Sep 23, 2020 2:00 PM

Meet Curly, the Curling Robot That Beats the Pros

Curling is one of the world's most precise sports. An ice-going, stone-throwing robot just mastered it.

I’ve spent enough time in robotics labs to confidently say that machines are far, far from replacing humans en masse in the workforce. They’re still too clumsy and stupid to work on their own, so they’re more likely to automate parts of your job. But I’ll be honest, I did not see this one coming: A robot named Curly just mastered the sport of curling, beating two Korean national teams. Building a robot to fire stones down a length of ice might sound eccentric, I’ll grant that, but rather than simply paving the way to the All Robot Winter Olympics of 2026, Curly is actually a big achievement in machine intelligence, one that could have implications for robotics beyond the rink.

Curling requires a mix of the physicality of bowling and the strategy of chess. Shoving off from a peg in the ice, a player called the “thrower” gently slides a puck made of granite, known as a stone, releasing it before a boundary called the “hogline.” The stone glides over 100 feet down the rink to the target, known as the “house.” An opposing team does the same, so both teams accumulate stones in the same house. At the end of the round, the team with a stone closest to the center of the target gets a point. If that team has extra stones closer to the center than the opposing team, those tally additional points.

The physicality of curling lies in the incredible precision required to set a stone on its course before releasing it at the hogline, giving it a spin to make it hook left or right while putting enough force behind it to land it in the house. The strategy comes from making sure your opponent doesn’t instead get their stones closest to the center of the house. For instance, a team can give a stone enough oomph to both bump out one of the other team’s stones, and with that collision halt their own stone on the target.

Video: Won/Science Robot

In other words, it’s one of the most delicate and methodical sports in the world. And it’s even trickier to teach a robot how to do it well. “All of that together is really a fantastic challenge that is not only special for curling, but it's I think a general challenge for all the AIs that go into the real world,” says Klaus-Robert Müller, a machine-learning researcher at Korea University, the Berlin Institute of Technology, and the Max Planck Institute for Informatics, and coauthor on a new paper in Science Robotics describing the system.

So, Curly has a lot of work cut out for it. The robot is shaped a bit like a flattened teardrop, with two wheels up front and one caster in the rear. (It’s difficult enough for a human to walk on ice, much less a humanoid robot, hence Curly’s wheels.) It’s equipped with two cameras, one that telescopes 7 feet high to give the robot a view of the house, and another on its face just above the front wheels to watch for the hogline. Curly grasps the stone between those front wheels using four smaller wheels arranged in a U shape. These are powered by a conveyor belt, which spins the stone, allowing the robot to curve its trajectory, just like a human player does. Spin clockwise and the stone will curl right; spin counterclockwise and it’ll go left.

That’s the bowling side of things—now for the chess. The researchers couldn’t just put Curly out on the rink and have it experiment with different throws, like a human would do to master the sport. This is, in general, a massive problem in robotics: It takes way too long for a machine to learn by trial and error in the real world, and it’s liable to hurt itself by trying some far-fetched maneuver and falling over. So instead, the team built a simulation of a curling game for a digital version of Curly to play around in. In this simulation, the researchers approximated the physics of the real world as best they could, but they were also missing some information because of the peculiar physics of curling.

Video: Won/Science Robot

“The unique thing about this sport is that the stones and the ice are actually changing all the time,” says Scott Arnold, head of development at the World Curling Federation, which was not involved in the research. (And which, by the way, does not seem to be threatened by our new robotic curling overlords.) “As a stone progresses through its life span, the bottom of it is actually polishing itself out all the time. So the interaction between the stone and the ice is changing every time you throw it.” Each stone—being a natural product made out of … stone—has its own quirks, too.

Imperfection is also built into the ice on purpose; what appears to be a smooth sheet of ice when you’re watching curling in the Olympics is in fact somewhat rough. A curling stone’s bottom surface is shaped like a cup, so if the ice was actually perfectly smooth, the stone would form suction and not move at all. Instead, crews sprinkle water on the frozen surface, forming tiny “pebbles.” When the stone sits on the pebbled surface, air gets under it and cancels that suction.

