Eagle Knights 2006: Four-Legged League

Eagle Knights 2007: Four-Legged League Alfredo Weitzenfeld1, Alonso Martínez1, Bernardo Muciño1, Gabriela Serrano1, Carlos Ramos1, and Carlos Rivera1 1 Robotics Laboratory Lab, ITAM, Rio Hondo 1, 01000 DF, Mexico [email protected] Abstract. In this paper we present the system architecture for our Four Legged RoboCup Soccer Team – Eagle Knights. We describe the system architecture: Vision, Localization, Sensors, Kinematics, Wireless Communication and Behaviors. Keywords: Four-legged, robocup, autonomous, vision, architecture. 1 Introduction RoboCup [1] is an international effort to promote AI, robotics and related field primarily in the context of soccer playing robots. In the Four Legged League, two teams of four robots play soccer on a relatively small carpeted soccer field. In addition to soccer playing, we have used our robots in other research projects, including biorobotics [2] and human-robot coaching [3]. Fig. 1. Four-Legged System Architecture. 2 RoboCup Four-Legged System Architecture The general four-legged system architecture includes the following modules to be explained in more detail in the following sections: • • • • • • • Sensors. This module receives information from the pysical sensors. We are particularly interested in vision and motor position feedback. Vision. The vision module receives a raw image from the camera, the main system sensor, and performs segmentation over the image. The module then recognizes objects in the field, including goals, ball, landmarks and other players. Motion. This module control robot movements, such as walk, run, throw the ball, turn, move the head, etc. It receives commands from the behavior module with output sent to the corresponding actuators representing individual leg and head motor control. Actuators. In addition to individual head and leg motor control, the system includes actuators for turning off and on head LEDs. Behaviors. This module makes decisions affecting higher level robot actions. It takes input from the sensors and the localization system to generate commands sent to the motion and actuators modules. Wireless Communication. This module receives all commands from the external Game Controller. The system transmits a data structure between all robots with information such as player id, location of ball if seen, distance to the ball, robot position and ball position. Localization. This module makes all the processing necessary to obtain a reliable localization of the robot in the field. In order to localize, our current model requires the robot to perceive at least two marks, either goals or landmarks. 2.1 Sensors This module receives information from the color camera and motor position feedback. The raw camera image is passed directly to the vision module while information received from motor positions are used in making certain movement decisions at the behavior module, such as when the robot is on its back. 2.2 Vision The vision system is inspired in previous work in the Small-Size League [4]. We divide the system into two modules: Calibration and Segmentation. Calibration The system is calibrated prior to the game. The regions of interest are selected in a HSI color space and then transformed to a YCrCb color space through a GUI that uses pictures previously taken. These regions are subsampled to reduce memory in the robot at 12.5% of the original usage. Segmentation and blob growing Using the previously defined values, the 7 colors of the game are assigned to objects in each image taken from the field. The segmented pixels are grouped by regions (blobs) for further processing. Fig. 2. A sample image classified using our calibration system. Real object images are shown on the right column, while classified objects are shown on the left column. Figure 2 shows sample output of the segmentation calibration process. Objects can be recognized after color regions are processed. Objects in the field must fulfill certain requirements in order to allow some confidence that the region being analyzed corresponds to the object of interest. For example, the ball must have green in some adjacent area with similar criteria used to identify goals. The identification of the landmarks is a little more complex, nevertheless after more elaborated comparison landmarks are identified. Finally, a data structure filled with object positions and additional objects characteristics. 2.3 Motion This module is responsible for robot motion control, including in particular walking and kicking. We have developed a number of different routines depending on team roles. For example, the goalie has different motions in contrast to other team players. This also applies to different head kicks and movements in general. 2.4 Actuators This module is responsible for turning off and on the head and tail LEDs. This module receives commands from the behavior module to indicate the particular action being currently performed by the robot. This module also displays the state of the Game Controller. Fig. 3. Basic Individual Goalie State Machine. 