World’s First! A Robot Takes on Tennis Players, Makes Extreme Saves, and Bounces Back at Lightning Speed
The Cutting-Edge Technology Behind the AstraTennis Moment
In 2016, AlphaGo defeated Lee Sedol, bringing artificial intelligence into the public spotlight for the first time in an almost astonishing way.
This human-machine contest, which went down in the history of AI, not only proved that machines could challenge the world’s top human players in complex games, but also became a landmark moment in AI’s transition into a new era.
Ten years later, in 2026, the “AstraTennis moment” for embodied intelligence has arrived.
On August 22, the World Humanoid Robot Games opened to great fanfare. At the opening ceremony, Galbot was invited to participate in a robot tennis match, competing on the same court as top tennis players. The event was broadcast live worldwide by China Media Group.
On the court, the humanoid robot autonomously ran toward, tracked, and hit tennis balls traveling at high speed, demonstrating a full range of tennis skills, including forehands, backhands, serves, returns, baseline rallies, and net volleys.
Against human athletes, the robot completed extended rallies. In doubles, it collaborated with human teammates in real time, dynamically adjusting its strategy based on the situation on court. It even made spectacular saves during high-speed offensive and defensive exchanges.
This was the first time in the world that a humanoid robot’s fully autonomous tennis match was presented in its entirety through a major live sports broadcast. On a real court, the robot perceived, made decisions, and acted autonomously, engaging in real-time competition and strategic interaction with human athletes.
Ten years ago, AlphaGo stepped onto the Go board. Ten years later, AstraTennis demonstrated the abilities of a tennis player on a real court.
From the board to the court, and from complex games in the digital world to perception, decision-making, and action in the physical world, AI is crossing a new capability boundary.
This is the AstraTennis moment for embodied intelligence. With this breakthrough, Galbot has created a landmark moment of original Chinese innovation in embodied intelligence.
An All-Around Humanoid Robot Plays Tennis: Physical AI Enters High-Speed, Real-Time Competition for the First Time
Tennis is one of the most challenging sports for embodied intelligence because it simultaneously tests a robot’s perception, decision-making, full-body motion control, and real-time competitive reasoning.
A high-speed incoming ball requires the robot to make a judgment within an extremely short window. Hitting the ball requires not only precise control of the arms and wrists, but also highly coordinated movement of the legs, torso, and upper body.
Once it enters an actual match, the robot must also adjust its strategy in real time according to the opponent’s movements: deciding how to run, how to hit, where to place the ball, and how to win the next point.
That is precisely why the globally televised exhibition match demonstrated a humanoid robot engaging in a complete tennis contest in a manner approaching that of a human athlete.
In terms of movement, Galbot displayed a highly human-like athletic state. Its serves, forehands, backhands, and returns were natural and fluid. It moved rapidly between the baseline, the net, and the corners of the court. When faced with shots landing in different locations, it continually adjusted its posture with quick, small steps and was able to make full-court saves in extreme situations.
But more important than “looking human” is “understanding tennis.”
Galbot has developed general-purpose tennis capabilities covering the entire match—from serves and forehand and backhand shots to baseline movement, net play, and control of shot trajectories and placement; from independent offense and defense in singles to positional coordination and tactical adjustments in doubles. It can autonomously choose how to hit and which strategy to adopt based on the rules and the situation on court.
This ability was also fully demonstrated during the intense competition. Facing world-class athletes, the robot completed extended rallies and continually adjusted its actions according to its opponents’ shots, thinking about how to create scoring opportunities.
In doubles, it was also able to collaborate with human teammates in real time, coordinating tactics based on the positions of its teammates and the movements of its opponents. This transformed the robot from merely a “ball hitter” into a genuine participant in the dynamics of the match. Even when it lost its balance during high-speed movement and accidentally fell, it could autonomously stand back up and continue playing, demonstrating the robustness required for real-world environments.
Even more noteworthy is that robots do not experience the psychological fluctuations common among human athletes. Whether leading or trailing, the robot does not become anxious or impatient, nor do mistakes affect its subsequent decisions. Instead, it remains focused on how to play the next ball and win the next point.
What impressed viewers most under the live cameras, therefore, was that Galbot had already demonstrated the complete set of abilities expected of a true tennis player: it can hit different types of shots, move and make saves, choose strategies based on the match, collaborate with humans, and think about how to win a real contest.
From AlphaGo to AstraTennis: AI Begins Challenging Complex Competition in the Real World
Ten years ago, AlphaGo stepped onto the Go board, overcoming the challenge of how AI could confront extraordinarily complex games in the digital world and ultimately defeat one of humanity’s top Go players. Ten years later, AstraTennis stepped onto the court, facing a real world that is more dynamic, open, and unpredictable:
A tennis ball flying at high speed, constantly changing trajectories, opponents whose movements shift in real time, and a humanoid body that must simultaneously run, maintain balance, swing the racket, and make contact with the ball.
Ten years ago, AlphaGo stepped onto the Go board. Ten years later, AstraTennis caught the more difficult tennis ball.
