Design and Implementation an Autonomous Mobile Soccer Robot Based on Omnidirectional Mobility and Modularity

Transcription

1 Design and Implemenaion an Auonomous Mobile Soccer Robo Based on Omnidirecional Mobiliy and Modulariy S. Hamidreza Mohades Kasaei and S.Mohammadreza Mohades Kasaei Absrac The purpose of his paper is o design and implemen a Middle size soccer robo o conform RoboCup MSL league. Due o heir advanages, omni-direcional mobile robos have found many applicaions especially in roboic soccer compeiions. However, omni direcional navigaion sysem, omni-vision sysem and kicking mechanism in such mobile robos have no ever been combined. This siuaion brings he idea of a robo wih no head direcion ino exisence, a comprehensive omni direcional mobile robo. Such a robo can respond more quickly and i would be capable for more sophisicaed behaviours wih muli-sensor daa fusion algorihm for global localizaion. Despie recen advances, effecive conrol and self-localizaion mehods of omni-direcional mobile robos remain as imporan and challenging issues. For his purpose, we uilize he sensor daa fusion mehod in he conrol sysem parameers, self localizaion and world modelling. Omnidirecional vision sensor of he vision sysem deals wih he image processing and posiioning for obsacle avoidance and arge racking. The boundary-following algorihm (BFA) is applied o find he imporan feaures of he field. We uilize he sensor daa fusion mehod in he conrol sysem parameers, self localizaion and world modeling. A vision-based self-localizaion and he convenional odomery sysems are used for robus self-localizaion. The localizaion algorihm includes filering, sharing and inegraion of he daa for differen ypes of objecs recognized in he environmen. In he conrol sraegies, we presen hree sae modes, which include he Aack Sraegy, Defense Sraegy and Inercep Sraegy. The mehods have been esed in he many Robocup compeiion fields of middle size robos. The mehods have been esed in he many Robocup compeiion field middle size robos. Some new ineresing mehods are described in he curren repor. Index Terms Mobile robo, Machine vision, Omni direcional movemen, Auonomous Sysems, Robo pah planning, Objec Localizaion. I. INTRODUCTION Robocup was esablished o promoe AI and inelligence roboic research, and is dream is o defea he FIFA World cup Champion eam by a humanoid robo eam. In RoboCup 2008, many eams realized eam plays as well as developed advanced hardware composed of a powerful kicking device, an omni direcional camera, and an omni direcional plaform. Our projec provides six robos of he same ype. Our aim is o have a cooperaive eam behavior while aaining a high individual robo performance in image Manuscrip received February 17, 2011; revised Sepember 28, The auhors are wih Young Researchers Club, Isfahan Branch, Islamic Azad Universiy, Isfahan, Iran. processing, decision making and moion conrol. Some sudied subjecs in his work are: "Arificial Inelligence includes Fuzzy Decision-Making and Behaviors Learning based on Sraegy concep, Advanced Jacobean MIMO Conroller, Image processing and high performance moor driver modules". Our robos are equipped wih hree-wheeled full omni direcional navigaion sysem and omni-vision sysem. We believe ha a reliable mechanical srucure design, robus pioneer algorihms and high performance sensor fusion are principles of a high eam performance in dynamic environmens [1]. The Middle Size League eam of he ADRO was founded in In 2007 we ranked 2 nd place Middle Size Soccer Robo League in 2 nd Inernaional Iran-Open Robocup Compeiions he Iran-Open is one of he, Asia s major RoboCup even. A he China-Open RoboCup 2007 as well as a he Iran Open RoboCup 2007 we ranked 2 nd place Middle Size Soccer Robo League, in 2008 we achieved he Firs Place in he 3 rd Inernaional Iran-Open Compeiions. The basis for our success was he robus and reliable hardware design, well-srucured sofware archiecure and efficien algorihms for sensor fusion and behavior generaion. Our main research ineres is boh, he developmen of learning robos and he developmen of improved sensor fusion and sensor inegraion echniques. Among ohers, several differen approaches have been invesigaed so far: Reinforcemen learning o learn inerceping a ball. Reinforcemen learning o learn dribbling and Pah Planning. Developmen of a compuaional efficien algorihm for self-localizaion. Developmen of esimaion procedures for robo and ball velociy Developmen of a vision sysem wih an omni direcional vision sensor and a fron camera o recognize ball posiion and find necessary informaion in field. In his paper, we will a firs describe he general hard- and sofware design of he ADRO Robocup Team, (secion 2) and afer ha focus on our scienic approaches in sensor fusion and learning (secion 3). Finally, secion 4 concludes his paper. II. HARDWARE ARCHITECTURE The ADRO Robo (Fig. 1) were designed and compleely buil in Roboic and AI cener from Islamic Azad Universiy-Khorasgan branch. The mechanical srucure of 616

