This weekend, in an arena packed with fans and competitors, a rather unusual soccer match took place. The players on the 9-by-6-meter field walked, passed, fell over, and even scored a few goals. No, this wasn’t the kindergarten championships—it was the 21st RoboCup, the international competition that pits robot soccer teams from universities around the world against each other.
There were several standouts for 2017, including the Rhoban Football Club from the University of Bordeaux and Bordeaux Polytechnic Institute, winners of “Best Humanoid”, the University of Bonn, and collaborators from The German Research Center for Artificial Intelligence and the University of Bremen.
This year, RoboCup landed back in Nagoya, Japan, site of the first competition. The event has grown in size and scope since then—it now features 15 competitions using various types of robots, including custom-built, off-the-shelf, and even some that are entirely virtual. But soccer, especially featuring humanoid robots, is the most important of them.
“Soccer is a very good research target, because everyone knows about soccer,” says Itsuki Noda, current president of RoboCup. “And also, soccer itself is a very complex and intelligent game, even for humans.”
Founders Minoru Asada, Yasuo Kuniyoshi, and Hioaki Kitano outlined the original goal for the project: To have the competition drive construction of a team of robot soccer players that can beat the human World Cup champions by 2050. In pursuit of this, the executive committee has been gradually upping the stakes, introducing new competitions every couple of years, and changing the rules and game design to push competitors into new territory.
“One of the big values of RoboCup is that it integrates many different AI challenges into a single system,” says Peter Stone, a professor of computer science at the University of Texas who runs the RoboCup teams there. “It’s not good enough to have a robot that can walk fast; it’s useless if it can’t also, with high reliability, see where the ball is, and figure out where it is on the field, and coordinate with its teammates.”
Last year saw a few significant rule changes—most importantly, a change from a bright orange to a regular-colored soccer ball—and teams responded by improving their entries’ computer vision. The Texas team finished second in the 2016 Standard Platform competition, says Stone, largely because of the success of their ball detection system. The Standard Platform competition requires teams to use the same hardware, so software is what makes a winning robot team in this event.
In addition to standard platform, competitors can enter humanoid leagues with three different sizes of purpose-built humanoid robots, ranging from about 16 inches high (won by Rhoban from Bordeaux) to full human size (won by the University of Bonn). Leagues of wheeled robots include small (7 inch diameter, won by Seer Robotics, a Chinese company with students from Peking and Zhejiang) or medium (square, about 20 inches on a side, won by the Beijing Information Science and Technology University) sizes, and have fewer restrictions on form.
Unlike BattleBots and other similar competitions, all of the robots in RoboCup are autonomous—the teams set them on the field and relinquish control to the software they programmed, which has to run not just the individual robots, but coordinate them as a team. The bots have to make decisions on their own and as a team, explains Stone. For example, it’s important for a robot to know where it is in relation to the field, the goal, the ball, and the other robots. But it can garner that information in several ways; it must balance it’s own comprehension—I took four steps this way, so I am four steps from the line—with visual input and what its teammates perceive of the field.
One of the most important changes in 2017 was the addition of a mixed-team challenge, says Joydeep Biswas, a former member of the wildly successful Carnegie Mellon robotics team, who brought a new team from the University of Massachussets-Amherst where he is currently an assistant professor of computer science. In the mixed-team challenge, teams were paired together without advance notice of who their teammates would be.
This has direct implications to real-world robotics. “As we move forward, we cannot expect all of the robots to be created by the same person or group,” says Biswas. “AI and software needs to be smart enough to play with team members which they have not programmed themselves.” This weekend, Biswas pointed out several new technical innovations that drove the competition forward over previous years, including changes to how the robots “kick” and the way they plan.
Also new in 2017 was the RoboCup@Home league, which features domestic robots trying to complete tasks such as fetching bottles and opening curtains. But these still had a secondary feel to the soccer bots.
Watching the humanoid soccer competition, it’s clear the robots have a ways to go. They often seem to move in slow motion. They waddle awkwardly and easily get turned around. But real progress is happening. Nowadays, the winning medium-size wheeled soccer team plays an exhibition match against the human trustees who run the competition. While the humans typically have their way with the robots, of late the robots manage to block some shots and get a few passes off of their own, though they are far from mounting successful offenses.
But it’s not farce. Roboticists can take away real lessons and practical knowledge from this game. Stone likens it to a grand challenge, like the space race or Deep Blue, the chess-playing computer. To accomplish a major goal that has little practical relevance itself requires a lot of technology that will be applicable in many other fields. To play soccer, the robots must perceive their environment, develop a plan or strategy, and then perform an action, like running, passing, or shooting.
The mixed-team challenge, points out Biswas, is a crucial step to getting robots by one manufacturer to work with robots from another. And perhaps most important, soccer is a game that requires real-time creativity—something that’s easy for people, and very hard for robots. Cracking that problem will make robots more useful in real-life situations, where robots have to react to changing conditions and virtually infinite scenarios. And doing so with humanoid robots has a particular benefit.
“In the near future, we need to collaborate with robots,” says Noda. “Humans understand each other by seeing the face, behavior, hand movement, and so on. So shape is very important for communication and interaction.”
Editor's note: This article originally misstated that RoboCup is in its 20th year. The competition is in fact in its 21st year. Smithsonian.com regrets the error.