■ Will we see autonomous humanoid robots thatplay (and win) soccer against the human soccerworld champion in the year 2050? This questionis not easy to answer, and the idea is quite vision-ary. However, this is the goal of the RoboCup Fed-eration. There are serious research questions thathave to be tackled behind the scenes of a soccergame: perception, decision making, action selec-tion, hardware design, materials, energy, andmore. RoboCup is also about the nature of intelli-gence, and playing soccer acts as a performancemeasure of systems that contain artificial intelli-gence—in much the same way chess has beenused over the last century. This article outlines thecurrent situation following 10 years of researchwith reference to the results of the 2006 WorldChampionship in Bremen, Germany, and discuss-es future challenges.

A World Championship of Soccer-Playing Robots

Starting with the vision of the RoboCup Feder-ation, we briefly discuss the major differencesfrom former AI challenges and outline the fed-eration’s statement. The last section consists ofsome statistics for the RoboCup competitions.

The VisionThe vision of robots playing against humanshas presented many challenges. Robots shouldbe able to handle situations without the needfor human assistance. They have to interpret agiven situation and use their skills accordingly.Given the nature of this dynamic environ-ment—which demands decisions in realtime—background knowledge is also essential.This requirement has been underestimated fora long time now and is especially needed ifrobots are one day to play against humans.Robots should act appropriately, too; that is,

although they can use their bodies, they mustnot cause harm to humans. The robot’s per-formance must therefore be restricted in orderto adapt to the situation of cooperating with ahuman, for example. The robot should havesoft surfaces rather than a hard metal or plasticbody, it should have roughly the same weightas a human, it should not be faster than ahuman, and so on. In a way, the situation islike conducting the Turing test while restrict-ing the computer to a certain computationalpower. If indeed we are able to create machinesof this kind, then we will also be seeing themin everyday situations. We may see robotsworking with the fire brigade, sitting in a streetcar, and so on. Technologically, these newrobots will require the development and appli-cation of new materials, sensors, and actuators.In addition, we must also address the issue ofenergy. RoboCup is thus an interdisciplinarylong-term project.

Many researchers in the community are con-fronting the question of whether we will actu-ally be able to achieve this vision one day. Theuncertainty is part of the attraction of the chal-lenge. Looking back, for example, to the air-craft industry of 100 years ago, researcherscould then only dream of the machines thatnowadays carry millions of passengers throughthe air every year. Back then, researchers alsohad to find new materials and new technolo-gies. Competition played a significant role andwas one of the driving forces behind the devel-opment.

Chess or Soccer?One of the primary questions RoboCup raises isthat of the nature of intelligence. Some 50years ago, we believed that intelligencedepended on the speed of computation. The

Articles

SUMMER 2007 115Copyright © 2007, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602

RoboCup: 10 Years of Achievements

and Future Challenges

Ubbo Visser and Hans-Dieter Burkhard

AI Magazine Volume 28 Number 2 (2007) (© AAAI)

Figure 1. The RoboCup Statement.

ability to play chess for example was consid-ered to be a key to understanding intelligence.Everyday intelligence—something everybodyis capable of—has been totally underestimated,and as a result, many of the dreams for AI fromthe 1950s failed. A comparison between chessand soccer reveals that the requirements forplaying soccer are closer to those needed ineveryday situations. It is also significantlyharder to meet these requirements when usingcomputer-controlled machines.

Chess has inspired the effort to create amachine with artificial intelligence, and therehave been various approaches to achieving thisgoal. In the end, the brute-force approach wassuccessful, resulting in the 1997 match inwhich Deep Blue beat Garry Kasparov. Thecomplexity of chess is undoubtedly high. Inthe end, however, search methods were able tohandle these requirements. Feuilletonistscould, however, comfort the shocked audi-ences with the statement that the computerwas not really intelligent as such.

Soccer-Playing Robots: RoboCupComputer programs that can play chess didnot, however, answer the question regardingthe nature of intelligence. Dealing with every-day situations, the combination of and inter-action between body and intelligence and theintegration and combination of various meth-ods were challenges in the mid-1990s.Researchers were seeking a new long-termvision, and thus the idea of RoboCup as a test-bed for AI and robotics was born (see figure 1).

Friendly competitions have often driventechnical advances—the automobile and air-craft industries provide prominent examples.Competitions have been used as test and eval-

uation platforms in the creation of new solu-tions.

Another area for autonomous robots is acatastrophe scenario in which robots locatevictims in areas that cannot be accessed byhumans (mostly due to dangerous situationssuch as gas, heat, collapsing buildings, and soon). RoboCupRescue offers a thorough testplatform for this kind of scenario. Anotherapplication area is that of everyday environ-ments such as a typical household. A new testbed that requires special skills was demonstrat-ed at RoboCup 2006: RoboCup@Home. Robotshad to fulfill tasks such as opening a door, fol-lowing a human, and so on. The vision alsoincludes taking care of the next generation.One of the areas of RoboCup is thereforedesigned for children and younger people:RoboCupJunior.

CompetitionsThe first RoboCup world championship washeld in Nagoya, Japan, in 1997. Annual com-petitions, accompanied by a scientific sympo-sium, have been carried out ever since, and thecommunity is still growing (see figure 2). Start-ing with 42 teams in 1997, by 2006 there wereapproximately 450 participating teams. In2006, the number of participants exceeded the2500 mark. With a total of 2561 participants—1492 in the senior and 1069 in the junior com-petitions—RoboCup is one of the largest (if notthe largest) robotic and AI events in the world.

Soccer-playing robots attract media, and theRoboCup Federation therefore also takes largesporting events into account when choosingthe locations for the annual event (for exam-ple, World Cups 1998, 2002, 2006, EuropeanCup 2004, and the Olympic games 2008).

Articles

116 AI MAGAZINE

RoboCup is an international joint project to promote AI, robotics, and related fields. It is an attempt to foster AI and intelligent robotics research by providing a standard problem where wide range of technologies can be integrated and examined. RoboCup chose to use soccer game as a central topic of research, aiming at innovations to be applied for socially significant problems and industries. The ultimate goal of the RoboCup project is ”By 2050, develop a team of fully autonomous humanoid robots that can win against the human world champion team in soccer.“ In order for a robot team to actually perform a soccer game, various technologies must be incorporated including: design principles of autono- mous agents, multi-agent collaboration, strategy acquisition, real-time reasoning, robotics, and sensor-fusion. RoboCup is a task for a team of multiple fast-moving robots under a dynamic environment. RoboCup also offers a software platform for research on the software aspects of RoboCup. One of the major application of RoboCup technologies is a search and rescue in large scale disaster. RoboCup initiated RoboCupRescue project to specifically promote research in socially significant issues.

The Robot LeaguesThere is still much to accomplish before we canhope to achieve the 2050 goal. In order to com-pete and indeed cope using the hardware andthe technologies from today, several leagueshave been created. In 1997 RoboCup had threeleagues, the Middle-Size League (MSL), theSmall-Size League (SSL), and the SimulationLeague (SL). Today, with the addition of theFour-Legged League (4LL) and the HumanoidLeague (HL), the number of soccer leagues isnow up to five.

