Modeling Observation Uncertainty for Soccer Playing Humanoid Robots

Stefan Kohlbrecher 1 , Oskar von Stryk 2

kohlbrecher@sim.tu-darmstadt.de 2 stryk@sim.tu-darmstadt.de

Abstract In recent years, humanoid robot soccer robots have shown increasingly robust and fast locomotion, making perception, world modeling and behavior control important. In this paper, we present the world modeling approach of the Darmstadt Dribblers humanoid robot team, which won the competitions in the RoboCup Humanoid KidSize League in 2009 and 2010. The paper focuses on modeling observation uncertainties originating from different contributing factors centrally in one module. This allows different state estimators to use this data in a consistent way, independently of the specic state estimation approach used.

I. I NTRODUCTION Playing soccer is a challenging problem, especially for small autonomous systems like those in the RoboCup and FIRA robot soccer leagues, as robust bipedal locomotion is a necessary prerequisite for this task. Despite this difculty, humanoid robots have shown increasingly reliable and agile bipedal locomotion abilities in recent years. This means perception, world modeling and behavior control are gaining importance, as the improved locomotion abilities can only be leveraged if the high level modeling and behavior system can make use of them. The contribution of this paper is a detailed description of an approach for modeling observation uncertainties in resource constrained humanoid robot systems that explicitly identies, considers and models different noise and error sources contributing to uncertainty in observations. It is shown how the improved modeling subsequently allows to achieve better results in state estimation modules based on parametric or sampling-based lters as well as a grid-based obstacle mapping approach. II. R ELATED W ORK Probabilistic methods for state estimation are state of the art and are used by most successful teams in robot soccer. Both parametric approaches like Kalman Filters [1] and their variants like Multi-Hypothesis Kalman Filters [2] are used as well as sampling based approaches like particle lters [3] providing advantageous properties when multimodality and nonlinearities prohibit the use of parametric techniques. RaoBlackwellized particle lters, representing a combination of parametric and sampling-based state estimation, are also used with success [4]. Recently, also constraint- and optimizationbased methods [5] are used and are a promising direction

for further research. This parallels research into the SLAM problem, where pose graph optimization methods [6],[7] are increasingly used for online applications in recent years. In robot soccer, different state estimation tasks generally have different demands in terms of complexity, scope and difculty. The two most prominent state estimation tasks that have to be fullled are estimation of the own pose of an autonomous soccer playing robot as well as estimation of the ball position. The self-localization task lends itself well to the implementation with a Monte Carlo Localization [8] approach using a particle lter due to its ability to represent ambiguous, multimodal state estimates as well as nonlinear motion and observation models. Ball Modeling on the other hand is often performed using a Kalman Filter or Multihypothesis Kalman Filter, as tracking of a single unambiguous object can be done well using these parametric lters [9],[10]. Both parametric and sampling-based lters rely on probabilistic motion and observation models in their prediction and update steps. Often [11],[12] the angular standard deviation of seen objects is a parameter that, while being congurable in many robot systems, stays xed unless human intervention changes it. While using such a model works well if the parameters are well tuned, xed noise estimates for sensors prevent leveraging the knowledge of the current robot state to achieve more accurate state estimation. For example, a goalie robot standing still inside the goal might generate much more accurate observations of the ball than his moving teammates. Not using this knowledge means that the robot generates state estimates with too high estimated uncertainty, effectively not using all the information available. III. H ARDWARE Our autonomous humanoid soccer platform is based on the HR30 platform by the Hajime research institute. With additions and modications the current model is denoted DD2010 (Fig. 1). It has 21DOF, all of which are actuated by Robotis Dynamixel servo motors (18xRX-28, 3xRX-64). Low level and hard real-time control are provided by a microcontroller board with a Renesas SH2 CPU [13], while high level cognition runs on a FitPC2 board in the back of the robot, featuring an Atom Z530 CPU. Both systems are connected by a serial (RS232) connection. Among the internal sensors of the robot are a 3 axis accelerometer board

Proceedings of the 5th Workshop on Humanoid Soccer Robots @ Humanoids 2010, Nashville (USA), 2010, December 7 ISBN 978-88-95872-09-4 pp. 35-40

