A curling robot has beaten humans at their own game.
Klaus-Robert MĂĽller at the Berlin Institute of Technology in Germany and his colleagues have developed a robot powered by artificial intelligence, called Curly, that beat teams of expert human athletes in curling matches.
Curling is a sport in which players slide heavy stones down an icy path towards a circular target. Players compete in two teams of four, with most players taking turns turn to “throw” a stone or to use brooms to sweep the ice in a stone’s trajectory.
Advertisement
The game requires both precise physical motion and strategic planning; points are awarded for stones closest to the centre of the target and a team wins by accumulating the highest score.
Purpose-built for the sport, Curly won three of four official matches against top-ranked Korean women’s curling teams and the reserve Korean national wheelchair curling team. The robot throws stones, but doesn’t sweep.
Mounted on wheels, the robot has a small crane-like neck with a video camera, which enables it to assess the position of stones, as well as a gripper that can rotate and release stones.
If Curly were to throw a stone at a certain speed and angle in simulation, the stone would theoretically travel to a specific point. But uncertainties in real life mean that on the ice the result is variable.
“The range where the stone hits has huge variance – more than 2 metres if you repeat this several times,” says Müller.
For that reason, the robot is powered by an AI known as a reinforcement learning algorithm, which takes into account real-world conditions, such as the position of other stones and the state of the ice, when deciding its next moves. In doing so, Curly achieved a top average shot accuracy of 1.3 metres from its target.
The initial condition of the ice depends on the person maintaining it, and it also changes after every throw, says Müller. Curly and the human players all had four preparation shots per game to gauge the initial state of the ice.
Once the game began, Curly continuously learned how to improve on its previous moves, based on the errors that arose from its preceding throws.
Science Robotics
