Localizing Objects During Robot SLAM in Semi-Dynamic Environments

Transcription

1 Proceedings of he 2008 IEEE/ASME Inernaional Conference on Advanced Inelligen Mecharonics July 2-5, 2008, Xi'an, China Localizing Objecs During Robo SLAM in Semi-Dynamic Environmens Hongjun Zhou Tokyo Meropolian Indusrial Technology Research Insiue, Japan Shigeyuki Sakane Chuo Universiy, Japan Absrac Mapping and localizaion play imporan roles for auonomous mobile robos. Since mos of he convenional mapping mehods assume saic environmen, he obained map lacks reliabiliy of localizaion when he assumpion does no hold in he real environmen. Addiionally, when a robo wans o do some asks (such as, open doors, push cabines, ec.) during SLAM process, he robo need localize he arge objecs. This paper deals wih mapping and localizaion for mobile robo in environmens including semi-dynamic objecs which change heir poses occasionally, such as mobile file cabines, chairs, and doors. We assume he semi-dynamic objecs are arge objecs for robo asks. We assume RFID ags wih unique ubiquious idenificaion code (ucode) are aached he objecs in place of barcodes currenly used. So he ag informaion allows us o recognize exisence of semi-dynamic objecs in he environmen when a mobile robo deecs he RFID ags. To perform mapping and localizaion for mobile robo, we propose SLAM-SD, an exended SLAM mehod o cope wih he semi-dynamic environmens. The mehod employs a framework of Dynamic Bayesian Nework and RBPF (Rao-Blackwellised Paricle Filer) which allows simulaneous localizaion of a robo, mapping of he environmen, and also localizaion of semi-dynamic objecs. The sysem updae he occupancy grid map properly when semi-dynamic objecs changed he poses. We conduced experimens using a mobile robo mouned a laser range-finder and an RFID ag anenna. The resuls show effeciveness of he proposed mehod. Index Terms mobile robo, mapping, localizaion, SLAM, semi-dynamic environmen, RFID ag I. INTRODUCTION Mapping and localizaion play imporan roles for auonomous mobile robos. Mapping requires localizaion of a robo while he localizaion requires he map of he environmen. So he inerdependence of localizaion and mapping is essenially a chicken-and-egg problem. In recen years, much progress has been made for simulaneous localizaion and mapping (SLAM) [1] [10]. A problem in mos of he SLAM mehods is hey assume saic environmens. Therefore, if dynamic objecs exis in he environmen, heir pose changes can lead o unreliable mapping and localizaion. Addiionally, The robo also need localize arge objecs of environmen during SLAM process. For example, he robo wan o open a door, bu i does no know he pose of he door, so he robo have o searching he door while i is doing SLAM. To clear he problem, we classify objecs of indoor environmens ino hree caegories: saic objecs, dynamic objecs, and semi-dynamic objecs by aking ino accoun of he frequency of heir pose changes. The saic objecs move rarely in he environmen, such as walls and heavy desks. Human is a dynamic objec which move quie frequenly. The semi-dynamic objecs move occasionally such as mobile file cabines, chairs, and doors. We also assume he semidynamic objecs is he arge objecs, and he robo need localize he objecs while i is doing SLAM. Some approaches [1], [2] have been proposed for mapping in dynamic environmens. A robo racks movemens of people and builds a map of he environmen simulaneously. Effecs of he movemens are canceled from he mapping resuls. During navigaion, he robo deecs he movemens and reduce errors in he localizaion. However, if he environmen has semidynamic objecs, hey may rarely change he poses when he robo performs mapping and he robo will find difficuly in deecing he pose change of semi-dynamic objecs. We proposes SLAM-SD, an exended SLAM