2 he players is layered and modular. Each layer can easily be replaced by an equivalen one. The componens in he lower layer, namely moors, encoders, ow ype wheels, baeries and an elecromechanical kicker [4], reaached o an aluminum plae placed 6 cm above he floor. The robos' lower layer consiss of a hree-wheeled movemen sysem. The wheeled movemen sysem has hree acuaed wheels and each is driven by one DC moor. Having hree independenly conrollable wheels, he robo can roae or move in eiher direc or curved pah. Using his sysem, robo can reach a maximum speed of abou 2 meers per second. The second layer conains he conrol elecronics. The hird layer conains a lapop compuer, an omni-direcional vision sysem, a fronal camera and an elecronic compass, all close o he maximum heigh of 80 cm. The players are capable of omni-direcional moion, based on hree omni-direcional roller wheels. Every fully auonomous robo of ADRO is equipped wih an omni-direcional vision sysem, a normal camera as fron vision, and an elecromagneic kicking device. The robo is conrolled by a noebook PC is demonsraed in figure 1. The main board uses AVR microconrollers as a digial processing uni ha enables he robo o have some onboard processing. The program for he microconrollers is wrien in AVR C language. The omni-direcional vision sysem consiss of a hyperbolic mirror, a Color Digial FireWire Camera (Basler Digial Camera) and on each robo we have a sandard lapop wih 2GHz CPU wih Windows Operaion Sysem. On his lapop he auonomous behavior of he robo is compued, based on he daa gahered from he omni-direcional camera. Fig. 1. Our robo is a 3-wheeled omni-direcional mobile robo wih an omni-direcional vision sysem, a normal camera and a kicking device, and is conrolled by a noebook PC A. Omni direcional wheels and robo chassis Figures Omni direcional robos usually use special wheels. These wheels are known as omni direcional poly roller wheel. The omni-direcional movemen sysem consiss of omni-direcional wheels, DC moors, a drive shafing sysem and a conroller. The hree omni-direcional wheels are he key componen of he robo, each wheel radially equipped wih dozens of small wheels. The hree-wheel drive robo can move a any direcion and a any momen. Alhough hree such wheels are sufficien for he robo o move omni-direcionally, a fourh wheel can provide redundancy in moion and conrol [2]. Our Robo srucure includes hree omni direcional wheels for moion sysem and hree small free wheels as feedback mechanism where shaf encoders are mouned on, as shown in figure 2. Fig.2. Three omni direcional wheels ac as acuaors while hree free wheels are for feedback. B. Omni direcional vision sysem Since he beginning of mobile robo, he map building was one of he mos addressed problems by researchers. Several researchers used omni direcional vision for robo navigaion and map building [3]. Because of he wide field of view in omni direcional sensors, he robo does no need o look around using moving pars (cameras or mirrors) or urning he moving pars [4]. The global view offered by omni direcional vision is especially suiable for highly dynamic environmens. The omni-direcional vision sysem consiss of a hyperbolic mirror, a fire wire color digial camera (Basler Digial Camera) and a regulaion device. The hyperbolic mirror, design by ourselves. The mirror can make he resoluion of he images of he objecs near he robo on he field consan and make he disorion of he images of he objecs far from he robo small in verical direcion. Searching hrough differen aricles and caalogues from various mirror-making companies; we found ha hey used he following hyperbolic curve for heir omni direcional vision mirror [5]. 2 2 x y = However, his equaion is suiable for he mirror wih large size and wide view. For our soccer player robo, we need an image wih a diameer of 4m on he field, so o achieve a compac mirror wih wide view, he above curve scaled down by a facor of 2.5 o yields: 2 2 x y = Fig.3. Profile of he overall mirror and he proposed mirror. Nex, a special hree sages process was considered for he mirror manufacuring: a. Curve fabricaion, b. Polishing, c. Coaing. In he firs sage he curve was fabricaed on seel 2045 wih CNC machining. Then, he work-piece was polished by a special process and finally Ni-P elecroless plaing was employed. 617

