Google develops a table tennis robot with a winning rate of over 40% against humans

According to news reports on August 11, 2024, DeepMind, a subsidiary of Google, recently announced that its research and development team has developed a table tennis robot that can reach the level of *** human table tennis players in competitions. The research and development team published a paper on the preprint website Arxiv introducing that this is the first learning robot agent in table tennis competitions that has reached the level of *** human players. Its main body is a six-axis robotic arm that can move back and forth and left and right through the bottom slide rail.

In 29 matches with humans, the robot won 13, with a winning rate of 45%. The opponents were all human players the robot had never seen before, and their skill levels ranged from beginner to advanced. The researchers said the robot lost all matches against advanced players but won all matches against beginner players and 55% of matches against intermediate players.

To achieve human-level speed and performance, the research and development team adopted a hierarchical and modular strategic architecture that enables the robot to not only master "low-level skills" such as forehand topspin and backhand push, but also formulate strategies through an "advanced controller" equivalent to the brain. During the game, the "advanced controller" can formulate the best skill plan based on the actual game situation, the robot's own ability, and the opponent's ability. After the game, the robot can also ***yze the battle data and continuously improve its skills.

The researchers said that this robot still has many shortcomings, such as a weak backhand game and being not good at dealing with fast balls, balls that are too high, too low, or strongly spinning. They will try to further improve the performance of the robot by improving the control algorithm and optimizing the hardware.