A Fuzzy Based Hierarchical Coordination and Control System for a Robotic Agent Team in the Robot Hockey Competition

A Fuzzy Based Hierarchical Coordination and Control System for a Robotic Agent Team in the Robot Hockey Competition Hani Hagras, Senior Member, IEEE, Rabie A. Ramadan, Mina Zaher, Hala Gabr and Hussien Fahmy Abstract This paper presents the system used by the team of the German University in Cairo (GUC) within the FESTO Hockey Challenge league that took place within RoboCup 2009. The goal of the FESTO Hockey Challenge is to have a competition between robotic teams where each team consists of three robots to compete in an Ice Hockey game. All robots are of the same mechanical, sensor and electronic capabilities so that the focus of the competition is to develop novel artificial intelligence techniques for robot control and coordination. The GUC team scored the 2 nd place in this competition after losing by penalty shoot outs in the final. The proposed approach for GUC team employed Hierarchical Fuzzy Logic Controllers (HFLCs) in which the low level behaviours are implemented using FLCs and the coordination between the behaviours is implemented by a high level layer. The coordination between the robotic agent team members is implemented by a situation based dynamic role allocation mechanism. The paper will describe the employed approaches and will report on the results achieved. T I. INTRODUCTION HE Festo Hockey Challenge league was introduced by FESTO company in the 2009 International Robotics Competition (RoboCup 2009) that took place in Graz, Austria, July 2009. The goal of the Hockey Challenge is to present a standard league where all the competing teams will be using the same robotic platform which has the same mechanical, sensor and electronic capabilities. Hence, the focus of the competition will be to develop novel artificial intelligence techniques that could advance the fields of robotic control and multi robotic agent coordination. The competition arena consists of a field of 4.5 m x 4.5 m. The goal area is marked green or blue at the back walls. The width of the goal is 50 cm. In front of the goals there are half circles (called the front line of the goal) of radius equal to 50 cm. These circle lines consist of metallic stripes which can be detected by inductive sensors. There are two additional metallic lines dividing the field in three equal parts one attacking area and two defensive areas. There are 4 bully-off points in the attacking area and 4 bully-off points in the defensive area. The bully points are marked by black circular disks of 10 cm diameter. Fig.1a shows the layout of the arena H. Hagras, A. El Molla, M. Zaher. H. Gabr, H. Fahmy are with the German University in Cairo, New Cairo City, Egypt (e-mail: hani@essex.ac.uk). and Fig. 1b shows a photo of the real arena. Each team consisted of three Robotino robots which have the same mechanical, sensing, communication and electronic platform. Each team can wear either a stripped green or blue t-shirt. Fig.1. The arena layout in which there is a division of the field into five lines, two goals, one kick-off point and 8 bully-off points [5].. A Photo of the real arena [5]. This paper will present the machine vision, robot control and the multi robotic agent coordination techniques for the GUC team. The employed techniques employ Fuzzy Logic Controllers (FLCs) and Hierarchical Fuzzy Logic Controllers (HFLCs). The GUC teams scored the 2 nd place in the competition after losing in the final match by penalty shootouts which verifies the power of the employed techniques. Section II, will provide an overview on the rules of the hockey challenge. Section III, will provide a high level overview on the Robotino robots [4], [5]. In Section IV, we will present the employed computer vision techniques. Section V will present HFLC approach. Section VI will present the robots behaviour structure. Section VII will present the robots intermediate behaviours. Section VIII will present an overview on the robotic agent team coordination system. In Section IX, we will present the results followed by conclusions and future work in Section X. II. THE FESTO HOCKEY CHALLENGE COMPETITION RULES The Festo Hockey Challenge rules follow the regulations of the International Ice-Hockey Federation (IHF) but are simplified for the competition. Each match has a referee. The regular playing time of a match is divided into three thirds of 5 minutes each. A goal will be only scored if the following is fulfilled: The puck has completely crossed the goal line and is

inside the goal area. The last player who has touched the puck must be in the defensive area of the corresponding goal. The puck was not touched by one player of the attacking team being beyond the front line of the goal. This means that no part of this player is between goal and front line of the goal. A game will be started at the beginning of each third or after a goal. The start will be done at the kick-off point. All team members must be in their downfield in front of their defensive line. The referee places the puck on the kick-off point. Only one player of a team is allowed to be nearby such that he can immediately catch the puck with the pushing device. After scoring a goal, the non successful team receives the right to start at the kick-off point. If there is an irregularity, the referee will start the game at one of the bully-off point and then only one player of each team is allowed to be nearby in order to catch the puck. The other team members must be in their downfield at least one meter away from the bully-off point. An offside will be counted if the attacking player is in the defensive area of the opponent team before the puck has completely crossed the defensive line. In the case of offside, the referee will stop the game and will restart the game with at the next bully-off point outside the defensive area. It is only allowed to have only one team member in their own defensive area, otherwise the referee will stop the game and give a penalty shoot. If the boundary of a goal is displaced then the game will be interrupted and penalty shooting will