A Multi-Robot System Based on A Hybrid Communication Approach

Transcription

1 Studes n Meda and Communcaton Vol. 1, No. 1; June 13 ISSN E-ISSN 35-88X Publshed by Redfame Publshng URL: A Mult-Robot System Based on A Hybrd Communcaton Approach Tngka Wang, Quan Dang & Peyuan Pan Faculty of Computng, London Metropoltan Unversty, London N7 8DB, UK Correspondence: Tngka Wang, Faculty of Computng, London Metropoltan Unversty, London N7 8DB, UK. Tel: E-mal: t.wang@londonmet.ac.uk Receved: March 4, 13 Accepted: March 1, 13 Avalable onlne: Aprl 11, 13 do: /smc.v11.14 URL: Abstract Communcaton plays an mportant role n a mult-robot system. In a large system wth many robots, t s dffcult for all robots to exchange nformaton at a tme because of ther lmted communcaton capactes. On the other hand, sometmes there s such a stuaton where no explct communcaton s allowed between robots. Therefore, effcent and relable communcaton together wth non-explct communcaton s crucal for a mult-robot system. Ths paper presents a hybrd communcaton approach for a mult-robot system, whch combnes the explct wth mplct communcatons va usng the predcton of robotc behavour and a fuzzy communcaton approach. The hybrd communcaton approach contans a robot performance rule base, a fuzzy nference engne, and a semantcs and grammar for communcatng. Based on the hybrd communcaton strategy, the avodance of collson wth multple robots n workng area and multple robots transportng a common object have been explored. The results of smulaton show that the mult-robot system can complete a cooperatve task successfully. Keywords: mult-robot systems, fuzzy logc, cooperaton, mplct, hybrd communcaton 1. Introducton Recently, there has been a growng nterest n mult-robot systems. Compared wth a sngle robot, a mult-robot system ncreases robustness by takng advantage of nherent parallelsm and redundancy (Km, et al. 8). Especally when dong the envronmental detecton, survvors searchng and other complex tasks, a mult-robot system s ncomparably superor. However, smply puttng multple robots together cannot consttute a mult-robot system, f they all try to functon ndependently of each other. It may lead to a conflct, or a crash among the robots n a dynamc workng envronment. How to organze the mult robots wth effectve communcaton between them s a major objectve of a mult-robot system. For ths reason a great attenton s gven to mult robots communcaton. Varous communcaton approaches for mult-robot systems have been developed n recent years (Aaron et al. 11, Tarque & Maram, 9 and Yan et al. 6), whch can be classfed as explct and mplct communcatons. Implct communcaton s usually wthout regard to the messages others receve. It could be based upon the envronment change or perhaps the behavour of other robots. Yan et al. (6) appled ths technque to coordnate a team of robots that have lmted explct communcaton. In the area of robotc search, the use of a rendezvous between two searchng robots at a pre-arranged spot has been studed (Roy & Dudek 1). Jelle et al. (5) appled non-communcatve strategy nto a mult-robot soccer team n dynamc envronments. Castelfranch et al., (1, 1) studed the theory of behavoural mplct communcaton. Caplupp at al., (13) modelled mplct communcaton n herarchcal Hybrd. Although the mplct communcaton approach for mult robots can fulfl some tasks, explct communcaton can sgnfcantly mprove the flexblty and adaptveness of a mult-robot system. Snce the advent of hgh-performance wreless local area network (WLAN) at relatvely low cost, ts use for wreless communcaton among mult robots has become a practcal proposton (Long et al. 4). However, for most systems wth a large number of robots, communcaton capacty s stll lmted. The study on the effcent and relable communcaton approaches s stll consdered a hot topc of research. Iqbal et al., (9) and Kashyap & Yan (7) proposed dynamc message nterpretaton archtecture for mult robots 91

