The sport of curling requires such precision and strategy that it’s sometimes referred to as “chess on ice.” Players push 40-pound stones across frozen sheets, rotating the stones just enough that they “curl,” and try to knock opposing teams’ stones out of central rings.
Subtle variables at play—tiny, ever-changing bumps in ice, the pressure exerted by one’s hand, the smoothness of the stone—all impact the outcome, so much that curling requires machine-like precision from its players.
So, it makes sense that an actual machine might have a shot at winning, if it could learn to strategize on its own. Enter Curly: a robot powered by artificial intelligence (AI) that recently competed against professional South Korean curling teams and won three out of four official matches.
Curly’s impressive feat is recounted in an article published this month in Science Robotics by researchers Seong-Whan Lee and Dong-Ok Won of Korea University and Klaus-Robert Müller of the Berlin Institute of Technology. The robot gave a top-ranked women’s team and a national wheelchair team a run for their money, the authors write, thanks to its “adaptive deep reinforcement learning framework.”
Curly actually consists of two robots that communicate with each other: a “skipper” that aims the stone and a “thrower” that pushes it across the ice, reports Brooks Hays for United Press International (UPI). It rolls on wheels and uses a conveyer belt to rotate the curling stone, reports Matt Simon for Wired magazine. One camera on Curly’s “head” is able to give the robot a view of the field, and another camera just above its front wheels watches the “hogline,” or the boundary on the ice where players are required to release the stone.
When Curly competes, it raises its white, teardrop-shaped head and extends its seven-foot-long neck to get a good view of the field. Then, not unlike its human opponents, the machine drops low and pushes the stone in a gentle, controlled move across the ice.
Researchers designed Curly to assess risk and judge uncontrollable environmental conditions, per UPI. In curling, the composition of the ice sheet changes with each throw, so Curly had to learn how to adapt and make corrections on each subsequent throw.
As Devin Coldewey reports for Tech Crunch, the achievement is remarkable because Curly is able to make decisions in real-time as it plays the game.
“The game of curling can be considered a good testbed for studying the interaction between artificial intelligence systems and the real world,” Lee, co-author on the study, tells UPI. AI machines often perform well in simulations but struggle to cope in the real world, a problem known as the “sim-to-real gap,” Hays notes.
This problem is particularly relevant to curling, because no two ice sheets are ever the same, reports Wired. Each time a stone is thrown, the ice’s bumpy surface will change. Researchers programmed Curly with physics models that simulate the ice sheet, and then trained Curly to use its test throws at the beginning of the match to adjust its models accordingly.
Then, when Curly’s camera rise up on its long neck to look at the field, the researchers programmed the robot to assess the riskiness of each possible 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,” co-author Müller tells Wired.
As Jenna West points out for Sports Illustrated, one of the trademarks of curling is “sweeping,” when the player’s teammates use brooms to strategically sweep the ice in front of a stone as it slowly glides forward. According to Smithsonian Science Education Center’s Hannah Osborn, sweeping helps melt the bumpy ice pebbles on the ice’s surface and reduces friction. If the stone needs to move faster along the surface, teams will sweep more; if they need it to slow down, they’ll hold off.
Curly isn’t designed for sweeping, West notes. To make sure it was a fair competition, the South Korean teams didn’t use sweeping when they competed against the robot.
“All the nuances that these guys are taking into consideration, it’s fascinating to be able to do it,” Scott Arnold, head of development at the World Curling Federation, tells Wired. “… Because our Olympic athletes are training, you know, 15, 20 years, just to understand this themselves.”