mehod for mobile robo in semi-dynamic environmens, in which he robo deecs he pose change of he semi-dynamic objecs and updaes he environmenal map. We aach an RFID ags o semi-dynamic objecs o disinguish i from saic objecs. SLAM-SD uilizes boh range scan daa obained by LRF and RFID ag informaion obained by RFID ag anenna. The mehod employees a framework of DBN (dynamic Bayesian nework) o represen dependency among various informaion: robo pose, odomery, pose of semi-dynamic objec, laser scanning daa, he RFID ag informaion, and he map (we use occupancy grids). The opimal map, robo pose, he semi-objec pose are esimaed simulaneously using RBPF [11], [12]. When he robo deecs pose changes of semidynamic objecs, he mehod properly updaes corresponding occupancy grids. II. PREVIOUS WORKS Denis and Wang [1], [2] have presened he soluions of SLAM for dynamic environmens. Since human (a dynamic objec) moves coninuously, he robo can rack he human movemen during he SLAM process and he robo can reduce localizaion error by canceling effecs of he movemens. However, if a semi-dynamic objec moves occasionally, i will be difficul o disinguish he saic and semi-dynamic objecs using he above soluion. Biswas e al. [17] proposed an EM approach o deec chairs (semi-dynamic objecs). Since he mehod requires an off-line learning phase, i can no deal wih he on-line SLAM and navigaion asks. Anguelov e al /08/$ IEEE. 595

2 [19] proposed o learn he laser daa and color image o deec walls and doors using an EM algorihm. Avos e al. [18] esimae saus of doors and localize he pose of he robo simulaneously. However, he approach does no deal wih he SLAM problem. Sachniss e al. [23] clusers local grid maps and uilize he maps wihin a RBPF for SLAM. A problem in using paricle filers o SLAM is curse of dimensionaliy since many variables such as maps and robo poses mus be esimaed simulaneously. Murphy e al. [11] and Douce e al. [12] proposed use of RBPF o cope wih SLAM problem. Hahnel [4] and Elianzer e al. [5] performed real robo experimens of simulaneous robo localizaion and mapping of occupancy grid using RBPF. Each paricle of he RBPF corresponds o a robo pose and is map. The joined probabiliy of he robo pose and is map is esimaed using he laser range finder scan score. Grisei e al. [6] proposed an approach o improve precision of gridbased SLAM using an accurae proposal disribuion. Based on he proposal disribuion he robo decreases re-sampling operaions o reduce he paricle depleion. However, none of hese SLAM mehods ake ino accoun he pose changes of semi-dynamic objecs. In conras, our SLAM-SD mehod exends he prior SLAM echniques o perform SLAM and localize he semi-dynamic objecs simulaneously. The environmen map can be updaed based on he deeced pose changes of semi-dynamic objecs. Hahnel [14] localizes he locaions of he RFID ags using paricle filer [15]. Since he sysem needs a previously buil grid map, i is differen from our approach. III. ASSUMPTIONS AND PRECONDITIONS In his paper, we assume an RFID ag is aached o a semi-dynamic objec. So a robo can disinguish he saic objecs and semi-dynamic objecs by RFID ag signal. The precondiions of he proposed approach are he followings: 1) The 3D geomerical informaion of a semi-dynamic objecs can be obained hrough inerne based on he ID (ucode [22]) of he RFID ag. 2) The RFID ag is aached o a semi-dynamic objec, and he relaive coordinae beween he objec and he RFID ag is also obained as in he previous condiion. 