Playing Robotic Soccer Based on an Explicit World Model

Home , RoboCup, Saarland

AI Magazine Volume 21 Number 1 (2000) (© AAAI) Articles The CS Freiburg Team Playing Robotic Soccer Based on an Explicit World Model

Jens-Steffen Gutmann, Wolfgang Hatzack, Immanuel Herrmann, Bernhard Nebel, Frank Rittinger, Augustinus Topor, and Thilo Weigel

Robotic soccer is an ideal task to demonstrate new reacting on mostly uninterpreted sensor input techniques and explore new problems. Moreover, as in pure behavior-based (Werger et al. 1998) problems and solutions can easily be communicat- or reinforcement learning approaches (Suzuki ed because soccer is a well-known game. Our et al. 1998), soccer seems to be a game that has intention in building a robotic soccer team and a structure that requires more than just react- participating in RoboCup-98 was, first, to demon- ing on uninterpreted sensor input. Our claim strate the usefulness of the self-localization methods we have developed. Second, we wanted to is justified by the fact that the two winning show that playing soccer based on an explicit teams in the simulation and the small-size world model is much more effective than other league in RoboCup-97 used this approach methods. Third, we intended to explore the prob- (Burkhard, Hannebauer, and Wendler 1998; lem of building and maintaining a global team Veloso et al. 1998). Further evidence for our world model. As has been demonstrated by the claim is the performance of our team at performance of our team, we were successful with RoboCup-98, which won the competition in the first two points. Moreover, robotic soccer gave the middle-size league. us the opportunity to study problems in distrib- Third, we intended to address the problem uted, cooperative sensing. of multirobot sensor integration to build a global world model and exploit it for coopera- obotic soccer is an interesting research tive sensing and acting. In the end, we identi- domain because problems in robotics, AI, fied more problems in this area than we solved. multiagent systems, and real-time rea- However, we believe that it is an interesting R topic for future research. soning have to be solved to create a successful team of robotic soccer players (Kitano et al. Although perception and sensor interpreta- 1997). Furthermore, it is an ideal task to tion were definitely the focus of our research, demonstrate the feasibility of new ideas and it was also necessary to develop basic soccer techniques and explore new problems. skills and forms of multiagent cooperation to We started to design a robotic soccer team show the advantage of our approach. Al- with the intention of participating in though this part certainly needs improvement, RoboCup-98 for three reasons: First, we intend- it was still effective enough to be competitive. ed to demonstrate the advantage of our percep- Furthermore, based on an accurate world mod- tion methods based on laser range finders el, our robots were much more reliable than (Gutmann et al. 1998; Gutmann and Nebel other teams. 1997; Gutmann and Schlegel 1996), which The rest of the article is structured as fol- make explicit world modeling and accurate lows: In the next section, we give a brief sketch and robust self-localization possible. of the robot hardware. We then describe the Second, we believe that soccer is a game, general architecture of our soccer players and where it is advantageous to base deliberation the soccer team. The next section focuses on and action selection on an explicit world mod- our self-localization approach, and then we el, and we intended to demonstrate that such describe our player- and ball-recognition an approach is superior to other approaches. methods that are needed to create the local Although it is possible to play robotic soccer by world model. The integration of these world

