A Historic Milestone in Embodied Intelligence
In April 2026, Nature magazine published a groundbreaking achievement that marks a significant milestone in the development of embodied intelligence: Sony AI's table tennis robot 'Ace' has successfully defeated multiple elite human players on a real physical table.
The Technology Behind Ace
Unlike the anthropomorphic robots often depicted in science fiction, Ace is an 8-degree-of-freedom (DOF) industrial robotic arm with a movable base and a custom racket assembly. This configuration provides greater flexibility than the human arm, allowing precise control over racket orientation and striking movements.
The system is equipped with what could be described as 'nine extra eyes' — nine high-speed cameras deployed around the court that precisely track the brand logo on the ball, calculating the ball's 3D position and complex spin data in real-time.
Reinforcement Learning Enables Real-Time Decision Making
Moving beyond traditional rigid trajectory programming, the Sony team infused the system with reinforcement learning (RL) capabilities. This allows Ace to make millisecond-level real-time decisions at the table tennis court, where ball speeds can be extremely fast and spin patterns highly complex.
The Journey to Victory
Ace began its competitive journey in April 2025, barely winning three out of five matches against elite amateur players in Japan. After nearly a year of intensive system iteration, the research team further enhanced the system's initial ball speed and offensive aggressiveness by late 2025, making ball placements closer to the table edges.
By March 2026, the system not only stabilized its serve return rate at 75% but also achieved extremely high offensive success rates. Ace consecutively defeated active T-League players Sone Sho and Ando Minami, as well as Kihara Miyuu, who consistently ranks in the world's top 25 women's singles.
Bridging Simulation and Reality
This achievement represents the first time that machines have crossed the massive gap between virtual simulation and real-world physics in the domain of human instantaneous reaction and dynamic decision-making.
However, Professor Peters from Darmstadt University of Technology in Germany points out that while dominating the table is impressive, this represents an extremely single-feedback task. There remains a long engineering barrier to overcome before truly solving core pain points such as fine grasping for general-purpose robots.