Parallel to the soccer leagues, other leagueswere also invented in order to tackle otherapplications. The RoboCupRescue League wasestablished in 2001 with robots and simulationdevoted to rescue scenarios, while the compe-titions in RoboCup@Home (established in Bre-men) demonstrate robots used for daily activi-ties. The RoboCupJunior leagues includecompetitions in soccer, rescue, and dance.They are designed to foster education inschools and are tailored for students aged 8–18years.

The Middle-Size League (MSL)In the middle-size league (MSL), middle-sizerobots must not exceed a diameter of 50 cen-timeters, a height of 80 centimeters, and aweight of 50 kilograms (see figure 3). Eachrobot has individual sensors and control pro-grams. Communication between robots can beachieved with a wireless local area network(WLAN). It is also possible to communicatewith an external computer.

Since 1997, several steps have been taken toaccommodate various new challenges. In 2002for example, the barriers around the field wereremoved and the field size was increased. Thelack of barriers also meant that certain sensorsused for localization (ultrasound or laser) couldno longer be used for that purpose. Nowadays,techniques based on vision (mostly omnidirec-tional cameras with 30 frames per second ormore) take on the task of localization. The sixrobots per team are able to use separate blue oryellow colored goals, colored landmarks, andwhite lines on a green field for orientation.Removing the barrier was a big step, given thatthe robots then had to control the ball to avoidkicking it out of the field.

MSL reports substantial technologicaladvancements. Some teams use big, powerfulrobots with strong kicking devices. Over theyears, however, the use of those types of robotshas faded. The tendency now is to build small-er, faster, and lighter robots. Despite othermodels in the past, an MSL robot nowadays isomniwheeled and omnivisional. The majority

of the teams have kicking devices that allowthem to kick high and pass over the opponentrobots. This makes a whole new range of tacti-cal options possible.

Most of the teams in the MSL have powerfullaptops on board. This also means that compu-tationally expensive methods can be used asopposed to a robot in the 4LL, for example.Thus, the most powerful image-processingmethods can be found in this league. Thedemand for environments less conformed forrobots has also now been met, for example,with regards to illumination. Since 1997, light-ing conditions have been defined (light, colors,ball, goals). As a consequence, methods haveimproved. In 2006 for instance, we already hadvariable lighting conditions with one half ofthe field receiving natural light from outside.

Kalman filter methods seem outdated for usein this league due to the fact that Monte Carlois gradually taking over localization in practice.The robots can localize various objects in realtime now using relatively inexpensive standardhardware. Also with regard to controlling therobots, for example, applying defending andattacking maneuvers, the teams are in goodshape. Some attacks, which combine dribblingand simultaneous ball control and high-speedobstacle avoidance, are breathtaking to watch,when measured by any standard. Multiagentscheduling, assignment, and role-picking prob-

Articles

SUMMER 2007 117

Figure 2. Number of Teams at the RoboCup World Championships.

0

125

250

375

500

1997199819992000200120022003 2004 2005 2006

lems are well understood and can be solved forstandard situations. Realistic simulationsenabling the performance of coarse-grainedtraining (and learning) have now become pos-sible. The developers of the champion Brain-stormers Tribots from Osnabrück (Germany)were already well recognized for their effectiveresults using machine-learning techniques inthe Simulation League.

For 2007 the aim is to play outside in theopen field. Another challenge will, however, bethe integration of mixed teams from differentlabs, which when combined on site must coop-erate in applying their soccer knowledge.

Small-Size League (SSL)Robots in the Small-Size League (SSL) must notexceed 15 centimeters in diameter. There is acentral computer that determines the actionsof the whole team, with each individual robotbeing controlled by radio communication. Twocameras (one for each half of the field) delivera bird’s eye view of up to 100 frames per sec-ond, making it possible for the exact positionsof the robots to be viewed as well as determin-ing the position and velocity of the ball. Giventhe central control of the team, good coordi-nation can be achieved.

Since 1997, the environment has undergoneenormous change. For example, the field sizeis now four times as large as it once was. Mostof these alterations took place in 2004—when,beside a change of the field size, the walls wereremoved and auxiliary lights banned. In thefirst year following this radical change, mostteams performed quite badly. Due to the highspeed of the robots and the power of most kick-ing mechanisms, a high level of control andintelligent game play were needed in order tokeep the ball on the field and thus keep the

game running. By 2006, many teams have suc-ceeded in reaching this level of play.

Increasing the field size also forced teams touse more than one camera. This brought onnew challenges concerning the developmentof techniques for sensor fusion and the effi-cient processing of growing amounts of imagedata.

Starting exclusively with the fact that robotshave different drives, robot design has evolvedincrementally. The current state-of-the-artdesign for a robot is one with four (in most cas-es self-made) omnidirectional wheels, enablingvelocities of up to 2.5 meters per second to bereached (see also figure 4). The rules concern-ing ball-handling mechanisms (kickers, chipkickers, and dribblers) have continually beenadapted over the years in order to prevent inco-herent and rough game play and to enforcemore intelligent team tactics.

In general, we have seen a high level of tac-tical game play by the majority of the partici-pating teams. Most teams were able to play pre-cise passes and to perform well-directed chipkicks and intelligent defense behaviors (forexample, one-on-one defense). Some of thetop teams even displayed capabilities known as“one-touch soccer” (“1-2-3 passing”), whichmeans that following a pass, the ball is directlyshot once again without any interference. Thesuccessful application of such capabilitiesdemands a high level of teamwork and robotcontrol. The 2006 tournament was significant-ly determined by these capabilities rather thanby the implementation of simple and straight-forward strategies.

Adaptive team strategies, estimation ofopponent strategies, and fast cooperative playsuch as passing and shooting have also beensuccessfully tackled in the small-size league.

Articles

118 AI MAGAZINE

Figure 3. Middle-Size League Robots.

Left: MSL robots from 1997. Right: Robots in the World Cup final 2006.

There is however as yet no overall solution,since dealing with a whole team of robots(from the design of the hardware to the devel-opment of the software) is challenging in itselfand requires many resources. However, withfast and accurate game play, the SSL gameshave been quite engaging and entertaining thisyear. This has also been reflected in the largeamount of positive feedback and enthusiasmreceived from the outside community.

In 2006, the winning CM Dragons’06(Carnegie Mellon University) team reached anew stage of performance with the use of pre-cise hardware and sophisticated software. Theteam’s robots were able to pass with a high ballspeed (up to 2.5 meters per second) over sever-al opponents directly to the “feet” of anotherteammate with a precision of several centime-ters.

One of the future challenges for this leagueis an 11 by 11 game on a larger field. Particu-larly concerning robot vision, important issuesare also color correction for lighting variationsacross different fields, multicamera fusion, androbust tracking for handling ball occlusionsand estimating flight paths of high kicks. TheSSL also must address latency modeling (a goodteam has a latency of approximately 110 ms)and prediction methods to account for latency.