2.5 Behaviors The behavior module receives information from sensors and localization, sending output to robot actuators. In defining our team robot behaviors, we specify three types of players: Attacker, Defender and Goalie. Each one has a different behavior that depends on ball position and Game Controller. Goalie basic behavior is described by a state machine as shown in Figure 3: • Initial Position. Initial posture that the robot takes when it is turned on. • Search Ball. The robot searches for the ball. • Reach Ball. The robot walks towards the ball. • Kick ball. The robot kicks the ball out its goal area. Attacker basic individual behavior is described by a state machine as shown in Figure 4: • Initial Position. Initial posture that the robot takes when it is turned on. • Search Ball. The robot searches for the ball. • Reach Ball. The robot walks towards the ball. • Kick Ball. The robot kicks the ball towards the goal. • Explore Field. The robot walks around the field to find the ball. Fig. 4. Basic Individual Attacker State Machine. 2.6 Wireless Communication This module receives commands from the Game Controller and passes them to the Behaviors module accepting connections using either TCP or UDP protocol. This module is used to cooperation between robots for have information about the state of the game and the world. 3 Localization An important part of playing soccer is being able to localize in the field in an efficient and reliable way. Localization includes computing distances to known objects or landmarks, use of a triangulation algorithm to compute exact positioning, calculate robot orientation angles, and correct any resulting precision errors. A block diagram for the algorithm is showin in Figure 5. In the next section we describe our Localization system based on 2006 rules. We are currently working on a new version that will adapt to the new rules of the league. The new Localization system is extended from the previous versions and will take into account probabilistic methods in conjunction with odometry data. 3.1 Distance to objects The first step in localizing is to obtain the distance between identified objects and the robot. It is important that the robot can distinguish at least two marks when it starts. After making several experiments, we developed a simple algorithm that computes distances to objects by using a cubic mathematical relationship that takes as parameter the object area and returns as result the distance to the object. In order to obtain this relation, we took a large number of measurements at different distances from the object that we are interested in. The distance range used went from 15 centimeters to 4 meters. Beyond four meters it became very difficult to distinguish between objects and noise. Fig. 5. Localization System block diagram. Figure 6 shows the area versus distance function together with the resulting standard deviation for this function. Fig. 6. Upper diagram shows relation between area (Y axis) and distance (X axis), while lower diagram shows the resulting standard deviation. We chose an interpolation function to match our data. Using Matlab we calculated the coefficients for the cubical segments of the interpolation (splines). Also, with this method, we only need to calculate the coefficients offline and load them in memory when the robot starts playing. When we want to calculate a distance to an object, we just have to evaluate a polynomial expression with the appropriate coefficients and according to the following equation: s ( x) = ax 3 + bx 2 + cx + d (1) We calculated the cubical function coefficients and tested them using different distances. Results from these computations are shown in Table 1. Table 1. Results of the interpolation function. 3.2 Triangulation Algorithm Following distance computation we apply a triangulation method from two marks to obtain the position of the robot on the field. Triangulation results in a very precise position of the robot in a two dimensions plane. If a robot sees one landmark and can calculate the distance to this landmark, the robot could be anywhere in a circumference with origin in the landmark, and radio equal to the distance calculated. While a single is not sufficient for the robot to compute its actual position, recognizing two landmarks can already help compute a specific location from the intersection of two circumferences, as shown in Figure 7. Note that the robot could be in one of two intersection points in the circumferences. Fig. 7. Triangulation with two landmarks. In Table 2 we show the results of the triangulation algorithm. To test the algorithm we put the robot in an arbitrary position in the field. Then we computed the average distance obtained from multiple measurements followed by an average error calculation. Note the large difference between the true position and the computed average. Table 2. Results of triangulation algorithm. Real Position [cm] (50,80) (160,90) (265,80) (65,185) (170, 190) (280, 