2 Degree of membershp Studes n Meda and Communcaton Vol. 1, No. 1; 13 communcaton whch has mproved the effcency n tme and storage. Ge Ran, et al. (7) presented an approach to mprovng the relablty of Wreless Sensor Networks, whch uses fuzzy logc to process the nformaton. Wang et al. (11) presented a predct-fuzzy logc communcaton approach and used the approach to mult-robot system competton. Ths paper, based on the paper (Wang et al. 11), further presents a hybrd communcaton strategy that ntegrates mplct and explct communcatons, and explores the strategy for mult robots cooperaton. The expermental and smulaton results of mult robots cooperaton demonstrate that the hybrd communcaton strategy s relable and effcency, and can be used to cooperate mult robots. The remander of the paper s organsed as follows. Secton presents the structure of a mult-robot system. The hybrd communcaton approach s studed n Secton 3. Secton 4 explores a case study and experments of multple robots transportng a common object. Fnally conclusons are gven n Secton 5.. The Hybrd Approach of Explct and Implct Communcatons The hybrd communcaton system contans a robot performance rule base, a fuzzy nference engne, and a semantcs and grammar for communcatng etc. Those components mplement both an explct communcaton approach - Fuzzy Logc Communcaton and mplct communcaton that s based on robot performance rule base, nference engne and sensor system..1 Explct Communcaton Approach - Fuzzy Logc Communcaton (Wang et al. 11) The Fuzzy Logc Communcaton system has followng components: a fuzzy nference engne, fuzzy logc rule base, and fuzzfcaton and defuzzfcaton, etc. The structure of the system s shown n Fgure 1. The fuzzy logc rule base s used to estmate the relablty of measurement and nference process. Measure nformaton Fuzzy logc rule base Fuzzyfcaton Robot performance rule base Fuzzy Inference engne DeFuzzyfcaton Fgure 1. The Fuzzy logc communcaton systems The explct communcaton wll follow the semantcs and a grammar, whch s defned as a fve-element vector as follows: C [ A, A, V, O, P] s where A s s an nteger whch represents the nformaton sent out by a robot, A r represents nformaton receved by a robot, V represents a verb or an nstructon, O represents the th object (also can be one of robots), P s a vector (real or fuzzy) whch represents the poston. For example, P=(X, Y), where X, Y can represent a real number, or a fuzzy word, such as near, medum or far. The membershp functons are shown n Fgure. The qualty of poston measurement s related to the dstance between the sensor and an object. The closer, the more accurate (or relable) the measurement s. Therefore, the relablty or accurate of the measurement depends on the measured dstance. r near medum far Dstance (metres) 1 9

3 Studes n Meda and Communcaton Vol. 1, No. 1; 13 Fgure. Fuzzfcaton functon: a dstance level Some of correspondence relatonshps between the number and the verb or nstructon have been gven n Table 1. The followngs are examples for two robots (robot 1 and robot ) communcatng wth each other: Table 1. The relatonshp of the number wth verb C1 [,1,,3,(,)] Number 1 Is at Where s 3 Speed 4 Stop 5 Acceleraton 6 Calculaton 7 Turn left 8 Turn rght C [1,,1,3,(1, far )] C3 [1,,3,3,(.7,.7)] Verb Verb or nstructon C 1 represents that robot asks robot 1 where s the poston of object 3, C and C 3 mean that robot 1 s replayng robot : the poston of object 3 s at the X-coordnate of 1 and Y-coordnate of far. And the speed of object 3 s.7 m/s n both X and Y drectons.. Implct Communcaton The robot performance rule base contans the robot codes of conduct, whch robots should follow. Based on the nformaton receved from the measurement (sensor) system, the nference engne can predct behavour of robots therefore acheve the mplct communcaton. The performance rules base contans the robot traffc rules, such as keepng on the left sde of road, speed lmts, and passng through cross roads, etc. Surely, dfferent robots can have a dfferent code of conduct. Typcal rules for negotatng the crossroads are as follows: If robot 1 s at the fork of a crossroad and t wll cross the route requred by robot, then robot wll wat for robot 1 to pass. If multple robots arrve at a crossroad at the same tme, the one at rght sde wll pass through the crossroad frst. If a statc obstacle n the path then the robot needs to start the new path plannng calculaton. If a movng obstacle n the path the robot needs to calculate the watng tme. And so on. Based on such performance rules, robots can predct each other s behavour (or mplct communcaton) and have tact cooperaton. The mplct communcaton s based on a sequence of reasonng. Therefore the nference engne, performance rules base and measurement nformaton are the three key elements. The mplct communcaton can sort out some of communcaton problems for a mult-robot system partcularly when t s dffcult to exchange nformaton at the tme or the nformaton s kept secret. However the mplct communcaton s not clear and fast enough sometmes. The explct communcaton s stll a commonly used communcaton approach. Therefore combnng the explct communcaton wth mplct communcaton, whle takng the advantages and avodng the dsadvantages of them, s an effcent and relable approach. In our mult-robot system, the hybrd strategy s: If one method s not workng, another s deployed; the mplct communcaton wll be used mostly to predct the behavour of other robots, and the explct communcaton wll be used mostly to gve nstructons. 93

