[Reading Science] Human vs AI... After Baduk, Table Tennis Next

Google DeepMind Unveils Table Tennis Robot Competing Against Humans
45% Win Rate Against Humans
Lost All Matches Against Experts
Significant That It Can Compete with Humans
Combining with Robots Is a Planned Step in AGI Evolution
Early Evolution into Humanoid Robots Seen as Physically Challenging

Google DeepMind, the creator of ‘AlphaGo,’ has returned with a ‘new weapon.’ DeepMind, which shocked humanity by winning against 9-dan Go player Lee Sedol, is once again challenging humans. This time, it is a robot that plays table tennis with humans. The robot athlete unveiled in conjunction with the 2024 Paris Olympics is still not quite ready to compete with humans, but attention is focused on how much it can improve by the 2028 LA Olympics four years from now.

◇ A robot playing table tennis with humans learns smashing through AI = Recently, Google DeepMind published a paper on a table tennis robot armed with artificial intelligence (AI). DeepMind’s table tennis robot is not humanoid. It consists of a robotic arm, a learning computer, and cameras. The robot arm, similar to those used in factories, is equipped with a table tennis racket. Cameras track the ball, and motion capture technology monitors the opponent’s movements. Through this process, the robot devises strategies and responds on its own. Using machine learning, it learns how to play table tennis and develops strategies to compete against humans.

A table tennis robot developed by Google DeepMind is playing a match against a human. Photo by Google DeepMind

Although the table tennis robot does not look like a player, its skills were formidable. The robot won 13 out of 29 matches against humans, achieving a win rate of about 45%. According to videos released by DeepMind, the robot’s performance was roughly at an amateur level. It won all matches against beginners but was helpless against more skilled players. However, the robot table tennis player was not powerless. It skillfully returned balls that barely grazed the net. Humans who lost to the robot sometimes showed disbelief, while those who won appeared confident, as if it was expected.

Although the overall win rate is below 50%, the fact that it defeated humans is regarded as a remarkable advancement not only in AI but also in robotics. Oh Sang-rok, director of the Korea Institute of Science and Technology (KIST) and a robotics expert, commented, "There were robots playing table tennis and volleyball 20 years ago, but now it is significant that robots can play table tennis as well as humans."

Why did DeepMind choose table tennis among many sports? They explain that table tennis involves judgment, physical ability, and various physical characteristics required to return the opponent’s shots. DeepMind trained the robot in various table tennis techniques. After mastering specific actions such as backhand, forehand, and serves, the robot learned algorithms to play comprehensive matches using these skills. It was trained to make decisions to choose the necessary offense or defense depending on the situation.

The most challenging part of implementing the table tennis robot was fast attacks. Physically and in terms of software, it was still difficult for the robot to handle humans’ rapid fast attacks. AI could not match humans’ ability to visually judge the ball’s spin. This means the robot cannot respond well when a skilled player puts spin on the ball or sends a ball without spin. It also struggled to react when the ball was tossed high or hit low for serves. Physically, it remains a reality that the robot cannot surpass humans’ quick movements.

To defeat skilled players, the robot must also develop the ability to execute unpredictable strategies. It needs to use tactics that catch humans off guard. The AI must also be able to analyze the opponent’s strategies?understanding which attacks or defenses the opponent excels at, their strengths and weaknesses, and how to counter different strategies. The more games played, and the more rematches with the same person, the higher the chances of winning become.

Panang Sanketh, who led this project, said he was surprised by the rapid development of the robot’s learning ability. He stated, "Just a few months ago, we expected the robot could not beat someone it had never faced before, but the results were different." Considering the time humans need to practice to become proficient at table tennis, the robot’s achievements cannot be ignored. He emphasized that table tennis, which requires executing various skills and strategies in action, is on a different level from strategic games like chess or Go.

The researchers claim that this experiment is significant as the first robot capable of enjoying sports with humans. They also argue it will become a milestone in robot learning and control. The science media outlet MIT Technology Review evaluated that robot developers have taken a step closer to creating robots that can safely work at home and in workplaces.

Noteworthy are the reactions of humans who competed against the robot table tennis player. Those who lost to the robot were likely frustrated, but advanced players who defeated the robot responded that it has potential as a practice partner. The generally positive reactions indicate that harmony between robots and humans is possible.

Professor Lrel Pinto of New York University explained, "This case is a fantastic example of robots working alongside humans."

◇ Completion through the combination of AGI and humanoid robots = The development of robots is accelerating in tandem with AI. Tesla has even set an ambitious plan to sell a humanoid robot called ‘Optimus.’ There are predictions that Elon Musk will displace humans in workplaces. Musk is developing AI through xAI, dreaming of the convergence of autonomous electric vehicles, robots, and AI. Jensen Huang, CEO of NVIDIA, also predicted the rise of robots, stating that AI will evolve into physical AI. At the GTX2024 event in April, Huang showcased advanced autonomous mobile robots (AMRs), and at the keynote speech at Computex 2024 in Taiwan in June, he staged a scene appearing to link arms with a humanoid robot, emphasizing that robots are part of NVIDIA’s future. This scene is regarded as a reminder that achieving artificial general intelligence (AGI) ultimately requires integration with robots.

Audiences at the 2024 WAIC event held in Shanghai, China, are watching Tesla's humanoid robot, Optimus.
[Image source=Reuters Yonhap News]

At the recent 2024 World Artificial Intelligence Conference (WAIC) held in China, the open-source humanoid robot Healthy Loong developed by China attracted attention. Compared to Western robots, Healthy Loong’s movements were slower, but its open-source nature allows anyone to develop software using this robot.

Wang Xingying, CEO of Unitree Robotics, argued, "To achieve AGI, physical robots that interact with the real world, participate in human activities, and imitate, learn, and understand human emotions are necessary."

However, it is expected to take considerable time before humans can play table tennis with humanoid robots. Director Oh said, "Compared to the speed of software development, physical development is slower," and predicted that it will still take a long time before humanoid robots that behave like humans appear. Musk has also repeatedly postponed the external sales date of Tesla’s humanoid robot Optimus.