Simulation League (SL)The Simulation league consists of a virtualplaying ground with the original soccer fieldsize and 11 virtual players per team. The mainchallenge here is to deal with the problems ofcooperation among 11 players. This league wasestablished to explore strategic and tacticalbehavior of a team. A soccer server simulates

the physical world including the bodies of theplayers. The server transmits interpreted sensorinformation to the programs of the players,that is, their “brains.” The server therefore per-forms simple perception actions. Players have arestricted view of the scene (adapted fromhumans), and the received information isnoisy. The players’ programs can generate aninternal view of the situation and can decideon the next action (for example, dash, kick,turn, each of them with a certain power anddirection). The server then takes these actionsand executes them in the virtual playingground. A soccer monitor can be attached inorder to view the game.

This league implements a multiagent sys-tem, and the scenario offers research potentialin various areas such as coordination and coop-eration, distributed planning, learning of vari-ous levels (skills for a single player and teambehavior), and opponent modeling. To preventcentral control, each player program has to actindependently. Communication with the useof short messages is only permitted using a say-command through the soccer server. Theseprograms are available online (sserver.source-forge.net) and can be downloaded. The leaguestarted with a two-dimensional simulation andis now progressing towards a three-dimension-al simulation containing a more realistic phys-ical simulation (ODE).

The best teams can now show real soccerlikecoordinated behavior. Successful attacks fromthe wings, offside traps, strategic positioning,and open-space attack instead of direct passingare just some examples of their improving skill.In two-dimensional simulation, cooperativedefense ability has improved over the years. Itseems that not only individual skills, such as

Articles

SUMMER 2007 119

Figure 4. Small-Size League Final.

(5dpo, Portugal versus CM Dragons, USA)

an interception or marking, but also coopera-tive behavior, such as strategic positioning,have been improved. As a result, the number oftie games has increased, and the differencesbetween teams in terms of goals scored havedecreased. To break such strong defenses, thetop-level teams have started implementingmore cooperative offense strategies, similar tothose seen in human soccer.

In the three-dimensional league (figure 5),the top teams had highly developed low-levelskills (ball interception, passing, dribbling,high and low goal shots). Some of the bestteams employed open-space attack strategiesinstead of direct passing. We also saw theincreasing use of high passes (especially fromthe wings to the goal area) and high goal shots(also due to increased goal height). The cham-pions of the 2D and 3D leagues, Wright Eagle(University of Science and Technology of Chi-na), and FC Portugal, played exciting games.

Coach CompetitionEach team is allowed to use a coach program,which can analyze the ongoing game using aglobal view. Such programs can provide adviceto its team during the breaks (for example,while the ball is outside the field lines), and cansubstitute players (a maximum of threetimes)—an important development given thatplayers get tired and have varying levels of skillwith regard to kicking, running, and power. Aspecial coach competition has been establishedto evaluate intelligent advice from the coaches.In addition, automated commentators existand are used to provide live commentary ongames in an appropriate manner.

The coach programs within the coach com-petition begin by analyzing the previous-gamelog files of a given fixed two-dimensional team.

The strategy used by the coach consists ofseveral patterns. In the first step (offline analy-sis), the coach has to detect specific patterns(simple behaviors) from the log files. In thenext step (online detection phase), the coachmust detect patterns that are activated in thefixed team. The coach then advises its teamduring several full matches against a fixedopponent. The coaches should then detect theplay patterns of the fixed-opponent in eachgame and report on them. The coach can senduseful advice to its team members to ensurethat the detected patterns are correctly fol-lowed. The performance of a given coach isbased solely on its ability to detect these pat-terns. The year 2006 marked the second year inwhich the coach competition was held usingthis format.

The winner of the coach competition in2006 was the MRL team from Iran. Comparing

the results of the 2005 and 2006 competitions,the teams’ abilities in modeling and the detec-tion of patterns had improved. There wasmarked progress in feature selection and gen-eration, and also in behavioral pattern model-ing and detection. However, there is still muchto be done.

The coach competition initially started in2001. The goal at the time was opponent mod-eling and team adaptation. Between 2001 and2004, the measure of the coaches’ performancewas based on their score differences. Merelyscoring more or receiving fewer goals did notforce teams to apply opponent modeling, sothey only used a static hand coded list of pre-defined amounts of advice. In 2005 and 2006the structure of the coach competitions waschanged so that teams had to detect differentplaying patterns activated in the opponentteam.

The next challenges in the SimulationLeague concern further improvements of long-term coordination as well as the adaptation ofopponents using the coach and automatedanalysis of earlier games. Players should be ableto synchronize their intentions. To get closer tothe real robots, new simulation efforts (forexample, simulating humanoid robots) areneeded.

Four-Legged League (4LL)Sony’s AIBO is the robot that makes this league.One of the reasons for the existence of thisleague is its platform independence. Everyresearch team has the same type of robot; thusthere are no hardware development issues.Thus the 4LL is, in fact, a software competitionbuilt for the 576 MHz robots. The AIBO robottype ERS-7 comes with a wireless LAN and has20 degrees of freedom, and the joints can becontrolled with a frequency of 125 Hz. A cam-era placed at the front of the head offers 30frames per second with a resolution of roughly300,000 pixels. Before ERS-7, the teams usedERS-110 and ERS-210 models. The newer typesprovided better on-board computer and cam-era performance but also have differing bodydesigns. Therefore, skills like walking and kick-ing have had to be redesigned. This change hasforced the teams to develop related tools usingmachine learning.

As is the case in other leagues, the rules ofthe game have changed over the years, alwayskeeping up with cutting-edge technology. Thefield size has increased to 6 x 4 meters, thewalls and the sight protection (see figure 6)have been removed, and the number of land-marks has decreased.

The robots are able to perform numerous

Articles

120 AI MAGAZINE

actions due to the number of joints and motors(three actively controlled DOF per leg). The ballcan be kicked in almost every direction that itis possible to carry the ball (but only for shortperiods of time as governed by the rules), andspectacular overhead kicks are also possible.This variation in skills supports the develop-ment of methods for planning and actionselection.

Begun in 1999, this particular league com-prises the most basic skills among the hardwareleagues. The 4LL also attracts spectators of allages and is therefore an important league inRoboCup. However, the announcement bySony in the spring of 2006 that it was stoppingthe production of AIBO portends a change inthis league. RoboCup Federation is still cur-rently seeking a hardware platform that canreplace AIBO—one that is just as ambitious interms of its perception and action capabilities.

Since every research team works on the samehardware platform, the focus of this leagueconcentrates on perceptions and on methodsof controlling the hardware. The robots are get-ting faster and faster every year due to theimproved controls, which are developed usingmachine-learning methods. The robots cannow reach up to 50 centimeters per second.This common hardware platform also fostersscientific exchange—it belongs too to the spir-

it of RoboCup that solutions and programs areexchanged between the various teams.