4 Studes n Meda and Communcaton Vol. 1, No. 1; 13 Based on the hybrd communcaton methods, multple robots can cooperate to complete a complex task. That s, the superorty of a mult-robot system over a sngle robot can be acheved. However, the desgn of a mult-robot system, the form of cooperaton and the requrements vary wth dfferent purposes. It s dffcult to fnd a unform cooperaton algorthm for all stuatons. 3. Communcatons Protocol The communcaton protocol for explct communcaton s mportant for a moble robots system. In our system, t conssts of two levels: the symbol transport and the message content. The symbol transport s basc for symbols transform. It provdes a channel between a sender and a recever over whch symbols are exchanged. The content level provdes formats of the symbols that can be understood between a sender and a recever therefore acheve the purpose of communcaton. The structure of the communcaton protocol s shown n Fgure 3 Symbol Transport Content Contents Layer Transport Layer Net Layer Lnk layer Physcal Layer 3.1 Symbol Transport Fgure 3. Structure of the communcaton protocol The symbol transport t s completely ndependent of the nformaton represented at the content level. Thus, the nformaton transmsson aspect of robot communcaton can be treated as an nstance of the more general stuaton of nter-computer communcaton. The symbol transport conssts of four layers, Physcal layer, Lnk layer, Net layer and Transport layer, whch are borrowed from the OSI model. Physcal layer: The physcal layer conssts of the basc networkng hardware transmsson technologes of a network. It s a fundamental layer underlyng the logcal data structures of the hgher-level functons n a network. The physcal layer defnes the means of transmttng raw bts rather than logcal data packets over a physcal lnk connectng network nodes. The bt stream wll be grouped nto code words or symbols and converted to a physcal sgnal that s transmtted over a hardware transmsson medum. Lnk layer: The lnk layer s the lowest layer n the Internet Protocol Sute, the networkng archtecture. It s the group of methods or protocols that only operate on a host's lnk. The servce on ths layer provdng to the net layer s a transmsson channel between neghbourng robots. Its functons nclude sendng and recevng data to and from neghbourng robots. Net layer: The net layer wll provde a routng of packets between any par of robots for the transport layer. To enable routng of packets, one very fundamental ssue s of addressng; each robot n the system should be unquely dentfed. IP addresses are used n the network. Each IP address s of 3 bts. In our robots system, the packet length s very small, and 8 bts are used for addressng. Transport layer: 94

5 Studes n Meda and Communcaton Vol. 1, No. 1; 13 Transport layer contaned the TCP/IP model that s the foundaton of the Open Systems Interconnecton (OSI) model of general networkng. Ths layer provdes end-to-end communcaton servces for applcatons. In a robot, there mght be multple applcatons runnng smultaneously, for example, the robot mght be detectng at the same tme when t s movng. To establsh a communcaton flow between two applcatons, transport layer offers the abstracton of ports. 3. Content Level At the content level contents layer should have the ablty to convert the receved symbols to nformaton whch robot could understand. The receved symbol conssts of a sequence of fve-element vectors, each of the fve-element vector s as follows: C [ A, A, V, O, P] The database of the relatonshp of the number wth verb (refer to Table 1) and related process functons are presented n ths layer. 4. Applcaton Case Study and Experments -Multple Robots Transport a Common Object s We now apply the hybrd approach of explct and mplct communcatons to a stuaton of mult robots cooperaton. The smplest example s two robots transportng a common object n a mult-robots workng envronment. It s requred that these two robots should keep the same spacng wthn a predetermned range throughout the whole process of moton and avod conflct wth other robots or obstacles. Accordng the robot s knematc model (Wang, et al. 1996), ths problem can be descrbed n theory for robot (=1,) as follows: r t x ( t) x () V ( t)cos ( t) dt (1) t y ( t) y () V ( t)sn ( t) dt () ( t) () t V ( t) tan ( t) dt L (3) D * ( x ( t) x j ( t)) ( y ( t) y j ( t )), (j,,j=1,) (4) or: D* ( x ( t) x ( t)) ( y ( t) y ( t)) j j, (j,, j=1,) (4') where x () and y () are the ntal stage (prmary) coordnates of the robot, φ s the steerng angle, V (t) s the speed, θ (t) s the angle between the robot and the x axs, L s the length, D* s the dstance between the supports of two robots (whch should be kept constant), and s the allowable range of the dsplacement between the two robots. Usually, L o >D*, where L o s the length of the common obstacle. Equatons (1)-(3) combne the knematcs model of the robots, and equaton (4) or (4') s a constrant (restrctve) condton for two robots moton. Ths knd of task can only be completed by cooperatve robots. There are two forms for transportaton of a common object: seral (.e., one robot n front of another) and parallel (.e., two robots run n lne abreast). If there s no obstacle n the workng space, these forms of transportaton are acceptable. However, n order to avod an obstacle, the seral transport form s preferable. In ths paper we wll only consder the form of seral as shown n Fgure 4. 95