No two ice sheets have ever been laid down the exact same way, both because of this random bumpiness, and because of the quality of the water used, which can affect the way the liquid hardens into a surface. Accordingly, human curlers are allowed to take practice throws before a match to get a feel for the ice. All that imperfection only grows more chaotic as the game goes on, as the friction from the stones modifies the ice.

What does this mean for an ice-going, stone-throwing robot? Well, if a human or Curly were to somehow throw two stones the exact same way, the stones still wouldn’t take the exact same path. “Stone throwing for humans or for robots is imperfect, and it's imperfect for various reasons,” says Müller. “One reason is that it's impossible to judge the ice quality.”

Video: Won/Science Robot

The equations in the researchers' curling simulation do their best to approximate the physics of ice, but they’re not perfectly correct. “And so there's a gap between simulation and reality,” says Müller. “One of the things that we first do is we train the robot, or the AI, to reflect the physics model, but the physics model is erroneous.” That is, a kind of uncertainty is built into the system to reflect the uncertainty that the robot will face in a real-world game. Humans deal with this uncertainty by taking those test throws before a match. Curly does this, too, essentially aligning its mathematical models with real-world experience. “We can also use the test throws to gauge or re-adapt the physics-trained model to the reality, overcoming this gap,” Müller adds.

Then there’s curling strategy. For every configuration of stones in the house, there’s a best move with your next throw—for instance, slamming the stone into one of your opponent’s stones, knocking it out, and replacing it with your own. In the simulation, the robot is given a scenario and considers different throws. “Then the question is, how risky are these throws?” Müller asks. “There's a certain uncertainty about them. So you have to estimate, what will this uncertainty be? And in light of these uncertainties, what is the best strategy?”

Upload that knowledge into the real robot’s brain and Curly can master curling without much exertion at all. With that telescoping camera, it looks over the house and contemplates its move. “So you detect the stones, you think about where to put the stone, then you compute all the possible throws with the physics model. Then you compensate and see where this stone would go, and what the possible variants would be,” Müller says.

Müller and his colleagues trained Curly so well that it beat pro Korean teams in three of four official matches. But keep in mind, this wasn’t an exact replication of what you’d see in the Olympics, as Curly can’t sweep. When you see a human thrower’s teammates scoot ahead of the stone with brooms, scrubbing the ice, they’re reducing friction and making the stone curve less, forcing it to tack a straighter course.

Still, even without sweeping, Curly has learned to parse a galaxy of other variables inherent in the sport. “All the nuances that these guys are taking into consideration, it's fascinating to be able to do it,” says Arnold, of the World Curling Federation. “And, especially, they said it only takes a minute or two to reprogram the robot to throw the next stone, which I thought was fascinating too. Because our Olympic athletes are training, you know, 15, 20 years, just to understand this themselves.”

In a weird way, Curly is adapting to the many variables of curling the same way a person would when firing a cannon in battle. “The problem is that these variations change constantly and are essentially impossible to measure,” says roboticist Ken Goldberg of the University of California, Berkeley, who wasn’t involved in the research. “The final position of the stone is undecidable. But as centuries of cannon, gun, and missile aiming have taught us, it is possible to compensate to some degree based on examples.”

Like a cannon operator missing the first shot, then readjusting a few degrees left or right and firing again, Curly can correct its mistake. The robot has an added challenge, because its previously thrown stones have altered the ice that subsequent stones will slide over—a cannon operator doesn’t have to worry about previous shells altering the air. Yet the robot can still correct its mistakes.

And that’s a critical skill for any robot. As machines grow more advanced and find more real-world applications, they’ll frustrate the hell out of us if they remain as unadaptable and rigid as bread sticks. Just as Curly learns from its practice throws, so too must other robots adapt to the chaos of a home, office, or street, and not just stubbornly try things that no longer work. There’s no Olympic gold medal for property destruction, after all.