For this reason, odometry error generally is modeled to be larger than commonly observed, in order to be robust against mentioned situations. B. Modeling Observation Uncertainty Small scale humanoid robots exhibit varying and potentially large uncertainties in vision-based observations for a multitude of reasons. First of all, camera systems might not provide exact timestamps of the images they generate. This is the case for our humanoid robots, as a commercial webcam (Philips SPC-1300) is used here. While industrial cameras generally support exact time stamping, they often only provide Bayer pattern data, which has to be preprocessed on the host CPU to retrieve a color image. To relieve the CPU of this task, we use a commercial webcam which provides YUV422 images through processing on a built in ASIC and streams these images via USB 2.0 with a resolution of 640x480 pixels at 30Hz. In comparison to the industrial cameras we tested, we also found the SPC-1300 webcam to have favorable sensitivity and robustness to low and varying lighting conditions. The camera uses a rolling shutter, meaning that images display warping if they are taken during motion of the camera. While methods for compensating for this warping using knowledge about the cameras movement exist [15], in our approach the effect can be subsumed in the additional angular variance of observations when the camera is moving fast. The intrinsic parameters of the camera are calibrated using a publicly available toolbox [16] and are used by the image processing pipeline. Our humanoid robots use measured joint angles of the kinematic chain from estimated foot on ground to the camera for estimating the extrinsic camera parameters with regards to the robot coordinate system. Using the intrinsic and extrinsic parameters, vectors of detected objects can be intersected with the ground plane for an estimate of their distance to the robot. For objects of known size like the ball, size-based distance estimation is also feasible and used. As the low weight camera sensor is mounted on a pan-tilt unit, the fast actuation of the pan and tilt servos is a major source of error, due to causing motion blur in camera images, as well as angular estimation errors due to errors in timestamp synchronization between read out servo joint values and image timestamps. This error is dependent on the angular rates of the pan and tilt servos of the robots camera head: head (t) = tilt (t) pan (t) (1)

Fig. 1: DD2010 humanoid robot ( c Katrin Binner)

as well as 3 single-axis angular rate gyros. The joint servos can be queried to report their measured joint angles. A more comprehensive overview is available in [14]. IV. M ODELING U NCERTAINTY For many applications in robotics, uncertainty in perception can be minimized by carefully selecting sensors and having very narrowly dened conditions and states the robot system might experience. Such approaches are leveraged in industrial robotics, but for the size- and weight-constrained systems in autonomous robot soccer such an approach is rarely feasible. For this reason, uncertainties have to be considered explicitly. A. Modeling Motion Uncertainty On the platform under consideration, odometry data can be estimated from the known desired footsteps that are generated on the microcontroller. This odometry information is impeded by signicant uncertainties, as actual displacement and slippage cannot be measured directly. Experiments and experience in competitions showed that slippage during turning of the robot was a major source of error and exhibited very irregular and nonlinear characteristics. For this reason, gyroscope data is used for the rotational part of odometry. Using this system, noise parameters for translational and rotational parts of odometry can be determined by experimentation. When optimizing parameters using a ground truth recording system, care has to be taken not to overt parameters, as certain situations in robot soccer matches like collisions of robots and wear in actuators might lead to higher odometry errors than commonly observed under ideal conditions.

Another major source of observation uncertainty is movement of the platform: As soon as the robot is moving, shaking of the platform leads to higher observation uncertainty. Due to the low available computational power and weight restrictions, approaches for image stabilization [17] are not feasible in our scenario, thus modeling this shaking motion as additional noise is the approach used in our system. The amount of noise due to movement of the robot is dependent on the movement state of the robot, e.g. the translational and angular velocities

Proceedings of the 5th Workshop on Humanoid Soccer Robots @ Humanoids 2010, Nashville (USA), 2010, December 7 ISBN 978-88-95872-09-4 pp. 35-40

of the robot on the playing eld: x robot (t) robot (t) robot (t) = y x robot (t)

(2)

Thus by considering the main contributing factors to errors in camera perception, we propose the following adaptive approach for modeling camera observation standard deviations using observation angle around the pitch axis as well as observation angle around the yaw axis:
base head loc robot (t)) (t) = + (head (t)) + (x base head loc robot (t)) (t) = + (head (t)) + (x

(3) (4)