6 Studes n Meda and Communcaton Vol. 1, No. 1; 13 D* Common object Robot 1 Robot Fgure 4. Two seral robots transport a common object From equatons (1)-(4), the requrement for two robots to transport a common object s to mantan the same speed and headng. Accordng to our mult-robot system structure desgn, ths can be acheved because of the good communcaton ablty. The leadng robot (robot 1 n Fgure 4) needs to avod a collson wth other robots n the same workng envronment based on the robot s performance rules or mplct communcaton approach and sends nformaton on ts speed V(t) and steerng angle φ(t) to the rear robot (robot n Fgure 4). On recevng any nformaton (nstructons) from robot 1, robot controls ts speed and drecton by ts own controller accordng to the nstructons and the desred moton (path plannng). Fgure 5 shows the structure of the requred control system, whch s based on robots communcaton. Communcaton a 1 (t), 1 t Path plannng X (t)*, Y (t)* Controller Robot Communcaton moton stuaton V(t), a (t), t... Path plannng X (t)*, Y (t)* 1 1 a (t), t Controller Robot X (t), Y (t) 1 1 X (t), Y (t) moton stuaton Fgure 5. Structure of control system Path plannng s somewhat dfferent from that of a sngle robot system. In a cooperatng system, the path plannng needs to consder the moton of another robot n the same workng envronment, n addton to the expected locus of the movng object. Therefore the path plannng approach should be event-based and need the hybrd communcaton approach to support. The basc dea of the event-based plannng and control theory s to ntroduce a new moton reference varable that s changng wth tme, but drectly related to the sensory measurement of the system, and the performance of move obstacle can then be predcted. Instead of tme, the desred system output s parameterzed by the new moton reference varable. The latter s desgned to effcently carry the sensory nformaton needed for the planner to adjust or modfy the orgnal plan to form a desred output (Wang et al. 1, 1999). As a result, for any gven tme nstant, the acton plan s a functon of the system output. Ths creates a mechansm to adjust and modfy the plan based on the sensory measurement. The path plannng of two robots transportng a common object can be descrbed as follows. Suppose that the expected locus of the object s E o (x,y), the loc of robot 1 and robot are E 1 (x,y) and E (x,y), then the path plannng can be descrbed as follows: * E ( x, y) f ( E, D, E ) 1 1 o * E ( x, y) f ( E, D, E ) o 1 where the map f 1 * satsfes the followng relatonshp: The dstance between E 1 (x,y) and E (x,y) s D, and the dstance between E 1 (x,y) and E o (x,y) s less than D, that s, (5) (6) 96

