co for po vis acc tra a str pre an dy Fu rob inc mu sen po cam De mo Wi soc So sho Facu  Abstrac  mplex moti r the movin osition pred sion is intr curacy and acking movi dynamic w rategy.   Thi ediction for nd the   Fuzzy  y namic envi  KeyWord  u zzy Contro The socce bots per a cludesmicro ulti agent re nsing in a u osition by p mera, and i ecision-Mak otion   speed Wireless Com The follo ccer robot occer Tourna There are ooting, drib  Positio for Sh  G  u lty of Ele  c t   :   The rob ion control ng object c diction algo roduced in robustness ing objects, window and is   paper   d r moving ba  y   Controller  ronment.  d s :   Soccer ol, Position P  1. Intr  e r robot sys team and orobotics, eal time con unique way processing it make a ch king system d   signal   o mmunication owing pictur system for ament)game several bas bbling and  o n Pred hooting  G yongIl Ry ectronics & Democ bot soccer has been w capture and orithm base n order to of the visio including a d an obstac deals   with all by using  r for shootin r   Robot,   K Prediction  r oduction  s tem consist Robot foo artificial ntrol and dy y. It has to an image f hoice of ne m, and then of   soccer n system. re shows th r Mirosot(( e. (Figure. 1 sic skills su interceptin  d iction o g Action  y ang, Myon & Automatio cratic Peop game base widely studi d shooting. ed on glob improve th on system f a   Kalmanfit  c le avoidan the   posito g   Kalmanfit  n g action in Kalman   filte  n  t s of 3 socc otball syste intelligenc ynamic visio estimate ba from a CC xt action in n it transm robots   in he structure Micro Rob 1) ch as movin ng in a rob  o f Ball a n in A So  n gSong Ch on,   Kim I  p le’s Repub ed ed A bal he for  t er , nce on  t er  n a er, cer em ce, on all CD n a mit a of bot ng, bot so  Sh  im ex re po be  v  re th a on ca  a nd Fuz occer R  oe, YongC  I l Sung   U blic of Kor occer game  h ooting . B mproved a h Figure The   kin xpressed as The   po epresented osition of c etween X-a  ω ,   are espectively. To desig he left and r desired po ne. [2-3] A Propo alculate the  z zy Con Robot Sy  hol Sin University, P rea e and the m y a good high accurac e 1. Structure nematic   mo follows.[2]  ⎪ ⎩ ⎪ ⎨ ⎧ = ⋅ = ⋅ = ω θ θ θ  & & &  sin cos  v y v x  o sture   of as   , , (  θ y x  c entral poin axis and fa linear   a gn Motion c right speeds osture   , (   d x  or tional Co e control fo  n troller ystem  P yongyang most importa motion co cy of action e of Soccer r odel   of   so ]  θ  θ soccer   ro )   ,   where nt and  θ   is ace of socc and   angul controller is s of robot so ) ,   d d y  θ   fro ontrol law[4 orce in prop  a nt skill is ontroller,it h n. robot system ccer   robot (1) obot   can ) , (   y x   is s a head an cer robot, a ar   velocit s to determ o that it mov om any init 4] is one t potion to er  the has is be the ngle and ties mine ves tial that rror
between a desired posture and a current one, and it is expressed as follows.  e e d r e e d l k d k v k d k v  θ θ θ θ ⋅ + ⋅ = ⋅ − ⋅ =   (2) The proportional controller have acheived zero error between target and current position, but not a desired head angle at the point. So control force has to be recalculated for turning robot to a desired one. A Fuzzy logic control law[5] is one that sub-controllers are the proportional ones and a entire control force is calculated by fuzzy logic combination of these sub-controllers. However,   the   high   game   performance depends   on   the   shooting   action,   and   it   is important to predict the ball position in real-time and move soccer robot to the predicted point correctly and quickly. Here,   we   formulate   the   following   two problems. First, the predictionof the ball postion in real time by using Kalman filter. Second, the design of the fuzzy controller to be movedsoccer robot to the correct position with keeping in the head angle of soccer robot smoothly  2. Prediction of ball position and Design of Fuzzy controller  1) Prediction of ball position In Mirosot(Micro Robot Soccer Tournament) game, the ball moves to any position continually without stopping, so it is important to predict the ball position in advance. The vision,however, is often used as the input sensor of the environment and several approaches have been proposed for positon estimation. But it is difficult to kick the ball exactly without prediction of ball position, because the speed of ball is fast in the game. As the image processing is more time-consuming, it is important to reduce the size of image to be processed for predicting the position in real time. We have introduced a Dynamic Window based object extraction as shown inFigure 2. Figure 2. A Dynamic Window The size of the square is defined as follows:  T v r h   ball ball   ⋅ ⋅ + ⋅ =   max 2 2   (3)  ball r   : the size of ball in image(Pixel), max  ball v   : the max speed of ball in the image(Pixel/s)  T   : sample time of the system (s) [Algorithm for Position Estimation]  Step 1 . Image capture: Image are captured in RGB in a   480 620   ×   resolution  Step 2. Choosing the dynamic window according to the position in last snap.  