The change taking place here is that AI is crossing the boundary from “digital intelligence” to “physical intelligence.” In the past, AI could understand information, reason, and make decisions on a screen. Today, it is beginning to have a body and bring that intelligence into the physical world, perceiving, judging, acting, and competing in an environment that is constantly changing.
AstraTennis therefore validates a much bigger proposition: Can AI truly enter the physical world and, as an embodied agent, perform complex movements or tasks?
Tennis is an exceptionally demanding test of this proposition. It requires a robot to simultaneously possess advanced perception, decision-making, motion control, and real-time competitive reasoning, and to integrate all these capabilities into a unified closed loop in the real world.
It also represents a genuine original technological innovation in embodied intelligence:
From the robot’s “brain,” which understands tasks and makes decisions, to its “cerebellum,” which controls the body in real time, Galbot has developed an independent technical path centered on the core challenges of embodied intelligence. It has brought these capabilities together and tested them in a task that is sufficiently complex and closely resembles genuine human intelligent activity.
This is the true value of the AstraTennis moment: physical AI developed independently in China has, for the first time, demonstrated a complete set of human-level intelligent capabilities in a high-speed, open, real-time competitive environment.
Combining the Brain and Cerebellum: The Technological Leap Behind AstraTennis
Tennis is a sport that pushes both “thinking” and “action” to their limits. A robot must understand the match and make decisions while also ensuring that its body executes those decisions precisely during high-speed competition.
Supporting all of this is Galbot’s independently developed embodied-intelligence foundation model, AstraBrain, which for the first time in the world integrates the brain, cerebellum, and neural control into a single system.
Within AstraBrain’s architecture, the “brain” is responsible for understanding and decision-making. The robot must determine the direction, speed, and landing point of the incoming ball; take into account the opponent’s position and the state of the match; decide how to hit the ball and where to place it; and determine how to organize its next offensive and defensive sequence.
In doubles and human-robot collaboration, it must also understand the states of teammates and opponents and make real-time tactical adjustments as the situation changes. In other words, it is responsible for thinking about “how to win the match.”
The “cerebellum,” meanwhile, solves the problem of translating these decisions into physical movements. A tennis shot takes place within an extremely short time window. The robot must maintain dynamic full-body balance while running at high speed, coordinating its legs, torso, arms, and even wrists to swing the racket with sufficient power and precision.
Driven by the cerebellum, the robot’s movements are no longer mechanical or fragmented command execution. Instead, they exhibit a more human-like, fluid, and highly dynamic athletic state.
But simply coordinating the brain and cerebellum is not enough to bring a robot to this level. The more fundamental question is: where does the robot learn these abilities?
This is where the deeper significance of Galbot’s technological breakthrough lies. AstraBrain can learn from imperfect human data and allow robots to learn from and reflect on human priors through tens of millions of iterations in a virtual tennis world. At the same time, multiple agents can compete and practice against one another, enabling the robot to achieve skill emergence and transfer those skills to the physical world.
Galbot’s core technology platform, “Galaxy Data,” serves two purposes. First, it can rapidly transform imperfect human data into large volumes of high-quality training data, allowing robots to learn efficiently from human experience. Second, it can continuously generate a wide range of high-quality data in the virtual tennis world, providing robots with a sustained and rich supply of training material.
Starting with the most basic returns, the robot gradually masters continuous rallies, baseline movement, shot placement, and other capabilities. It continually practices and competes against agents of different skill levels, accumulating far more skills in the virtual world than would be possible under real-world conditions.
In a sense, this is also the most intriguing technological parallel between AstraTennis and AlphaGo.
AlphaGo continuously improved its playing strength through extensive game data and self-play. Galbot, by contrast, has brought these capabilities into the physical world.
The robot is no longer facing static black and white stones on a Go board, but a physical world filled with high-speed motion and uncertainty. It must simultaneously solve problems involving perception, decision-making, motion control, and complex physical interaction.
When a robot can autonomously complete a complex athletic contest on a real court, what is truly being unlocked is the broader potential of embodied intelligence.
Today, the robot faces high-speed tennis balls, constantly changing opponents, and rapidly shifting match situations on the court. Once it moves beyond the court, AstraBrain will face a physical world that is more complex, more open, and much closer to the realities of everyday life.
At home, it will need to understand people’s needs and perform service tasks that continuously change;
In factories, it will need to adapt to different processes, environments, and forms of collaboration;
In hospitals, retail settings, and a growing range of public-service scenarios, it will likewise need to handle real-world situations that cannot be exhaustively anticipated and independently make judgments and take action as conditions change.
These tasks may appear entirely different from tennis, but they point to the same central question: when intelligence truly has a body, can it understand the world and continue acting within it?
AstraBrain offers an answer grounded in a real sporting arena.
Beginning with a tennis ball flying at high speed, AstraBrain has put perception, decision-making, action, and competitive reasoning through a concentrated real-world test. In the future, when these capabilities move beyond the court and into homes, factories, hospitals, and other real-world settings, embodied intelligence will transform far more than a single match.