a Toshiba notebook LIBRETTO 70CT running LINUX. The robot is controlled using SAPHIRA (Konolige et al. 1997), which comes with the PIONEER robots. Finally, to enable communication between the robots and an off-field computer, we use the WAVELAN radio ethernet. In addition to these components, we added PLS200 laser range finders manufactured by SICK AG to all our robots. These range finders can give depth information for a field of view with an angular resolution of 0.5 degrees, and an accuracy of 5 centimeters to a distance of 30 meters. Handling the ball with the body of the PIO- NEER 1 robot is not an effective way of moving the ball around the field or pushing it into the opponent’s goal. For this reason, we developed a kicking device using parts from the MÄRKLIN METALLBAUKASTEN. Furthermore, to steer the ball, we used flexible flippers that have a length of approximately 35 percent of the diameter of the ball. Although these flippers led to some discussions before the tournament, it was finally decided that the use of such flippers does not violate the RoboCup rules. In fact, we believe that taking the idea of embodiment seriously, such a ball-steering mechanism is necessary to play soccer effectively and authentically. In fact, without the flippers, it is almost impossi- ble to retrieve the ball from the wall, which means that the referee must relocate the ball, Figure 1. Three of Our Five Robots: Two Field Players and the Goal Keeper. which is annoying for everyone—in particular, for spectators. Furthermore, without the ball- steering mechanism, the ball is easily lost when models into a global model and the problems running with the ball. we encountered are described in the next section. We then sketch the behavior-based control of the soccer agents and show how a basic General Architecture form of multiagent cooperation is achieved. Our robots are basically autonomous robotic Finally, in the last section, we describe our soccer players. They have all sensors, effectors, experience participating in RoboCup-98 and and computers on board. Each soccer agent has present our conclusions. a perception module that builds a local world model (figure 2). Based on the observed state of Robot Hardware the world and intentions of other players communicated by the radio link, the behavior-based Because our group is not specialized in devel- control module decides what behavior is activat- oping robot platforms, we used an off-the-shelf ed. If the behavior involves moving to a partic- robot—the PIONEER 1 robot developed by Kurt ular target point on the field, the path-planning Konolige and manufactured by ActivMedia. In module is invoked, which computes a colli- its basic version, however, the PIONEER 1 robot is sion-free path to the target point. hardly able to play soccer because of its limited To initialize the soccer agents, start and stop sensory and effectory skills. For this reason, we the robots, and monitor the state of all agents, had to add a number of hardware components we use a radio ethernet connection between (figure 1). the on-board computers and an off-field com- On each robot, we mounted a video camera puter (figure 3). connected to the Cognachrome vision system If the radio connection is unusable, we still manufactured by Newton Labs, which is used can operate the team by starting each agent to identify and track the ball. For local infor- manually. A large number of the other teams mation processing, each robot is equipped with in the middle-size league used a similar

38 AI MAGAZINE Articles

Communication Behavior- Effectors Perception based Sensors control

Path- planning

Player

Figure 2. Player Architecture. approach (Asada and Kitano 1999). Unlike other teams, we use the off-field computer and the radio connection for realizing global sensor integration, leading to a global Global Graphical world model. This world model is sent back to Sensor User all players, and they can use this information to extend their own local view of the world. Integration Interface Thus, the world model our players have is similar to the world model constructed by an over- head camera, as used in the small-size league by teams such as CMUNITED (Veloso et al. 1998). Commu- nication Self-Localization Off-field Computer We started the development of our soccer team with the hypothesis that it is an obvious advantage if the robotic soccer agents know Radio Ethernet their position and orientation on the field. Based on our experience with different self- localization methods using laser range finders (Gutmann et al. 1998), we decided to use such Goal Player 1 Player 2 Player 3 Player 4 a method as one of the key components in our Keeper soccer agents. A number of different self-localization methods exist based on laser scans (Gutmann and Schlegel 1996; Weiß and von Puttkamer 1995; Figure 3. Team Architecture. Lu and Milios 1994; Cox 1990). However, these methods are only local; that is, they can only more costly from a computational point of be used to correct an already-existing position view. For these reasons, we designed a new self- estimation. Thus, once a robot loses its position, it will be completely lost. Furthermore, all localization method that trades off generality the methods are computationally demanding, for speed and the possibility of global self-local- needing 100 milliseconds to a few seconds on ization. Our method first extracts line segments a modern computer. Global methods are even from laser range scans and matches them

SPRING 2000 39 Articles

Scan with extracted line segments

Robot

Position hypotheses

RoboCup field model

Figure 4. Scan Matches Lead to Position Hypotheses.

