I, Robot. You, Ball.
Can a group of Austin computer scientists train a team of machines to play soccer without blowing a fuse?
On the third floor of UT-Austin’s Gates Dell Complex, Katie Genter is standing on a six-by-nine-meter piece of turf holding a toddler-size robot under one arm. Genter has enormous blue eyes and a blond ponytail; the robot, a 58-centimeter-tall humanoid machine designed and built in France and programmed to play soccer, is made of shiny white plastic. His name is Gouda Daniels, and he has two fingers and a thumb on each hand, skiboot-shaped feet, glowing eyes, ear holes like small saucers, and a serious “injury”: his sonar, lodged in his plastic chest, is malfunctioning. To move the ball around the field he’ll have to rely solely on vision.
When Genter, who is working on a Ph.D. in computer science at UT, powers Gouda on and places him on the turf, he adopts an athletic stance. He bends his knees, lifts his hands to shoulder height, and swivels his head. His teammates—Alison Brie, Arnold Schwarzen-cheddar, Gene Parmesan, and Colby Jack Bauer (the cheesy names are an inside joke)—are turned off. So are his opponents, a squadron of extra robots in numbered jerseys. The idea is to give Gouda a chance to focus on his individual rather than team skills.
This is more complicated than it might seem. With unchallenged vigor, Gouda kicks the ball forward, past the immobile opposing team, which appears to be a good enough start. Unfortunately, once the ball comes to a stop behind the other team’s players, Gouda can’t see it. His head starts to spin. He walks in a befuddled circle. Then he sets out in the opposite direction, toward his own goal.
“Oh, boy,” Genter says. “I might help him out.” She moves onto the field and nudges the ball into view.
This has the desired effect. Gouda stops, spins, and begins to stride purposefully toward the ball. But then, as if he’s overeager to show off his moves, he suddenly totters dangerously and then falls over, before he can even get in a kick. “This is just sad,” Genter says, resisting the urge to help him up. The problem, she explains, is that Gouda’s walking program was recently altered so that he takes longer strides. That’s helpful for moving faster but unhelpful for maintaining balance. Motion, Genter says, is the biggest challenge for soccer-playing robots. Gouda has mastered incredible feats of cooperation, vision, and localization (the robot’s ability to determine where it is in relation to stationary and moving objects). But walking requires constant recalibrations to deal with holes and bumps in the turf or slight changes in gradient. These bots, it seems, weren’t made for walking.
Now flat on his back, Gouda begins the laborious process of getting up. First, he props up his hips with his fists. Then he kicks his boots in the air for momentum, plants his feet, and hoists himself into a standing position. Vertical once again, he takes a few more steady strides. Then he teeters and falls down again.
Once he’s back on his feet, Genter follows him around the way you’d follow a toddler who’s just learning to walk: a little bent over, ready to catch him. This time, though, Gouda manages to maintain his footing and weaves his way past his opponents, kicking the ball several more times before, finally, scoring a goal.
This small victory took place in late June, just a few weeks before Genter and Gouda headed out to Hefei, China, for the 2015 Robot Soccer World Cup, an international robotics competition. The team they belong to, UT Austin Villa, is a mainstay in the Standard Platform League, one of RoboCup’s five soccer leagues. Each league has a different focus and different rules; games in the Simulation League, for instance, take place entirely in virtual reality. In the Standard Platform League, teams of five robots play soccer against one another, and no remote control is allowed; the humans cannot assist their robots during play.
Austin Villa—programmed by seven UT graduate students and a post-doctoral scholar under the guidance of Peter Stone, a computer science professor who founded the franchise, in 2003—is usually regarded as a force to be reckoned with. In 2012 the team won the RoboCup in Mexico City, defeating B-Human, a German team that had gone 49-0 before the final match, by a commanding 4–2 score. In 2013 Austin Villa came in a very respectable third. But a year later, after reprogramming some of its code, Austin Villa struggled from the start, failing even to advance beyond pool play.
This year, the team is hoping to regain its former dominance. But in the weeks leading up to departure, the players are struggling with setbacks ranging from physical damage to faulty vision algorithms. “Yesterday we had a meeting and we started panicking,” says Genter, who also serves as the chair of the Standard Platform League’s Technical Committee. “I think it’ll come together, but it’s not peaceful fine-tuning right now.”
The ultimate goal of the RoboCup competition is to create, by 2050, teams of humanoid robots that can beat the winning team of the human World Cup. Watching Gouda struggle to keep himself upright, that goal seems a little far-fetched. But in 1997, when the first competition was held, it was a feat just to set a single robot in motion and a huge accomplishment to get multiple robots walking at the same time. “They ran into walls, fell over, and in one case, even caught on fire,” Stone says. Now, not twenty years later, teams of five robots are not only moving but moving quickly while keeping track of other tasks.