3) Two objecs wih he same shape do no exis simulaneously in he deecable field of an RFID anenna. The precondiion (1) is expeced in fuure as in he proposal [20] which uilizes disribued knowledge robo asks obained via inerne. The precondiion (2) will help he robo o esimae he pose of he semi-dynamic objec using he esimaed RFID ag s pose. The precondiion (3) is inroduced o simplify he problem so ha we can esimae he pose of he semidynamic objec in he sensing area of he RFID anenna using he geomerical informaion. If here are wo objecs have he same shape, he pose esimaion of he objecs will be difficul. IV. SLAM-SD SLAM-SD uses RBPF o esimae he robo pose, he environmen map and he pose of he semi-dynamic objecs Fig. 1. U-1 D-1 D D+1 U U+1 X-1 X X+1 Z-1 Z Z+1 M-1 M M+1 A dynamic Bayesian nework (DBN) model of SLAM-SD. simulaneously. Fig.1 show a DBN which represens he dependency relaionships beween he environmen map (M), he robo pose (X), he pose of he semi-dynamic objecs (D), and sensor daa(z). The node X and he node D indicae he robo pose and he pose of he semi-dynamic objec a ime, respecively. The node Z is he sensor daa obained a ime. Z includes boh range scan daa obained by LRF and RFID ag informaion obained by RFID ag anenna. The odomery daa and he environmen map are indicaed by he node U and he node M, respecively. A. Pose Esimaion of Robo and Semi-dynamic Objecs Eq.1 shows he condiional probabiliy of he hidden nodes X and D when he observaion node Z and U T are given. We can rewrie he Eq.1 o Eq.1. P (X 1:,D 1:,M Z 1:,U 0: ) = P (M X 1:,Z 1: ) P (X 1:,D 1: Z 1:,U 0: ) (1) P (M X 1:,Z 1: ) of he Eq.1 is he probabiliy of he occupancy grid map which is buil by he sensor daa based on he esimaed robo pose [13]. P (X 1:,D 1: Z 1:,U 0: ) is he poserior probabiliy of he pose of he robo and semidynamic objec when Z 1: and U 0: are given. The probabiliy will be feedbacked o P (M X 1:,Z 1: ) for he environmen mapping. The pose calculaion of he robo and he semidynamic objec consiss of wo seps: predicion and updae. The robo esimaes he joined probabiliy of he robo and semi-dynamic objec pose (X,D ) based on he sensor daa which is obained from ime 1 o ime 1 in he predicion sep. Predicion: P (X,D Z 1: 1,U 0: 1 ) = P (X X 1 ) P (D D 1 ) X 1 D 1 P (X 1,D 1 Z 1: 1,U 0: 1 ) (2) We replace he P (X 1: 1,D 1: 1 Z 1: 1,U 0: 1 ) wih B 1. Then, Eq.2 will be rewrien o he following form: 596

3 200cm 5 95 objec 180cm (a) 150cm 300cm (b) 300cm Fig. 2. (a) Deecable field of he RFID anenna. (b) An approximae sensor model of he RFID anenna Fig. 3. Robo laser beam Evaluaion of he obained range profiles using ray racing. = P (X X 1 ) P (D D 1 ) B 1 (3) X 1 D 1 In he updae sep, he prediced probabiliy (Eq.2) is muliplied by he prior probabiliy of he sensor daa o calculae B. Updae: B = P (X 1:,D 1: Z 1:,U 0: ) = η P (Z,U X,D ) P (X,D Z 1: 1,U 0: 1 ) = η P (Z,U X,D ) P (X X 1 ) P (D D 1 ) B 1 X 1 D 1 (4) η is a normalizing consan for he probabiliy and P (Z,U X,D ) is he prior probabiliy of he sensor daa. The expansion of he prior probabiliy is he following: P (Z,U X,D )=P(Z D ) P (Z,X ) = P (I D ) P (L D ) P (L,X ) (5) In Eq.(5), I indicaes exisence of RFID ag, which is deeced by he RFID anenna. L indicaes range scan daa obained by LRF. P (I D ) and P (L D ) are he prior probabiliies of he LRF and RFID ag, respecively, when he pose of he semi-dynamic objec is given. P (L,X ) indicaes he prior probabiliy of he LRF when he pose of he robo is given. We call prior probabiliy of he sensor daa P (Z,U X,D ) as sensor model. B. Sensor Model The prior probabiliy of RFID anenna (P (I D )) is calculaed based on acual measuremens. In fron of he RFID ag, we se a 200[cm] 200[cm] space for RFID ag measuremen. The space has been divided ino 10[cm] 10[cm] cells, he ag anenna lisens he RFID ag for 100 imes a each cells. Fig. 2 shows frequency of sensed RFID ag signals. The dark color area indicaes he anenna can receive he RFID ag signal more imes han he ligh color area. The ellipse of Fig.2(b) (200[cm]x180[cm]) shows an approximaed area in which he ag anenna can receive he RFID ag signal. We