Here, base and base are the base angular standard deviations that are always affecting perceptions performed through the image processing system of the robot. Errors/noise due to imperfect calibration and limited accuracy of the vision system are subsumed here. On our system, we found that the noise the object detections by the image processing system in image coordinates is low compared to noise introduced by the effects due to motion of the camera. head head (head (t)) and (head (t)) represent added noise due to head camera motion. In our system, this added noise is modeled by simple linear functions: head (head (t)) = ctilt tilt (t) cpan pan (t) (5)

Fig. 2: Approximation of measurement covariance in state space from observation angular standard deviations and

with ctilt and cpan being constant factors determined through experimentation. Our system always waits for updated joint values after receiving an image, so proper extrinsic camera parameters can be determined. From the difference between joint values from preceding to current image, approximate angular rates of the head servos head (t) can be determined. loc loc (xrobot (t)) represent added noise due (xrobot (t)) and to locomotion of the whole robot. Experiments have shown that any movement of the legs leads to increased error in observation angles, with no distinct correlation between different movement patterns and different noise characteristics in observation that could be leveraged. For this reason, the same added standard deviation due to movement of the robot is used regardless of the exact movement pattern: robot (t)) = loc (x 0 moving robot (t) is (0, 0, 0)T , if x otherwise. (6)

as well as velocity of the ball in cartesian robot coordinates: px (t) py (t) (7) x(t) = vx (t) vy (t) Ball state estimation is performed using robot coordinates, so it is not affected by errors in the robots own pose estimate. To account for robot motion, the ball state is transformed by odometry data, which is sufciently accurate for the timeframes that are important in the highly dynamic robot soccer scenario. The observation update can be performed by considering two different sources of information: The robots own camera as well as ball state estimates received from other robots by wireless communication. A. Ball Model Update from Camera Data Using the robots own camera, the ball state is only partially observable, as only the position and size in the image can be detected, while the velocity remains unknown. In the case of the ball, two approaches for estimating the distance are feasible: Intersecting the ball perception vector with the ground plane as well as size-based measurements. The projectionbased method is shown in Fig. 2. Here, the ball perception vector is projected onto the ground along with the angular standard deviation around the pitch axis and around the yaw axis to construct an approximate observation covariance matrix in state space. The standard deviation of the projectionbased distance estimate is: dp = hc hc tan(p + ) tan(p ) (8)

While this approach works well for our humanoid robots, other systems might exhibit strong correlations between different movement schemes and resulting noise in observations which could be leveraged by the use of more sophisticated functions robot (t)). for the approximation of loc (x V. BALL M ODELING State estimation of the ball is performed by employing a Kalman Filter. The ball state estimate incorporates the position

Proceedings of the 5th Workshop on Humanoid Soccer Robots @ Humanoids 2010, Nashville (USA), 2010, December 7 ISBN 978-88-95872-09-4 pp. 35-40

This distance estimate and the accompanying standard deviation alone can be used for the observation update. However, as our image processing system can also provide a size-based distance estimate, we also leverage this information when it is available. It can be more accurate than the projectionbased distance estimate, especially if the ball is far away and the robot is moving. Both measurement methods have very different error characteristics, so we consider them to be independent given the ball state and fuse both sizeand projection-based estimates using optimal fusion of two Gaussian estimates. The projection-based distance estimate (dp , dp ) and the size-based estimate (ds , ds ) are fused to get a fused estimate (df , df ). The optimal gain is: kf = dp 2 dp + ds 2
2

B. Ball Model Update from Team Data In the RoboCup soccer scenario, robots may communicate among each other and exchange their ball state estimates. This is used extensively our team. Unlike the observation from camera frames, communicated ball estimates from team members provide a full state estimate including ball velocity. If a robot did not detect the ball with its own camera, it will update the ball state estimate from the best (i.e. lowest uncertainty) team estimate available after the uncertainty of the own estimate has surpassed a threshold value. As mentioned earlier, ball state estimation is performed in robot coordinates. For this reason, estimates have to be transformed into world coordinates to be useful for other robots. This is performed by multiplying both the top left and lower right 2x2 matrices of the state covariance with a rotation matrix representing the robot orientation as well as transforming the mean vector into world coordinates. To account for various sources of error (uncertainty in own and other robots pose estimate, imperfect synchronization of timestamps) the teammates covariance is multiplied by a factor to subsume the various sources of added uncertainty. This covariance is then used for the Kalman update of the ball state estimate. In practice, the fast exchange of ball state estimates by all three robots on the eld implies that the ball position and velocity are known to the team at nearly all times, as the reader might verify by looking at videos of the RoboCup 2010 nal game showing internal world model data [18]. VI. S ELF L OCALIZATION We use a Monte Carlo Localization [8] approach for pose estimation of our soccer playing autonomous robots. The approach as described in [19] is made more robust by employing sensor resetting with a preceding ltering step which we call Particle Maturing MCL. Here, new particle candidate (template) poses are generated from single incoming observations of landmarks, for example perceptions of goal poles or eld line crossings. These template poses do not get inserted into the particle set right away but have to mature by getting updated from consistent observations of different features over time before they become candidates for insertion into the particle set by sensor resetting. This greatly enhances robustness of the system, as particle templates generated from erroneous perceptions will get purged in the maturing step, before they are inserted in the particle set used for pose estimation. For determining the particle weights in the observation update step of the lter, we utilize the observation standard deviations described earlier. For each particle m, the angular differences between the current measurement and the state estimate provided by the particle xt [m] along the pitch and vertical axes are used for determining the particle weight:
[m] = |p xt [m] |

