The world’s first AI robot football match shows how far we still have to go
At first glance, it looked like a slow-motion parody of the world’s most beloved sport. On a small indoor pitch in Beijing, teams of humanoid robots wobbled across the turf, scanning the field with glowing blue eyes, tripping over their own feet, and occasionally collapsing in awkward heaps. At one point, two machines had to be stretchered off after failing to stand up again.
Yet for the engineers behind the event, this wasn’t a failure – it was a breakthrough.
Staged in June 2025, China’s first fully autonomous robot football match marked a milestone in the development of embodied artificial intelligence. Each of the robots, including several built on the well-known Nao platform, played without any remote control or manual input.
They made their own decisions based on what they could see, hear, and compute. In short: they played football – on their own.
But what the match also revealed, more clearly than ever, is just how hard it is to build a robot that can actually move like a footballer. Despite dramatic advances in artificial intelligence, physical movement remains one of the most difficult frontiers in robotics.
The complexity of movement
It’s easy to underestimate how difficult walking is—until you try to program a machine to do it. Human movement, especially athletic movement like running or kicking, involves a staggering level of complexity.
A robot must process visual information from cameras, understand its own position using inertial sensors, predict the consequences of its motion, and then make rapid adjustments to stay balanced and stable. All this needs to happen in real time, in a changing environment, with no external assistance.
“It’s not just a computing challenge,” says Cheng Hao, CEO of Booster Robots, one of the teams that participated in the Beijing tournament. “It’s a motor skills challenge. Even when the AI makes the right decision, the body might not be able to execute it.”
And this is where robot football hits its limits. A robot may ‘see’ an open teammate and decide to pass – but if it can’t lift its foot properly or falls in the process, the decision is meaningless.
The actuator bottleneck
The real-world limits of robot motion are often rooted in hardware. Most humanoid robots today use electric servo motors to move their joints. These are precise, but limited. They struggle with speed, force, and the kind of dynamic flexibility humans use to stay upright under pressure.
“Current actuators aren’t good enough to replicate human-level motion,” says Dou Jing, organizer of the World Humanoid Robot Games, which will take place in Beijing this August.
“If we want robots to run, pivot, or recover from a fall like a person, we may need entirely new types of actuators – perhaps inspired by biology.”
Alternative technologies exist – such as hydraulic systems or variable-stiffness actuators – but they tend to be bulky, expensive, or fragile. Soft robotics offers promise, but lacks the torque needed for high-intensity movement.
Without better “muscles”, robot footballers will always be at a disadvantage, no matter how smart their software gets.
The promise of AI – and its limits
On the software side, progress is happening fast. AI models can now learn from thousands of football matches, real or simulated, and develop strategies for passing, positioning, and shooting.
Reinforcement learning has enabled robots to teach themselves behaviours through trial and error, without relying on human programming.
But the gap between simulation and reality (the so-called “sim-to-real” problem) remains a major obstacle. What works in a virtual environment often fails in the messy, unpredictable physical world.
This is especially true for embodied systems, like humanoid robots. Small differences in friction, lighting, or joint response can send carefully trained agents tumbling.
“AI can help with decision-making, but it doesn’t magically solve the physical coordination problem,” says Dr Xiaolin Wu, a robotics researcher at the Chinese Academy of Sciences. “You still need the right mechanical design, and that’s often the harder part.”
In short: an AI might know what to do. Getting its robot body to actually do it is another story.
Meet the Nao robot
Perhaps no robot better illustrates this challenge than the Nao, a 58-cm tall humanoid widely used in research, education, and, most famously, robot football.
Originally developed by Aldebaran Robotics in France, the Nao became the official platform for the RoboCup Standard Platform League in 2008.
Since then, it has starred in countless university tournaments around the world. Its compact frame, sensor suite, and open programming interface made it an ideal testbed for robot learning.
But the Nao’s limitations are just as well known. Its movements are slow, its balance is tenuous, and its ability to recover from falls is poor. In the Beijing tournament, several Nao-style robots struggled to stay upright for more than a few seconds during active play.
In a twist of irony, the company that created Nao – later acquired by SoftBank and renamed SoftBank Robotics Europe – was placed into administration in 2022. Though Nao robots continue to be used, their future as a commercial product remains uncertain.
Still, for researchers, the Nao remains a symbol of how far we’ve come – and how far we still have to go.
What would it take?
So, can robots really play football?
The answer, at least for now, is: sort of.
They can follow the rules, scan the field, and make decisions using AI. They can walk – sometimes. They can kick, though not very hard. And they can lose their balance and crash to the ground with surprising comic timing.
What they can’t do yet is play the game with anything close to human speed, agility, or resilience. To get there, experts say, we’ll need:
- Better actuators that mimic biological muscles
- More robust hardware that can handle impact and recover from falls
- Smarter integration between AI decision-making and real-time control
- Richer simulations and improved sim-to-real transfer
- Possibly, entirely new robot designs tailored to athletic performance
Some researchers have proposed training robot athletes the same way we train humans – by drilling core motor skills, using game footage for strategy, and developing “muscle memory” through repetition.
Others suggest skipping humanoid forms entirely and designing new bodies better suited to robotic sport.
The long game
Robot football isn’t just a novelty. It’s a serious research field with major implications. The same challenges that plague a robot footballer – balance, coordination, adaptability – are the ones we’ll need to solve to build robots that can navigate disaster zones, assist in elderly care, or work in factories alongside humans.
That’s why tournaments like the one in Beijing matter. They bring these problems out of the lab and into public view, forcing designers to contend with real-world unpredictability.
For now, robots remain clumsy athletes. But step by step – and fall by fall – they’re learning to play the game.
And someday, they might even score.