3 Fig. 4. The ypical panoramic image capured. The regulaion device can adjus he heigh of he omni-direcional vision sysem and he disance beween he hyperbolic mirror and he camera. I is able o shoo very fas, 10m/s when abou 800 urns and a curren of 60[A] are applied. I is raher small (lengh is abou 20cm and abou 5cm diameer) and lighweigh (2[kg]). Only a ransformer, a capacior, some resisors and a swich is used so i is in heory very reliable. And mos imporan shooing power can be varied by varying he ime of he applied curren. The disadvanage of he use of a solenoid is ha i operaes a a high volage and curren, so i can be quie dangerous. This can be solved by hiding dangerous pars in a black box. I also uses a lo of power for a really shor ime, so a capacior is needed o supply high volage and curren. We used a DC-DC converer (Boos regulaor) for geing differen currens o have differen power of shooing. Fig. 5. The Schemaic diagram of he omni direcional vision sysem. C. Ball handler and kicking sysem One of he mos essenial pars of a soccer robo in Middle size and Small size classes in he kicking sysem, his sysem is in charge of kicking he ball upon he command issued by he processor of robo. Almos every eam develops heir own unique shooing device. In his secion, we are going o describe designed and developed our muli power kicking sysem ha enables loop and vary shooing power. The principle used for our kicking devices is self-inducance. By sending a curren rough a urn of wire a magneic field can be build. As he number of urns or curren increases, he magneic field increases oo. Wih magneism ferromagneic maerials can be araced or repulsed. This phenomenon is used in a solenoid. For kicking device he solenoid has o be really fas, because i ravels 10cm in abou 10ms when shooing a 10m/s. To design a good solenoid and o obain maximum velociy of ball some parameers like: inducance, response ime, resisance, force, dimensions and core-maerial should be balanced carefully. This design akes advanage of he propery ha a solenoid has a ferromagneic core which is araced ino he coil cenre. The piece of nylon which is aached o he iron bar is a non-ferro and shoos ouwards and his he ball. Fig. 7. Ball handler mechanism. III. THE SOLENOID CONTROLLING CIRCUIT Through his circui, he ime needed for firs charging of he condenser o 450V, is 13 seconds. This volage will be abou 150v afer a sroke wih he maximum force. The ime need for recharging he capacior from 100 vol o 450 vol is abou 9 seconds. Parallel o he capacior is he solenoid, modeled wih a coil and a resisance wih he values of he solenoid. The solenoid is conneced o a ransisor which is conrolled by a pulse source. The ransisor can handle high currens (IRG4PC50FD).When he source-signal is high he ransisor is closed and solenoid is acivaed, when low i is open and solenoid is idle. The ransisor opens in 380[ns] and closes in 70[ns]. Parallel o he solenoid is a diode wih a resisance o "cach" he back-curren generaed when urning of he solenoid. The solenoid designing for he normal condiions is no able o reurn o he firs place afer he sroke, because afer imposing and cuing he volage from Solenoid, is shaf will be sopped in he las possible place. For reurning i o is firs condiion, we should use anoher mechanism. This reurns he solenoid o is firs place afer he sroke. An easy and cheap soluion for he above problem is using a spring a he end of he solenoid. Fig. 6. Solenoid mechanism sysem. Fig. 8. The solenoid conrolling circui design. 618

