The concept of powerful supercomputers going head to head with the human race is something typically seen in science fiction. In 1997, it became reality after the IBM computer Deep Blue beat chess grandmaster Garry Kasparov in a legendary 3 ½-2 ½ series. This would only continue in 2012 when the software program Zen beat Masaki Takemiya at Go with a five- and four-star handicap, and in 2013 when Crazy Stone beat Yoshio Ishida with a four-stone handicap. DeepMind’s AlphaGo would later make history in 2015 by beating European Go champion Fan Hui without a handicap.
While defeat in a game of chess doesn’t necessarily signal the end of humanity as we know it, it has introduced more possibilities for artificial intelligence (AI) in the realm of sports—one of them being curling.
Curling is often regarded as “chess on ice.” Played as a team sport, human players use a stick to maneuver and aim 40-pound pucks across the ice to glide toward a red-blue target. Human players are generally excellent at determining the physics of how these large stones cross the ice at different speeds and directions. Recently, a curling-playing robot known as Curly managed to beat an elite team of South Korean curling players.
Curly: The Ultimate Curling Contender
Curly is a small white, turtle-shaped robot developed by a team of roboticists at Korea University in Seoul. Before its latest win, machines were typically used in sports that could be played in digital environments. Curly’s victory is the first time a physical robot was able to go against human players in the real world.
Physical robots usually fail to take into account unexpected circumstances. In the case of curling, abrupt changes in friction, humidity, temperature or ice melt can significantly affect the puck’s performance. Machine-learning expert Klaus-Robert Müller, one of the principal researchers on the Curly study, shared that the main challenge with curling is that the ice changes every time the puck slides across the ice. As a result, it is virtually impossible to gather sufficient data capable of considering all possible permutations. Thus, if a robot only used its knowledge from simulations, it wouldn’t be able to design a strategy during the actual event.
According to the researchers, even elite players usually have a margin of error by an average of 3 or 4 feet, which Curly was able to match.
The game ran for four sessions. To win, the machine used data gained from both simulated plays and physical games. This allowed Curly to adjust its strategy accordingly with each throw. Typically, machine AIs need to be run through a simulator before each turn. However, this would have broken the rules of curling and consume too much time.
For one of the sessions, Curly was able to scan the ice and develop its winning strategy. One of its stones was already within the inner red circle while the opponent’s was sitting just a few feet outside the blue outer circle, blocking a clear shot for Curly. The robot was able to successfully analyze the predicament and timed the release of the stone it was carrying to make its winning shot right smack in the center. It helps that the robot is also equipped with anti-slip software, preventing it from keeling over.
However, lead author Seong Whan Lee has been quick to express that the goal was not showing that robots could do it better but to encourage the use of collaborative robots.
“It is not about humans versus robots but rather humans and robots cooperating,” he said.
Deep Reinforcement Learning and Foosball
We are reminded of the German research team that created KIcker (see The KIcker Story: Foosball and Deep Reinforcement Learning, October 22, 2020) and what may be the ultimate foosball contender. KIcker is the brainchild of Bosch Rexroth and DXC Technology. This automated foosball system integrates various tech systems, including cameras, industrial PCs, controllers, servomotors and a neural network. This allows it to operate and control four player rods.
According to Michael Krause from Bosch Rexroth, it is virtually impossible to program every single combination of moves in foosball. This is where the neural network does its magic. The neural network is particularly significant because it allows the machine to make decisions on how to move the players.
“That was our mission: no programming,” Kraus said. “Just neural networks controlling the foosball game.”
While KIcker has had its fair share of games and opponents, the machine is no longer being used to beat humans in a game of foosball. Instead, KIcker is now being used as an educational tool for students to become interested in automation. In fact, the intention behind creating the machine was never for it to become the foosball champion of the world.
“We initiated our foosball study because we wanted to know how to apply artificial intelligence, especially machine learning, in industrial automation,” Krause said.
So far, Bosch Rexroth and DXC Technology have successfully demonstrated that simulation-based deep reinforcement learning can be an effective process in industrial automation. So where does that leave robots and artificial intelligence in the world of sports?
It’s Not a Competition
It’s unlikely that robots will be replacing human players anytime soon. While they do show an impressive capacity in some of the most challenging human feats, it’s important to remember that their “talents” often surpass human capability. Take for example the table tennis robot Forpheus, which can counteract spin shots. It is equipped with two cameras that allow it to track the ball while simultaneously tracking its opponent with two other cameras. However, that isn’t necessarily stopping the world of science and engineering from dreaming up ways to beat humans at their own game (quite literally) anyway.
AI and robots are going to be a mainstay in sports, but that doesn’t necessarily mean it’s all competition. The innovative processes behind the likes of Curly show potential in understanding the science behind these games, subsequently allowing human players to further improve their capabilities. Then Tibbets of the Brooklyn Lakeside Curling Club in New York has shared that Curly’s development presented the same problems they were attempting to solve, such as the changing conditions on the ice. Various sports are already using AI robots to practice and enhance their skills. However, these technologies can also make way for innovations that go beyond just the game.
Similar to KIcker, AI robots have allowed for relevant insights in the field of machine learning and automation that have contributed significantly to industrial and manufacturing applications. In addition to that, these technologies have opened up even more possibilities in how to better improve efficiency and optimize production. KIcker might not be beating anyone in foosball anytime soon, but it will continue to promote the technology behind it.
The Curly study was published in Science Robotics.