China's original innovation has hit the "crucial game-winning shot", and Galaxy Universal has forged the AstraTennis moment.

On August 22, at the National Speed Skating Oval "Ice Ribbon", a tennis ball crossed the net and flew towards the opposite side at high speed.

Standing on the other side was a humanoid robot.

It fixed its eyes on the incoming ball, moved its steps to adjust its position, then turned its body and swung its racket. The tennis ball was hit back to the half of the court where the human athlete was, starting a multi-shot rally between human and robot.

This is a scene at the opening ceremony of the 2nd World Humanoid Robot Games. Galaxy General Robotics competed on the same field with top tennis players in the world, completing autonomous running and hitting under the live broadcast lens of China Media Group. In the singles session, it needs to judge the ball path independently; after entering the doubles, it also collaborates with human teammates in real time, dynamically adjusts strategies according to the situation on the court, and even makes extreme saves in high-speed offense and defense.

This live broadcast of this large-scale event has achieved a milestone breakthrough on a global scale, fully presenting the fully autonomous tennis rally of humanoid robots. It is also pioneering to realize that the robot can perceive, make decisions and act autonomously on the real court, and carry out real-time confrontation and collaborative game with human athletes.

Ten years ago, in 2016, AlphaGo defeated Lee Sedol on the chessboard, making artificial intelligence enter the public view in a shocking way for the first time. Ten years later, Galaxy General Robotics autonomously returned a tennis ball from a world-famous player on the court, presenting the world with the AstraTennis moment of embodied intelligence.

And this time, this iconic moment was created by China, which has become the most hardcore declaration of China's original technological innovation.

1. Being similar in appearance is far from enough, tennis has become a hardcore big test for embodied intelligence

This year's Humanoid Robot Games is a grand event. According to the official event data, a total of 2056 robots participated in the conference. 51 events including running, martial arts, football and table tennis are all testing the sports limits of humanoid robots.

Tennis was not included in the official competition events of this Games, but presented in the form of an opening ceremony exhibition match.

This is not because this sport is too early for robots — on the contrary, it is because Galaxy General is currently the only player in the world that can bring fully autonomous tennis confrontation to the real competition court. The setting of the rules has to wait for enough participants to complete the match list. And Galaxy General chose to give the only opponent on this match list to a real human tennis player.

On the football field, robots have time to adjust their positions; the hitting area of table tennis is relatively fixed, which mainly tests the upper body. In contrast, tennis requires the robot to catch up with the high-speed incoming ball while running all over the court, keep balance and swing the racket accurately, and adjust the strategy in real time according to the opponent's changes, which puts almost harsh requirements on the whole-body control and autonomous decision-making.

In other words, humanoid robots have been able to complete a whole tennis game in a way close to that of human athletes, and what the audience can intuitively feel most on the court is the full sense of realism it shows.

When an incoming ball is pressed to the corner, it will not stand in place and swing hastily, but move laterally first, then use a series of small steps to find the hitting position. After the body's center of gravity stabilizes, the trunk turns accordingly, and the racket swings towards the incoming ball. Facing a more tricky landing point, it will even run across most of the court to complete the save in a nearly unbalanced state. Serving, receiving serve and switching between forehand and backhand are all connected into a set of natural movements.

"Being like a human" is only the first step, and more importantly, "understanding tennis".

During the baseline rally, the robot will control the landing point of the return ball according to the position of the opponent; when an opportunity ball appears, it can move to the net to change the hitting method. In singles, it independently organizes the offense and defense of each shot; after entering doubles, it can observe the position of its teammates, judge who should handle the ball, and fill the gap in the court in time.

What's more interesting is that robots do not have psychological fluctuations like human athletes. No matter whether the score is leading or falling behind, it remains unhurried, and will not affect subsequent judgments due to mistakes, but always focuses on how to play the next ball and how to win the next point.

The live broadcast lens further magnifies the gold content of this technology. Demonstrations in the laboratory can be shot repeatedly, but the incoming balls on the real court will not strictly follow the training script. When facing world-class athletes, the robot has to give a quick response to every shot.

This is the AstraBrain Latent technology released by Galaxy General Robotics in March this year, which first demonstrated the high-dynamic, long-term tennis rally between humanoid robots and real humans. This video once made the overseas tech circle doubt their own eyes. Tesla CEO Elon Musk also responded on the social platform, and Andrej Karpathy, former head of AI at Tesla, said that when he first saw the video, he even suspected that the picture was generated by AI.

Today, on the global event live broadcast, Galaxy General has responded to all doubts with an excellent performance.

2. Big brain and cerebellum collaborate, the technological leap behind AstraTennis

Anyone who has played tennis knows that tennis is a sport that pushes both thinking and action to the limit, and the same is true for robots: it must not only understand the game and make decisions, but also make its body execute accurately in high-speed confrontation.

Supporting all this is the self-developed embodied intelligence large model of Galaxy General, AstraBrain, which is also the world's first embodied large model that integrates brain, cerebellum and neural control.

In the system of AstraBrain, the "big brain" is responsible for understanding and decision-making. It observes the incoming ball and understands the game situation, decides what kind of return strategy to adopt, and thinks about how to win the game. The "cerebellum" solves the problem of how to coordinate the whole body within milliseconds to make the robot complete movement, balance and swing.

Tennis hitting occurs in a very short time window. The robot needs to complete the whole-body dynamic balance while running at high speed, so that the legs, trunk, arms and even wrists form a high degree of coordination to complete the swing with sufficient explosive force and accuracy. Driven by the cerebellum, the robot's movements are no longer mechanical and fragmented instruction execution, but can present a more anthropomorphic, smooth and highly dynamic motion state.