(9)

resulting in the fused distance mean df = dp + kf (ds dp ) and the fused distance variance df 2 = (1 kf )dp 2 (11) (10)

Using this approach, the error models for both size- and projection-based distance estimation can be adjusted and learned independently. It remains to be shown how the directional standard deviation in observation space is projected into state space to arrive at a standard deviation y in state space that models the directional uncertainty around the vertical axis. For this, again referring to Fig. 2 we use: y = tan( )df (12)

Knowing y and df , an approximate covariance matrix Rbase for the Kalman Filter observation update is: Rbase = df 2 0 0 y 2 (13)

For readability, the example in Fig. 2 shows a scenario where the directional angle to the ball percept p is 0 (i.e. the ball is directly in front of the robot). For arbitrary angles p , Rbase has to be transformed by the following rotation matrix: M ( p ) = cos(p ) sin(p ) sin(p ) cos(p ) (14)

The transformed covariance matrix used for the Kalman update is then given by: Rrotated = M (p )Rbase M (p )
T

(15)

The observation mean has also be transformed to reect the described transformations. This transformation is straightforward and omitted here for brevity.

(16) (17)

[m]

= |p xt [m] |

Proceedings of the 5th Workshop on Humanoid Soccer Robots @ Humanoids 2010, Nashville (USA), 2010, December 7 ISBN 978-88-95872-09-4 pp. 35-40

These are then used for computing the particle weight: wt

(18)

Here, (t) and (t) are the observation standard deviations as dened in Equations (3) and (4).

(a) Ground truth

(b) Sector based obstacle model

is tracked without explicit consideration of measurement uncertainty. While being sufcient for local obstacle avoidance, updating the model by odometry is only possible for short distances moved by the robot due to coarse discretization of the state space in conjunction with the used polar coordinate system. As the model does not use a probabilistic representation of occupancy, there is no model of free space and no information about non-occupied areas: It remains unknown if the sectors marked as unoccupied contain free space, or if they just have not been observed for a longer time (and are occupied). As can be seen in Fig. 3b), the width of some obstacles is overestimated due to noisy or erroneous direction estimates of the panning camera. In contrast to this, a grid-based method currently in development utilizes an occupancy grid map representation [20] of the environment commonly used in SLAM. Our image-based obstacle detection system is capable of detecting free space and obstacles inside the eld of view of the camera, which makes it possible to use similar sensor models like those used for mapping with ultrasonic sensors. The obstacle perceptions taken with the onboard camera are again subject to angular noise (t) and (t) as given in Equations (3) and (4). These noise characteristics are used by the sensor model for updating the obstacle map. Fig. 3c) shows an dynamic occupancy grid map learned in a real world situation. It is worth noting that the robot not only has a detailed map of occupied (black) areas, but a just as detailed map of free as well as unknown areas. Thus, using the occupancy grid mapping approach, an occupancy estimate of every location on the playing eld is available. These estimates can in turn be used for path planning, behavior control and gaze control of soccer playing humanoid robots, which are directions of ongoing work. VIII. R ESULTS The proposed approach was tested with ground truth data as well as during the RoboCup competition in Singapore. A. Quantitative Evaluation For evaluation of the proposed system, an approach for evaluating the overall system performance of both self localization and ball modeling at the same time is used. For measuring state estimation quality when the ball is stationary we place the ball in the middle of the center circle. The ground truth of the ball position is thus precisely known. Our metric for the quality of state estimation is the RMSE of the ball position estimate in world coordinates given ground truth. This metric relies on both self localization as well as ball modeling and therefore is well suited to determine overall system performance. The robot walks a predened trajectory while all relevant sensor data are recorded. This dataset can then be used to vary parameters like the observation uncertainties and and produce repeatable results. For determining state estimation quality when the ball is moving, the ball was kicked towards a goalie robot and the data was recorded as explained before.