use he ellipse o represen he sensor model of he anenna. We se he likelihood of he dark area is 95, and he ligh area is 5 (Fig.2(b)). P (L D ) of Eq.5 is he prior probabiliy of he LRF. The probabiliy is calculaed by he evaluaion of LRF scan score and he shape of he objec. We simulae he LRF daa from each candidaure pose of he robo o he surface of he objec using ray racing. Then evaluae differen beween he simulaed sensor paern and obained LRF sensor paern. The evaluaed score will be he likelihood P (L D ). The robo also performs he scan mach [6] for he robo pose esimaion. P (L X ) is he likelihood of he scan score, and he score will be he prior probabiliy of he LRF sensor daa for he robo pose. C. Esimaion using RBPF To be precondiion, he robo can ge he geomerical informaion of he semi-dynamic objec and relaive coordinae beween RFID and he objec hrough ID of RFID ag. The pose of he objec can be esimaed from he pose of he RFID ag. The pose of he RFID ag has hree parameers: x, y and he normal direcion of he RFID ag. The robo pose also has hree parameers: x, y and he direcion of he robo head (θ). The pose of he robo and objec are esimaed by Eq.4. We esimae he Eq.4 using sequenial imporance sampling. The sae ransiion probabiliy of he robo pose and he semi-dynamic objec are P (D D 1 ) and P (X X 1 ), respecively. q(x 1:,D 1: Z 1: 1 )=P(D D 1 ) P (X X 1 ) P (D 1: 1,X 1: 1 Z 1: 1 ) (6) q(x 1:,D 1: Z 1: 1 ) is he proposal disribuion of he robo pose and he objec pose based on he observed sensor daa is given a ime 1. The disribuion will updae he B using he flowing weigh parameer. ω = P (X 1:,D 1: Z 1: ) q(x 1:,D 1: Z 1: 1 ) P (Z,U X,D ) (7) Esimaion of he environmen map, he pose of he robo and he semi-dynamic objec are performed by he flowing ieraion. 597

4 SLAM SD Algorihm 1 Iniializaion ( =1) N paricles are generaed in he RFID anenna sensing area using uniform disribuioni =1,..., N, and se ime =2. 2 P redicion of P aricle Transi he paricles from he sae of ime 1 o ime. The ransiion probabiliy is shown as following: D (i) X (i) D (i) P ( X (i) P ( (i) D D (i) 1 ) (i) X X (i) 1 ), indicae he predicaed paricle sae (i) based on he ransiion probabiliy (P ( D D (i) 1 ) (i) and P ( X X (i) 1 )). Paricle evaluaion based weigh. ω (i) P (Z,U D (i) = P (I D (i),x (i) ) ) P (L D (i) ) P (L,X (i) ) The above formulaion is he likelihood for paricle updae via RFID anenna and LRF sensor. Normalize he calculaed likelihood. 3 Updae of P aricles: Resample he paricle based on he calculaed likelihood using Residual Resampling algorihm [15]. Calculae he parameer o evaluae he effecive of he paricles. N eff = 1 N i=1 (ω(i) ) 2 The esimaed pose of he robo and he RFID ag will be he pose of he paricle which hold he bigges weigh, moreover, he seleced paricle mus be in he paricle se which has bigges N eff. 4 Map Building: Generae he map M (1) which corresponds o each paricle using probabiliy P (M (1) X (i),z 1: ) Se he ime from o +1. As shown in he sep 1 of SLAM-SD algorihm, iniially, he robo generaes he paricles of he probabiliy P (X 1,D 1 ) based on he uniform disribuion. P (X 1,D 1 ) is he joined probabiliy of he robo and semi-dynamic objec pose. Since he robo and objec pose are hree-dimensional variable respecively, he paricle which represens he robo and objec pose is six-dimensional vecor. I is very difficul o esimae a six-dimensional hidden variable using sequenial imporance sampling. However, we know he iniial pose of he robo and sensing area of he anenna, he iniial paricle will be generaed in he sensing area of he anenna. We can esimae he six-dimensional variable in he limied area. As shown in Fig.4, wo RFID anennas are mouned on he robo, if he robo receive he RFID ag signal, he paricles will be generaed uniformly in he ellipse (200[cm] 180[cm]) Fig. 4. A