Robot Ecosystems
The Trials and Topics of RoboCup
ASADA Minoru--About six years ago KITANO Hiroaki,[*1] KUNIYOSHI Yasuo[*2] and myself started a project called "RoboCup." It was a new sort of challenge to the approaches to AI (Artificial Intelligence) prevalent at that time in the West. Lately it has become quite well known, but it emerged from a series of workshops, actually. Robotics scientists use the phrase "behavior skills" to distinguish embodied intelligence (intelligence which emerges from the interactions of the hands, legs, head and other physical body parts) from the intelligence of symbolic reasoning. There was a debate among roboticists about whether such embodied intelligence could rightfully be called "intelligence," or whether new aspects of intelligence could be extracted from such approaches. In 1993 we were looking at ways to express it.
Mr. KITANO had just come back from a robot contest sponsored by the AAAI (American Association of AI). He said it was boring. They were doing things like holding time trials to see how quickly robots could take a paper cup off a desk. Their robots would first observe their environment with a video camera, and reconstruct world models based on these images. They would sit perfectly motionless for 10 minutes doing this, and then move 10 cm or so and then repeat the process. Classic symbolic reasoning, but not really a spectator sport, nothing to capture the viewer's imagination. That's where we came up with the idea of having the robots play soccer, and the result was RoboCup.
At the time my research group had been developing shooting robots, a departure from the industrial robots made to serve in factories, something a bit more fun. We were interested in instilling some of these new concepts in them, and both Mr. KITANO and Mr. KUNIYOSHI had similar concerns. In the fall of 1993 we decided to hold the first RoboCup in Nagoya, in 1997. We had four years to prepare. Now, this may look like only so much fun and games, but as a research theme it was quite meaningful, really defining a lot of priorities for us, even among complex issues like multiagent hierarchies.[*3] It is even being proposed as a standard problem.[*4] Just as winning at chess remained a standard problem for years for symbolic reasoning in AI, so might RoboCup become a standard problem for embodied intelligence. Something which AI researchers everywhere can attempt to solve. While as one of its organizers I believe that the RoboCup contest is an interesting general theme, personally I expect that they become used as a basis for developing the model of cognitive robots, including learning, evolution and cognition processes. Chess was a standard AI problem for 40 years, but once IBM's Deep Blue had beaten world champion Garry KASPAROV, it had achieved one level of completion. It is important that our new objectives develop a contrasting set of thematic concerns.
Chess was static, dealing entirely with perfectly organized information, in a one-to-one battle of the wits. RoboCup, on the other hand, deals with multiagent hierarchies. What is decisive about this is, I suppose, the existence of multiagents making autonomous decisions towards a common purpose. This is a problem for which there is no final goal. I think that this is the essential difference, and solving it is our mission. We need to discover how to create the robots so that they achieve a sense of society--as surely they must if they are to work towards a common goal--including ways of regulating themselves and cooperating with others. We need to explore exactly how far they can evolve from emergent behavior based solely on the design of their internal structure. There are three leagues in RoboCup. The first is the "simulation league," for those who haven't the financial, human, architectural or other resources to actually build working robots. The games are played out in software on a "soccer server," so it's a league that's easy for even computer scientists to contribute to! (laughs) Each team consists of 11 programs (one for each player), so a total of 22 programs run simultaneously in a competition between two teams. It looks like a video game, though because there are 11 programs in contest with 11 other programs, it involves issues of complete dispersal, flocking behavior, and other problems associated with how to gain coherent, or at least coordinated learning behavior. The simulation agents are made to have characteristics like the physical robots with their 90 degree visual fields, using video technology, meaning that things become more vague with distance, more precise up close. They can exchange information with their teammates, though their opponents can also overhear them. Because they're "live," they can try and fool their opponents, or engage in noise battles. If they rely on too much transmission power, however, they naturally expend their energy, and fatigue more easily, just as they will fatigue if they expend too much energy too early on in the game. The simulation league reproduces these kinds of "realistic" conditions, including having the players adhere to "soccer-like" movements.
The other two leagues use actual robots, one for "little," and one for "middle" sized units. The "little" robots are roughly 15 cm in diameter. Equipping them with sensors, video apparatus and CPU's was beyond the scope of the technology at that time--though it is possible today--so we created instead five robot bodies operated by one CPU, on a playing field with one observation camera, affording a continuous view of the entire playing field, mounted on the ceiling. We're using a ping pong table as our field, and playing with a golf ball.