(c) Grid based obstacle model

Fig. 3: Comparison of obstacle modeling approaches: (a) shows a real world scene featuring multiple obstacles. (b) shows the robot-centric obstacle model used by the Darmstadt Dribblers in the last 3 years, while (c) shows the grid-based obstacle model currently in development.

VII. O BSTACLE M ODELING In the 2009 and 2010 RoboCup competitions, our robots utilized a sector-based, robot centric obstacle model. Using this model, the obstacle distance in 72 sectors around the robot

Proceedings of the 5th Workshop on Humanoid Soccer Robots @ Humanoids 2010, Nashville (USA), 2010, December 7 ISBN 978-88-95872-09-4 pp. 35-40

As can be seen in the results in Table I, the approach using adaptive observation uncertainties copes well with both situations, while xed values produce inferior results in one of the scenarios.
Observation model Fixed , = 5 Fixed , = 15 Adaptive , RMSE static ball 25.3 cm 15.7 cm 15.8 cm RMSE moving ball 9.4 cm 16.6 cm 10.3 cm

R EFERENCES
[1] R. E. K alm an, A new approach to linear ltering and prediction problems, Transactions of the ASME Journal of Basic Engineering, vol. 82 (Series D), pp. 3545, 1960. [2] P. Rybski and M. M. Veloso, Handling diverse information sources: Prioritized multi-hypothesis world modeling, Computer Science Department, Pittsburgh, PA, Tech. Rep. CMU-CS-06-182, May 2006. [3] S. Thrun, Particle lters in robotics, in Proceedings of the 17th Annual Conference on Uncertainty in AI (UAI), 2002. [4] C. Kwok and D. Fox, Map-based multiple model tracking of a moving object, in RoboCup 2004: Robot Soccer World Cup VIII, ser. Lecture Notes in Articial Intelligence, D. Nardi, M. Riedmiller, C. Sammut, and J. Santos-Victor, Eds., vol. 3276. Springer, 2005, pp. 1833. [5] D. G ohring, H. Mellmann, and H.-D. Burkhard, Constraint based world modeling in mobile robotics, in Proc. IEEE International Conference on Robotics and Automation ICRA 2009, 2009, pp. 25382543. [6] K. Konolige, G. Grisetti, R. Kummerle, W. Burgard, B. Limketkai, and R. Vincent, Sparse pose adjustment for 2d mapping, in IROS, Taipei, Taiwan, 10/2010 2010. [7] M. Kaess, A. Ranganathan, and F. Dellaert, iSAM: Incremental smoothing and mapping, IEEE Trans. on Robotics, TRO, vol. 24, no. 6, pp. 13651378, Dec 2008. [8] F. Dellaert, D. Fox, W. Burgard, and S. Thrun, Monte carlo localization for mobile robots, in IEEE International Conference on Robotics and Automation (ICRA99), May 1999. [9] A. Ferrein, E. Ferrein, L. Hermanns, and G. Lakemeyer, Comparing sensor fusion techniques for ball position estimation, in In Proc. of the RoboCup 2005 Symposium, 2005. [10] J. Silva, N. Lau, J. Rodrigues, J. L. Azevedo, and A. J. R. Neves, Sensor and information fusion applied to a robotic soccer team, in RoboCup, ser. Lecture Notes in Computer Science, J. Baltes, M. G. Lagoudakis, T. Naruse, and S. S. Ghidary, Eds., vol. 5949. Springer, 2009, pp. 366377. [11] T. R ofer, T. Laue, M. Weber, H.-D. Burkhard, M. J ungel, D. G ohring, J. Hoffmann, B. Altmeyer, T. Krause, M. Spranger, O. von Stryk, R. Brunn, M. Dassler, M. Kunz, T. Oberlies, M. Risler, U. Schwiegelshohn, M. Hebbel, W. Nistico, S. Czarnetzki, T. Kerkhof, M. Meyer, C. Rohde, B. Schmitz, M. Wachter, T. Wegner, and C. Zarges, Germanteam 2005, HU-Berlin, U-Bremen, TU-Darmstadt, U-Dortmund, Tech. Rep., 2005. [12] T. R ofer, T. Laue, J. M uller, A. Burchardt, E. Damrose, A. Fabisch, F. Feldpausch, K. Gillmann, C. Graf, T. J. de Haas, A. H artl, D. Honsel, P. Kastner, T. Kastner, B. Markowsky, M. Mester, J. Peter, O. J. L. Riemann, M. Ring, W. Sauerland, A. Schreck, I. Sieverdingbeck, F. Wenk, and J.-H. Worch, B-human team report and code release 2010, 2010, only available online: http://www.bhuman.de/le download/33/bhuman10 coderelease.pdf. [13] D. Scholz, M. Friedmann, and O. von Stryk, Fast, robust and versatile humanoid robot locomotion with minimal sensor input, in Proc. 4th Workshop on Humanoid Soccer Robots at the 2009 IEEE-RAS Int. Conf. on Humanoid Robots, Paris, Dec. 7 - Dec. 10 2009, p. to appear. [14] (2010) The Darmstadt Dribblers website. [Online]. Available: http://www.dribblers.de/ [15] S. P. Nicklin, R. D. Fisher, and R. H. Middleton, Rolling shutter image compensation, in RoboCup, ser. Lecture Notes in Computer Science, G. Lakemeyer, E. Sklar, D. G. Sorrenti, and T. Takahashi, Eds., vol. 4434. Springer, 2006, pp. 402409. [16] J. Y. Bouguet, Camera calibration toolbox for Matlab, www.vision.caltech.edu/bouguetj. [17] M. Irani, B. Rousso, and S. Peleg, Recovery of ego-motion using image stabilization, 1994. [18] (2010) The video analysis of Darmstadt Dribblers playing the nal at robocup 2010 on YouTube. [Online]. Available: http://www.youtube.com/watch?v=puL3XH So2k [19] S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics (Intelligent Robotics and Autonomous Agents). The MIT Press, 2005. [20] A. Elfes, Using occupancy grids for mobile robot perception and navigation, IEEE Computer, vol. 22, no. 6, pp. 4657, 1989.