4 IV. ROBOT SOFTWARE We have developed a sofware sysem o fully uilize he hardware abiliies. In his secion inroduces sofware pars conain: image processing algorihm, posiion conroller archiecure, world model consrucion, arificial inelligence, rajecory, and nework and eam sraegy from a viewpoin of sofware sysem. Three acions are alloed o he robo: aack, suppor, and defense. The aack is realized hrough he following process: firs, he robo acquires he ball. The robos coninuously ry o ge he ball. Nex, he robo face o ball and argeing o opponen goal hen dribbles and shoos he ball ino he opponen goal. During he dribble, he robo adjuss is direcion oward he opponen goal and dribbles wih he fases possible speed. One of he advanages of our robo is a srong kicking device; herefore he robo can shoo a loop-ball ino he opponen goal before he opponen defender comes close o he robo. The suppor robo akes a posiion behind and near o he robo wih he ball. The suppor robo feches he ball only when he ball is near o he suppor robo. The defence robo is locaed beween he ball and own goal. The defence robo doesn acively approach when he ball is far. In our eam sraegy hree saes are alloed o he eam robos: aack, defense, and inercep. The robos auonomously choose o acivae each of he roles. Fig. 9. Vision sysems, objec deecion (goals, flag spos, ball ) and line deecion for self localizaion algorihm. A. Image processing algorihm Uilizing a digial camera, each ime he compuer on each robo performs he processing of he curren frame and calculaes he posiion, direcion and velociy of he robo. I also deermines he posiion and velociy of he opponen robos as well as he posiion and velociy of he ball. The image-processing algorihm firs filers he image by using a able for labeling he colors hen recognizes he coniguous regions hrough eiher a BFS or a DFS search algorihm and finally exracs he posiions by looking in he Image o ground map able. The server gives he necessary commands o he image processing compuers. The algorihm used o find objecs is opimized o process he maximum number of frames. Firs i searches he pixels by swiping hem wih cerain seps, when i finds a desired one and deecs ha objec, saves is coordinaes so he nex ime i can sar back wih he same poin abou. Someimes for beer image manipulaion he RGB color space is convered o HLS (Hue, Sauraion and Luminance). To recognize a cerain color, a combinaion of condiions on Hue, Sauraion and RGB is used. This procedure makes he color recogniion independen from he change of brighness and oher unprediced condiions. We are rying o evaluae new mehods o find some kinds of objecs based on paern recogniion o reduce he effec of changing he colors on algorihm. The image processor receives is daa hrough fire wire por conneced o a Basler digial video camera wih he speed of 20 o 30 frames per second. B. World model consrucion Alhough each agen ries o exrac he real world map as accurae as possible, bu noisy daa and non-global opimized algorihms reduce he reliabiliy of processed daa. The world model module receives differen daa ses from every agen. Each daa se conains differen environmenal informaion like self, ball and opponens posiions. Each daa carries a confidence facor; a larger confidence facor means a more reliable piece of informaion. The mos recen daa ses are hen chosen for daa fusion, in which he following rules and facs are applied: Closer objec are of more accuracy. Objecs furher han a specific disance could be said o be oally inaccurae. (This disance is experimenally known) An objec in he field canno move faser han an exreme value. Fig 10. A schemaic represenaion of he robo s hardware and sofware modules and heir inerconnecions 619

5 Wih respec o he above fac, he module filers unwaned duplicaes of objecs, (many opponens close o each oher seen by differen agens), calculaes he bes approximaion for ball and opponens posiions wih firs order Kalman filering, gives every objec a confidence facor, applies a low pass filer on daa and finally consrucs a complee world model. This new world model conains informaion abou he objecs which may no have been seen by each agen correcly and also enhances approximaions of all environmenal informaion. The consruced world model is hen sen back o all agens so hey will have a beer view of he world around hem [7]. The ineracion beween he modules on differen machines is provided by a communicaion proocol which bundles commands and parameers generaing command packes and inerpres he incoming packes for oher modules. In he following, each layer, is inerface and parameers will be discussed in deails. C. Arificial inelligence roles or behaviors. Then AI Core sends a se of roles o Role Engine o be assigned o he robos. Because here are insances in which he image-processing uni canno see he ball, a memory is implemened in he AI Core for he posiion of ball ha specifies which robo owns he ball. Since here is a relaionship beween new roles and old roles, roles are changed in a manner ha robos never experimen sudden changes in