Prof. Lee Seong-whan and his team’s paper published in Science Robotics
The team developed a robot AI technology capable of adapting to and outperforming in new ice environments.




연구진 소개

▲ From left: Dr. Won Dong-ok (first author), Professor Klaus-Robert Müller (adjunct professor at KU, co-author), and Professor Lee Seong-whan (corresponding author).

A research paper by Professor Lee Seong-whan and his student, Dr. Won Dong-ok (Department of Artificial Intelligence), was published in the world-renowned science journal Science Robotics on September 23.
* Research title: An adaptive deep reinforcement learning framework enables curling robots with human-like performance in real world conditions

The published paper, An adaptive deep reinforcement learning framework enables curling robots with human-like performance in real world conditions reports on adaptive deep reinforcement learning methods for training curling robots to quickly adapt to new ice environments without re-learning. As a result, curling robots can stably adapt to uncertain ice environments and show performance on the level of experienced human players.

An official Olympic sport, curling is sensitive to irregular changes in the quality of ice sheets affected by stadium temperature, humidity and ice surface status. Thus, it takes years of training for players to capably read the ice conditions in order to successfully deliver the stone to the desired location. However, the curling AI robot developed in this study was able to establish an optimal throwing strategy after only three to four days of learning and training. Its throw force and direction and stone curl rotation were carefully controlled, resulting in similar outcomes to those of an experience curling player.

Professor Lee Seong-whan, the corresponding author of the paper said, “The world’s first AI curling robot, Curly, is the fruit of advanced convergence technology, including the AI technology required for training to adapt to various ice conditions and the robotics technology required for training to throw the stones to desired locations.” He added, “This paper introduces the core technologies of robot AI that enables the robot to actually play curling at the performance level of experienced players. While the existing machine learning-based methods were restricted to solving and verifying problems in virtual or laboratory environments where the conditions are very stable, this paper studies the real ice conditions with high uncertainties that no one has attempted to challenge before. As a result, we have developed a core technology for AI for robots equal to the intelligence of a highly skilled human. This is a remarkable achievement, considering that the limits we faced with the machine learning-based AI technology can now be conquered with intelligence on the level of a skilled human in real conditions.”

This research was conducted as a part of the Artificial Intelligence Graduate School Support Project and ICT Convergence Industry Core Technology Development Project supported by the Ministry of Science and ICT/Institute of Information & Communications Technology Planning & Evaluation. 

[Description of Figures]

▲ Image 1. An adaptive deep reinforcement learning framework for the proposed curling robot AI.

▲ Image 2. Curling experiment and results in an actual curling alley.

▲ Image 3. The AI curling robot system (Curly) composed of pitching/skipping robots and curling AI.