be counted against the team whose robot displaced the goal. If one player pushes a player of the opponent team to force him in a non regular position then the referee will stop the game and there will a penalty shoot against his team. Penalty shooting will be performed where one player of the attacking team will be selected to perform the penalty shot. One player of the defensive team will be selected as goalkeeper. The player of the attacking team starts moving with the puck at the kick-off point if the referee releases the penalty shot. The player has only one chance to hit the puck in the goal. The goalkeeper can try to stop the attacking player but no wheel of the robot is allowed to cross the front line of the goal. III. AN OVERVIEW OF THE ROBOTINO ROBOT The mobile robot system Robotino (shown in Fig.2a ) [4], [5] is driven by 3 independent omnidirectional drive units. They are mounted at an angle of 120 to each other. The three omnidirectional drive units of Robotino define the robot as being holonomic, meaning that the controllable degrees of freedom equals the total degrees of freedom of the robot. The chassis is protected by a rubber bumper with integrated switching sensor. The robot diameter is 370 mm and its height including housing is 210 mm. Each of the 3 drive units consist of DC Dunker motors with nominal speed of 3600 rpm. The robots have an incremental encoder with a resolution of 2048 increments per motor rotation. The robot is equipped with 9 Infrared Red (IR) distance measuring sensors which are mounted in the chassis at an angle of 40 to one another. The sensors are capable of relative distance measurements to obects at distances of 4 to 30 cm. The anticollision sensor is comprised of a switching strip which is secured around the entire circumference of the chassis. The robot has also an inductive proximity sensor which serves to detect metallic obects on the floor. In addition, the robot has also diffuse sensors which consist of a flexible fibre-optic cables connected to a fibre-optics unit which works with visible red light. The robot is also equipped with a colour webcam. There are numerous communication interfaces on board including 2 Ethernet ports, 2 USB ports, 2 RS232 ports, parallel port, VGA port and Wireless LAN Access Point following the standards 802.11.g and 802.11.b. The access point can be switched into a client mode. There is a Linux operating system with real time kernel running on the embedded PC 104 pulse with AMD LX800 processor. The main part of the controller is the Robotino server, a real time Linux application. It controls the drive units and provides interfaces to communicate with external PC applications via W-LAN. There is an API with libraries which allow to create applications for Robotino in numerous programming languages including C++ und C which were used by the GUC team. IV. THE EMPLOYED COMPUTER VISION TECHNIQUES The computer vision system employed by our team processed the images captured by the each robot camera to provide information about the orientation and depth (distance) from the puck, the two goals, the team mates and the opponents as well positions of different arena landmarks. We needed very robust computer vision systems that can provide very fast updates for the dynamic hockey game. We used only one camera for each robot which meant that we could not benefit from stereo vision algorithms and we did not use panoramic vision. (why did not we use the panoramic vision? ) This makes calculating depths and orientations of different obects quite challenging so we had to improvise. The application also needed real-time performance, so we had to focus on algorithm efficiency as we could not afford any delays in the robot reaction and we had to process three video streams on the same server computer. Furthermore, the robots had a complicated shape which made their identification harder. We also had to adapt to different lighting conditions and above all the arena did not have enough reference points. To overcome these problems, the vision system was mainly divided into four phases: image analysis and region formation, extracting region information, obect identification and pattern recognition and finally calculating

different obect locations. These phases are discussed in the following subsections: A. Image Analysis and Region Formation This phase was mainly inspired by the open source library CMVision developed in Carnegie Mellon University [3] although we made a lot of maor modifications to make it suitable for our application. The purpose of this phase is to extract different connected regions from the image each having a certain colour. This can be applied using colour segmentation. At first, each pixel is assigned a certain colour of interest using 2 bitwise AND operations and a single table lookup as in [2]. After that, primary regions are formed by applying a Run Length Encoding on the image grouping all consecutive pixels with the same colour together in one region. Finally, larger regions are formed by merging primary regions on 8-connectivity basis using a Union-Find algorithm with path compression and union by rank heuristic which is proved to be linear for all practical values of the problem. Unlike the CMVision library, we used the HSV colour space instead of the YUV colour space because it gave us more power in capturing the different colours we needed and in making the system highly adaptable to different lighting conditions. We also performed 8- connectivity in region grouping instead of 4 connectivity which gives more accurate results and we did that without having to iterate on the whole image but on the formed regions only. B. Extracting Region Information. After performing the segmentation in phase one, we have a set of regions with different colours in different parts of the image. This phase is responsible for collecting information about each of these regions. For each region produced in the previous phase we get the bounding box, the region colour, area and centroid. C. Pattern Recognition and Obect Identification In order to identify a region as a particular obect, we use some heuristics