Google AI robot beats humans at table tennis, wins 45% of matches

Google DeepMind has introduced the first AI-powered robot table tennis player that can compete at an amateur level.

The system consists of an industrial robotic arm, the ABB IRB 1100, integrated with custom AI software from DeepMind.

The robot's performance was tested in 29 matches against unseen human players of various levels. It won 45 percent of the matches overall. It lost to all advanced players, but won 100 percent against beginners and 55 percent against intermediates, demonstrating a solid amateur level.

“This work takes a step toward that goal and presents the first learned robotic agent that achieves amateur human-level performance in competitive table tennis,” researchers said in the study.

Matching human skills

Despite recent advances, achieving human-level accuracy, speed and adaptability of robotics remains a challenge.

Unlike strictly strategic games like Go or Chess, table tennis requires the acquisition of advanced low-level skills and strategic moves. It often favors bold, feasible actions over ideal plans.

According to researchers, table tennis is therefore an ideal testing ground for the development of robotic skills such as fast movements, instantaneous decision-making and system architecture for direct human competition.

Playing a full competitive match against human opponents is an important test for the development of robots in both physical and strategic tasks. No research has been successful in addressing this difficulty.

Adaptive AI in table tennis

The team's approach resulted in a robot that can play competitively at a human level and that people also enjoy playing against.

This was achieved through modular policy design, methods to transfer skills from simulations to reality, real-time adaptation to new opponents and user studies where the robot was tested in real matches.

The robot’s architecture combines a high-level controller that determines the optimal skill for each situation with a library of low-level skills. Each low-level talent focuses on a single table tennis action, such as backhand aiming or forehand topspin.

According to researchers, the system collects information about the pros and cons of these skills, allowing the experienced controller to choose the best course of action based on the opponent's skills and game statistics.

The robot is first trained through simulation using reinforcement learning, and after repeated cycles of real-life play, its performance improves. By using a hybrid strategy, the robot can improve its skills and adapt to more challenging gameplay.

“It’s really amazing to see how the robot plays players of all levels and styles. Our goal was to get the robot to an intermediate level. Amazingly, it did that, all the hard work paid off,” said Barney J. Reed, a professional table tennis coach who has worked with DeepMind, in a proposition.

The team played competitive matches against 29 table tennis players of different levels, including beginners, intermediate, advanced and advanced+, as determined by a professional table tennis instructor, to assess the level of our agent.

The robot was physically unable to serve the ball, so the human and the robot played three matches using modified versions of regular table tennis rules.

The robot won 46 percent of games and 45 percent of matches against all opponents. Furthermore, the robot won 55 percent of matches against intermediate players, lost all matches against advanced and advanced+ players and won all matches against novices.

According to the researchers, the results clearly show that the agent performed at an average level of human play during rallies.

Furthermore, participants in the study rated the robot highly for “fun” and “engaging,” regardless of their skill level or the outcome of the game. Most wanted to play again.

They played an average of 4:06 of every 5 minutes freely. Although more experienced players took advantage of some of the robot's shortcomings, they still had fun and saw potential in the robot as a dynamic practice partner.

Interviews revealed that the robot struggled with underspin, which was supported by the fact that its performance dropped significantly as spin increased.

Researchers claim the problem offers specific targets for future training developments due to the challenges of dealing with low balls and identifying spin in real time.

The details of the team research were published in the magazine arXiv.