210) (70,290) (186,300) (285,300) (65,380) (165,400) (290,350) Average [cm] (80.06,139.28) (163.86, 104.44) (278.08,108.39) (95.04, 195.87) (175.25, 196.10) (301.64, 222.44) (80.93, 303.52) (195.75, 313.52) (301.17, 315.51) (68.53, 414.64) (180.04, 421.87) (312.87, 362.08) Error [cm] (30.06,59.28) (3.86, 14.44) (13.08,28.39) (30.04, 10.87) (5.25, 6.10) (21.64, 12.44) (10.93, 13.52) (9.75, 13.52) (16.17, 15.51) (3.53, 34.64) (15.04, 21.87) (22.87, 12.08) 3.3 Angle Calculation Once we find the robot position we need to find its orientation to complete localization. We refer to two vectors whose origin is the robot location and the end points are the coordinates of the marks that we use as references for the triangulation, as shown in Figure 8. Fig. 8. Calculation of robot orientation. 3.4 Correction Algorithms While testing our algorithm with a moving robot, we noticed that in many occasions our data was not consistent between two contiguous frames. To fix the problem we added a correction algorithm taking historical data from positions already calculated by the robot in obtaining the average of these measurements. We reduced the variation of the output signal for the triangulation algorithm by using the following average filter function: n s ( x) = ∑ x(i) i =0 n (2) Figure 9 shows sample output from this filter correction. Our original signal produced variations of approximate 10%. By applying this filter we managed to reduce this variation to less than 3%. See [5] for more details. Fig. 9. Historical Average Filter. After applying this correction algorithm, we tested the system and obtained good results without affecting the performance of the system. These results are shown in Table 3. Table 3. Final results for the localization system. Region 1 2 3 4 5 6 7 8 9 10 11 12 Real Position [cm] (50,80) (160,90) (265,80) (65,185) (170, 190) (280, 210) (70,290) (186,300) (285,300) (65,380) (165,400) (290,350) Average [cm] (69.6, 124.8) (160.27, 92.95) (270.83, 100.35) (81.76, 188.81) (170.47, 193.41) (294.86, 216.79) (75.09, 293.53) (189.9, 303.53) (289.36, 303.19) (65.63, 413.74) (170, 415.22) (305.8, 355.25) Error [cm] (19.6,44.8) (0.27,2.95) (5.83, 20.35) (16.76, 3.81) (0.47, 3.41) (14.86, 6.79) (5.09, 3.53) (3.90, 3.53) (4.36, 3.19) (0.63, 33.74) (5.33, 15.22) (15.80, 5.25) Note that errors were computed by field regions, where the complete field was divided into twelve similarly sized areas, as shown in Figure 10. In some regions errors were larger due to changes in illumination. Yet, when compared to the 20cm approximate AIBO size, worst errors were a bit less than a full body length. Current work focuses in dividing the field into differently sized regions depending on required localization precision. Fig. 10. Localization results by field region. 4 Conclusions We have presented the system architecture for the Eagle Knights Four-Legged team with special emphasis on our real time localization system. The system calculates distances with the help of an interpolation function producing good results and in real time. The algorithms used allowed us to estimated a reliable position for the robot without using probabilistic methods like other teams do. We are currently incorporating localization information as part of our game playing strategy while adapting the algorithm to specific regions in the field to produce better qualitative game playing results as opposed to costly numerical accuracy. Our team started competing in 2004 and has since then continuously participated in regional or world events. This work is part of broader research we are pursuing in the robotics laboratory at ITAM. One of the related areas is that of human-robot interaction in the context of social cognition where we are using soccer coaching as the application domain, see [6] for sample videos. More information can be found in http://robotica.itam.mx/. Acknowledgments. This work is supported by French-Mexican LAFMI, UC MEXUS CONACYT, CONACYT grant #42440, and “Asociación Mexicana de Cultura, A.C.” in Mexico. References 1. RoboCup Technical Committee. “Sony Four Legged Robot Football League Rule Book.” Mayo 2004. RoboCup. Official Web Site URL:http://www.robocup.org 2. 3. 4. 5. 6. Flores Ando, F., and Weitzenfeld, A., 2005, Visual Input Compensation using the Crowley-Arbib Saccade Model, Proc. International Conference on Advanced Robotics ICAR, Seattle, USA, July 17-20. Weitzenfeld A and Dominey PF, Cognitive Robotics: Command, Interrogation and Teaching in Robot Coaching, RoboCup Symposium 2006, June 19-20, Bremen, Germany. Martínez-Gómez, L.A., and Weitzenfeld, A., 2004, Real Time Vision System for a Small Size League Team, Proc. 1st IEEE-RAS Latin American Robotics Symposium, ITAM, Mexico City, October 28-2. Martínez-Gómez, J.A, Weitzenfeld, A., 2005, Real Time Localization in Four Legged RoboCup Soccer, Proc. 2nd IEEE-RAS Latin American Robotics Symposium, Sao Luis, Brasil, Sept 20-23. Dominey PF and Weitzenfeld A, Videos for command, interrogate and teach AIBO robots, ftp://ftp.itam.mx/pub/alfredo/COACHING/