Home Artificial Intelligence The AI-Powered Robot That Learnt Curling Using Adaptive Deep Reinforcement Learning

The AI-Powered Robot That Learnt Curling Using Adaptive Deep Reinforcement Learning

by Bernard Marr
0 comment

In curling, a sport that’s been referred to as “chess on ice” because of the strategy and precision involved, a robot named Curly beat Korean national teams in three out of four official matches. Robots have certainly come a long way, but they are still fairly clumsy, and most lack the dexterity of the human body. So Curly, the robot that mastered curling, is quite impressive.https://www.linkedin.com/embeds/publishingEmbed.html?articleId=8468261694260346636

What is Curling?

To fully appreciate this feat of technology, it’s important to understand the sport of curling. Curling requires the physicality of bowling as players push a 40-pound stone down a sheet of ice from a boundary called a hog line toward a target that’s 100 feet away. The target for the stone is called the house that has concentric circles—the closer you get to the target, the more points you get.

In curling, you compete against a team whose players also attempt to hurl their granite puck closer to the target or knock out yours to earn the most points. Curling strategy is about figuring out how to keep your opponent’s stone away from the house by bumping it out of position while doing so with enough finesse that your stone aligns in the house in an optimal position. The trick is that the friction of the stone and ice makes the elements faced by the competitors always shift throughout the match. Curling is no easy feat for man and an incredible accomplishment for a machine.

Curly and Its Creators

Klaus-Robert Müller at the Berlin Institute of Technology in Germany and his colleagues are behind Curly’s creation. Curly is powered by artificial intelligence, specifically an adaptive deep reinforcement learning framework. The robot has two wheels in the front and a caster wheel in the back. It has a telescoping camera that reaches 7 feet in the air to help the robot see the house and another one right above the front wheels for it to spot the hotline. Along with four smaller wheels shaped in a U and powered by a conveyor belt, the robot grasps the stone with its front wheel. It’s the U-shaped wheels that enable the robot to spin the stone, the curl that makes the stone spin right or left, a critical technique in the sport.

To help Curly learn the strategy of curling, the development team created a simulation of a curling game that Curly could compete with and learn from. The challenging thing to simulate was the ever-changing conditions that happen in each match—the ice conditions the stone’s polish and other physics of the sport. Human competitors must continually adapt to changing conditions. As a result, there was a gap between the simulation and reality.

Before a match begins, competitors are allowed test throws to learn more about the current conditions. Curly also did the test throws and then needed to align real-world experience with the mathematical models it learned from. It was programmed to compare the current conditions experienced in the test throws against the training model and adjust as necessary.

Additionally, during the match, Curly had to learn the best moves for the next throw, depending on the position of the competitor’s stones. In the simulation, Curly was given various scenarios and considered different throws, ultimately gauging the risk of each type. Using the knowledge it gained in training and then adjusting to the real-world conditions and progression of the match, the robot adapted its plan accordingly to achieve success. 

It’s important to note that these matches between Curly and the Koreans weren’t a replica of the sport since there was no sweeping—the process where teammates sweep a broom in front of the stone to scrub the ice to reduce friction and make the stone travel a straighter course—done by Curly or the Korean competitors in these matches.

Ultimately, Curly’s training and development resulted in success showing that artificial intelligence can adapt to real-world conditions. Olympic-level curling competitors learn the nuances of their sport over 15 to 20 years. It was truly remarkable that a robot powered by AI achieved so much in such a short amount of time and successfully adapted to the many variables that are part of a curling competition. Curly showed that even if there is a gap between physics-based simulators and real-world conditions, artificial intelligence can overcome it. This will be important as other artificial intelligence systems are developed.

You may also like