, we have he following seing: The moion model and measuremen model of all robos are idenical which conforms o (1) and (2). Each robo has individual ranslaional velociy v =.2 m/s and roaional ω =. randn rad/s, he randn is a normally disribued random number. The iniial pose error(m,m,rad)=[.1,.1,.1], he iniial acual posiion(m,m,rad)=[ randn, randn, 2 randn] and he measuremen error(m,rad)=[.1,.1]. Under he above seing, he posiion of one robo calculaed from our approach, decenralized EKF(De-EKF) and he dead-reckoning() are shown in Fig. 2. We can see ha, he and De-EKF have similar performance in he beginning. However, as he disance increases, esimaor from have a cumulaive error, while ha from De-EKF accompanies he groundruh consanly. Then, we compare he performance of he following approaches: i) which esimaes he sae depending on odomery only. ii)l-ekf ha one robo esimaes is sae by four known landmarks. iii)our approach() ha five robos esimaes heir saes mainly uilize he relaive measuremen. We employ RMSE o indicae he accuracy of hese approaches. The resul of one robo is depiced in Fig.. As expeced, he has he wors performance, while he L-EKF has he bes performance because he known landmark wih absolue posiion can reduce he uncerainy
.12 RMSE orienaion (uni:rad) RMSE posiion (uni:m). L-EKF..2.2.1.1. L-EKF.1.8.6..2 1 2 6 7 8 9 1 1 ime sep (. s) 2 6 7 8 9 1 ime sep (. s) (a) (b) mini PC which is a Iner NUC on our Turlebo2. The experimen environmen is shown in Fig. 6. Five Turlebo2 navigae in his environmen wih he moion sraegy as we menioned previously. The experimen resul is shown in Fig. 7 and Fig. 8. We can see he sae of five robos calculaed by our approach approximaes o he groundruh. The RMSE of posiion and orienaion is less han.m and. rad respecively. TABLE I H ARDWARE U SED IN E XPERIMENT Fig.. Comparison of,, L-EKF: (a) RMSE of robo posiion; (b) RMSE of robo orienaion; Hardware Type Descripion Turlebo2 base YMR-K1-W1 Provide Odomery Full HD Camera Microsof LifeCam 12 Relaive pose deecion Kinec Microsof Kinec V1 Relaive pose deecion rigid body marker - Unique feaure.12 Aprilag Tag6h11 Unique feaure.1 Inel NUC Ki J76-9 Daa processing Opirack Camera Prime 1 Provide ground ruh.2 9% Z 7% Z % Z 2% Z 1% Z RMSE posiion (uni:m).18.16.1.8.6 TABLE II S OFTWARE U SED IN E XPERIMENT..2 Sofware Version Descripion Opirack Moive 2. Sysem driver ROS camera driver 1 2 6 7 8 9 1 ime sep (. s) Fig.. Comparison of he performance wih insufficien measuremens of robo efficienly. Our approach is less accurae han he LEKF, since he porable landmarks which are saionary robo or esimaed robos have heir own uncerainy. However, our approach do no need he known landmark. And i s capable for feaureless environmen. Addiionally, he accuracy of our approach is close o ha of L-EKF. In order o es he sabiliy and robusness of our approach, we simulae he condiion wih insufficien relaive measuremen. The performance of our approach wih 1%9% measuremen in each ime sep is compared in Fig.. I depics ha wih over % available measuremen, our approach has normal performance as usual. Even hough here are 2% available measuremen, our approach sill has a moderae performance ha he RMSE in posiion is approximaely.12m. usb cam..6 Openni camera driver 1.11.1 ROS Kinec driver Aprilags2 ROS.9.8 Provide relaive pose ROS Kineic Robo operaing sysem Mocap Opirack 2.8. Opirack driver Fig.. Srucure of Turlebo2 B. Experimen Fig. 6. Experimen environmen 8 6 posiion y (uni:m) An experimen wih five modified Turlebo2 is performed o validae he performance of our approach. The hardware and sofware used in experimen are specified in able I and II. Turlebo2 is a low-cos personal wheeled robo ki wih open-source sofware. The srucure of modified Turlebo2 is illusraed in Fig.. Each Turlebo2 is equipped wih encoder o acquire odomery. The exerocepive sensors are srucure ligh camera kinec and full HD camera locaed in fron and back of his robo o acquire relaive measuremen. The acual informaion of robo posiion and roaion are provided by moion capure sysem(opitrack) wih error less han.mm and. respecively. Two feaures are applied on Turlebo2. One is he rigid body marker, used o be deeced by OpiTrack. Anoher is Aprilag, used o be deeced by camera. The daa of each robo is processed in 2 groundruh -2-2 6 8 1 posiion x (uni:m) Fig. 7. Trajecory of five robos 12 1