against an a priori model of the soccer field. To robot position is then updated, taking into ensure that extracted lines really correspond to account that the robot has moved since the field-border lines, only scan lines significantly scan was taken. longer than the size of soccer robots are consid- Our hardware configuration allows five laser ered. Then, the correspondence problem be- scans a second, using only a few milliseconds tween scan lines and lines of the a priori model for computing position hypotheses and the is solved by backtracking over all possible pair- position update. Although a laser scan can ings between scan lines and model lines—sim- include readings from objects blocking the ilar to the method described by Castellanos, sight to the field borders, we did not experience Tardós, and Neira (1996). Successful matchings any failures in the position-estimation process. lead to position hypotheses, of which there are In particular, we never observed the situation only two if three field borders are visible (figure that one of our robots got its orientation wrong 4). and “changed sides.” After the brief sketch of the matching algorithm, one might suspect that the worst-case run time of the algorithm is exponential in the Building the Local World Model number of model lines. However, a closer After the self-localization module matched a inspection reveals it runs in cubic time because range scan, the sensor data are interpreted to of geometric constraints (Weigel 1998). More- recognize other players and the ball (figure 6). over, we expect this algorithm to be almost lin- Scan points that correspond to field lines are ear in the number of model lines in “natural,” removed, and the remaining points are clus- settings such as office environments. Our self- tered. For each cluster, the center of gravity is localization algorithm is implemented in a computed and interpreted as the approximate straightforward way (figure 5). position of a robot (figure 7). Inherent to this From a set of position hypotheses generated approach is a systematic error depending on by the scan-matching algorithm, the most the shape of the robots. plausible one is selected and fused with the For ball recognition, we use a commercially odometry position estimate using a Kalman fil- available vision system. If the camera sees an ter. The Kalman filter returns the optimal esti- object of a certain color (a so-called blob), the mate (the one with the smallest variance) for a vision system outputs the pixel coordinates of given set of observations (Maybeck 1990). The the center of the blob and its width, height,

40 AI MAGAZINE Articles

RoboCup field model

Scan Set of position Laser hypotheses scan Matching

Plausibility Most plausible Check position

Robot position Kalman new estimation estimated by Filter of robot position odometry

Self-Localization Module

Figure 5. Self-Localization Module.

RoboCup field model From global sensor integration

Laser range Matched finder Self- scan localization Robot Odometrie position

Player Player World World recognition positions modeling model

Ball Ball Vision recognition position Sonar

To global sensor integration

Figure 6. Perception Module.