that try to match the properties of that region extracted from the previous phase with the real properties of the obect in question. For example: the puck should be the only red region in the scene with some minimum area in order to avoid noisy red parts, the green goal is a rectangular green region with certain geometric ratios and so on. The robots were a bit tricky because they had an irregular shape and they appeared on the segmented image as a group of separate black regions. This problem was solved by performing nearest neighbour clustering on these black regions and grouping all the little regions near to each other into a single cluster to represent one robot. D. Calculating Different Obect Locations Getting the obect depth with one camera was one of the trickiest parts. We were able to develop a simple yet powerful method to calculate the real depth of obects accurately in terms of camera parameters like height, tilt and resolution (Fig.2b shows the parameters involved in calculating depth and orientation for our vision system). This method can get the depth of any obect on the ground or above the ground with a known height with an accuracy of 90-95 %. The idea depends on observing the phenomena that obects lying on the same vertical level but different distances from the camera have different heights in the image taken with that camera. In our case, the relation between the real depth (d) of the obect and the height (h) of that obect in the image is given by the equation: d = h * tan ( ) (1) It is ust one line of code, which is much faster than the typical approaches that use stereo vision. The orientations can also be calculated using the same concept. Fig.3a shows the real image obtained from the robot camera and Fig.3b shows the corresponding segmented image obtained from the vision system which shows the accuracy of finding the various obects of interest. E. Capabilities of the Computer Vision System By the end of the four phases, the vision system was capable of performing the following using only one camera: Identify all obects of interest with very high accuracy even from long distances. Calculate depths and orientations of different obects with an accuracy of 90-95 % on a 320 240 resolution (higher resolutions would yield better results). Process a video stream of 15 frames per second from each of the three robots using a 5 % utilization of a 3 GHz processor running Windows XP. Adapt to different lighting conditions. Locate the absolute location of a robot in the arena. Fig. 2. Image of Robotino of Festo Didactic GmbH & Co.KG [4], [5]. The parameters involved in calculating depths and orientations for our vision system Fig.3. The real image captured from the robot camera. The segmented image obtained by our vision system.

V. THE ROBOTS HIERARCHICAL FUZZY LOGIC CONTROLLERS (HFLCS) Single rule base FLC suffer from the serious limitation that the number of rules increases exponentially with the number of variables involved [10]. If we have a robot with only eight input sensors, if we represented each input by only three fuzzy sets then for a single rule base we need to determine 3 8 =6561 rules, which is very difficult to design, also this huge rule base translates directly to slower controller response. To cope with the rule explosion problem and its effects on the design and real-time operation of the FLCs, we will hierarchically decompose the control problem by breaking down the input space for analysis by sharing it amongst multiple low level fuzzy behaviours. Each behaviour responds to specific types of situations, and then integrating the recommendations of these behaviours via a high level fuzzy coordination layer. Each behaviour is an independent and self contained FLC with a small number of inputs and outputs and a small rule base and it serves a single purpose (e.g. edge-following, obstacle avoidance, goal seeking, etc ) while operating in a reactive fashion. The behaviours will typically (but not necessarily) map different inputs sensors to common actuators outputs [9]. Such primitive behaviours are building blocks for more intelligent composite behaviours, i.e. their capabilities can be combined through synergistic coordination by a high level fuzzy coordination layer to obtain an overall coherent behaviour that achieves the intended task(s). Fuzzy coordination allows the ability to express partial and concurrent activations of behaviours and the smooth transition between behaviours. The hierarchical fuzzy systems have a nice property that the total number of rules increases linearly rather than exponentially as in the single rule base FLC [10]. For example, we divide the robot controller into four cooperating behaviours namely obstacle avoidance, goal seeking, left and right edge following [6]. If we represented each input using three fuzzy sets (as in the case of a single rule base FLC) then the obstacle avoidance behaviour, using three forward facing sonar sensors, will have a rule base of 3 3 = 27 rules. The left edge following behaviour, using two left side facing sonar sensors, will have a rule base of 3 2 = 9 rules, the right edge following behaviour will have the same number of rules. The goal seeking behaviour, using a single goal detection sensor will have only a rule base of 3 rules. Thus the total number of rules in the low behaviours is 27 + 9 + 9 + 3 = 48 rules and the total number of rules in the coordination layer is 4 (number of behaviours), thus we need a small number of rules which are much easier to be determined than 6561 rules in the case of the single rule base FLC. There are many papers reporting implementations of Hierarchical Fuzzy Logic Controller (HFLC) which had produced good results [1], [6], [7], [8], [9]. Our work in this paper confirms the efficiency of the HFLC. The following section will explain the behaviours structure for the robots. VI. THE BEHAVIOURS STRUCTURE In our HFLC, each low level reactive behaviour will be an FLC. Each low level behaviour will receive a subset x h of the total crisp inputs x available to the HFLC, all behaviours will produce preferences to the same common outputs, which are the