The robots from the "middle" league are around 45 cm in diameter, and because these units can each carry their own perceptual gear, the ceiling cameras have been forbidden (although they were permitted in part of the first meet in Nagoya). These robots each act on their own, relying purely on the information received from their video apparatus. The fields in this league are nine times that of a ping pong table, or about one fifteenth that of an actual soccer field. At present, they aren't able to make "throw-ins" and so we have the field surrounded by a wall. It's actually more like ice hockey, but there just aren't as many hockey fans as there are for soccer (laughs), so we're calling it "soccer."
The reason that we didn't select baseball is that most of the action occurs between the pitcher and the batter, and that the rules are so complex. In terms of simplicity of instantaneous attack and protection, and sharing of space, one other game in consideration was basketball, but dribbling was too difficult for most robots. Soccer was just the simplest available.
SAKURA Osamu--When I was researching chimpanzees in Africa, the local kids used grapefruits as soccer balls. You're right about soccer being a simple sport. Maybe that's part of its attraction.
ASADA--Both for us and for the robots, the rules are simple and easy to work with. We've already got enough on our hands, keeping 11 robots in competition with each other. Our stated final goal is to be able to take 11 humanoid robots to the world cup in 2050, and defeat the champions. In other words, the solution of a new standard problem, much like defeating the world's grandmaster in chess, was the solution for a standard problem for some 40 years.
SAKURA--Originally they estimated that computers would be able to defeat humans in chess by the 1960s, though it actually took 30 years longer than expected. So while we might say that soccer with humans is a goal for the end of the 21st century . . . .
ASADA--I'd be 97 years old in 2050, so I have no assurance that I'll even be alive at that time. Take for example the time frame for evolution from the Wright brothers' invention of flying machines to the commercial application of the jumbo jet. Predicting the potential for growth in scientific technologies is extremely difficult, even for things previously deemed impossible. The computing power that used to fill an entire room now fits in one's pocket. Some people say that considering scientific advances like this, the idea of a team of robotic humanoids beating, if not the world champions, then at least a team of ordinary mortals, does not seem that out of the question.
SAKURA--I find the fact that you, Mr. KITANO and Mr. KUNIYOSHI each independently came upon the same ideas at the same time quite interesting. Throughout history, at the beginning of important movements we often find places where several people come upon strikingly similar ideas at the same time. Take for example the great Hungarian scientist John Von NEUMANN (1903-1957). From the Hungary of his youth also came the POLANYI brothers (Karl, 1886-1964, and Michael 1891-1976), Gyorgy LUKACS (1885-1971) and many others. For whatever reason, there are many synchronicities making prewar Hungary a sort of "smart spot" on the map. DIAGHILEV's (1872-1929) Ballets Russes must have been a similar phenomenon, having NIJINSKY and KARSAVINA dancing to music written by STRAVINSKY or RAVEL, in front of sets designed by PICASSO or to a script by Jean COCTEAU. In many genre we find that there are periods and places that have attractors, if you will.
One other thing that I thought was interesting was how you mentioned that chess was the standard problem in the West, whereas Japan, being a latecomer to the field, was in effect exempt from this issue, and in fact may have gained a kind of "late justice" in having a fresh position to envision the next generation of issues from.
ASADA--I had never really thought about it that way. It was more a matter of Mr. KITANO coming around and discussing the idea with me when I had already begun researching soccer playing robots. Mr. KUNIYOSHI was onto very similar sorts of things.
SAKURA--And none of you had any idea that the others were doing this, right?
ASADA--I'm sure that we weren't completely ignorant of each others' existence, but we certainly didn't know of such projects, no. To be frank, Mr. KITANO was researching it from an AI point of view, and Mr. KUNIYOSHI from a robotics standpoint, but the ideas that they were able to explore just weren't all that interesting. They were looking for a project that would be more productive. That's really what it came out of. Mr. KITANO's thinking doesn't really mesh that well with most Japanese anyway, he really just wants research themes that are engaging to himself, but that ingrates him more to Americans and Europeans than it does Japanese. I do think that your perception of our synchronicity is interesting, though.
Next