SPRING 2000 41 Articles

global sensor-integration module. Because soccer players and balls tend to move slowly (< 1 meter a second), a simple greedy algorithm can be used to track objects. Furthermore, friends and foes can be identiﬁed by comparing sensed object positions with the positions of team members determined using the self-localization algorithm. Knowing who and where the team members are is, of course, helpful in playing a cooperative game. Other information that is useful is the global ball position. Our vision hardware recognizes the ball only to a distance of 3 to 4 meters. Knowing the global ball position even if it is not directly visible enables the soccer robot to turn its camera into the direction of where the Figure 7. Line Segments Are Extracted from a ball is expected, avoiding a search for the ball Range Scan and Matched against the Field Lines, and by turning around. This information is impor- Three Players Are Extracted from the Scan. tant in particular for the goal keeper, which might miss a ball from the left while it searches for the ball on the right. and area size. From these pixel coordinates, we It should be noted, however, that because of compute the relative position of the ball with the inherent delay between sensing an object respect to the robot position by mapping pixel and receiving back a message from the global coordinates to distance and angle. This map- sensor integration, the information from the ping is learned by training the correspondence global world model is always 100 to 400 mil- between pixel coordinates and angles and dis- liseconds old; thus, it cannot be used to control tances for a set of well-chosen real-world posi- the robot behavior directly. However, apart tions and using interpolation for other pixels. from the two uses spelled out earlier, there are To improve the quality of the position estima- nevertheless a number of important problems tion, we use the sonar sensors as a secondary that could be solved using this global world source of information for determining the ball model—and we will work on these points in position. the future. First, the global world model could From the estimated position of the player, be used to reorient disoriented team members. the estimated position of other objects, and the Although we never experienced such a disori- estimated position of the ball—if it is entation, such a fallback mechanism is certain- visible—the soccer agent constructs its own ly worthwhile. Second, it provides a way to local world model. By keeping a history list of detect unreliable sensor systems of some of the positions for all objects, their headings and soccer agents. Third, the global world model velocities can be determined. To reduce noise, could be used for making strategic decisions, headings and velocities are low-pass filtered. such as changing roles dynamically (Veloso et Position, heading, and velocity estimates are al. 1998). sent to the multirobot sensor integration module. In addition to objects that are directly observable, the local world model also contains Behavior-Based Control and information about objects that are not visible. First, if an object disappears temporarily from Multiagent Cooperation the robot’s view, it is not immediately removed The soccer agent’s decisions are mainly based from the world model. Based on its last-known on the situation represented in the explicit position and estimated heading and velocity, world model. However, to create cooperative its most likely position is estimated for a few team behavior, actual decisions are also based seconds. Second, information from the global on the role assigned to the particular agent and world model is used to extend the local world on intentions communicated by other players. model of a player. Although the control of the execution can be described as behavior based, our approach Global World Model differs significantly from approaches where behaviors are activated by uninterpreted sensor The global world model is constructed from input, as is the case with the ULLANTA team time-stamped position, heading, and velocity (Werger et al. 1998). In our case, high-level fea- estimates that each soccer agent sends to the tures that are derived from sensor input and

42 AI MAGAZINE Articles the communication with other agents deter- decision on which mechanism to use and in mine what behavior is activated. Furthermore, which direction to turn is made according to behaviors can invoke significant deliberation, the current game situation. such as planning the path to a particular target The mapping from situations to actions is point. implemented in a decision tree–like manner. It The behavior-based control module consists should be noted that details of tactical deci- of a rule-based system that maps situations to sions and behaviors were subject to permanent actions. In the current version, only a few rules modifications even when the competition in (less than 10) are needed, and all of them have Paris had already started. As a reaction to teams been designed by hand and improved over that would just push the ball and opponents time after gathering new experiences from over the field, we modified our behavior to not playing games. Even during the competition in yield in such situations. Paris, we refined some of them. The rules are If all the soccer players would act according evaluated every 100 milliseconds so that the to the same set of rules, a swarm behavior would module can react immediately to changes in result, where the soccer players would block the world. Depending on whether the agent each other. One way to solve this problem is to fills the role of the goal keeper or a field player, assign different roles to the players and define there are different rule sets. areas of competence for these roles (figure 8). The goalie is simple minded and just tries to If these areas were nonoverlapping, interfer- keep the ball from rolling into our goal. It ence between team members would not hap- always watches the ball—getting its informa- pen, even without any communication tion from the global world model if the camera between players. Each player would go to the cannot recognize the ball—and moves to the ball and pass it on to the next area of compe- point where the robot expects to intercept the tence or into the goal. In fact, this was our ini- ball based on its heading. If the ball is on the tial design, and it is still the fallback strategy left or right of the goal, the goalkeeper turns to when radio communication is not working. face the ball. To allow for fast left and right There are numerous problems with such a movements, we use a special hardware setup rigid assignment of competence areas, howev- where the “head” of the goalie is mounted to er. First, players can interfere at the border lines the right, as shown in figure 1. If the ball hits between competence areas. Second, if a player the goalie, the kicking device kicks it back into is blocked by the other team, broken, or the field. removed from the field, no player will handle The field players have a much more elabo- balls in the corresponding area. Third, if a rate set of skills. The first four skills concern sit- defender has the chance to dribble the ball to uations where the ball cannot be played direct- the opponent’s goal, the corresponding for- ly, and the two last skills address ball handling: ward will most probably block this run. For Approach-position: Approach a target position these reasons, we modified our initial design carefully. significantly. Even during the tournament in Go-to-position: Plan and constantly replan a Paris, we changed the areas of competence and collision-free path from the robot’s current added other means to coordinate attacks as a position to a target position and follow this reaction to our experiences from the games. path until the target position is reached. Path If a player is in a good position to play the planning is done using the extended visibility ball, it sends a clear-out message. As a reaction graph method (Latombe 1991), which is fast to receiving such a message, other players try to enough to be executed in each execution cycle. keep out of the playing robot’s way (figure 9), Observe-ball: Set the robot’s heading such helping to avoid situations in which two team- that the ball is in the center of focus. Track the mates block each other. In other words, we also ball without approaching it. rely on cooperation by communication, as the Search-ball: Turn the robot to find the ball. If UTTORI team did (Yokota et al. 1999). However, the ball is not found after one revolution, go to our communication scheme is much less elab- home position and search again from there. orate than UTTORI’s. Based on communicating Move-ball: Determine a straight line to the intentions, areas of competence can be made goal that has the largest distance to any object overlapping, as shown in figure 8. Now, the for- on the field. Follow this line at increasing wards handle three-quarters of the field, and velocity and redetermine the line whenever attacks are coordinated by exchanging the appropriate. intentions. Shoot-ball: To accelerate the ball, either turn We do not have any special coordination for the robot rapidly with the ball between the defensive moves. In fact, defensive behavior flippers, or use the kicker mechanism. The emerges from the behavior-based control

