Why table tennis challenges robotic masters
When most engineers think of automation, they picture repetitive assembly lines, not the lightning‑fast, spin‑laden world of table tennis. Sony AI chose this sport precisely because it strains perception, planning and control in a real‑world arena. Unlike video games that run in predictable digital realms, a ping‑pong match demands instant decisions, razor‑sharp physical execution and constant adaptation to an opponent’s unpredictable shots.
Eyes everywhere: nine cameras and three gaze‑systems
To track a 40‑millimetre ball traveling at 20 metres per second, the robot – nicknamed Ace – relies on a constellation of nine high‑speed cameras positioned around the table. These lenses capture the ball’s location 200 times each second, delivering a three‑millimetre‑accurate read‑out with only a ten‑millisecond lag. Yet position alone tells half the story; spin dictates how the ball will bounce and react to the paddle.
A second vision suite, dubbed “gaze control systems,” employs three specialized camera assemblies with tilting mirrors and a flexible tele‑lens. By keeping the tiny logo on the ball in sharp focus, a neural network extracts rotational speed from the dancing image, allowing Ace to predict bounce trajectories with human‑like intuition.
Learning by playing against virtual foes
Rather than hard‑coding every possible stroke, the system taught itself through endless simulated matches. In a physics‑rich digital arena, Ace practiced against a variety of AI opponents, learning to fuse sensor data, anticipate ball flight, and plan limb movements. Over time it compiled a library of “policies” – distinct playing styles such as aggressive topspin, deceptive backspin, or wall‑hugging placements.
During a live rally, the robot selects a new policy every 32 milliseconds, computes a trajectory for its eight‑degree‑of‑freedom arm, and checks the path for collisions with the table or its own structure before execution.
Hardware built from the ground up
The mechanical arm boasts eight freedoms: two linear rails for sliding motion and six rotary joints for fine articulation. This arrangement mirrors the flexibility of a professional player’s wrist, elbow and shoulder, enabling Ace to accelerate the ball up to roughly 20 m/s – the speed of a hard drive from a top‑ranked human.
Because official table‑tennis rules require a two‑handed toss, the robot incorporates a small cup at the paddle’s tip to hold the ball during service, compensating for its single‑arm design.
From simulation to competition
In April 2025 Ace faced seven human competitors, including five elite athletes with over a decade of intensive training. The robot’s adaptability and split‑second decision‑making allowed it to hold its own in best‑of‑three matches, showcasing a leap forward for embodied AI in sports.
Peter Dürr, director of Sony AI in Zurich, emphasizes that mastering table tennis is more than a novelty; it proves that machines can navigate complex, fast‑changing physical environments with the same fluidity once thought exclusive to humans.