We can also now claim new and advancedtechnological skills. In 4LL, topics andadvancements include low-level behaviors,learning, motion detection, and motion mod-eling. This year, the old question of whether touse a Kalman filter or a particle filter for local-ization was clearly won by the Kalman filter,since both finalists employed this method.Interestingly, 4LL differs from the Middle-SizeLeague, in which Monte Carlo–based methodsare preferred. A step towards collaborativeworld modeling was demonstrated in the pres-entation of the Microsoft Hellhounds where ablind robot was controlled by two seeing ones.The German Team showed that AIBO robotscould actually play soccer with a black-and-white ball instead of an orange one. This ischallenging because the on-board camera of anAIBO is rather limited. TsinghuaHephaestus, ateam from China, demonstrated the ability tolocalize on the field without colored land-marks.

A demonstration game with 11 AIBO robotsper team on a middle-size field was held in Bre-men for the first time at a RoboCup WorldChampionship. The existing programs fromthe 4LL had been adapted to the larger fieldsize. The experiment demonstrated individual-

Articles

SUMMER 2007 121

Figure 5. Current View of a Three-Dimensional Simulation League Game

istic styles of play, since no efforts had beentaken to improve cooperation with more play-ers. Nevertheless, localization and ball han-dling were performed as well as on the smallfield.

The champion NUbots (University of New-castle, Australia) team was superior due to itsprecise perception and action. Most of itsmatches ended with very clear results. It isinteresting to note, too, that in all real robotleagues the results of most matches appear tobe very clear-cut, for example, 6:0 or evenmore in only 20 minutes. This occurs evenbetween very good teams. These results are pos-sibly explained by the small fields on whichteams can often exploit small advantages in astraightforward way. In contrast, matchesbetween above-average teams in the Simula-tion League often result in draws.

Humanoid LeagueA humanoid robot has humanlike proportionsand appearance: one head, one body, twoarms, and two legs. Two different robot sizesexist—the child-sized robot with a maximumheight of 60 centimeters and the teen-sizerobot, which has a maximum size of 120 cen-timeters. The robots are also restricted to amaximum foot size, and a minimum height ofthe center of mass or the length of the arms.

The quick technological advancementswithin the RoboCup community can be bestdescribed in this league. In 1998, during theRoboCup World Championships in Paris, Hon-da’s P2 (predecessor of ASIMO) kicked the ball,while the world watched. The robot was

remotely controlled and should not have fallendown under any circumstances.

The RoboCup Humanoid League started offin 2002 with demonstrations of single skillssuch as walking or a penalty shoot-out. Thesetechnical challenges still belong to the compe-tition; however, since 2005 we have also seen 2versus 2 games on a field that is similar to the4LL field in terms of size and colored land-marks.

The robots on the field have to actautonomously as in all RoboCup soccer leagues.Human intervention is not allowed, and evenwhen robots fall down (which happens veryoften), they must get up by themselves (other-wise they are penalized). More than 30 percentof the kid-size robots were able to accomplishthis in 2006. Goalies were able to jump or fallinto one of the corners of the goal in order tocatch the ball (first seen in 2004, two years afterthe start of the league). Most robots were alsoable to kick the ball with force without loosingbalance. Some teams implemented multiplekicking behaviors applied to different situa-tions. The first teams also implemented stabi-lizing reflexes such as stop walking, which is atechnique used to attempt to regain balancewhen instability is detected. Some teams alsoimplemented protective behavior if a fall wasseen as unavoidable.

Most teams used motion macros with briefstops when changing walking direction and forcapturing images. In contrast, team NimbRo(University of Freiburg, Germany) used omni-directional walking, which enabled its robotsto change walking direction and walking speed

Articles

122 AI MAGAZINE

Figure 6. A Scene from a Semifinal of the Four-Legged League 2006.

rUNSWift (Australia) Versus Microsoft Hellhounds (Germany).

smoothly while continuously receiving visualfeedback.

In contrast to the other real robot leagues,there has been, up to now, no convergence tostandard solutions. The construction ofhumanoid robots differs from team to team.Some teams use low-cost construction kits.Other teams purchase more expensive com-mercial humanoid robots. However, mostteams construct their robots themselves. Whilemost robots were actuated with the help ofsmall servomotors, the 140 centimeter robot ofPal Technology (figure 7, the black robot on theright in the background) used harmonic drivegears. The teen-size robot, Lara of DarmstadtDribblers, was constructed to use antagonisticshape memory wires as muscles (see figure 8,the large robot on the left).

Perception (especially vision) of the gamesituation is realized using methods that havebeen adapted from other leagues. However,they are not able successfully to avoid colli-sions and contact with the ground. Whilesome teams used an omnidirectional camera asthe main sensor, others relied on directed wide-

angle cameras, and the remaining teams used amovable narrow-angle camera to keep track ofthe ball, the other players, and the field land-marks. No special lighting was used for thefields, and only the ceiling lights of the hallwere turned on. Most vision systems were ableto tolerate significant additional daylight com-ing through the windows. The 2 versus 2 finalswere played on the center court, where thevision systems had to adapt very quickly to dif-ferent lighting conditions. Complex mechani-cal compositions are responsible for new chal-lenges in perception. One of the problems isthe determination of the camera position andits direction.

The 2006 champion was once again the win-ner from last year, Team Osaka from Japan withthe robot Vision. It was a dramatic final—thefirst half had a clear result—4:0 for Team Nim-bRo from Freiburg—but Team Osaka was ableto reach a draw in the second half and finallyto win the game by 9:5 in over time.

Future developments will at first concernmore reliable and flexible skills. Given the his-torically fast developments in RoboCup, we

Articles

SUMMER 2007 123

Figure 7. Humanoids at RoboCup 2006.

These robots are dominated by constructions based on aluminum, carbon, and servomotors. Prototypes with alternative concepts like arti-ficial muscles (the gray in the left background) are exceptions and were not competing in this case.

should see substantial progress during the nextfew years. The next step will be the integrationof developments from other leagues. Neverthe-less, for competitions to reach the level ofhuman players, a long period of developmentis needed—if indeed this is possible at all. Newmaterials and training methods must be inves-tigated, with the current perception still on abasic level. Thus, RoboCup will remain a biglaboratory for investigation and integration ofdifferent fields.

Increasing overall system robustness is also amajor challenge. On the mechanical side, bet-ter actuators are needed that can tolerate shockloads, stronger skeleton structures, and softprotective covers. Professional electronicdesign and more emphasis on system integra-tion and testing is also needed, as is an overallincrease in reliability. On the software side,adaptiveness and learning could be used toimprove system robustness.

Most of the institutions participating inRoboCup will probably be unable to sustainthe construction of a complete team of sophis-ticated humanoid robots over a long period oftime. As a result, a combination of differenthumanoids may be developed at various insti-tutions, bringing together one team for thefuture. Such cooperation will be a challenge initself, let alone the numerous hardware andsoftware issues that must also be addressed.

RoboCupRescue LeaguesDisaster rescue is a serious social issue, whichin the case of RoboCupRescue involves a largenumber of heterogeneous agents in a hostileenvironment. There are two leagues operatingin this arena: the Rescue Robot League and theRescue Simulation League. The primary goal ofthe RoboCupRescue League is to promoteresearch and development in this socially sig-nificant domain at various levels, includingmultiagent teamwork coordination, physicalrobotic agents for search and rescue, informa-tion infrastructures, personal digital assistants,a standard simulator and decision support sys-tems, evaluation benchmarks for rescue strate-gies, and robotic systems that are all integratedinto comprehensive systems for the future.Integration of these activities will create thedigitally empowered international rescuebrigades of the future.