SPRING 2000 43 Articles

right forward

right defender goalkeeper

left defender left forward

Figure 8. Roles and Areas of Competence.

described earlier. When the ball enters our half and came in third at RoboCup in Stockholm.) of the field, our defenders go to the ball and The key components for this success are block the attack. Surprisingly, this simple most probably the self-localization and object- defensive strategy worked quite successfully. recognition techniques based on laser range finders, which enabled us to create accurate and reliable local and global world models. Conclusion and Discussion Based on these world models, we were able to Participating in the RoboCup-98 tournament implement reactive path planning, ﬁne-tuned was beneﬁcial for us in two ways. First, we got behaviors, and basic multiagent cooperation, the opportunity to exchange ideas with other which was instrumental in winning. Finally, teams and learned how they approached the our kicker and the ball-steering mechanism problems. Second, we learned much from play- certainly also had a role in playing successful ing. As pointed out at various places in the arti- robotic soccer. cle, we redesigned tactics and strategy during Acknowledgments the tournament, incorporating the experience This work has been partially supported by we attained during the games. Deutsche Forschungsgemeinschaft as part of The performance of our team at RoboCup- the graduate school on Human and Machine 98 was quite satisfying. Apart from winning Intelligence; Medien- und Filmgesellschaft the tournament, we also had the best goal dif- Baden-Württemberg mbH; and SICK AG, ference (12:1), never lost a game, and scored which provided the laser range finders. Fur- almost 25 percent of the goals during the thermore, we would like to thank ActivMedia tournament. This performance was not acci- and Newton Labs for their timely support, dental, as demonstrated at the national Ger- resolving some of the problems that occurred man competition VISION-RoboCup-98 on 30 just a few weeks before RoboCup-98. September to 1 October 1998 in Stuttgart. Again, we won the tournament and did not References lose any game. (Endnote: In 1999, our team Asada, M., and Kitano, H., eds. 1999. RoboCup-98: again won the German VISION-RoboCup Robot Soccer World Cup II. Lecture Notes in Artificial

44 AI MAGAZINE Articles

Clear Clear out out

Clear out Clear Clear out out Clear out

Clear out

A B

Figure 9. Cooperation by Communication. A. Player gets ball and notiﬁes teammates. B. Player can run.