roles (for example he role never changes from defense o aack in nex cycle). Role Engine receives a se of roles from AI Core and provides he Behavior Engine wih a se of behaviors for robos. Twin or riple roles are implemened so ha he robos really cooperae wih each oher o do heir roles. Behaviors are he building blocks of he robo's performance which includes simple acions like roaing, or geing he ball and ec. The Behavior Layer is he lowes layer in our archiecure. This layer receives a sequence of behaviors along wih some parameers from he upper layer (Role Engine) and execues he essenial subrouines in order o accomplish a cerain behavior. These subrouines use world model informaion and rajecory daa in order o perform necessary movemens. and PD H respecively), a plan ransfer funcion ( P H which is obained from he sysem dynamics) and a self-localizaion ransfer funcion (as a feedback funcion ha only senses he robo s posiion). A noise node, N, is also included ha has an addiive effec on he sysem posiion inpu. The inpu of he sysem is considered o be a sep funcion and he oupu is he robo posiion and orienaion. Fig 12. Conrol diagram of he fully auonomous robo. E. Robo Self-Localizaion Our self localizaion mehod is based on deecion of whie lines in field. Because according o MSL rule no flag and no color goals exis in field since Robocup 2008, now he whie line poins are he only visual informaion ha could be used as landmarks for robo s self localizaion. So our vision sysem racks all whie lines ha exis only in he region field color (green) and robos use a digial compass (MTi-sensor) for he robo heading reference. Afer his secion we ry o conver he acquired whie line poin ino he real world disance map. We employ a Mone Carlo Localizaion (MCL) mehod [1]. For our algorihm, he field model is a Caresian coordinae sysem wih he origin a he cener of he field. The robo s sae is represened by a vecor X = [ x, y, θ ] T which consiss of a posiion ( X, Y ) and θ an orienaion. We provided he algorihm, which deecs only orienaion, made he posure θ ingredien known in MCL, and planned he dimension reducion of he sae vecor. The orienaion deecion is explained furher below. For localizing, we have o consruc he poserior densiy p ( x y1... y ) from he sae of a robo x and he sensor daa y a he curren ime. In he paricle filer mehods, a probabiliy densiy is represened by a se of N random samples (Paricle). The mehod proceeds in wo phases. In he firs phase we predic a curren sae of he robo. Tha is specified as a condiional densiy p ( x... x 1 u 1) from he previous sae x 1 and a conrol inpuu 1. The predicive densiy is obained by he following inegral. For our algorihm, we se he conrol inpu u 1 as odomery daa and add i o each paricle. p ( x y1... y ) = p( x x 1, u 1) p( x 1 y1... y 1) dx 1 Fig. 11. World model consrucion and arificial inelligen srucure. D. Posiion conroller archiecure Figure 12 shows he overall block diagram of he conrol sysem. Thse omni direcional robo conrol loop conains a PID and a PD conroller (wih he ransfer funcion PID H Fig. 13. The field lines deecion and self localizaion. 620

6 In he second phase we updae he densiy according o he sensor daa y. The likelihood of y a sae x is represened as ( ) y x using Bayes heorem. p p. The poserior densiy is obained ( x y... y ) 1 p = ( y x ) p( x y1... y 1) p( y y... y ) Sensor daa y is disance o he field line. The sae is compared o y and he likelihood is updaed of each paricle. Afer ha, weighed paricles are normalized and re-sampled. Re-sampling is done according o he weigh of each paricle: new paricles are generaed around he paricles ha have high likelihood. In he RoboCup soccer field mos consiuens are sraigh lines or perpendicular segmens of lines. The robo s orienaion is deeced by searching for inclinaion of he sraigh line ingrediens in he circumference seen from he robo. This approach is simple, high-speed, and he derivaion of he robo orienaion becomes efficien. Our ask is o use MCL in he RoboCup environmen. I is possible o falsely deec orienaion because he form of he field is symmeric agains he cener of he field. We solved he false deecion by using compass sensor for his problem. Our self localizaion algorihm is a one of he very fas and effecive algorihm o rack robo s localizaion, and i only akes several milliseconds o finish he localizaion compuaion for one frame image. Experimens show ha he posiion error of robo s self localizaion can be less han 50 cenimeer. ADRO