Rescue Robot LeagueThe competition field itself now has many gapsand debris and requires cooperation among therobots. In previous years the robots did nothave to cooperate, as there was only one robotin the arena. In the Rescue Robot League (fig-ure 8), robots now cooperate with each other

in order to maximize the overall number ofidentified victims. Difficulties in cooperationbetween robots from various teams arise due tothe differences in hardware and software.Achieving cooperation between robots in thisleague is therefore a significant step forward.

A new subleague in the Rescue Robot Leaguewas introduced in 2006: the Virtual RobotsCompetition. This competition is based on theUrban Search and Rescue (USAR)-Sim codedeveloped at the University of Pittsburgh,which in turn is based on the game enginefrom the commercial computer game UnrealTournament. USARSim allows high fidelitysimulations of multirobot systems. It currentlyoffers the possibility to simulate commercial aswell as self-developed robot platforms. USAR-Sim complements the rescue league in an idealway, with a realistic physical simulation ofteams of robots operating within collapsedbuildings. On the one hand, it offers the possi-bility to simulate search and rescue scenarios inwhich every agent has capabilities comparablewith those found on real robots, such as sens-ing with laser range finders or heat sensors. Onthe other hand, it opens the door to investi-gating aspects of autonomous multirobotcooperation on the “sensor level” withinunknown and unstructured domains. The lat-ter aspect is of particular interest given thephysical robots in the rescue league, whichmore and more allows for the development ofautonomous systems.

The virtual robot competition offers a widerange of challenges since each team can choosethe number of robots, the type of robots, aswell as whether they have an operator or runrobots completely autonomously. As has beenseen this year however, the difficulty of run-ning a large number of robots that face nearlythe same problems as real robots at the sametime, such as simultaneous location and map-ping (SLAM), exploration, and coordination,seems to be enough of a challenge. Some teamsfor instance have in particular focused on teamcoordination with more than 10 robots and theonline merging of maps. Simulation environ-ments are significant for the development ofphysical robots in every league. Almost allteams develop simulators over years. This yearhowever, teams in the Rescue Robot Leaguehave begun using simulators to validate theircontrol algorithms.

Rescue Simulation LeagueThe rescue domain represents a real multiagentscenario since a single agent cannot solve mostof the encountered problems. Fire brigades, forexample, depend on police forces to clearblocked roads in order to extinguish fires (see

Articles

124 AI MAGAZINE

figure 9). Moreover, the task is challenging dueto the limited communication bandwidth, theagent’s limited perception, and the difficulty ofpredicting how disasters will evolve over time.Therefore, this domain requires a high level ofteam cooperation and coordination, and so farno general solution has been found.

The purpose of the infrastructure competi-tion within the Rescue Simulation League (RSL)is to foster the development of software com-ponents for simulation, such as the simulationof the spread of fire. This is necessary becausethe teams actually compete against a disastrousenvironment rather than against each other, asis the case in other leagues. During the infra-structure competition this year, human-in-the-loop interface extensions for the simulator ker-nel were introduced. These become particularlyrelevant in utilizing the simulation system forreal disaster mitigation, such as the training ofincident commanders and disaster data inte-gration.

In the RSL, three major breakthroughs haveemerged over the last two years. First, a new,more realistic fire simulator has been intro-duced, enabling greater agent challenges. Forexample, buildings can now be preemptivelyextinguished. Other improvements include theincrease in the Kobe city map scale from 1:10to 1:4, plus a revised and more advanced pres-entation style of simulation. This new formatpresents the simulation on three separatescreens—one large screen displays the run ofeach team on the map in three-dimensionalformat, and on two smaller screens, a two-dimensional view and statistical informationare displayed.

Furthermore, locomotion in three-dimen-sional rescue simulation, for example, reach-ing other floors by using stairs or climbing ran-dom steps, present further challenges—challenges that form the basis for real robotteams to operate autonomously within the realworld in three dimensions. The simulationcompetition also opens the door for researchersto overcome the enormous technical hurdles ofrunning real robots as a team and having toprepare their robot architectures, even in largescale environments.

A special outdoor demonstration (the RescueRobot Field Test) was organized as a specialevent at RoboCup 2006 in Bremen, wheremobile robots and fire brigades worked togeth-er in a simulated hazardous material accident(see figure 10). As an opening attraction ofRoboCup, the Rescue Robot Field Test con-tributed to the general success of the champi-onship by demonstrating the advantages ofrobot technologies and their potential applica-

tions in daily life circumstances. In addition toits value as an effective demonstration tool forthe general public, as well as being an appeal-ing scenario for both the press and television,the Rescue Robot Field Test has served as a testbed for the development of advanced robots byvarious academic institutions and, to a largeextent, students and other young scholars.

RoboCup@HomeRoboCup@Home, first begun in 2006, focuseson real-world applications and human-machine interaction with autonomous robots.The aim of the league is to foster developmentof useful robotic applications that can assisthumans in everyday life. The league consists ofa series of fixed tests and an Open Challenge.The ultimate scenario is the real world itself. Togradually build up the required technologies, abasic home environment is provided as thegeneral scenario. In the first years it will consistof a living room and a kitchen (see the setup infigure 11), but soon it will also cover otherareas of daily life, such as a garden or park, ashop, a street, or other public places. Figure 12shows a living room and a dining room with arobot during the competition in Bremen. Thistest will become more advanced over time andserve as an overall quality measurement in thedesired areas. The tests should feature human-machine interaction and be socially relevant,application oriented, scientifically challenging,easy to set up, cost effective, simple, userfriendly, time efficient, and, lastly, be interest-ing to watch.

Articles

SUMMER 2007 125

Figure 8. A Robot from the Rescue Robot League Finds Its Way through the Step Fields

In the open challenge, freely chosen robotabilities can be displayed and proposals forfuture tests can actually be presented to theleague. According to the criteria ofRoboCup@Home, a jury then determines thescore and rankings. All robots participating inthe RoboCup@Home competition have to beautonomous. During the competition, humansare not allowed to directly (remote) control therobot, but natural interaction is allowed. So forexample, joystick, mouse, and keyboard con-trol are not allowed, but speech and gesturecommands are. For the early years, headsetswill be permitted. Decentralized computingwill also be allowed but may be difficult toachieve given general communication prob-lems that could occur during the course of thecompetitions. The real world presents a highdegree of uncertainty, dynamic changes, andvariation. In RoboCup@Home environments, arobot has to deal with all these factors. The lev-el of uncertainty will of course too increaseover the years in order to adapt to the currentenvironment.

Eleven teams participated at the 2006 com-petitions in Bremen. Five of them were new toRoboCup and had a background in human-

machine interaction. Human-machine interac-tion was demonstrated with the use of naturallanguage commands. Communication toobetween the scientists and the audience duringthe competition has shown to be a successfulway to provide an understanding for the moti-vation behind the robot technology presented.During the competitions, robots were able tocope well with the natural lighting conditionsand environment. The robots were able to nav-igate in unstructured and stochastic environ-ments and to manipulate real objects.