outputs of the HFLC, so each behaviour will map different inputs to common outputs which are the omni-drive rotation magnitude and the omni-drive side movement input whose fuzzy sets are shown in Fig.4. The low level FLCs outputs will approximate the centroids of output fuzzy sets to represent the preferences from the perspective of the low level behaviour goals. Fig. 4. The Fuzzy sets for the behaviours FLCs and HFLC outputs. The omni-drive rotation magnitude. The omni-drive side movement input. We have used four low level behaviours which are obstacle avoidance, goal seeking, left and right edge. In the following subsections, we will explain further each of these low level behaviours. These basic FLCs compose the basics of our robots low level, intermediate and high level behaviours. A. The Obstacle Avoidance Behaviour For the Obstacle Avoidance (OA) behaviour, we used 7 IR sensors from the 9 that surround the robot s casing. To limit the rule base size we grouped those sensors into three groups. The first group detect obstacles facing the robot, this group consists of the front sensor and its two neighbours from the right and left. The OA FLC was fed with the minimum of those three sensors to represent the sensor reading for obstacles located infront of the robot. The other two groups represented obstacles that are located on the robot s right and left sides. These two groups share a sensor with the front sensors group in addition to two more sensors located on the right or left of the robot s case. Again, the OA FLC was fed with the minimum of those three sensors on the right and left. Hence, the OA receives three inputs where each represents the minimum distance from the front, left and right sides respectively. Each of the FLC inputs were represented by two fuzzy sets Near and Far as shown in Fig.5a. Due to the IR sensors short range, the obstacle avoidance behaviour also takes machine vision inputs into consideration detecting other robots as obstacles from far distances. Hence, the OA behaviour receives also inputs representing the orientation (represented by three fuzzy sets as shown in Fig. 6a) and depth (represented by two fuzzy

sets as shown in Fig.6b) from the opponent robot which are viewed as obstacle. Hence, the OA behaviour has a rule base of 2*2*2*3 = 24 rules. Fig. 5. The fuzzy sets used each sensor group for OA behaviour FLC. The fuzzy sets for the inductive sensors for the goal keeper behaviour. B. The Left and Right Edge Following Behaviours For the left edge following behaviour using two left side facing sonar sensors and representing each sensor by three fuzzy sets, this will lead to a rule base of 3 2 = 9 rules, the right edge following behaviour will be the same. C. The Goal Seeking Behvaiour For the Goal Seeking (GS) behaviour, the target can be the red puck or the opponent goal or the team own goal or the opponents robots. The inputs for the goal seeking behaviour are obtained from the machine vision block and they represent the depth and the orientation from goal target (wither the target is the red buck or the opponent goal or the team own goal or the opponent robot). For the goal seeking behaviour, the orientation input is represented by three fuzzy sets (as shown in Fig. 8a) while the depth (distance) from the goal is represented by two fuzzy sets (as shown in Fig. 8b). Hence, this will lead to a rule base of 6 rules. the gripper. The first one is exactly behind the gripper while the second is 15 cm behind the gripper. These sensors can detect the silver line that forms a semi circle in front of the goal marking the goal keeper area. When the sensor is located on top of the middle of the silver line the reading ranges from 0.4 to 3.0 as the sensor moves forward or backward to scan the edge of the silver line the readings range from 3.0 to 6.0 as the sensor can not detect the silver line it returns readings above 8.0. Using the two inductive sensors the goal keeper behaviour implements the line following behaviour to follow the silver line. Together with the puck orientation, the robot can be guided to move on the line of the goal area and protect the shooting angles. The two inductive sensors inputs are both expressed using three fuzzy sets (as shown in Fig.7b) and the puck orientation is expressed using three membership sets resulting in 3 3 = 27 rules. When the robot can not locate the puck the goal keeper will replace the puck orientation input with the orientation of the nearest opponent robot. If for a several seconds the puck did not re appear, the robot will start looking around for it while staying tight in goal. On the other hand, if the puck depth becomes within a certain range the goal keeper will burst and grab the puck to hinder any shots being made. E. The High level Coordination Layer Fig.7 shows the HFLC architecture, at the high level there is high level FLC which is responsible for the coordination of the low level FLCs based behaviours. Each low level behaviour has a context of activation C representing when it should be activated. The context is represented by fuzzy set as shown in Fig.8 to handle the linguistic and numerical uncertainties associated with these contexts. Each behaviour is activated with a strength given by the truth value of the context, i.e. the degree of firing of the fuzzy set. Fig.6. The fuzzy sets for the GS FLC for The orientation input. The depth input. D. Goal Keeper The Goal Keeper is considered a key player of the team as a good goal keeper is a maor strength to the team defence capabilities. The goal keeper behaviour is another FLC which takes as input two from the inductive sensor readings and inputs from the machine vision block. The machine vision block provides the puck orientation and depth. According to the puck orientation, the robot will move left or right to block the shooting angle. In order to block the far post angle the robot needs to stand outside from the goal and move in a semi circle in front of the goal. In order to do this the inductive sensors were used. We equipped each robot with two inductive sensors. Those sensors are located behind Fig. 7. The HFLC architecture.