has developed and implemened hree separae omni direcional wheels coupled wih shaf encoders placed 60 apar of he main driving wheels. Free shaf roaion and he flexible connecion o he srucure ensures minimum slippage and firm conac of hese wheels o he ground, all hese resul in a grea improvemen in oupu precision. In order o avoid he remaining cumulaive error, odomery sysem parameers can be iniialized every ime he vision could calculae he posiion reliably. F. Robo Pah Planning wih Arificial Poenial Field Since he moion rajecory of each robo is divided ino several median poins ha he robo should reach hem one by one in a sequence he oupu obained afer he execuion of AI will be a se of posiion and velociy vecors. So he ask of he rajecory will be o guide he robos hrough he opponens o reach he desinaion. The rouine used for his purpose is he poenial field mehod (also an alernaive new mehod is in progress which models he robo moion hrough opponens same as he flowing of a bulk of waer hrough obsacles)[7][8]. In his mehod differen elecrical charges are assigned o our robos, opponens and he ball. Then by calculaing he poenial field of his sysem of charges a pah will be suggesed for he robo. A a higher level, predicions can be used o anicipae he posiion of he opponens and make beer decisions in order o reach he desired vecor. In our pah planning algorihm, an arificial poenial field is se up in he space; ha is, each poin in he space is assigned a scalar value. The value a he goal poin is se o be 0 and he value of he poenial a all oher poins is posiive. The poenial a each poin has wo conribuions: a goal force ha causes he poenial o increase wih pah disance from he 1 1 goal, and an obsacle force ha increases in inverse proporion o he disance o he neares obsacle boundary. In oher words, he poenial is lowes a he goal, large a poins far from he goal, and large a poins nex o obsacles. U ( q) = U ( q) U ( q) Goal + Obsacle If he poenial is suiably defined, hen if a robo sars a any poin in he space and always moves in he direcion of he seepes negaive poenial slope, hen he robo will move owards he goal while avoiding obsacles. The numerical poenial field pah planner is guaraneed o produce a pah even if he sar or goal is placed in an obsacle. Fig. 14. Goal force (Aracive poenial o he goal) Fig. 15. Obsacle force (Repulsive poenial) If here is no possible way o ge from he sar o he goal wihou passing hrough an obsacle hen he pah planner will generae a pah hrough he obsacle, alhough if here is any alernaive hen he pah will do ha insead. For his reason i is imporan o make sure ha here is some possible pah, alhough here are ways around his resricion such as reurning an error if he poenial a he sar poin is oo high. The pah is found by moving o he neigh boring square wih he lowes poenial, saring a any poin in he space and sopping when he goal is reached. G. Nework The nework physical layer uses he ring opology. The UDP (User Daagram Proocol) nework proocol is used for he sofware communicaion layer. The daa flow of he nework is as follows: A half field daa (he daa represening he posiion and saus of he robos, opponens, goals and he ball) is ransmied o he server from each clien compuer of 621

7 robos, he server combines hem, consrucs he complee global localizaion field hen sends he appropriae daa and commands (indicaing which objecs each robo should search for) back o he cliens. When he daa is compleed i is passed o he AI uni for furher processing and o decide he nex behavior of he robos. research focus in muli-robo sysem. As he MSL environmen be-coming more and more complex, and he field larger and larger, for he field of view of robo is limied, he occlusions of imporan objecs such as ball is very common in dynamic compeiion, and he inconsisen world model of every robo brough by percepion noise will also affec he eam s sraegy. So we are ineresed in building he accurae and consisen world model of muli-robo eam by cooperaive percepion such as cooperaive objec localizaion and robo s cooperaive localizaion. Fig. 16. Goal Force + Obsacle force VI. CONCLUSION The performance of our robo eam in Iran-Open Robocup compeiions 2008 (1s place) showed ha he combinaion of mehods and echniques described in his paper are led o a successful soccer player eam. In our robo, omni direcional navigaion sysem, omni-vision sysem and a novel kicking mechanism have been combined o creae a comprehensive omni direcional robo. The idea of separaing