mobile robo and sensors (RFID anenna, laser range-finder) used in he experimens. Fig. 5. (lef), An RFID ag used in he experimens. (righ), A card which embeds an RFID ag. area shown in Fig.2(b) We use a six-dimensional vecor (I x,i y,i θ,x x,x y,x θ ) o represen he paricle. X x,x y,x θ is he coordinae of he robo pose, and I x,i y,i θ is he coordinae of he RFID ag which is aached on he semidynamic objec. For example, if K paricles are used for he robo pose esimaion, and corresponding o each robo pose, M paricles are used for he RFID ag localizaion, hen N = K M paricles are necessary for he pose of he robo and he RFID ags. As shown in he sep 2 of SLAM-SD algorihm, The robo predics he paricle sae of he nex ime based on he sae ransiion probabiliy. Then he robo resamples he paricle based on he weigh in he sep 3. Afer he resampling sep, he robo calculaes he N eff o evaluae he effecive of he paricles. The esimaed pose of he robo and he RFID ag will be he pose of he paricle which hold he larges weigh, moreover, he seleced paricle mus be in he paricle se which has he larges N eff. The pose of he semi-objec will be calculaed by he pose of he RFID ag. Finally, we can esimae he map using he pose of he robo and semi-dynamic objecs based on he probabiliy P (M X 1:,Z 1: ) of Eq.1. V. EXPERIMENTS To evaluae he effecive of our sysem, we conduced experimens of he SALM-SD mehod using he log daa 598

5 obained by a real mobile robo. We used a mobile robo, Pioneer3 (made by AcivMedia) and RFID reader/wrier wih wo anennas and RFID ags (953MHz, made by Fujisu). A LRF (LMS200, made by SICK) is mouned on he robo (Fig.4). The maximum disance deecable from he RFID anenna is 2[m], and he RFID ag has 256 byes memory including 192 byes user area. PC used for he experimens has a CPU of AMD Ahlon XP and 1 GB memory. The SLAM-SD algorihm is implemened based on GMapping developed by Grisei e al. [21]. A. Experimens in a room environmen Fig.6 shows iniial experimens of SLAM-SD in which a robo performs SLAM in an environmen in which a mobile file cabine exiss. If he robo does no deec he RFID ag, he robo goes forward and performs SLAM. Ten paricles were used for he SLAM of he robo. When he robo deeced he RFID ag, he paricles are generaed based on an uniform disribuion funcion in he sensing area of he RFID anenna (Fig.6(a)). To localize he RFID ag, 100 paricles are used o represen possible robo poses. The oal number of paricles for his SLAM-SD is N = K M= = Through he ieraion of he SLAM-SD algorihm (form sep 1 o sep 4), he paricles converge ino he posiion of he RFID ag (Fig.6(b)). Afer paricle convergence, we use a recangle o represen he pose of he cabine (Fig.6(c)). The posiion and orienaion of an arrow show he posiion and direcion of a normal vecor of he RFID ag. The iniial experimens show ha he robo can localize he semi-dynamic objec (cabine) and perform SLAM simulaneously. There are wo mobile file cabines in he environmen (Fig.7(a)). The robo localizes he second cabine using he above mehod (Fig.7(b),(c)). Afer localizaion of he wo cabines, he robo urns back and localizes wo cabines again. Since he pose of he firs cabine was changed (Fig.7(c)), when he robo finds he pose of he localized firs cabine was changed, he value of he grids which are occupied by he firs cabine will be updaed. The value of he grids which are occupied by he original cabine pose will be se 0, and grid value of he curren pose will be added. In his experimen, he size of he occupancy grid is 1[cm]x1[cm]. The compuaional ime of one SLAM-SD cycle (from sep 2 o sep 4) calculaion is abou 1.2 second. The experimenal resuls show SLAM-SD can updae he map appropriaely when he robo find he pose change of he semi-dynamic objec. B. Experimens in a corridor environmen The mobile robo localizaion