Intelligence. New York: Springer-Verlag. New York: Springer-Verlag. Burkhard, H. D.; Hannebauer, M.: and Wendler, J. Gutmann, J.-S., and Nebel, B. 1997. Navigation 1998. AT HUMBOLDT—Development, Practice, and The- Mobiler Roboter mit Laserscans. In Autonome Mobile ory. In RoboCup-97: Robot Soccer World Cup I, ed. H. System 1997, eds. P. Levi, Th. Bräunl, and N. Oswald, Kitano, 357–372. Lecture Notes in Artificial Intelli- 36–47. Informatik aktuell (Navigation of Mobile gence, Volume 1395. New York: Springer-Verlag. Robots Based on Laser Scans). Stuttgart: Springer-Ver- Castellanos, J. A.; Tardós, J. D.; and Neira, J. 1996. lag. Constraint-Based Mobile Robot Localization. Paper Gutmann, J.-S., and Schlegel, C. 1996. AMOS: Com- presented at the International Workshop on parison of Scan-Matching Approaches for Self-Local- Advanced Robotics and Intelligent Machines, 2–3 ization in Indoor Environments. In Proceedings of April, Manchester, United Kingdom. the First Euromicro Workshop on Advanced Mobile Cox, I. J. 1990. BLANCHE: Position Estimation for an Robots, 61–67. Washington, D.C.: IEEE Computer Autonomous Robot Vehicle. In Autonomous Robot Society. Vehicles, eds. I. J. Cox and G. T. Wilfong, 221–228. Gutmann, J.-S.; Burgard, W.; Fox, D.; and Konolige,

