Smarter than AlphaGo, Curly’s first competition with humans

Research team led by Professors Lee Seong Whan  and Sull Sanghoon develops AI curling robot



At 2:30pm on Thursday, March 8th Korea University demonstrated the capabilities of the world’s first artificial intelligence (AI) curling robot, which was developed with support from the Ministry of Science and ICT (Minister Yoo Young-min). A two-end match was held between the AI curling robot named ‘Curly’ and a high school team from Gangwon (Chuncheon Mechanical and Technological High School) at the Curling Center of Korea Paralympic Committee Icheon Training Center (Icheon, Gyeonggi).

In April 2017, Korea University was selected as a member of the AI curling robot developers’ consortium by the Ministry of Science and ICT. In its first year, the research team used AI technology to develop the AI curling software application ‘CurlBrain’, which creates strategies for optimal stone throws, and the AI curling robot ‘Curly’.
* The consortium consists of 60 researchers from eight organizations including Korea University, Seoul Korea Curling Federation, DGIST, UNIST, and NT Robot, among others.

Curly is capable of viewing the prevailing situation in a game through its camera, of devising strategies by employing Deep Learning, and of moving around the ice in order to play the game.
* Curly’s Deep Learning database is based on the records of a huge number of previous international curling games (1,321 games, 11,000 ends, 160,000 throws).

* When the skip robot sees the situation of a game through its camera, CurlBrain devises strategies for achieving the best throws and sends the resulting data to the thrower robot. Then, the thrower robot executes the correct throw in terms of power and direction as well as stone rotation.



As an additional attraction at the demonstration, an AI curling software contest was held, in which the top four developers from the preliminary contest competed. The champion took on CurlBrain, which had emerged victorious at the International AI Curling Software Contest in Japan last November.



In the second year of the consortium’s research (2018), sweeper robots will be developed and their performance will be constantly improved until they reach the level of human curling experts. The hope is that the technology will in time be implemented in Deep Learning-based AI games, in computer vision devices dealing with motion, and in the more precise control of AI robots. Furthermore, it will serve as a basis for increasing the competitiveness of curling by helping both beginners and experts in their training.



Professor Lee Seong Whan  from the Department of Brain and Cognitive Engineering at Korea University, which is a major research institution, said, “There have been efforts in the past to apply AI to the world of physics, but this is a very rare case in which an AI program successfully controls robots which require human-like advanced thinking.” He added, “I hope Curly can help the Korean national curling team win medals at the next Winter Olympics in Beijing”. The research team also plans to broaden the sphere of application of the technology behind the curling robots, by employing it in various fields such as automated manufacturing and unmanned transport services.




Another primary developer of the curling robot, Professor Sull Sanghoon (School of Electrical Engineering, College of Engineering, Korea University) said, “Unlike AlphaGo, which considers only a finite number of possible stone locations, Curly has to take into account a wider range of variables, including an infinite number of possible stone locations, and a huge range of possible stone impacts, ice quality, and gaming performance.” He continued, “Along with theoretical input, the technology must also demonstrate practical expertise. The performance of the technology will be greatly improved once sweeper robots are developed in the second half of the year.” He thanked his fellow researchers for devoting so much of their time to working in the cold conditions of the curling center at which the research took place.