7 Studes n Meda and Communcaton Vol. 1, No. 1; 13 E ( x, y) E ( x, y) D E ( x, y) E D 1 o (5') Smlarly, the map f * satsfes the relatonshp as follows: E ( x, y) E ( x, y) D E ( x, y) E D o (6') These equatons have multple solutons. For smplcty, equaton (5) and equaton (5') can be replaced by the followng: That s, Thus equatons (1)-(6) gve a mathematcal descrpton on the trajectory plannng problem where two robots transport a common object. The sgnfcance of equaton (5") s that only one of two robots needs to consder the constrant condton E 1 (x,y)-e o <D, when t s plannng a path. Therefore t can stll use the sngle robot path plannng approach (Wang et al. 1). The other robot's path plannng should be based on dstance D. D can be chosen as a new moton reference varable or moton stuaton. In fact, the path plannng and control of robot 1 (leadng robot) uses a fuzzy algorthm, whch can ensure a sutable path and control result. Robot (auxlary robot) can also use the fuzzy algorthm, but t should pay more attenton to the moton stuaton and hence the rule control s dfferent from that of robot 1. Some of the rules are: If V ( t) V ( t) and D D* /, then mantan speed. If If If If and D > D*, then speed up. and D < D*, then mantan speed. and D > D*, then mantan speed. and D< D*, then reduce speed. where D ( x ( t) x ( t)) ( y ( t) y ( t)), V ( t ) and V ( ) are the veloctes of robot 1 and robot. The steerng control rules for robot are as follows: If 1 and 11, then ; If 1 and 11, then after T seconds; If 1 and 11, then ; If 1 and 11, then after T seconds; V ( t) V ( t) V ( t) V ( t) V ( t) V ( t) V ( t) V ( t) 1 1 * E ( x, y) f ( E, D) 1 1 E1( x, y) Eo D Where φ 1 and φ are the steerng angles of robot 1 and robot, φ 11 s φ 1 before samplng, and. A smlar approach can be appled to transport a common object by a group of (n+1) robots n cooperaton. One robot (leader) sends all nformaton about ts planned speed V(t) and steerng angle φ(t) to the others accordng to the desred moton (path plannng). Its path plannng control can be based on the sngle robot's path plannng approach. The other n robots wll control ther speed and drecton respectvely by controllers accordng to the desred moton (path plannng) and the nstructons. The nstructons and the dstances between other robots are the new moton reference varables for n auxlary robots for whch path plannng and control s event-based. 97 o * D T V( t) (5")

8 Studes n Meda and Communcaton Vol. 1, No. 1; 13 Robot 4 5 B Goal pont 4 Robot 5 3 Obstacle Obstacle 1 Robot 3 Start pont A Robot Fgure 6a. Smulaton result of two robots transportng a common object n mult-robotcs workng envronment Smulaton: Suppose that the length of each robot s L 1 (or L ) = 1 meter, the loadng capacty s 5 klograms, the length of object s L o =5 meters, and ts weght s 7 klograms. The range of the object's allowable devaton between the two robots s =.5 (meter). The dstance between the supports of the two robots D* s 4 meters. Fgure 6a shows the smulaton result and Fgure 6b shows the dstant error curve devaton between the two robots. From the Fgure 6b, t s clear that the error s largest when the robots turn a corner. However the error never reaches more than.4 meter n the control process. Clearly, ths cooperaton method works successfully. Fgure 6b. The error curve (devaton between robots) How to avod a conflct wth other robots n the mult-robots workng envronment s a key for the whole system to succeed. In the smulaton process, the mplct communcaton, or robot traffc rules, has sorted out the problem successfully: robot 4 wated at pont B (n Fgure 6a) when robot 1 and have ndcated that they are gong to turn to rght. At pont A (n Fgure 6a), robots 1 and need to wat robot 3 to pass, then they wll contnue movng. If there s a rule: the lght-loaded robot should avod heavy-loaded robot n the workng envronment, then robot 3 wll wat at pont A untl robots 1 and have passed. The robot traffc rules can have dfferent contents n dfferent workng envronments. 5. Conclusons A mult-robot system has been wdely used n auto-warehouse and large workng place. However achevng the cooperaton n a mult-robots team s a very complex task. It covers a wde range of dscplnes and technologes, ncludng dstrbuted artfcal ntellgence, computer communcaton, control system, cooperaton and 98