The fact that these robots are programmed to act both adversarially and cooperatively is one of the unique components of RoboCup. Competitors in most robot competitions, such as the International Aerial Robotics Competition, act alone. In 1993, when Stone was a graduate student at Carnegie Mellon, he read “On Seeing Robots,” an article by University of British Columbia professor Alan Mackworth that proposed the idea of a robot soccer competition. For Stone, a competitive soccer player himself, that sounded like an inspired notion. At the time, much of the research in robotics focused on static planning, such as programming a robot that can get to the airport assuming the world around it is perfectly static. Stone was more interested in the interactions between systems: a robot that can get to the airport while navigating around other vehicles. What better way, he thought, to improve robots’ cooperative capacities than to train them for soccer?
Shortly after Mackworth’s paper was published, two groups in South Korea and Japan began putting together an international robot soccer tournament. Stone quickly contacted them, and when the first RoboCup was played four years later, in Japan, Stone was there. Now he’s a vice president of RoboCup and one of two people who have participated in every RoboCup since the competition began.
And his predictions about RoboCup have been borne out. According to Stone, RoboCup research has contributed to countless advances in fields such as machine learning and robot vision. Spinoff technology from RoboCup has been put to use in disaster rescue robots and industrial robots. Amazon has acknowledged that much of the technology behind the robots it employs in its warehouses draws on ideas first tested at RoboCup. In fact, RoboCup has often been compared to the Apollo program: there’s no immediate value in putting a man on the moon or in training robots for soccer, but both have fostered innovation that has tremendous practical use.
And every year the competition becomes fiercer, and new challenges are added to the tournament. This year in Hefei, Genter explains, the goalposts will be white rather than the customary yellow.
That’s a big hurdle for every team. By league rules, all the robot players are white; now that the goalposts are the same color, the robots keep mistaking opponents for goalposts.
It’s this difficulty that’s causing Genter the most anxiety as she prepares for the team’s trip across the Pacific Ocean. “We’re trying to fix it,” she says. She turns Gouda off and sits him down on a table, beside a row of computers. “I think we will. But that’s the panic point at the moment.”
The RoboCup soccer tournament takes place in mid-July in Heifei’s Anhui Conference and Exhibition Center, a huge, warehouse-like space. Twenty-nine Standard Platform League teams from nineteen countries take turns playing on five SPL fields, loud with simultaneous matches.
In the quarterfinals, Austin Villa—one of four teams from the United States—is playing a nail-biter against Chile. They’ve gotten this far after a series of near-miss victories, including a comeback against the Austrian Kangaroos and a triumph over China’s TJArk team following an all-nighter spent reviewing code and recalibrating the goalie’s pose.
Now, though, Austin Villa is increasingly apprehensive. Its whistle-detection algorithm is too sensitive for the crowded conditions in the Exhibition Center; the players keep detecting starter whistles from other fields and begin playing too soon, drawing penalties. And because the Internet connection here is slower than the one it’s used to back home, the team communication systems are malfunctioning.
These are sizable challenges, but Chile is struggling with the same issues, so it isn’t until the seven-minute mark, when Chile scores a goal from midfield, that Stone and his team decide that it’s time to change their strategy and call a time-out. As soon as play stops, Stone and his crew bring their robots to the sideline and turn off their whistle detection software, which has caused so much confusion in this noisy cavern. They know this will cause their players to start playing about fifteen seconds after each whistle, but it will also prevent penalties once play has gotten going.
Exactly fifteen seconds after the next whistle blows, announcing the resumption of play, Chile scores another goal, just as Gouda and his teammates are starting to move. The time-out, Stone and his team members realize, was a terrible mistake.
When halftime arrives, the humans of Austin Villa run onto the field and carry their players to the recharging station. Once all the players are plugged in, the humans huddle up and decide to turn the whistle function back on. When play resumes, this proves to be the right move; in the second half, though the team doesn’t score any goals, it holds Chile scoreless. Still, the loss is disappointing. “It’s painful to think that one poor decision may have cost us the quarterfinal,” Genter says.
A few days after the tournament, Genter still seems tired; those long nights of recalibrating algorithms take their toll. Even so, she and her crew are already discussing their goals for RoboCup 2016, to be held in Leipzig, Germany. Their priority over the next year will be to give the robots the ability to communicate without Wi-Fi, so that a slow Internet connection won’t affect their performance. Vision and whistle algorithms are important to RoboCup play, but communication is the key to everything.
Even if Gouda, Brie, Schwarzen-cheddar, Parmesan, and Bauer learn to run and kick with elan—or if they simply manage to avoid falling down and catching on fire— they’ll still be playing soccer. And soccer, as any coach will be quick to tell you, is very much a team sport.