RoboCup‐97: Robot Soccer World Cup 1

Citation

Andrew, A.M. (1999), "RoboCup‐97: Robot Soccer World Cup 1", Kybernetes, Vol. 28 No. 5, pp. 72-75. https://doi.org/10.1108/k.1999.28.5.72.2

This is the proceedings of the First World Cup Soccer Games and Conference held in Nagoya, Japan, in conjunction with the International Joint Conference on Artificial Intelligence IJCAI‐97. It is part of an international initiative termed RoboCup, whose ambitious long‐term goal is stated as:

to build a team of robot soccer players, which can beat a human World Cup champion team.

The idea of robots playing soccer was first mentioned by Professor Alan Mackworth (University of British Columbia) in a paper in 1992. Independently of this, a group of Japanese scientists organised a Workshop on Grand Challenges in the same year, and decided that robot soccer provided a potentially profitable Grand Challenge. This was followed up a year later by the decision to launch a robotic competition, initially planned to be within Japan. However, the news spread and interest was expressed by workers in other countries, and RoboCup was established as an international project. The first public announcement was made in 1993, and regulations were drafted.

An event called Pre‐RoboCup‐96 was held during an International Conference on Intelligent Robotics and Systems (IROS‐96) in Osaka in November 1996. Eight teams competed in what was termed the “simulation league” and there were demonstrations of robots suitable for the “middle size league” (see below).

In the 1997 event, a total of 40 teams participated (real and simulation combined) and over 5,000 spectators attended, and the event was judged a great success.

Needless to say, the robotic devices, real or simulated, fall far short of what would be needed to achieve the ambitious long‐term goal. The robot players, in the versions where they are physical devices, in fact run on wheels. (There is, in the book, a paper discussing the use of legged players, but suggesting quadrupeds for the time being. There seems to be no suggestion of settling for this and calling the game polo!) There are three “leagues” in which real robots can compete. One is the small‐size league, played on a field the size of a ping‐pong table, and another is the middle‐size for which the linear dimensions of the table are three times as great. For both of these, there are five robots in a team. For the small‐size league it is expected that control will be due to global viewing of the field, but for the middle‐size entrants the robots should have on‐board vision systems and be governed by what is in their field of view. The third league for real robots is the “expert robot league” for which the conditions are left open, allowing the demonstration of special features.

To allow competition in the “simulation league”, the competing programs must connect to a “soccer server” which manages the simulation and display. There must be a separate control for each player, though possibly using copies of the same program for each; overall global control is not allowed. Each competing program is able to send messages to its own player, selected from a repertoire of instructions, and “sensory data” from the player is conveyed in the other direction. The soccer server can also let a program play against eleven humans, where the latter control their respective virtual players through appropriate interfaces.

The competitions clearly have entertainment value, but their significance goes well beyond this. The stimulus provided by the Grand Challenge is real. The Preface and also the back‐cover notes emphasise that 1997 was the year in which a computer beat a world master in chess, and also that in which an autonomous vehicle performed successfully on the surface of Mars. These achievements could well have been the aims of earlier Grand Challenges.

It is clear that in the work to date on RoboCup, major problems of balance and dynamics have been bypassed by the use of wheeled robots, but the remaining problems of multi‐agent operation and of opponent modelling are enough to be going on with. At the end of one of the papers, in which the outcome of the 1997 event is reviewed, it is argued that a deficiency of the schemes tried was that they did not learn on‐line, so that the behaviour did not change radically during a match. Future entries will presumably learn, or at least hone their skills, so as to adapt to their opponents′ special characteristics. Another feature that was said to be missing was communication between members of the same team, like the verbal communication between human players.

The papers in this first publication show that the project has already stimulated valuable effort. They are placed under three main headings of Overview papers (four in number), Technical papers (14), and Team descriptions. The last group is further subdivided into four papers on small‐size teams, five on middle‐size ones and no less than 18 on simulation teams. A Team description submission was required for each entry in the competitions, but many of these are effectively technical papers emphasising what are felt to be the special features of the respective entries.

Included in the Technical papers, in addition to that on quadruped robots already mentioned, there is one on a system using the currently‐popular Java programming language for the controlling software, allowing easy modification. Others review general aspects of multi‐agent operation, and of the real‐time processing of sensory data, and generation of the required multimedia presentations. Another considers the value of project‐based education in robotics, of which RoboCup could be held to be an extreme example. There is also a rather surprising contribution, in the form of a careful analysis of 2,402 football articles that have appeared in the Times newspaper, all of them from the 1996‐1997 football season. The idea seems to be that this will be useful in deciding about the rules and characteristics of future robotic football, so as to preserve the features that make live football interesting.

In the Team descriptions, quite complex mathematical arguments are used in presenting the playing algorithms used. The currently‐fashionable topic of reinforcement learning is invoked in several presentations, not surprisingly in the context of this kind of delayed‐reward successive‐choice activity. Subsumption architecture is also mentioned, again not surprisingly since the requirement is to pursue fixed objectives (it is dangerous to use the word “goals”) which are interrupted to perform the constituent activities of kicking or dribbling the ball or getting into a favourable position or intercepting an opponent. One of the papers by a Japanese author (“Development of self‐learning vision‐based mobile robots for acquiring soccer robots behaviors” by Takayuki Nakamura) combines reinforcement learning with an interesting form of self‐organisation amounting to automatic partitioning of the state space to suit the control requirement.

Another paper describes the development of playing skills by a genetic programming method, where the necessary evolution took place prior to the competition. The evolutionary process is described as being fascinating to watch, since it began with policies such that most of the players never touched the ball during a game, and gradually improved until players could be seen taking on distinct roles as attackers and defenders.

All of the treatments illustrate the shortcomings of what is referred to in the final paper as GOFAIR, or Good Old Fashioned Artificial Intelligence and Robotics. Early work on robotics, still relevant in industrial applications and elsewhere, tended to assume the environment was unchanging or slowly changing, except when operated on by the robot itself. There are many interactive situations, amounting in fact to much of real life, where the GOFAIR approach is inadequate, and the RoboCup environment is a prime example and a useful testbed.

The book, and the whole project, are extremely interesting and potentially valuable. The plan is announced in the book to hold RoboCup‐98 in July 1998 in Paris during the Third International Conference on Multi‐Agent Systems (ICMAS‐98), and no doubt details of this will appear in a later volume. The third event, RoboCup‐99, is announced for Stockholm, as part of the International Joint Conference on Artificial Intelligence (IJCAI‐99).