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

Abstract

Most artificial intelligence (AI) based learning systems act in virtual or laboratory environments. Here we demonstrate an AI-based curling robot system named 'Curly' that competes on a real-world curling ice sheet. Curly encompasses (1) an AI-based curling strategy and simulation engine under consideration of the high 'icy' uncertainty, (2) the thrower robot enabled by autonomous driving with traction control, and (3) the skip robot that allows to recognize the curling field and stone configuration based on vision technology. The Curly performed well both: in classical game situations and when interacting with human opponents, namely, the top-ranked Korean amateur high school curling team.

An adaptive deep reinforcement learning framework enables curling robots with human-like performance in real-world conditions

“…Last, we report on a series of tournament games that are unique events against human opponents, which present a real practical challenge for our robot, Curly. In the coming sections, we will conduct a comparison of the DRL adaptation, rule-based adaptation (33), no adaptation on simulated curling (see the "Simulated curling" section and Fig. 4), and a real curling ice sheet (see the "Test throw experiment on a real curling ice sheet" section and Fig.…”

“…To practically validate the resulting extended DRL framework (denoted later as adaptation DRL), we adapt the PG method (13,46) and also consider a simpler algorithm denoted as rule-based method as a baseline (see the Supplementary Materials for details) (33). We have introduced the system with the rule-based method (33).…”

“…To practically validate the resulting extended DRL framework (denoted later as adaptation DRL), we adapt the PG method (13,46) and also consider a simpler algorithm denoted as rule-based method as a baseline (see the Supplementary Materials for details) (33). We have introduced the system with the rule-based method (33). The main difference between the proposed adaptation and the rule-based method is whether they have dealt with the test bed environment with changes and full of uncertainty.…”

“…S8). Curling has a considerably higher number of allowed moves because the game progresses in a continuous environment (2,3,33). Because of stones' collisions, curling modeling requires a complex and timeconsuming physics-based simulation process to accurately describe and simulate the possible allowed moves for the next step (31,34).…”

A game strategy model in the digital curling system based on NFSP

“…Hong et al [15] conducted an in-depth study on changing the camera attitude to obtain curling images. Won et al [16] developed a curling robot based on curling strategies, which defeated the top-ranked amateur team in Korea.…”