TABLE I: RMSE of ball state estimate from ground truth using different observation model settings. The rst two observation models use xed noise estimates, while the third one uses the proposed adaptive approach.

B. Qualitative Evaluation During the RoboCup 2010 competition, the proposed approach was used on all autonomous humanoid soccer robots used by the Darmstadt Dribblers team for the soccer matches in the Humanoid KidSize League. Over the course of the competition, our robots managed to score 74 goals. The majority of shots towards the Darmstadt Dribblers goal were repelled by the goalkeeper robot, with only 2 goals scored against the Dribblers. As onboard data was recorded during all matches for subsequent debugging and evaluation purposes, a thorough evaluation of world modeling and state estimation of our robots in competitive conditions is possible. This proved very useful as many faults cannot be detected, much less explained, without having ground truth data (video) synchronized with recorded internal robot data. Ground truth for the data recorded at RoboCup 2010 was not annotated so far, so a statistical evaluation cannot be presented. Video footage of the nal match in the Humanoid KidSize League synchronized with internal robot data is publicly available online [18]. IX. C ONCLUSIONS In this paper, we presented the world modeling approach used in the Darmstadt Dribblers autonomous soccer playing robots with a focus on modeling observation uncertainties caused by different contributing factors. The proposed approach reduces the amount of parameters that need to be estimated in robot systems, as the observation standard deviations are estimated in a principled way in a single module and can subsequently used by all other state estimation modules. As examples for such modules we have shown how the proposed approach can be leveraged in parametric lters like a Kalman Filter used for ball state estimation, non-parametric lters like the particle lter approach we use for robot pose estimation, as well as in sensor models for grid-based mapping. ACKNOWLEDGMENT The authors would like to thank past and current team member of the Darmstadt Dribblers for providing the hardand software base for this work.

Proceedings of the 5th Workshop on Humanoid Soccer Robots @ Humanoids 2010, Nashville (USA), 2010, December 7 ISBN 978-88-95872-09-4 pp. 35-40