and navigaion in an indoor environmen is effeced by he door sae change sricly. In his experimen, we change he saes of he doors and dus boxes o verify our algorihm. Iniially, we conduced an experimen in a shor corridor (10[m] 3[m]). There are wo doors and wo dus boxes. The pose of he red sick shows he pose of he door which is localized by he robo. And he pose of he red recangle shows he pose of he dus boxes. The gray arrow indicaes he pose of he RFID ag. The direcion of Fig. 7. An experimen o deec a cabine and o updae he map when i was moved o oher place. he gray arrow indicaes he normal direcion of he RFID ag. Fig.8(a,b,c) shows he phoos of he semi-dynamic objecs, he mapping resul of hese objecs are shown in he porions which are encircled by he cycles. The mapping resul verifies our algorihm can deal wih he pose change of he door (opening and closing) and dus box. To verify he generalness of SLAM-SD algorihm, we carry ou a experimen in a longer corridor. The environmen and he robo are shown in Fig.9(A), and he resul of he experimen is shown in Fig.9(D). There are 3 ypes of he semi-dynamic objecs, large size door, small size door and dus box in he environmen. The number of he large door, small door and dus box are 6, 7 and 11, respecively. The porions which are encircled by he cycles are zoomed in. In he upper por of Fig.9(D), zoomed porions and he phoos of he semidynamic objec which corresponds o he zoomed porions are shown. Fig.9(B) shows he mapping resuls of he middle of he corridor. A phoo shows he mapped door and a dus beside he image. As shown in Fig.9(B), he mapping resul can no be piled up on he pose of he door, i was piled up on he ouside of he dus box. The reason of error is ha he door and dus box are in he RFID sensing area, and anenna can deec RFID ags which are aached o he door and he dus (b) (c) (d) 599

6 (a) (b) (c) (d) Fig. 6. Experimens o deec a mobile file cabine and updae he map of a room when he cabine was moved. (A) (B) (C) (D) Fig. 9. An experimen o deec semi-dynamic objecs, dus boxes and doors, in a long corridor. box a same ime. In addiion, he shape of he door is similar o he ouside of he dus box, he robo can no disinguish beween he door and he ouside of he dus box by evaluaion of laser range paern which is generaed by ray racing and he range scan daa obained by LRF. Experimen resuls are shown in Table I. The small door localizaion has wo errors, and he oher objecs localizaion are very successful. The grid size of he generaed map is 2[cm]x2[cm], he compuaional ime for one SLAM-SD cycle (from sep 2 o sep 4) is abou 0.8 second. VI. CONCLUSION In his paper, we proposed SLAM-SD mehod for mobile robo in semi-dynamic environmens. The robo SLAM 600

7 (a) Fig. 8. An experimen o deec semi-dynamic objecs (dus boxes and doors) in a shor corridor. TABLE I RESULTS OF THE EXPERIMENTS TO DETECT SEMI-DYNAMIC OBJECTS (DUST BOXES AND DOORS) INALONGCORRIDOR. big size small size dus door door box Number of objecs Correcly localized objecs number and arge objecs localizaion are performed simulaneously. The mehod uilizes RFID ags o localize he semi-dynamic objecs in he environmen and he map can be updaed based on he localized poses of he robo and he semidynamic objecs. To validae effeciveness of he algorihm, we conduced experimens o deec semi-dynamic objecs such as mobile cabines in a room, dus boxes and doors in a corridor environmen. The boh experimenal resuls show ha he SLAM-SD works successfully for mobile robo in he semi-dynamic environmens. Fuure work will include he following issues: (1) use of he RFID ag informaion o access 3D models of he semidynamic objecs via inerne and (2) sensor fusion of image daa and range daa o reliably localize he semi-dynamic objecs. (b) (c) [7] M. Monemerlo, S. Thrun, D. Koller, B. Wegbrei, FasSLAM: A facored soluion o he simulaneous localizaion and mapping problem, In Proc. of he AAAI Naional Conf. on Arificial Inelligence, [8] M. Monemerlo, S. Thrun, D. Koller, B. Wegbrei, FasSLAM 2.0: An improved paricle filering algorihm for simulaneous localizaion and mapping ha provably converges, In Proc. of he Sixeenh In. Join Conf. on Arificial Inelligence (IJCAI), [9] S. Thrun, An online mapping algorihm for eams of mobile robos, In. Journal of Roboics Research, 20(5), pp ,2001. [10] J. Djugash, S. Singh, G.A. Kanor, and W. Zhang, Range-only SLAM for robos operaing cooperaively wih sensor neworks, In Proc. of he IEEE In. Conf. on Roboics and Auomaion, May, [11] K. Murphy, Bayesian map learning in dynamic environmen, Neural Info. Proc. Sysems (NIPS), [12] A. Douce, N. de Freias, K. Murphy and S. Russell, Rao-Blackwellised paricle filering for dynamic Bayesian neworks, Proc. of Conf, on Uncerainy in Arificial Inelligence (UAI), [13] H.P. Moravec, Sensor fusion in cerainy grids for mobile robos, AI Magazine, pages 61-74, Summer, [14] D. Hahnel, W. Burgard, D. Fox, K. Fishkin, and M. Philipose, Mapping and localizaion wih RFID echnology, In Proc. of he IEEE In. Conf. on Roboics and Auomaion (ICRA), [15] J. Liu and R. Chen, Sequenial Mone Carlo mehods for dynamical Sysems, Journal of he American Saisical Associaion, Vol. 93, pp , [16] J.S. Liu, Meropolized independen sampling wih comparisons o rejecion sampling and imporance sampling, Sas. Compu. 6: pp , [17] R. Biswas, B. Limkekai, S. Sanner, S. Thrun Towards objec mapping in dynamic environmens wih mobile robos, In Proc. of he IEEE/RSJ In. Conf. on Inelligen Robos and Sysems (IROS), [18] D. Avos, E. Lim, R. Thibaux, and S. Thrun, A probabilisic echnique for simulaneous localizaion and door sae esimaion wih mobile robos in dynamic environmens, In Proc. of he IEEE/RSJ In. Conf. on Inelligen Robos and Sysems (IROS), [19] D. Anguelov, D. Koller, E. Parker, and S. Thrun, Deecing and modeling doors wih mobile robos, In Proc. of he IEEE In. Conf. on Roboics and Auomaion (ICRA), [20] N.Y. Chong, H. Hongu, K. Ohba, S. Hirai, K. Tanie, A disribued knowledge nework for real world robo applicaions, Proc. of he IEEE/RSJ In. Conf. on Inelligen Robos and Sysems (IROS), pp , [21] G. Grisei; C. Sachniss; W. Burgard, GMapping, he source code is available a : hp://svn.openslam.org/daa/svn/gmapping. [22] Ubiquious ID Cener, hp:// [23] C. Sachniss and W. Burgard, Mobile Robo Mapping and Localizaion in Non-Saic Environmens, Proc. of he Naional Conference on Arificial Inelligence (AAAI), REFERENCES [1] Denis Wolf, e al., Online simulaneous localizaion and mapping in dynamic environmens, In Proc. of he In. Conf. on Roboics and Auomaion (ICRA), [2] C. Wang, C. Thorpe, and S. Thrun, Online simulaneous localizaion and mapping wih deecion and racking of moving objecs: Theory and Resuls from a Ground Vehicle in Crowded Urban Areas, Proc. of he IEEE In. Conf. on Roboics and Auomaion (ICRA), [3] J. Leonard and H. Feder, A Compuaionally efficien mehod for largescale concurren mapping and localizaion, In Roboics Research: The Ninh In. Symp. Snowbird, UT: Springer Verlag [4] D. Hahnel, W. Burgard, D. Fox, S. Thrun, A highly efficien FasSLAM algorihm for generaing cyclic maps of large-scale environmens from raw laser range measuremens, In Proc. of he IEEE/RSJ In. Conf. on Inelligen Robos and Sysems (IROS), [5] A. Elianzar and R. Parr, DP-SLAM: Fas, robus simulaneous localizaion and mapping wihou predeermined landmarks, In Proc. of he. In. Conf. on Arificial Inelligence (IJCAI), 2003 [6] G. Grisei, C. Sachniss, and W. Burgard, Improving grid-based SLAM wih Rao-Blackwellized paricle filers by adapive proposals and selecive resampling, In Proc. of he IEEE In. Conf. on Roboics and Auomaion (ICRA),