As the tasks in the Open Challenge and thefinals are not predefined, a jury had to evaluatethe performances by use of defined evaluationcriteria. This criteria included presentation, rel-evance to daily life, human-robot interaction,applicability, ease of setup, autonomy, difficul-ty, originality, success, and scientific value. Thewinner of the first RoboCup@Home competi-tion was the AllemaniACs team, fromRheinisch-Westfälische Technische-HochschuleAachen, Germany, which used one of its MSLrobots to perform in this league. The jury wasparticularly impressed by the precise, fast, androbust navigation displayed, including dynam-ic path planning and obstacle avoidance.

Articles

126 AI MAGAZINE

Figure 9: Scene from the Rescue Simulation League.

It is, however, worth mentioning that 45percent of the registered teams within thisleague actually had a background in human-machine interaction, with the others stem-ming from the robotics area. Communicationand exchange between these teams and theteams already active in RoboCup will on onehand help them to foster an understanding ofthe special demands of a robot competition(with respect to robustness of the robot systemsand the organization of teams) while on theother hand will allow for a gaining of newinput from the field of human robot interac-tion (which, up until now, has not directlybeen addressed in the RoboCup initiative).Issues encountered during the RoboCup@Home Competition include intuitive, human-machine interaction, manipulation of physicalobjects, and people recognition and tracking.

RoboCupJuniorMore than 1,000 young people in 239 teamsfrom 23 countries participated in theRoboCupJunior competitions in soccer, rescue,and dance. Soccer is played with one or tworobots per team, and the playing ground has

different gray scales enabling orientation to beperformed using simple light sensors. The ballemits infrared signals, and simple light sensorscan localize the ball. The rescue competitionconsists of line-following with obstacle avoid-ance and identification of objects. The coursealso includes an inclined plane. The dancecompetition is a free-style competition, inwhich the robots can perform dancing or act-ing according to a particular story. Often theperformances include both humans and robots(see figure 13). A jury then decides on the win-ner based on various criteria such as technicaldesign, performance value, and aesthetics.

The junior teams may use commercial kitsbut are also able to use their own designs. Themarket for related kits is growing in both sizeand technical aspects (following the develop-ments of technology), and thereforeRoboCupJunior is of great interest to commer-cial firms, particularly in Asia. Numerousawards are given to encourage and motivatedevelopment within this area. Recognition isparticularly given to those mixed teams fromdifferent countries, which were actuallyformed during the competition.

Articles

SUMMER 2007 127

Figure 10. Mobile Robots Work Together in a Simulated Hazardous Material Accident.

Team Germany 1 is a joint German RoboCup rescue project from University of Osnabrück, University of Hannover, and Fraunhofer IAIS. Theteam demonstrated two platforms at the outdoor event. The robot shown on the picture is based on a commercial platform from Telerob.

RoboCup Competition and Research

The combination of competition and sciencewithin the title “RoboCup Championship andSymposium” highlights the scientific claimthat the competition is one form of evaluationin which ideas and methods are exchanged. Infact, RoboCup is a platform for (mostly young)researchers, and its success and fast technolog-ical advancements over the last 10 years haveshown that the concept is effective, intelligent,and extremely relevant. The concept is basedon a number of points, which are discussednext.

Long-term goals are written down and aremaintained from year to year with a roadmap,which is based on the vision of 2050 but alsoconsiders current technical restrictions (forexample, battery life, camera resolution, CPUpower, and so on) of the market. In this way,the environmental conditions become moreand more challenging until they actuallyresemble the real conditions.

RoboCup events are very costly and by nomeans comparable to a “normal” scientificconference. One of the issues of the RoboCupFederation is therefore to decrease the expens-

es by allowing robot play in ordinary environ-ments such as sports halls (as seen in 2005within the German AI conference) and outsideon the green (planned in 2007 for the Middle-Size League).

Ideas, solutions, and also complete softwareprograms have been published providing newteams the chance to obtain the best methodsfor use on their robots. A prominent example isthe code of the German Team (4LL), whichwon the championships in 2004. A number ofteams downloaded the software from the Ger-man Team’s web page and were then oppo-nents at the same level in 2005. The winner’sdisadvantage in publishing their secrets is out-weighed by the references and citations and,therefore, their enhanced scientific reputation.This is one of the goals of the federation. Inaddition, the community as a whole benefitsfrom the exchange—the roadmap can bealtered towards the 2050 vision as the majorityof the robots advance technologically.

The technical challenge can be seen as thetechnological boundary of RoboCup. Therobots have to deal with various technical chal-lenges such as rough terrain or walking withhumanoids. One thing that all challenges havein common, though, is that the specific

Articles

128 AI MAGAZINE

Figure 11. The Setup of RoboCup@Home.

Source: RoboCup@Home rulebook.

10.00

3.004.00

Kitchen

Garden

Living Room

4.50

6.50

3.50

requirements are in fact needed in order toadvance to 2050 and as yet are not part of theregular game yet. The roadmap can be changedonce a technical challenge is successfully com-pleted by a number of teams. Colocated eventssuch as IJCAI, AAMAS, DARS, and ACTUATORSunderline the close relationship to the scientif-ic community.

RoboCup Techniques and Methods

As in other autonomous systems that deal withdynamic environments, a robot must have sen-sors in order to perceive its environment, actu-ators in order to manipulate its environment,as well as decision-making algorithms for theselection of appropriate actions. This dynamicenvironment requires a fast perception ofchanges (numerous cycles per second), whichmeans that there is only little or no time for acomplex processing of all the sensor data. Inthe early days, researchers in the MSL workedwith ultrasound sensors, which have sincebeen replaced by laser sensors. These days,most teams work with methods that are basedon camera images (mostly omnidirectionalcameras). An important issue is the integrationof various sensor data with the time compo-nent, which is commonly achieved using prob-abilistic methods (for example, Kalman filtersor particle filters).

In the early days of RoboCup, perceptionhad to be supported by strictly defined lightingconditions and colored landmarks and objectsin all leagues. The reduction and stepwise elim-ination of these means has been primarily test-ed in technical challenges. Recently, MSL, SSL,and the Humanoid League have been able toachieve play within ordinary illumination sit-uations in indoor halls.

The situation is different in the 4LL since thecamera of the AIBO has only a limited sensibil-ity. It has a visual angle of only approximately57 degrees in width and 42 degrees in height.Therefore, the robots still make use of the col-or-coded landmarks, which can actually getconfused with other objects in the background(for example, people’s clothing that has thesame colors as the landmarks). Teams try toavoid confusion by taking the horizon intoaccount. The position of the horizon in a par-ticular picture can be determined using the cin-ematic chain spanning the feet to the headwith the camera, but it is subject to error givenquick movements, and so on. Alternativesinclude the use of the field lines, which havealready been demonstrated during the chal-lenges between some teams, and actually a fur-

ther reduction of landmarks has been an -nounced for 2007.