odomery sensors from he driving wheels was successfully implemened. Three separae omni direcional wheels coupled wih shaf encoders placed apar of he main driving wheels. The resul was reducing errors such as slippage in he ime of acceleraion. Combinaion of odomery and vision led o a more accurae and reliable self-localizaion algorihm. A new hyperbolic wih a special coaing echnique was used o creae a wide and compac omni vision mirror. Fig. 17. Poenial a every poin; i is highes in he obsacles and lowes a he goal. Elsewhere i is generally higher farher from he goal and near obsacles. V. CURRENT RESEARCH FOCUS Our curren main research focuses are lised as follows: a. Robus robo vision: The final goal of RoboCup is ha he soccer robo eam defeas human champion, so robos will have o be able o play oudoors and ge rid of he color-coded environmen sooner or laer. We will go on developing our robo vision sysem o make ha he robo can work well in he environmen wih highly dynamic illuminaion and even in oally new field wihou any off-line calibraion. We are also doing furher research on he arbirary FIFA ball recogniion mehod by our omni-direcional vision sysem based on he developmens we have achieved [6]. We are also ineresed in recognize and disinguish he robos belonging o differen eams and oher generic objecs by using he advanced paern recogniion echniques. b. Muli-robo cooperaion: Muli-robo cooperaion holds an imporan place in disribued AI and roboics field. We have designed a good muli-robo cooperaion mechanism and also realized several wo-robo cooperaive behaviors. Now we have o do deeper research o develop our robo s cooperaion abiliy by involving more robos and more complex cooperaive behaviors in his mechanism. c. Muli-robo cooperaive percepion: Muli-robo cooperaive percepion is anoher imporan Fig.18. ADRO Middle size soccer robo eam from Islamic Azad Universiy Khorasgan Branch REFERENCES [1] S.H.Kasaei, S.M.Kasaei, S.A.Kasaei, M.Taheri, S.A.Monadjemi " Modeling and Implemenaion a Fully Auonomous Soccer Robo Based on Omni-direcional Vision Sysem", Indusrial Robo: an Inernaional Journal, Emerald Group[ISSN X] 37/3 (2010) [2] S.H.Kasaei,S.M.Kasaei, S.A.Kasaei,M.Taheri, "Design and Implemenaion a Fully Auonomous Soccer Player Robo" Proceedings of WASET, Volume 39, ISSN: , Hong Kong, China, March 23-25, [3] S.H.Kasaei, S.M.Kasaei, S.A.Kasaei, M.Taheri, "Effecive Mecharonics Models and Mehods for Implemenaion an Auonomous Soccer Robo" 17h IEEE Iranian Conference on Elecrical Engineering, Tehran, Iran, May 12-14, [4] S.H. Kasaei, S.M.Kasaei, S.A.kasaei, M.Taheri Design and Developmen of Novel Solenoid-Based Kicking Sysem. 9h ICME, Birjand, Iran, March 3-5, [5] Hyun-Sik Shim, Yoon-Gyeoung Sung, Seung-Ho Kim and Jong-Hwan Kim, "Design of Acion Level in a Hybrid Conrol Srucure for Vision Based Soccer Robo", Proceeding of he 1999 IEEE/RSJ Inernaional Conference on Inelligen Robos and Sysems, pp [6] Lu, H., Zhang, H., e al.: Arbirary Ball Recogniion Based on 622

8 Omni-direcional Vision for Soccer Robos. RoboCup 2008: Robo Soccer World Cup XII, pp (2009). [7] S.H.Kasaei, S.M.Kasaei, S.A.Kasaei, M.Taheri, " Modeling and Implemenaion of Omni-direcional Soccer Robo wih Wide Vision Scope Applied in Robocup-MSL ", 2nd inernaional Conference on Man-Machine Sysem 2009, Ocober 2009, Penang, Malaysia. [8] S.H.Kasaei, S.M.Kasaei, S.A.Kasaei, M.Taheri, pah planning for mobile robos Navigaion Applied o Dribbling via arificial Poenial Field, 1s Iranian Conference on sofware developmen, Khorasgan Azad Universiy (Isfahan) may 2007, Isfahan, Iran. Seyed Hamidreza Mohades Kasaei was born in Isfahan, Iran in He received he B.S. degree in Compuer engineering from Khorasgan Islamic Azad Universiy (Isfahan) in 2009, and he coninuous educaions in M.S. degree in compuer engineering field of Arificial Inelligen in Universiy of Isfahan. Currenly he is manager of hree Robocup eams in Roboic and Arificial Inelligence Cener of Khorasgan Islamic Azad Universiy, Isfahan, Iran. His eams work in Middle size soccer robo league, Humanoid soccer robo (Kid size) league and 2D soccer simulaion league ha hey can obain differen ranks in roboic compeiion. He has published over 40 research papers including one Bes Paper and one Innovaion Awards from Emeral Group Publicaion (England). His main research ineress are in Arificial Neural Nework, Muli agen sysems, Agen collaboraion, Roboic engineering, Machine vision and Inelligen conrol sysems. 623