9 Studes n Meda and Communcaton Vol. 1, No. 1; 13 competton strategy, etc. Some solutons to certan types of problems are even unknown to humans. Because of ts dversty, t s hard to make a comparson of dfferent approaches due to lack of commonly accepted test standards and procedures. The research platforms used dffer greatly, as do the key assumptons used n dfferent approaches. Ths paper presented the hybrd communcaton strategy of combnng the explct wth mplct communcatons for mult-robot system. A fve-element vector s used to descrbe the semantcs and grammar for robot communcaton. Robot performance rule base (or traffc rules) s ntroduced nto the system that can predct behavour of robots and then acheve the mplct communcaton. It sgnfcantly reduces overheads for explct communcaton amongst robots. Based on the hybrd communcaton strategy, avodng a collson among multple robots n a workng area and multple robots transportng a common object have been explored n a mult robots workng envronment. The results of smulaton show that the mult-robots can cooperate successfully to complete tasks based on the hybrd communcaton approach presented. References Aaron B. St. Clar, Amn Atrash, Ross Mead, & Maja J. Matarć. (11). Speech, Gesture, and Space: Investgatng Explct and Implct Communcaton n Mult-Human Mult-Robot Collaboratons, Refereed Workshop 11 Amercan Assocaton of Artfcal Intellgence Sprng Symposum on Mult-Robot Systems and Physcal Data Structures (SS-11-9). Castelfranch, C., et al., (1). Behavoral Implct Communcaton (BIC): Communcatng wth Smart Envronments. Internatonal Journal of Ambent Computng and Intellgence (IJACI), (1), Castelfranch, C., et al., (1). Behavoral Implct Communcaton (BIC): Communcatng wth Smart Envronments va our Practcal Behavor and Its Traces. Innovatve Applcatons of Ambent Intellgence: Advances n Smart Systems, 1-1. Caplupp M., & Segala, R. (13). World Automata: a compostonal approach to model mplct communcaton n herarchcal Hybrd, 3rd Workshop on Hybrd Autonomous Systems 13, 17 March 13, Rome, Italy. Long, X. L., Jang, J. P., & Xang, K. (4). Towards Mult-robot Communcaton, proceedng of IEEE Internatonal Conference of Robotcs and Bommetcs, ROBIO, Iqbal, J., Yousaf, M. M., & Awas, M. M. (9). A scalable approach of message nterpretaton by demonstratons for mult-robot communcaton. Proceedng of IEEE 13th Internatonal Multtopc Conference, INMIC 9. pp1-6. Jelle, R. K., Matthjs, T. J., & Spaanand, N. V. (5). Non-communcatve mult-robot coordnaton n dynamc envronments. Robotcs and Autonomous Systems, 5(-3), Kashyap, S., & Yan, M. (7). Communcaton-Effcent Dynamc Task Schedulng for Heterogeneous Mult-Robot Systems. Proceedngs of the 7 IEEE Internatonal Symposum on Computatonal Intellgence n Robotcs and Automaton, pp3-35, USA, June -3. Km, S. C., Shn, K. H., Woo, C. W., Eom, Y. S., & Lee, J. M. (8). Performance Analyss of Entropy-based Mult-Robot Cooperatve Systems n a MANET. Internatonal Journal of Control, Automaton, and Systems, 6(5), Ran, G., Zhang, H., & Gong, S. (1). Improvng on LEACH Protocol of Wreless Sensor Networks Usng Fuzzy Logc. Journal of Informaton & Computatonal Scence, 7(3), Roy, N., & Dudek, G. (1). Collaboratve robot exploraton and rendezvous: algorthm, performance bounds and observatons. J. Autonomous Robot, 11(), Tarque, H., & Maram, Y. (9). A Fuzzy Approach to Energy Optmzed Routng for Wreless Sensor Networks. The Internatonal Arab Journal of Informaton Technology, 6(). Wang, T., Mehd, Q., & Gough, N. (1996). Knematcs models of autonomous guded vehcles and ther applcatons. Proc. of ISCA 5th Internatonal Conference, Reno Nevada, USA Wang, T., Mehd, Q. H., & Gough, N. E. (1999). An ntegrated navgaton system for AGVs based on an envronment database. Internatonal Journal of Computers and Ther Applcatons, 6(1), Wang, T., Dang, Q., & Pan, P. (1). A Path Plannng Approach n an Unknown Envronment. Internatonal Journal of Automaton and Computng (IJAC), 7(),

10 Studes n Meda and Communcaton Vol. 1, No. 1; 13 Wang, T., Dang, Q., & Pan, P. (11). A Predct-Fuzzy Logc Communcaton Approach for Mult Robotc Cooperaton and Competton. Journal of Communcatons, 6(3), Yan, M. J., Nckerson, V., & Jng, G. (6). Mult-robot Aggregaton Strateges wth Lmted Communcaton. Proceedngs of the IEEE/RSJ Internatonal Conference on Intellgent Robots and Systems, Bejng, Chna. Ths work s lcensed under a Creatve Commons Attrbuton 3. Lcense. 1