calculated according the following equation: Fig. 8. The fuzzy sets for the contexts. The high level FLC receives the crisp inputs to the contexts d (=1,..H), where H is the total number of behaviours. These crisp inputs are then fuzzified by matching each input to its context fuzzy membership function. When the crisp inputs are fuzzified against the fuzzy context membership functions we obtain for each crisp input d a membership value ( d ) for each context C. C The high level coordination FLC has a coordination rule base which contain coordination rules that describe in a fuzzy way when each behaviour should be activated to influence the operation of the robot at each moment, the coordination rules take the following format: IF d is C THEN Behaviour Where d is the crisp input to the context total number of behaviours. B =1,..H (2) C and H is the B is the behaviour output fuzzy set. Note that we have a coordination rule for each behaviour, thus we have a total of H coordination rules. From Equation (2) we see that that each behaviour output should be activated with a strength given by the truth value of its context C. In the inference engine of the high level FLC, we use the product t-norm for the implication operation and maximum t- conorm to aggregate the fuzzy outputs of the various rules. The Height defuzzification is employed for the high level FLC. However for the Height defuzzification, we need to compute the centroid of the output fuzzy sets from each low level behavior. However if in each control cycle, we calculate the centroid of the combined output fuzzy set of each behavior which again could be approximated by calculating the Height defuzzification for each low level behavior. At each control cycle, each low level FLC based behaviour will receive a subset x h of the total crisp inputs x available to the HFLC and will generate a preference from the perspective of its goal which is represented by the defuzzified output of each behavior (which approximates the centroid of the behaviour output fuzzy set). The consequent of each context rule is the behaviour output fuzzy set B and its centroid is approximated by the defuzzified output of the behavior output, k=1, c. The crisp output for each HFLC output y k. Where k is HFLC ytk k (k=1, c) is yt k = H 1 H 1 k y The crisp input d 1 to the obstacle avoidance behaviour context is the minimum value of the front sonar sensors and the nearest depth to obstacles obtained from the vision system. The context C for the obstacle avoidance behaviour 1 is in the form of a fuzzy set as shown in Fig. 8 which defines that the obstacle avoidance behaviour should be active when the robot path is obstructed by an obstacle and the closer the robot gets to the obstacle, the higher will be the activation of the obstacle avoidance behaviour. d 1 is fuzzified using the fuzzy context membership functions of C to a membership 1 value ( d ). The crisp input d 2 to the left edge following C1 1 behaviour context C is the minimum value of the left side 2 sonar sensors. The crisp input d 3 to the right edge following behaviour context C is the minimum value of the right side 3 sonar sensors. The crisp input d 4 to the goal seeking behaviour context C is d 4 1,. The context C for the goal 4 seeking behaviour is in the form of a fuzzy set as shown in Fig. 8 which defines that the goal seeking behaviour should be active when the robot path is clear, the clearer the robot path the higher will be the activation of the goal seeking behaviour. d 4 is fuzzified using the context membership functions of C to a membership value ( d ). After all 4 C C C 4 4 (3) the crisp inputs d (=1,..4) are matched and fuzzified against their fuzzy contexts, the fuzzified values are then fed to the inference engine which determines which rules (and hence which behaviours) are fired from the coordination rule base. The coordination rule base contains a coordination rule for each behaviour which relates the contexts to the behaviours. In case of an attacking situation, the robot can coordinate OA, GS and the left and right edge following behaviours. As we have four behaviours then we will have four coordination rules as follows: IF d 1 is LOW THEN Obstacle Avoidance, IF d 2 is LOW THEN Left Edge Following, IF d 3 is LOW THEN Right Edge Following, IF d 4 is HIGH THEN Goal Seeking. The system is capable of performing very different tasks using identical behaviours by changing only the context rules and co-ordination parameters. For example, in the Goal defensive situation, the robot will be coordinating the goal keeper behaviour mentioned above with the GS behaviour which seeks the puck and opponent and when the puck is within a given safe distance, the robot can activate the goal seeking behaviour to get the puck. We can eliminate the unneeded behaviours from the context rules according to the robot s mission.