6 7 8 9 1 RMSE posiion (uni:m).....2.2.1.1. ime sep (. s) (a) Turlebo2 posiion RMSE orienaion (uni:rad)......2.2.1.1. Turlebo2 orienaion 6 7 8 9 1 ime sep (. s) Fig. 8. experimen resul(a) RMSE of robo posiion; (b) RMSE of robo orienaion; (b) VI. CONCLUSION AND FUTURE WORK In his paper, a decenralized approach wih improved EKF for cooperaive localizaion is proposed. Compared wih oher muli-robo localizaion mehods, he proposed approach is advanageous in high accuracy, scalabiliy and robusness of sae esimaion, as well as low hardware cos and compuaional complexiy. This approach provides a soluion for localizing a group of robos in feaureless environmens. The experimen resuls show ha he proposed approach can yield similar localizaion accuracy as he landmark-based mehod does, wihou using any pre-known landmarks. For five robos, he RMSE of posiion and orienaion is less han.m and. rad respecively. When he measuremens are insufficien, he proposed approach sill can yield a moderae localizaion performance. However, he sysem performance would decline when he robos move a high speeds. Fuure work will be focused on addressing his problem hrough using high sampling raes and faser esimaion algorihms. REFERENCES [1] I. J. Cox, BlancheAn Experimen in Guidance and Navigaion of an Auonomous Robo Vehicle, IEEE Transacions on Roboics and Auomaion, vol. 7, no. 2, pp. 19 2, 1991. [2] S. Saeedi, M. Trenini, M. Seo, and H. Li, Muliple-Robo Simulaneous Localizaion and Mapping: A Review, Journal of Field Roboics, vol., no. 1, pp. 6, 216. [] J. Borensein and L. Feng, Umbmark: A mehod for measuring, comparing, and correcing dead-reckoning errors in mobile robos, 199. [] H. R. Evere, J. Borensein, and L. Feng, Sensors for Mobile Robos, no. November 198, 199. [] R. Masinjila and P. Payeur, Consisen mulirobo localizaion using heurisically uned exended Kalman filer, Proceedings - 217 IEEE h Inernaional Symposium on Roboics and Inelligen Sensors, IRIS 217, vol. 218-Janua, pp. 297, 218. [6] A. Howard, M. Maark, and G. Sukhame, Localizaion for mobile robo eams using maximum likelihood esimaion, IEEE/RSJ Inernaional Conference on Inelligen Robos and Sysem, vol. 1, pp. 9, 22. [Online]. Available: hp://ieeexplore.ieee.org/lpdocs/ epic/wrapper.hm?arnumber=1128 [7] E. D. Nerurkar and S. I. Roumeliois, Asynchronous muli-cenralized cooperaive localizaion, IEEE/RSJ 21 Inernaional Conference on Inelligen Robos and Sysems, IROS 21 - Conference Proceedings, no. i, pp. 2 9, 21. [8] L. Luf, T. Schuber, S. I. Roumeliois, and W. Burgard, Recursive decenralized localizaion for muli-robo sysems wih asynchronous pairwise communicaion, Inernaional Journal of Roboics Research, pp. 1 16, 218. [9] R. Madhavan, K. Fregene, and L. E. Parker, Disribued heerogeneous oudoor muli-robo localizaion, in Roboics and Auomaion, 22. Proceedings. ICRA 2. IEEE Inernaional Conference on, vol. 1. IEEE, 22, pp. 7 81. [1] A. Marinelli, F. Pon, and R. Siegwar, Muli-robo localizaion using relaive observaions, Proceedings - IEEE Inernaional Conference on Roboics and Auomaion, vol. 2, no. April, pp. 2797 282, 2. [11] A. Howard, Muli-robo simulaneous localizaion and mapping using paricle filers, The Inernaional Journal of Roboics Research, vol. 2, no. 12, pp. 12 126, 26. [12] E. D. Nerurkar, S. I. Roumeliois, and A. Marinelli, Disribued maximum a poseriori esimaion for muli-robo cooperaive localizaion, Proceedings - IEEE Inernaional Conference on Roboics and Auomaion, pp. 12 19, 29. [1] A. Franchi, G. Oriolo, and P. Segagno, Muual localizaion in mulirobo sysems using anonymous relaive measuremens, Inernaional Journal of Roboics Research, vol. 2, no. 11, pp. 12 122, 21. [1] A. Prorok, A. Bahr, and A. Marinoli, Low-cos collaboraive localizaion for large-scale muli-robo sysems, in Roboics and Auomaion (ICRA), 212 IEEE Inernaional Conference on. Ieee, 212, pp. 26 21. [1] L. Luf, T. Schuber, S. I. Roumeliois, and W. Burgard, Recursive decenralized collaboraive localizaion for sparsely communicaing robos. in Roboics: Science and Sysems, 216. [16] R. Kurazume, S. Nagaa, and S. Hirose, Cooperaive posiioning wih muliple robos, in ICRA, vol. 2, 199, pp. 12 127. [17] R. Kurazume, S. Hirose, S. Nagaa, N. Sashida, and K. Nakahara-ku, Sudy on Cooperaive Posiioning Sysem, Inernaional Conference on Roboics and Auomaion, no. April, pp. 121 126, 1996. [18] R. Kurazume and S. Hirose, Sudy on cooperaive posiioning sysem: opimum moving sraegies for cps-iii, in Roboics and Auomaion, 1998. Proceedings. 1998 IEEE Inernaional Conference on, vol.. IEEE, 1998, pp. 2896 29. [19], An experimenal sudy of a cooperaive posiioning sysem, Auonomous Robos, vol. 8, no. 1, pp. 2, 2. [2] T. R. Wanasinghe, G. K. Mann, and R. G. Gosine, Disribued collaboraive localizaion for a heerogeneous muli-robo sysem, in Elecrical and Compuer Engineering (CCECE), 21 IEEE 27h Canadian Conference on. IEEE, 21, pp. 1 6. [21], Disribued leader-assisive localizaion mehod for a heerogeneous muliroboic sysem, IEEE Transacions on Auomaion Science and Engineering, vol. 12, no., pp. 79 89, 21. [22] B. E. Nemsick, A. D. Buchan, A. Nagabandi, R. S. Fearing, and A. Zakhor, Cooperaive inchworm localizaion wih a low cos eam, Proceedings - IEEE Inernaional Conference on Roboics and Auomaion, pp. 62 6, 217. [2] S. Thrun, W. Burgard, and D. Fox, Probabilisic roboics. MIT press, 2.