There is something slightly strange about watching a humanoid robot run after a football.
It is not strange because robots are playing football. We have been imagining machines doing things like this for decades. What feels strange is that we are now watching engineers deliberately put robots into situations where they can fall over, miss the ball, bump into each other, make the wrong decision and sometimes look completely lost.
If the objective were simply to show that a robot can walk, there would be much easier ways of doing it.
You could put it on a flat floor, give it a carefully prepared route and ask it to walk from one point to another. The robot would probably look impressive. But that demonstration would tell us relatively little about whether the machine is actually becoming useful.
A football field is a very different place. The ball doesn't wait for the robot. Other players don't wait either. The robot has to look around, understand what is happening, decide what it wants to do and then move its body quickly enough to make that decision matter. And if something unexpected happens, there is no human standing beside it telling it what to do next.
That is what makes these competitions interesting.
The football, tennis, table tennis and running events involving humanoid robots may look like entertainment from the outside, but there is a serious engineering problem hiding underneath all of them: can a machine learn to deal with a world that refuses to behave exactly as expected?
A factory is easy compared with a football field
Modern industrial robots are already extremely good at certain jobs.
If a robotic arm has to pick up the same component from the same location thousands of times, engineers can create an environment where almost everything is predictable. The robot doesn't need to wonder where the object might be. It doesn't need to negotiate with another robot that suddenly changes direction. It doesn't need to understand a human walking unpredictably through its workspace.
The environment has been designed around the machine.
The real world doesn't work that way.
A person might leave a box in the wrong place. A door might be half open. Someone might walk directly in front of a robot. An object might fall. The floor might be uneven. Another machine might suddenly stop working. A robot that is useful only when everything happens exactly as expected is not particularly useful outside a controlled environment.
Sports create a small version of this problem.
A tennis ball doesn't care where the robot expected it to go. Neither does a football.
That forces the machine to continuously update its understanding of the situation. It has to see something, interpret it, predict what may happen next and then decide how its body should respond.
Humans do this so naturally that we rarely notice how complicated it actually is.
Watch someone return a tennis serve. They don't calculate the ball's trajectory consciously. They don't stop to identify the ball, measure its velocity and then separately calculate where their arm should move. Their brain combines all of that information almost instantly.
For a robot, reproducing even a simplified version of this ability is a major challenge. And that is why a game can sometimes be a better test than a carefully staged demonstration.
The ball is only part of the problem
There is another reason sports are useful for robotics research, and it has less to do with physical movement than most people realise.
A robot needs to understand movement. Seeing a ball is not enough.
If a camera tells the robot that the ball is currently two metres away, that information becomes almost useless if the robot doesn't understand where the ball is going. The machine needs to estimate its speed, direction and likely trajectory, and it needs to do that while its own body is moving.
This becomes particularly obvious in table tennis.
The ball is small, fast and constantly changing direction. A successful return requires much more than recognising a white object against a table. The robot has to understand the motion of the ball and position itself accordingly. That may sound like a very narrow skill, but the underlying ability is much broader.
A robot working in a warehouse also needs to understand moving objects. A robot assisting people at home needs to anticipate human movement. A machine operating in a disaster zone needs to deal with objects and people whose positions are constantly changing.
The sport is different. The underlying problem is surprisingly similar.
This is one reason I think the most interesting development in robot sports isn't necessarily that robots are becoming better athletes. It is that they are gradually being pushed from recognising the world to interpreting the world. Those are two very different things.
Football is where things get complicated
A single robot playing against a ball is already difficult enough. Put several robots on a field and the problem changes completely.
Now the machine doesn't just need to know where the ball is. It needs to know where its teammates are, where its opponents are, what everyone else appears to be doing and what it should do in response.
Imagine a robot standing near the goal while another teammate has possession of the ball.
The obvious instruction—"go toward the ball"—may actually be the wrong one.
The robot might be more useful if it moves into an open position. Perhaps its teammate is about to pass. Perhaps an opponent is approaching. Perhaps another robot has already decided to move toward the ball.
The machine needs some understanding of the larger situation rather than simply reacting to the nearest object.
That is where robot football becomes more interesting than it looks.
It starts raising questions about cooperation.
If humans are eventually going to work alongside groups of robots in warehouses, factories, hospitals or emergency situations, those machines cannot all behave like independent individuals. They will have to divide tasks, communicate, recognise what others are doing and adjust when plans change.
Football is obviously not a perfect simulation of those environments.
But it is a convenient one.