Step 3. Segment the dynamic window into small square, the size of block is decided by the ball size in the image(usually   ball r   /4) choosing the center of the square as a seed.  Step 3. Recursive algorithm to expand the area. Calculating the center of the region and filtering of the area including outside of the ball.  Step 4. Image Map to Global World The least square method is adopted to fit the map
⎩ ⎨ ⎧ ⋅ + + + = ⋅ + + + =  v u b v b u b b y v u a v a u a a x  3 2 1 0 3 2 1 0   (2) , where   ) , (   y x   is the ball position in the global world, and   ) , (   v u   is the ball position in image, and   3 , , 0 , ,   L = i b a   i i   are coefficients of approximation function. The Kalman filter can be described by the following equation to make prediction the ball position at next time.  T x z h x x   n n n n   / ) ˆ ( ˆ   1 1 1   − − −   − + =   & &   (5) ) ˆ ( ˆ   1 1 1   − − −   − + =   n n n n   x z g x x   (6)  n x &   : predicted estimate of ball velocity at time   n   , 1 ˆ   − n x &   : estimation of ball velocity at time   1 − n  and all preceding times, 1 − n z   : sensor reading at time   1 − n   ,  n x   : predicted estimate of ball position at time  n   , 1 ˆ   − n x   : estimation of ball position at time   1 − n  and all preceding times,  T   : sample time in vision system,  h   and   g   are filter parameters. 2) Design of Fuzzy controller for kicking up In soccer robot system, the Shooting action is   very   important   and   it   depends   how   well motion controller acts so that soccer robot move to target correctly and quickly from any posture. By using the preceding controllers, especially propotional and pure fuzzy one, the robot moves only forward. Therefore, it has to recalculate the control law for turning and back movement near to the target in order to improve the accuracy of control. However, in the game, the robot have to start from any posture and move to the target as   soon   as   quickly   and   correctly.   For   this problem, it contains not only forward movement controller,   but   also   turning   and   backward direction one. So, we have proposed the Fuzzy controller combining   with   forward,   backward   turning movements by using Fuzzy logic. The following picture shows the fuzzy input variables to be selected. Figure 3. fuzzy input variables  ①   Forward Direction Controller  [Algorithm   forForward   Direction Movement] Step   1 :   Determining   the   control   gains: 12 . 0 , 85 . 0   = =   a d   k k  Step 2 : Calculating postion and angle errors. 2 2   ) ( ) (   R d R d   y y x x d   − + − =   (7) ) , ( 2 tan   R d R d T   x x y y a   − − = θ   (8)  R T e  θ  θ  θ   − =   (9)  Step 3 : Calculating the control law.  e a d forward R e a d forward L k d k V k d k V  θ θ ⋅ + ⋅ = ⋅ − ⋅ =   (10) ,where   forward R forward L   V V   ,   are   left   and   right speed of soccer robot respectively.  ②   Backward Direction Controller  [Algorithm   forBackward   Direction Movement]
Step   1 :   Determining   the   control   gain: 12 . 0 , 85 . 0   = =   a d   k k  Step 2 :.Calculating postion and angle errors. 2 2   ) ( ) (   R d R d   y y x x d   − + − =   (11) ) , ( 2 tan   R d R d T   x x y y a   − − = θ   (12)  R T e  θ  θ  θ   − =   (13) If   e  θ   is more than 180° ormore less -180°, then   the   angle   is   changeable   to   value between -180° and 180°.  Step 3 : Calculating the control law.  R Backward R e a d R L Backward L e a d L V V k d k V V V k d k V  − = ⋅ − ⋅ = − = ⋅ + ⋅ =  , ,  θ θ   (14)  ③   Turning Controller  [Algorithm for Turning Movement] Step 1:   Determining the control gain:   2 . 0 = a k  Step 2: .Calculating angle errors.  R d e  θ  θ  θ   − =   (15) , where   d  θ   is the desired angle, and   R  θ   is the current one. If   e  θ   is more than 180° ormore less -180°, then the angle is changeable to value between -180° and 180°.  Step 3 : Calculating the control law.  e a Turn R e a Turn L k V k V  θ θ ⋅ = ⋅ − =   (16) The figure 4 shows the fuzzy input partition, and the fuzzy rules are as follows.  [Fuzzy Rules]  Rule 1: If   《 the position is   Near 》 and 《   R  θ   is not 0 》  Then   ) 0 , ( ] , [   R R L   l TurnContro V V  θ =  ... ... ...  Rule   26: If   《   R  θ   is 180 》 and   《   T  θ   is 180 》  Then ) , , , , ( ] , [   T d d R R R R L   y x y x trol ForwardCon V V  θ  θ , =  [Algorithm for a proposedFuzzy Control] Step 1 : For Vision system, Estimation of the current position of ball soccer robots.  Step 2: Prediction of ball position and speed at next time by Kalman filter.  Step   3: Calculating   distance   and   angle   error between the predicted point and the current one.  Step 4: By using the proposed fuzzy controller, moving robot to the just back point of ball so that it can be convenient for kicking up.  Step   5: Determining   the   shooting   direction taking consideration of the position of opposite goalkeeper robot.  Step 6: If the ball is very near to the robot, turning the head of robot to above   shooting direction.  Step 7: Moving the robotto opposite goal along the straight linequicly for a very short time. Figure 4. Fuzzy Input Partition  3. Experiments  To verify the proposed approach, we have the   following   experiment.   Based   on   the calculation   of   the   executive   time   of   vision system, we have determined the sample time as 35ms and allocated 3 robots per a team, and it
tak res the mo pe pro orb G po soc the inc rap usi en red kes 5 minut The follo sult of game e preceding ore   goals rformance oposed met bit of socce N umber o Goals(precedin Figure 5 It shows osition by us ccer robot h e ball corre convenient pidly by usi In this pa ing   global nvironment i First, we duce   the  es per a gam owing table e between t g one. As y and   it   m of   shootin thod and t er robots . table. Num f round ng one)   5( 5. Prediction o that it have sing kalman have   mov ectly. And environme ing the prop  Con  a per, an app l   vision   s is presented e introduced time   of  me. e shows the the proposed you know, i means   that ng   action b the Figure mber of Goals 1   2 (4)   6(5) of ball and Sho e been pred nfilter   corr ved the posi it shows th ent   ,   it   h posed fuzzy  n clusions  p roach for system   in d. d a dynami image   pro  e comparisio d method an it records th improve   th by   using   th 5 shows th s 3   4 3(2)   7(4 ooting action dicted the ba rectly and th ition shootin hat under th have   move controller. ball shootin a   dynam ic window ocessing   an  on nd he he he he 4 4) all he ng he ed ng mic to nd pr co w po th ba ag ex [1 C pp [2 20 [3 M Sh In [4 M S M pp [5 of C [6 E W C pp  redicted the ontinually in Second, which can m osture smo he   basic   m ackward dir Lastly, t greement of xperiment. 1]   YongCh Control   》   , p.595 2] Viki Lee, 002, pp.13- 3] Ping Zho Mobile Rob hooting 》 , ndustrial Inf 4]   G.Klan Modelling imulator   》  M odelling o p.137-150 5] Peter Jam f Control a Central Quee 6] Martin Se Estimation a Wheeled Mo Conference p.1120-1129 e ball positio n real-time. we propo move robot t othly, uses motion   cont rection and the presente f the propo  Re  h ol   Sin,  KIM   IL  ,   《 Robot S 35 ou, Alihua Y ot for Mov IEEE Inte formatics 20 ncar,   《   R  and   Valid ,   Mathem of Dynamic mes Thomas and Strateg ensland Uni eyr, 《 Propri and Slip   C obile Robo on Robotic 9 on which m . sed the fuz o target star fuzzy com troller   such turning one ed simulatio osed method  e ferences  《   Theory  SUNG   uni Soccer 》 , Y Yu,   《 Moti ving Object ernational C 006, pp.136 Robot   Socc dation   in matical   an cal Systems s, 《 Evolutio gies in Rob iversity 200 ioceptive Na Control for ots 》 , IEEE cs and Auto moved 8 rob zzy control rting from a mbination w h   as   forwa e. on shows go d with the r of   Intellig iversity   20 Yujin Robot ion Control t Capture a Conference 69-1374 cer   Collis Multi-Ag nd   Compu s 2003, vo onary Learn bot Soccer 03, pp.89-11 avigation, S Autonomo E Internatio omation 20 bots ller any with ard, ood real gent 06, tics l of and on ion gent uter l.9, ing  》 , 15 Slip ous onal 06,
[7] J. Campbell, 《 A Robust Visual Odometrya nd Precipice Detection System Using Consumer-grade Monocular Vision 》 ,IEEE on Robotics and Automation 2005, pp.3421-3427 [8] G.Cao, 《 A novel local invariant descriptor adapted to mobile robot vision 》 , American Control Conf 2004, Vol 13,pp.2196-2201 [9] S. Lim, 《 Optical flow estimation using high frame rate sequences 》 , Image Processing Conf 2001,Vol 2, pp.925-928 [10] Sheng Fu,   《 SLAM for Mobile Robots Using Laser Range Finder and Monocular Vision 》 , IEEE 2007, pp.13-18