VII. INTERMEDIATE BEHAVIOURS Besides the basic FLC and HFLCs behaviours mentioned above, the robots have intermediate behaviours as explained in the following subsections. A. The Attacking Behaviour The attacking behavior makes use of the HFLC coordinating the OA and GS (The robots seek the opponent goal) behaviours. In case the robot can not locate the goal it will rotate around the puck in the gripper while performing obstacle avoidance. This means that the axis of rotation is a normal vector to the puck s surface. As the distance and angle from goal becomes optimum for shooting, the HFLC will be disabled and the shooting behvaiour will be triggered. According to the situation different shooting techniques take place. We have three different shooting mechanisms. The first one is shooting from long distances. This one will be triggered from a far distance from goal as it takes about 1.5 meters to execute. The robot simply rotates while moving forward in a certain direction. As the robot s side faces the goal it will suddenly rotate to the goals direction. Another interesting shooting mechanism is done when approaching the goal from a dead angle. The robot simply leaves the puck from the gripper and moves diagonally leaving the puck the right or left of the gripper. The second shooting mechanism involves the robot bursting to the right or left accordingly for about 1 meter drifting with the puck to the gripper s side. The robot would end this diagonal movement with a sudden rotation in the opposing direction rotating about 340 degrees hitting the puck to the facing direction. The third shot is the basic one which is similar to the second one but without the drifting part. Another attacking mechanism involves the edge following behaviours. This approach involves sneaking to the opponent defensive area. If the robot with the puck detects that it is still far away from opponent s goal and opponent s robots are between the robot and the goal, this behaviour will be enabled. The robot would keep track of its orientation from the goal. Then the robot would rotate 90 degrees or -90 to face any of the side walls. The robot would then move towards the wall until the IR sensors detect a certain range. The robot would then maintain a certain distance from the wall using the edge following behaviour while moving towards the opponent s defensive area. As the robot crosses the silver line marking the opponent s defensive area, or detects an obstacle in the way that might be an opponent s robot, the attacking robot would rotate in the opposite direction until it locates the opponent s goal. As the robot detects the goal it will continue with the normal attacking behaviour mentioned above. The advantage of such an approach is that the robot s gripper can totally hide the puck from the side view, and together with the wall the puck is totally hidden from opponent s robots. Finally, as the robot with the puck shoots, the game situation changes and this will be managed by the robotic agents team coordination module mentioned below. B. Opponent Obstruction Opponent obstruction is a behaviour designed to take one robot of the opponent s team out of play. The motivation behind this behaviour is that while our robot has the puck, the opponent team robots are all obstructing our attacker. For this reason it is a good choice to start obstructing one of them to enhance the attacker s odds to reach the goal. Opponent Obstruction behaviour simply makes use of the GS behaviour. This time the goal is one of the opponents robots. The behaviour consequents are modified slightly in the case of opponent obstruction in order to make the tackles with more speed to produce biggest impact possible. The rule of the game states that only one team robot can be present in its own defensive zone per instance. If another robot goes in its defensive area to retrieve the puck or obstruct another robot the other team is rewarded a penalty. To avoid those penalties, robots that are not assigned to be the goal keeper keep track of inductive lines all the time. Indicating an inductive line means that the robot is moving into one of the teams defensive areas. If the robot can detect any of the two goals, the robot can directly take the decision to enter the defensive area to pursue its goal or not. This can be done by checking the depth of the goal that enables the robot to know which defensive area it is about to enter or leave. On the other hand, if none of the goals is in sight the robot would have to check its situation. The robot would stop rotate until it locates one of the two goals, figure out its position with reference to that reference point and take the decision accordingly. C. Defence Behaviour The defence behaviour is a precautionary action to secure our defensive area when the puck is loose or with the opponent. The defence behaviour makes use of the GS FLC. However, this time the inputs are our team s goal. Furthermore, the FLC is modified to reduce speed in order to avoid entering our restricted defensive area. When the robot reaches a certain depth from our goal it will stop and start rotating trying to locate the puck in the field. As the robot locates the puck it will start tracking it. Note that the role assigned to that robot according to the situation is to only track the puck and do not interfere. This is because another robot is currently trying to grab the puck because it has a better chance to grab it. Also note that if the puck grabbing robot looses sight of the puck while the defence behaviour robot has the puck in its sight the roles will suddenly switch. The puck seeking robot that failed to retrieve the puck will take the defence role and the defending robot will burst to grab the puck leaving its position to the former robot. Finally, if all three robots loose track of the puck the game situation will change to the last situation to be discussed in the next section.