The rules are simple enough to understand, while the environment is unpredictable enough to expose weaknesses in a robot's decision-making.
A machine can be brilliant at walking in a straight line and still have no idea what to do when three moving objects suddenly appear in front of it.
A football field finds that weakness very quickly.
This is also why the dancing robot criticism is a little complicated
There is a common reaction whenever videos of humanoid robots dancing, singing or performing some unusual stunt appear online.
People ask a reasonable question: What is the point?
If the world needs better robots for factories, logistics, construction and elderly care, why spend so much time making robots dance?
I don't think there is a single answer to that. Some of these demonstrations are clearly designed to attract attention. There is nothing wrong with that. Robotics companies need investors, customers and public interest, and a robot performing a dance is considerably easier to turn into a viral video than a robot spending eight hours moving components around a warehouse.
But entertainment can also become an early market.
That part is easy to overlook.
A technology doesn't always become cheaper because engineers suddenly discover a brilliant new invention. Sometimes it becomes cheaper because more people start buying it, companies start producing it at larger volumes and manufacturers become better at making the individual components.
Humanoid robots are full of expensive components.
Motors, joints, sensors, batteries, actuators and computing hardware all have to work together. If only a few hundred machines are being produced, many of those parts remain expensive.
If the industry eventually reaches hundreds of thousands or millions of machines, the economics could look very different.
So a robot that entertains people today might not be commercially irrelevant simply because it isn't working in a factory.
It may be helping create the market and manufacturing experience that eventually makes more serious applications affordable.
Of course, that doesn't mean every dancing robot is secretly a breakthrough.
Sometimes a dancing robot is just a dancing robot.
The important point is that we shouldn't assume the entertainment market and the industrial market have nothing to do with each other.
China has a particularly interesting position
This is where the discussion becomes bigger than robot competitions themselves.
China has spent decades building one of the world's largest manufacturing ecosystems. It has enormous supply chains, large industrial markets and a huge number of companies capable of producing mechanical and electronic components at scale.
For humanoid robotics, that matters.
Building an impressive prototype is one thing. Building thousands of reliable machines is another. Building millions of them at a price that businesses can actually afford is an entirely different challenge.
The companies that eventually succeed may not simply be the ones with the most impressive robot videos.
They may be the ones that figure out how to manufacture the machines efficiently, maintain them, supply replacement parts, standardise components and integrate them into existing workplaces.
China's manufacturing infrastructure gives its robotics industry an obvious advantage in that particular part of the race.
But there is no reason to pretend the race is already finished.
Humanoid robotics still has some very basic problems.
Battery life remains an issue. Robots can be expensive. Software has to become much more reliable. Machines need to operate safely around people. And perhaps most importantly, engineers still have to figure out how to make a robot useful for many different tasks rather than extremely good at one carefully designed demonstration.
That last problem may turn out to be the hardest one.
The million-robot question
There is a simple idea in manufacturing that becomes extremely important when discussing humanoid robots: scale changes economics.
Making one complicated machine by hand is expensive.
Making a million machines using standardised components, automated production and a mature supply chain is a completely different proposition.
The first version might cost a fortune. The millionth version could cost a fraction of that. But getting from one to a million is not simply a matter of pressing a button.
Companies need reliable designs. Suppliers need to produce compatible parts. Engineers need standards. Software needs to work across large numbers of machines. Batteries need to become more efficient. Repairs need to be straightforward.
This is another reason robot competitions could matter more than they initially appear to.
Competitions create environments where different companies and research teams are forced to solve similar problems under similar rules.
How large should a robot be?
How should it communicate?
How should batteries be designed?
What happens when a robot falls?
How should performance be measured?
These questions may sound boring compared with watching a humanoid score a goal, but standards are exactly the sort of boring infrastructure that can determine whether an industry grows slowly or rapidly.
The smartphone industry didn't become enormous simply because someone built a phone with a touchscreen.
An entire ecosystem had to develop around it. Humanoid robotics is going to need something similar.
The hardest part begins when the stadium disappears
This is probably the part of robot sports that interests me the most. A football field has rules. A tennis court has boundaries. A race has a starting point and a finish line. The real world has none of these.
A robot working in someone's home won't know exactly what will happen five seconds from now. A robot in a factory may encounter a problem that engineers never anticipated. A machine sent into a dangerous environment may have to make decisions without a human controlling every movement.
That is the real test.
Can a robot take the abilities it has learned in controlled environments and use them somewhere messy, unpredictable and unfamiliar?
We are not there yet.