The removal of the walls (which is now thecase in all leagues) has changed the require-ments of perception and action, but with dif-fering results. With the use of walls, a strategyas used in ice hockey was useful. Alternatively,the robot could try to push the ball along thewalls into the opponents’ goal. In fact, thegame became often stuck along the walls, sinceseveral players tried to push the ball in oppo-site directions. On the other hand, the removalof the walls created apprehension that thegame would be often interrupted because theball would leave the field. Hence theHumanoid League and the 4LL introduced spe-cial restart points instead of a literal throw-in.The ball is placed at a restart point inside of thefield, which gives the other team a smalladvantage.

MSL rules require a throw-in at the placewhere the ball leaves the field. In the begin-ning there were many repeated throw-insbecause robots failed to carry this task out cor-rectly. Nowadays, however, the robots havewell developed skills, and the occurrence ofthrow-ins has decreased.

In comparison, the Humanoid League hadno walls from the very start of the league for-mation. Teams were also able to use better-per-forming cameras. Robots using omnidirection-al cameras could use it in the same way as theMiddle-Size League. Other robots also benefit-ed over the AIBOs due to the higher position oftheir cameras.

As far as the actuators placed on the physicalrobots are concerned, there are now more fastand flexible maneuvers. There is a clear prefer-

Articles

SUMMER 2007 129

Figure 12: RoboCup@Home.

Robots deal with everyday situations such as a living room environment. They areable to open doors, follow humans, and manipulate real objects.

ence for smaller robots in MSL due to their dis-played superiority in the games over the lastyears (despite the fact that the larger, heaviersystems are more powerful). The various kick-ing devices—from less successful blowingdevices to powerful spring-based devices—didnot dominate, as expected, during the games.Until 2005, the ball was played flat on theground. Since then, and especially since 2006,the majority of the teams can kick the ball inthe air, enabling the team to play more effi-ciently and—in terms of soccer—in morehumanlike fashion. This is the case for bothSSL and MSL.

The development of the humanoid league isstill in its infancy; however, we have alreadyseen rapid progress in terms of the softwareand construction of humanoid robots. Onlythree years ago, walking without falling wasconsidered a success. The next significant stepwas a goal-oriented kick. In 2006, the partici-pating teams made good progress in imple-menting key soccer skills for their robots. Over-all, the walking ability of the robots improved.Walking speed increased, and walking behav-iors become more flexible and stable. Keepingbalance, especially in critical situations, is stillan issue. Balance requires excellent coordina-tion of sensors and actuators. There is a needto think about new ways of dealing with thevarious problems, to try out new materials(artificial muscles, artificial skin), and to devel-op new methods in addressing the energyproblem. Cooperation between scientists fromvarious fields is also necessary and highlyadvantageous.

Another distinct kicking skill worth men-tioning is the bicycle kick in the 4LL, whichwas first played by the German Team in 2002.Roughly 30 different kicks have been devel-oped for the AIBO (for all models), rangingfrom powerful kicks with the head and thebody to kicks with the legs in different direc-tions. Further skills also include dribbling andrunning. The robots must, of course, at alltimes obey the rule that holding the ball forlonger than three seconds is not allowed.

The two-dimensional simulation league usesa simplified simulation of ball handling. Theplayer can determine the direction and powerof a kick in the kick command, while the soc-cer server computes the resulting speed of theball compared to the last ball’s speed as well asthe position of the ball relative to the player.Additional background noise is also added.Therefore, the setting is reasonably complicat-ed, and good kicking and dribbling behaviorare not easy to perform. Interception of a mov-ing ball is also complicated given that percep-tion of the environment is challenged by noiseinterference. Varied approaches of machinelearning have been used to develop efficientskills for kicking, passing, dribbling, and run-ning. A number of these skills are available in alibrary for general use.

The three-dimensional simulation league isbased on a physical simulation by means ofODE. To date, only very simple player shapesare used. More realistic designs are under devel-opment in cooperation with the HumanoidLeague. However, the performance of comput-ers is actually somewhat restricting. Again,good skills can, however, be developed usingmachine learning.

A comparison between requirements for arobot in the RoboCup environment and arobot in a traffic scenario yields differingresults. Perception is easier in RoboCupbecause the robots act in a defined environ-ment. Primitive actions, such as braking, accel-erating, and steering, however, are easier in thetraffic scenario. This means that reactivebehavior is possible for the robot, and if thereis doubt, the car can actually be stopped. Inrobotic soccer, however, even primitive actionshave significantly more options (particularlywith the legged robots) given the differentnumber of degrees of freedom (such as drib-bling with a four-legged robot).

Cooperation between robots is still a chal-lenging problem. The only league achievinggood results has been the Simulation League.We can learn from this that classical planningalgorithms do not perform well in real time orin dynamic environments and that the time

Articles

130 AI MAGAZINE

Figure 13: Performance during the Junior Dance Competition.

issue is critical—a robot has to perceive, todecide, and to act in less than 100 ms. Solu-tions for these kinds of problems can also berelevant and of interest to robots in the trafficscenario.

The experience in RoboCup over the last 10years proves that the need for cooperation inrobot soccer largely depends on the size of thefield and the number of players in a team. Sim-ilar to human soccer teams, the SimulationLeague has a great need for high-developedcooperation strategies. It is interesting to note,too, that the development of these strategiesfollowed the same path as that of human soccerhistory—it started with static positions of theplayers, then the positions were dynamicallyadapted, and now we see increased movementcombined with related complex strategies.

In the leagues of real robots, the smallerfields did not force long-term strategic behav-ior. Only basic cooperation skills like passingwere used, and it seems that to a great extentthey are sufficient. However, an experimentperformed with the 11 by 11 demonstrationgame in the Four-Legged League has shownthat such strategies do not necessarily corre-spond well across different leagues. Thus, theleagues will soon require increased cooperationwhen playing larger fields, which will thenenable the strategies already developed for theSimulation League to be implemented.

Another difference between real and simu-lated robots relates to the control architecture.Both robots use layered architectures to a largeextent; however, while the real robots requirehigh-output efforts in order to manage boththe lower reactive control and short-term goals,the simulated robots are able already to man-age advanced levels of control using long-termintentions.

In all areas of RoboCup we have to solveoptimization problems within large parameterspaces as well as with the challenge of incom-plete and questionable data. Among this, too,are perception problems, such as the determi-nation of ball velocity or opponent modeling,primitive skills such as dash, kick, dribble, aswell as complex behavior such as team strategyand tactics. Machine-learning methods havebeen tested in all of these areas and haveproven to be extremely valuable.

Machine learning has played a major rolefrom the very beginning of RoboCup and wasfirst used in the Simulation League. This leaguestill provides a lot of data for machine learningand in addition is often used as a benchmark inthe world outside of RoboCup. Machine-learn-ing methods include reinforcement learning,evolutionary methods, support vector

machines, neural networks, case-based reason-ing, plus many others, often used in combina-tion. They appear on different layers, startingwith low-level skills (such as kicking or drib-bling), including positioning as well as passingbehavior, and ranging up to the level of strate-gic behavior. Opponent modeling appears onceagain on the different levels and plays a keyrole for the coach competitions as discussedpreviously.