SPRING 2000 45 Articles

K. 1998. An Experimental Comparison of Kawabata, K.; Kaetsu, H.; and Asama, H. Knowledge Representation and Reasoning Localization Methods. In Proceedings of 1999. Cooperative Team Play Based on (KR’92) and will serve as the program chair the International Conference on Intelligent Communication. In RoboCup-98: Robot Soc- for IJCAI’01. In addition, he is a member of Robots and Systems (IROS’98), 736–743. cer World Cup II, eds. M. Asada and H. the editorial boards of Artificial Intelligence Washington, D.C.: IEEE Computer Society. Kitano, 333–347. Lecture Notes in Artificial and AI Communication and an associate edi- Kitano, H.; Asada, M.; Kuniyoshi, Y; Noda, Intelligence. New York: Springer-Verlag. tor of the Journal of Artificial Intelligence I.; Osawa, E.; and Matsubara, H. 1997. Research. His research interests include RoboCup: A Challenge Problem for AI. AI Jens-Steffen Gutmann knowledge representation and reasoning, Magazine 18(1): 73–85. is a Ph.D. student in the with an emphasis on semantics, algo- rithms, and computational complexity, in Konolige, K; Myers, K; Ruspini, E. H.; and graduate school of Hu- particular in the areas of description logics, Saffiotti, A. 1997. The SAPHIRA Architecture: man and Machine Intel- temporal and spatial reasoning, constraint- A Design for Autonomy. Journal of Experi- ligence at the University based reasoning, planning, belief revision, mental and Theoretical Artificial Intelligence of Freiburg. He received and robotics. His e-mail address is 9(1): 215–235. his diploma in computer [email protected]. Latombe, J.-C. 1991. Robot Motion Planning. science from the Univer- Dordrecht, The Netherlands: Kluwer. sity of Ulm in 1996. His research interests are robust mobile robot Frank Rittinger is cur- Lu, F., and Milios, E. E. 1994. Robot Pose navigation and robotic soccer playing. His rently a graduate student Estimation in Unknown Environments by e-mail address is [email protected]. at the University of Matching 2D Range Scans. In IEEE Comput- Freiburg, Germany. Be- er Vision and Pattern Recognition Conference Wolfgang Hatzack re- fore joining the Artificial (CVPR), 935–938. Washington, D.C.: IEEE Intelligence Group in Computer Society. ceived his diploma in computer science (Di- Freiburg, he studied for Maybeck, P. S. 1990. The Kalman Filter: An plom-Informatiker) from one year at the Universi- Introduction to Concepts. In Autonomous the University of Bonn ty of Sussex at Brighton, Robot Vehicles, eds. I. J. Cox and G. T. Wil- in 1996. Currently, he is Great Britain. His e-mail address is frit- fong, 194–204. New York: Springer-Verlag. a Ph.D. student in the [email protected]. Suzuki, S.; Takahashi, Y.; Uchibe, E.; Naka- Artificial Intelligence muro, M.; Mishima, C.; Ishizuka, H.; Kato, Research Group at the Augustinus Topor is a T.; and Asada, M. 1998. Vision-Based Robot University of Freiburg, focusing on air-traf- graduate student at the Learning toward RoboCup: Osaka Universi- fic control and robotics. University of Freiburg, ty “TRACKIES.” In RoboCup-97: Robot Soccer Germany. He is currently World Cup I, ed. H. Kitano, 305–319. Lec- Immanuel Herrmann is working on his thesis on ture Notes in Artificial Intelligence, Volume path planning in dy- 1395. New York: Springer-Verlag. currently working on his Master’s thesis in mathe- namic environments. Veloso, M.; Stone, P.; Han, K.; and Achim, S. matics at the University 1998. The CMUNITED-97 Small Robot Team. of Freiburg. He is also a In RoboCup-97: Robot Soccer World Cup I, ed. graduate student in com- Thile Weigel received his H. Kitano, 242–256. Lecture Notes in Artifi- puter science. He partici- diploma in computer sci- cial Intelligence, Volume 1395. New York: pated successfully in dif- ence in 1999 for his thesis Springer-Verlag. ferent contests such as on self-localization, path Weigel, T. 1998. Roboter-Fußball: Selbst- the Association of Computing Machinery planning, and reactive lokalisation, Pfadplanung und Basis- International Collegiate Programming control in the RoboCup fähigkeiten (Robotic Soccer: Self-Localiza- Contest. context. His e-mail ad- tion, Path Planning, and Basic Skills). dress is weigel@informa Diplomarbeit, Fakultät für Angewandte Bernhard Nebel is a tik.uni-freiburg.de. Wissenschaften, University of Freiburg, professor in the Depart- Germany. ment of Computer Sci- Weiß, G., and von Puttkamer, E. 1995. A ence at Albert-Ludwigs- Map Based on Laserscans without Geomet- University Freiburg, ric Interpretation. In Intelligent Autonomous Germany, and director of Systems (IAS-4), eds. U. Rembold, R. Dill- the AI Research Lab. He mann, L. O. Hertzberger, and T. Kanade, received a Ph.D. (Dr. rer. 403–407. Dordrecht, The Netherlands: IOS. nat) from the University Werger, B.; Funes, P.; Schneider Fontan, M.; of Saarland in 1989 for his work on knowl- Sargent, R.; Witty, C.; and Witty, T. 1998. edge representation and belief revision. He THE SPIRIT OF BOLIVIA: Complex Behavior is a member of the International Joint Con- through Minimal Control. In RoboCup-97: ferences on Artificial Intelligence, Inc., Robot Soccer World Cup I, ed. H. Kitano, board of trustees and a member of the grad- 384–356. Lecture Notes in Artificial Intelli- uate school of Human and Machine Intel- gence, Volume 1395. New York: Springer- ligence at Albert-Ludwigs-University. Verlag. Among other professional services, he Yokota, K.; Ozaki, K.; Watanabe, N.; Mat- served as the program cochair for the Third sumoto, A.; Koyama, D.; Ishikawa, T.; International Conference on Principles of

46 AI MAGAZINE