VIII. HIERARCHICAL COORDINATION OF THE ROBOTIC AGENTS TEAM From our perspective we have three main game situations. Situation one is when one of our robots has the puck. Situation two is when our team does not have the puck but one or more of our robots can locate the puck in the field. Situation three is when none of our robots can see the puck which directly implies that none of our robots have it either. The following subsections will explain each of these situations and how our team handles the given situation. A. First Situation: A Team Member has the Puck When one of the team members has got a puck in its gripper, then this will mean low distance for the front medium IR and low depth as obtained from the machine vision system. In such a situation, the robot with the puck will have to attack and will try to reach shooting positions. The second robot will return to be goal keeper or will stay as goal keeper if was previously assigned to be goal keeper. Finally, the last robot will obstruct opponents robots. B. Second Situation: Team does not have the Puck but One or More of Our robots Can Locate the Puck The second situation states that none of our team members has the puck; however, one or more of our team robots can see the puck. Behaviours of that situation are as follows: the robot that can locate the puck would do the goal seeking behaviour. This time the goal of the robot is the puck itself. The robot closest to our team s goal would return to goal keeping position, or stay in goal if was already assigned goal. The third robot would do the defence behaviour. It is important to point out that our game situations do not differentiate between a loose puck and an opponent s having the puck. In both case the game situation is set to the second one, however, the puck grabbing procedure referred to as Goal seeking of situation two differs in actions depending on the context as it is explained in the upcoming subsections. As mentioned in the second situation one robot performs GS to grab the loose or opponent s puck. A special GS behaviour which can be referred to as puck seeking or puck grabbing. Basically, this behaviour is based on the GS FLC where the GS FLC is fed with the puck depth and puck orientation. The behaviour is also fed with the locations of opponent s robots around the puck if any can be located. Note that this time OA is only activated if the puck depth is too far. Furthermore, if the located puck is close to an opponent s robot the controller outputs an excessive speed when the puck depth is near. This ensures that if the puck is near the opponent s robot or even in its gripper the excessive speed would knock the puck away from the opponent robot and out of its gripper if it was inside the gripper. Note that if the robot looses track of the puck and another robot locates it the same situation will still apply, however, the roles will change. The robot that can locate the puck will do the puck seeking behaviour and the other robot will probably do the defence behaviour. C. Third Situation: None of Our Robots can Locate the Puck In this situation no robot can locate the puck. For this reason highest defensive precautions are made. The robot most suitable (namely the robot with the least depth to our team s goal and most probably is already assigned to be goal keeper) for being the goal keeper will return to goal and enable goal keeper behaviour. The other two robots will do the defence behaviour. They will return to nearest post outside of our defensive area and start searching for the puck. Note that as any of the three robots detect the puck, the situation will change to situation two. If certain time duration passes and the puck was not detected then it is feared that the opponent s team are hiding the puck. The roles would change by switching the two defending robots to opponent obstruction behaviour for a duration of time allowing them to expose the hidden puck. Then they would return back to defence behaviour. D. The Team Hierarchical Coordination Mechanism In the previous sections the three game situations were explained together with the behaviours associated with them. Our coordination system is dynamic depending on the situations mentioned above and if one of the robots is missing due to technical failures or due to suspension from the game the behaviour with the least importance is neglected. For instance, assume that one robot is missing and game situation is situation two. This means that at least one of our team s robots can locate the puck, however, puck is loose or with an opponent s robot. According to which robot is more suitable for the role, one robot would grab the loose puck and the other would return to protect the goal by being assigned the goal keeper role. The third role is neglected as only two roles are available. The team hierarchical coordinator can be viewed as the high level coach that gathers information from the robots and determine the game situation. According to that game situation the high level coordinator would figure out the different roles to be assigned and their priorities. Then the coordinator would assign the role with the highest priority to the fittest robot for that role. Then it would assign the second important role to the more suitable robot from the other two. Finally, it would grant the last robot the least important role as it was not fit enough to fit the more important roles. This hierarchical coordination technique allows room for dynamic role allocation. In other words, if a robot malfunctions during the game it will not respond to the coordinator with updates. This indicates that the mentioned robot is the least suitable for any of the roles to be assigned, hence, giving the chance for the coordinator to assign the roles to the other functioning robot. Fig. 9 shows the structure of the team hierarchical coordination which assembles from each robots its own the distance from puck and from the team goal. The team

