First they came for chess, then Go... and now, oh for crying out loud, AI systems can beat us at curling

Video Machines have been able to beat human players at chess for years and now they have trounced their creators at "chess on ice," as curling practitioners sometimes refer to their sport.

Curling involves sliding hefty polished stones on a sheet of ice toward a target area, sometimes with spin so the stone curls. It's similar to various boules games and shuffleboard, but may involve an additional element: a team member "sweeping" the ice to shift the stone's trajectory toward its target.

A robot dubbed Curly recently played four matches, without sweeping, against a top-ranked Korean women’s curling team and the Korean national wheelchair curling team.

It – or they, since Curly consists of a thrower unit that gathers visual data near the target area, and external computers – played wheelchair-style, because sweeping is complicated for people in wheelchairs and for robots. And it won three of its four matches.

So long, Top Gun... AI software waxes US F-16 pilot's tail 5-0 during virtual dogfight drills

READ MORE

As described in an article [subscription required] published on Wednesday in the journal Science Robotics, Curly's creators managed this feat by training the robot to adapt to variable environmental conditions on the fly using a deep reinforcement learning (DRL) algorithm.

"Curly performs adaptive actions that can respond to the environment changes that occur continuously with every shot," explain authors Dong-Ok Won, Klaus-Robert Müller, and Seong-Whan Lee in their paper. "These changes have a notable influence on the performance if not compensated appropriately in a continuous manner."

Creating Curly required the system's designers to model environmental factors (like temperature, humidity, and friction), internal factors (the robot's ability to accurately carry out commands), and changes in the playing area (like the presence of other stones).

The Curly system includes an AI model that plans strategy (where to aim the stone), an adaption DRL model to adjust to environmental changes, and the two robot units – the thrower and the skip robots.

The skip robot – the "skipper" who stands in the "house" (target area) and directs the thrower – acquires the coordinates of the stones on the ice sheet and transmits them to the AI model, which passes the data to the curling simulator to formulate an optimal throw. The throw strategy gets passed to the adaptation DRL model for any necessary dynamic adjustments. Then the chosen target coordinates get passed to the thrower, which ideally will push its stone to the calculated spot, as you can see here:

Youtube Video

At about 1.3m accuracy, Curly is competitive with the results posted by national wheelchair curling teams (0.8m to 1.3m) in the Paralympic Winter Games of 2018. It's not quite as good as the 0.2m to 0.8m accuracy at the Olympic Winter Games 2018, where sweeping was allowed.

Even so, the numbers suggest that human players ought to have prevailed and the paper's authors offer several theories about why they didn't. They suggest that either that the players were too relaxed – Curly didn't make them feel competitive – or were too nervous, giving the robot an edge.

They also speculate that the AI strategy module may have offered better guidance than what human players came up with because the AI considers rare events and uncertainties in a way that people don't.

Regardless, the boffins hope their research will help other robot designers create systems that can incorporate real-time feedback into their operations. ®