However, only the collaboration of big brain and cerebellum is not enough to make the robot reach such a level. The more critical question is: Where on earth did the robot learn these capabilities?

When a person learns tennis, he usually practices swinging and footwork first, and then forms a sense of the ball through a lot of rallies, but it is difficult for robots to obtain a sufficiently complete set of learning materials.

One way is to record professional athletes playing, but this requires high-precision motion capture in a large-scale venue. When swinging the racket quickly, the subtle changes of the wrist and racket surface must also be recorded accurately. Even if complete human movements are collected, there are still differences between the robot's joint structure and motion range and that of humans, making it difficult to replicate directly.

The special feature of AstraBrain is that it can learn from imperfect human data, and allow the robot to learn and reflect through tens of millions of human priors in the virtual tennis world, and conduct multi-agent game and sparring at the same time, so that the robot can achieve the emergence of skills and migrate to the physical world.

After these "imperfect" data enter Galaxy General's core technology platform "AstraWorkshop", they will continue to be amplified in the virtual world. The platform can convert scattered human movements into data that robots can learn, and continuously generate training tasks under different ball speeds, landing points and opponent conditions.

According to Wang He, founder and CTO of Galaxy General, the training time accumulated by robots in the virtual tennis world is equivalent to humans playing tennis continuously day and night for decades.

What's more interesting is that multiple virtual robots can also spar with each other. They try different running routes and hitting strategies in a large number of rounds, and continue to adjust according to the results. The robot thus gradually masters continuous hitting, baseline mobilization and landing point control, and also begins to learn how to create scoring opportunities in the game.

In other words, human data tells the robot what the movement should look like roughly, and reinforcement learning helps it find an implementation method suitable for its own body.

3. AI begins to challenge the complex game in the real world

In a sense, AstraTennis and AlphaGo form a technical echo spanning ten years, and also reflect the difference between the two generations of artificial intelligence paradigms.

AlphaGo relies on massive chess games and self-game to complete strategic game in a closed, deterministic digital chessboard space, proving that AI can beat top human players in a world with clear rules, opening the era of deep game AI. But within the chessboard, everything is limited to symbols and deduction; while AstraTennis pushes the game intelligence from the virtual digital space into the real physical world full of noise, disturbance and uncertainty.

Robots no longer only do deduction on paper, but must simultaneously complete environment perception, situation decision-making, whole-body dynamic control and complex physical interaction. Intelligence no longer stops at thinking, but needs to complete the closed loop of "cognitive decision-making + whole-body execution". This double test of highly bound thinking and body seems to be the basic requirement of sports scenarios, but it is the core barrier that has long been stuck in front of the industrialization implementation of embodied intelligence, and it is also a technical problem that the global industry has always been difficult to break through.

If we want robots to truly adapt to the real physical world, we must not only overcome the algorithm problem of intelligent decision-making, but also solve the implementation pain points of precise body execution and dynamic environment adaptation, which also makes the industry trapped in various realistic bottlenecks for a long time: high-quality physical interaction data is scarce, complete human movements cannot be directly replicated to the robot body, and a large number of solutions still stay in scripted demonstrations in the laboratory, which is difficult to withstand the real test of public live broadcast.

Aiming at the implementation dilemma of the whole industry, Galaxy General's technical path gives a key breakthrough idea: enlightenment relying on imperfect human prior data, completing skill emergence through tens of millions of multi-agent games on the virtual tennis court, and then completing the migration to reality; the integrated architecture of big brain and cerebellum opens up the complete link from high-level cognitive decision-making to millisecond-level whole-body motion control, fundamentally solving the industry pain points of disconnection between decision-making and execution and poor adaptability to real scenarios.

When a humanoid robot can independently complete a complete set of tennis games in a large-scale event live broadcast, its value has long gone beyond the act of playing tennis itself. This technical base tempered by high-dynamic confrontation has general migration potential: a large number of sample learning reduces the data threshold of complex tasks; self-game learns the logic of collaboration and confrontation; virtual training allows robots to accumulate massive experience before landing in reality; the collaboration of big brain and cerebellum makes intelligence truly transformed into stable and reusable behaviors. This means that a large number of unstructured tasks that highly rely on human experience, such as home assistance, industrial operation, and medical assistance, are entering the solvable territory of embodied intelligence.

Objectively speaking, compared with professional athletes, robots still have room for improvement in hitting explosive power and tactical diversity. However, realizing continuous autonomous confrontation with human athletes on an unscripted public court has marked that global humanoid robots have crossed a key technical inflection point: embodied intelligence is no longer just a single-point demonstration, and begins to have the ability to continuously complete long-time-sequence complex tasks in real, uncontrollable scenarios.

Looking at the global track, most teams still follow the splicing route of large model plus off-the-shelf controller, while Galaxy General chooses to start from the bottom of embodied intelligence, fully build a native technical chain from "big brain" task understanding and game decision-making to "cerebellum" whole-body real-time motion control, and choose tennis, a real confrontation scenario with extremely high difficulty, to complete centralized verification.

This is the industry weight of the "AstraTennis moment". It is not only a single-point technological achievement of an enterprise, but also represents that the Chinese team has embarked on an independent exploration path of original innovation on the next-generation AI core track of embodied intelligence.

The highlight moments of AI in the past ten years mostly came from overseas laboratories, but today, local Chinese teams are producing original breakthroughs with global discussion attention in the field of physical intelligence, pushing the capability boundary of AI from the closed chessboard to a broader and more complex real world.

AstraTennis has thus become an embodied intelligence milestone presented by China's original technology and witnessed by the whole world.