Curly: An AI-based Curling Robot Successfully Competing in the Olympic Discipline of Curling

Dong-Ok Won, Byung-Do Kim, Ho-Jung Kim, Tae-San Eom, Klaus-Robert Müller, Seong-Whan LeeIJCAI 2018 · DOI

Introduction

While Artificial Intelligence has achieved superhuman performance in "perfect information" games like Chess and Go, transitioning these capabilities into the physical world remains a formidable challenge. Real-world environments are often unpredictable, noisy, and dynamic. Curling, an Olympic sport, serves as an ideal test bed for physical AI because it requires both high-level strategic planning and precise physical execution in a highly uncertain environment.

The Curly robotic system, featuring the Skip robot (left) and Thrower robot (right) on a standard curling rink.

The primary difficulty in curling lies in the "slippery" nature of the ice. The surface is covered in "pebbles"—tiny frozen water droplets—that change friction and shape based on temperature, humidity, and previous play. This creates a significant "sim-to-real" gap: a strategy that works in a static simulator may fail on the actual ice sheet.

System Architecture and Methods

The Curly system is a multi-layered architecture designed to perceive the environment, simulate outcomes, and execute precise physical actions.

Hardware: Skip and Thrower Robots

The system utilizes two specialized mobile robots. The Skip robot is positioned at the far end of the rink. It uses vision-based sensors to recognize the configuration of stones in the "house" (the target area). The Thrower robot is responsible for the physical delivery of the stone. It features autonomous driving with traction control optimized for ice and a specialized gripper to manage the rotation and release of the curling stone.

The Curly system workflow: vision-based stone detection feeds into a physics-based simulator and ice calibration model to calculate optimal throw parameters.

The Sim-to-Real Bridge

To overcome environmental uncertainty, the researchers developed a physics-based simulator that models stone trajectories and collisions. A critical component is the Ice Calibration step. By analyzing the actual trajectories of stones thrown during a game, the system updates its internal physics parameters in real-time to match the current state of the ice.

Strategic AI

The decision-making "brain" uses reinforcement learning and tree search algorithms. Unlike traditional AI that might aim for a single point, Curly's AI accounts for an uncertainty area. It evaluates potential moves based on a probability distribution, selecting "stable" strategies that are likely to succeed even if the ice conditions vary slightly from the model.

Experimental Results

The system was tested against a top-ranked Korean amateur high school curling team. Curly demonstrated the ability to adapt to real-world ice conditions and successfully translate complex strategic plans into physical actions. The robot performed competitively in both standard game scenarios and direct matches against human opponents, proving that AI can effectively navigate continuous physical spaces where the rules of physics are dynamic.

Limitations and Future Work

Despite its success, Curly currently lacks a deep "understanding" of its failures. Future research aims to integrate Explainable AI (XAI) to help the system analyze why specific shots missed, allowing it to learn in a manner more analogous to human intuition. Additionally, the current system does not include "sweeping" (altering the ice in front of a moving stone), which is a vital component of professional curling. Integrating robotic sweepers remains a complex future objective.

Summary updated

BibTeX
@inproceedings{won2018curly,
  author = {Dong-Ok Won and Byung-Do Kim and Ho-Jung Kim and Tae-San Eom and Klaus-Robert Müller and Seong-Whan Lee},
  title = {{Curly: An AI-based Curling Robot Successfully Competing in the Olympic Discipline of Curling}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/870},
}
Figure