coordinator then analyses the situation of the robot depending on the assembled information. After this and depending on the game situation, the coordinator assigns the roles to the various robots depending on their location to the puck and to the team goal. Fig. 9. An overview of the structure of the team hierarchical coordination. Fig.10. The GUC (in blue) game against the University of Osnabrueck (in green) in the final game IX. RESULTS In this section, we will summarise the results achieved by the GUC team in the Festo Hockey Challenge held within Robocup 2009. These game results serve as comparison with the computer vision, control and the agent coordination systems followed by the other teams. In the first game, the GUC team played against ESTI Tunis (Tunisia) and the GUC team won 2-0. In the second game the GUC played against Polytech'Lille (France) and the GUC won 6-0. In these two games, the GUC team played only with one robot against three opponent robots (as two of GUC robots became faulty on this day). This shows the power of our techniques in term of defence and attacking and in spite of playing outnumbered the robotic team compensated with better vision, control and artificial intelligence systems. The GUC team then played against the University of Osnabrueck (Germany) who employed a panoramic vision approach which allowed them 360 o vision of the arena. However the GUC team won 4-0 due to the better fuzzy control and coordination systems employed. The GUC then won against HEIG Yverdon (Switzerland) 5-0 before loosing 2-4 from HHT Budapest (Hungary) due to sudden hardware failure in the GUC robots. The GUC then qualified to the semi final to play against Polytech'Lille (France) two games where the GUC won in the first game 3-1 and in the second game 5-0. The GUC then played in the final game against University of Osnabrueck (Germany) where the game finished 2-2 and the GUC lost by penalty shootout 1-0. Fig.10 shows a screenshot of the final game. X. CONCLUSIONS AND FUTURE WORK In this paper, we have show a description of the vision system, the control and team coordination mechanisms used by the GUC team for the Festo Hockey Challenge held within RoboCup 2009. We have shown the employed vision techniques which provided robust and accurate results using only camera. In addition, we have used a fuzzy logic based approach to implement the low level behaviours within the robots. We have also employed fuzzy context based coordination for the high level coordination of behaviours to form HFLCs. The robotic agent team coordinator involved a situation based hierarchical coordination which enabled the dynamic role allocation to the various robotic which allowed to handle the dynamic environment of the game. The GUC developed systems has enabled to gain the 2 nd place in the competition after loosing by penalty shootout in the final. For our future work, we will aim to employ type-2 fuzzy systems which are able to better handle the high levels of uncertainties encountered in this dynamic game. We will also investigate the application of reinforcement learning techniques to enable the robots to adapt to the changing environment and circumstances. ACKNOWLEDGEMENT REFERENCES [1] A. Bonarini, Anytime learning and adaptation of hierarchical fuzzy logic behaviours, Adaptive Behavior Journal, Vol. 5, No. 3-4, pp. 281-315, 1997. [2] J. Bruce, T. Balch, and M. Veloso, Fast and inexpensive color image segmentation for interactive robots. Proceedings of Proceedings.2000 IEEE/RSJ International Conference on Intelligent Robots and Systems ( IROS2000) Japan, October 2000, pp. 2061-2066. [3] CMvision: Open Source Library http://www.cs.cmu.edu/~bruce/cmvision/ [4] Festo Didactic GmbH & Co.KG: Technical Documentation of Robotino, Denkendorf 2007, http://www.festo-didactic.de [5] Festo Didactic GmbH & Co.KG: http://www.festo-didactic.de [6] H. Hagras, V. Callaghan and M. Colley, "Prototyping design and learning in outdoor mobile robots operating in unstructured outdoor environments," IEEE Magazine on Robotics and Automation, Vol. 8, No.3, pp.53-69, September 2001. [7] H. Hagras, M. Colley and V. Callaghan, " Learning and adaptation of an intelligent mobile robot navigator operating in unstructured environments based on a novel Online Fuzzy-Genetic System," Journal of Fuzzy Sets and Systems, Vol. 141, No. 1, pp. 107-160, January 2004. [8] A.Saffiotti, The uses of fuzzy logic in autonomous robot navigation, Journal of Soft Computing, Vol. 1, No. 4, pp. 180-197, 1997. [9] E. Tunstel, T. Lippincott and M. Jamshidi, "Behaviour hierarchy for autonomous mobile robots: Fuzzy behaviour modulation and evolution," International Journal of Intelligent Automation and soft computing, Vol. 3, No. 1, pp. 37-49, 1997. [10] L. Wang, Analysis and design of hierarchical fuzzy systems, IEEE Transaction on Fuzzy Systems, Vol. 7, No. 5, pp. 617-624, October 1999.