Besides the Simulation Leagues in soccer andrescue, a wide range of tools used for the devel-opment, programming, and testing for each ofthe different real robot leagues exist. Manyteams have in addition developed their owntools based on simulations of their team robotsin a simulated environment. Since work withreal robots is time consuming and expensive,the use of such tools has many benefits. Someof these tools are complete simulations of thesoccer games; such is the case in the 4LL. It ispossible to test different skills and strategies invirtual reality, and of course they are also veryhelpful for debugging.

An important utilization of such simulationtools is that of machine learning. Skills in the4LL and in the Humanoid League need to beoptimized in a high dimensional parameterspace. The large degrees of freedom allow formany different solutions, and the performanceof the robots can therefore be improved everyyear. Simulation is used for first explorations,which later must then be evaluated and poten-tially refined and adapted to the real robots.Furthermore, a comparison of new candidates(for example, of new individuals in evolutionor of search directions using hill climbing) withexisting solutions allows for a preselectionprocess prior to beginning the actual experi-ments.

In contrast, the experiments with the realrobots lead to improvements of the simulationtools, with the ability to forecast by simulationimproving each time. Several tools are builtaround physical simulations resulting in a sub-stantial increase in performance. Nevertheless,the general experience in the RoboCup com-munity demonstrates that simulation is still faraway from total substitution of experiments inreality.

The most popular programming languageused is C++, although other languages, such asProlog, are also used by some teams within theSimulation League. Even Java, a language thatwould not perhaps be an obvious choice whentalking about real-time issues and robotics, isused successfully in the Simulation League.Apart from these pure programming languages,other unique languages are used, such as the

Articles

SUMMER 2007 131

coach language in the SimulationLeague.

ConclusionRoboCup was invented to tacklegeneral AI challenges. Now, after10 years, continued developmenthas resulted in significantimprovements particularly acrossthe areas of construction, percep-tion, cooperation, and interac-tion. The integration of all theseoutcomes and improvements intoa complete running system has ofcourse presented various chal-lenges. However, the large num-ber of well-performing teams nowin existence demonstrates that agreat deal of progress has in factbeen achieved during this rela-tively short period of time. Theactual integration of various fea-tures and rational behavior withrestricted resources is a key prob-lem in understanding (artificial)intelligence at all.

Nevertheless, there is room fordevelopment with the largemajority of issues in RoboCup,specifically relating to scientificquestions as well as technicalsolutions. Sometimes a gapbetween the two is taken into con-sideration, separating low-levelfeatures like basic perception andaction on the one side and high-level thinking on the other. Thiscould also be considered as thedifference between subsymbolicand statistical approaches versus

symbolic ones. Past experiences inRoboCup have, however, clearlyshown that none of these approachescan solve all issues alone (as is stillsometimes claimed). Teams from dif-ferent leagues have, for example, beenable to demonstrate that high-levelsymbolic approaches lead to advancedbehavior when combined with tech-niques that are based on low-leveldata. Higher-level approaches describesituations qualitatively. This is alsouseful in other domains like trafficanalysis or cell tracking for cancercells.

Most of the teams combine theirwork with other challenges from out-side RoboCup in terms of addressingscientific questions as well as technicalsolutions. The evaluation and theexchange of ideas during the competi-tions and the symposium is an effec-tive and worthwhile foundation fortheir future work. To achieve the 2050vision—to see robots come head to headwith humans in a soccer match—is notin truth really as important, but it is,indeed, an inspiring goal.

AcknowledgmentsWe are grateful for the compilation ofthe results from the various leagues,which formed a substantial basis ofthis article. The results were compiledby the organizing committee membersof RoboCup 2006 and in particular bySven Behnke, Andreas Birk, JoschkaBoedecker, Carlos Cardeira, AlexanderKleiner, Tim Laue, Paul Plöger,Thomas Röfer, and Thomas Wisspeint-ner. In the end, however, all of thiswork is possible only because of theparticipation of the thousands ofresearchers within the RoboCup com-munity all over the world.

Technical PapersThis article has been written as areport and therefore does not refer toany specific articles. However, allpapers presented on the RoboCupsymposia, listed by year of publica-tion, have been published as follows:

Bredenfeld, A.; Jacoff, A.; Noda, I.; andTakahashi, Y. 2006. RoboCup-05: Robot Soc-cer World Cup IX. Lecture Notes in Comput-er Science Vol. 4020. Berlin: Springer.

Nardi, D.; Riedmiller, M.; Sammut, C.; and

Santos-Victor, J. 2005. RoboCup-04: RobotSoccer World Cup VIII. Lecture Notes inComputer Science Vol. 3276. Berlin:Springer.

Polani D.; Browning, B.; Bonarini, A.; andYoshida K. 2004. RoboCup-03: Robot SoccerWorld Cup VII. Lecture Notes in ComputerScience Vol. 3020. Berlin: Springer.

Kaminka, G.; Lima, P.; and Rojas, R. 2003.RoboCup-02: Robot Soccer World Cup VI. Lec-ture Notes in Computer Science Vol. 2752.Berlin: Springer.

Birk, A.; Coradeschi, S.; and Tadokoro, S.2002. RoboCup-01: Robot Soccer World Cup V.Lecture Notes in Computer Science Vol.2377. Berlin: Springer.

Stone, P.; Balch, T.; and Kraetzschmar, G.2001. RoboCup-00: Robot Soccer World CupIII. Lecture Notes in Computer Science Vol.2019. Berlin: Springer.

Veloso, M.; Pagello, E.; and Kitano, H. 2000.RoboCup-99: Robot Soccer World Cup III. Lec-ture Notes in Computer Science Vol. 1856,p. 802. Springer, Stockholm.

Asada, M., and Kitano, H. 1999. RoboCup-98: Robot Soccer World Cup II. Lecture Notesin Computer Science , Vol. 1604. Berlin:Springer.

Kitano, H. 1998. RoboCup-97: Robot SoccerWorld Cup I. Lecture Notes in Computer Sci-ence , Vol. 1395. Berlin: Springer.

Ubbo Visser is an assis-tant professor at theCenter for ComputingTechnologies (TZI), Uni-versity of Bremen, Ger-many. He holds a MSc inLandscape-Ecology, aPh.D in geoinformatics,and a Habilitation in

computer science. His area of expertise is AIwith the focus on knowledge representa-tion and reasoning. The application areasare twofold: multiagent systems (RoboCup)and the semantic web.

Hans-Dieter Burkhardis head of the ArtificialIntelligence Group inthe Institute for Com-puter Science at theHumboldt University ofBerlin. He has studiedmathematics in Jena andBerlin and has worked

on automata theory, petri nets, distributedsystems, VLSI diagnosis, AI, socionics, andcognitive robotics. He is a fellow of ECCAIand vice president of the RoboCup Federa-tion.

Articles

132 AI MAGAZINE

Please Join Us Please Join Us in Vancouver forin Vancouver for

AAAI-07!AAAI-07!