Watching China’s humanoid robots lately,
it’s easy to get the impression:
Robots are almost catching up to humans.
They can run.
They can punch.
They can dance.
Some can even perform very complex full-body movements.
After watching these videos,
the natural next thought is usually:
“If they can already do these things, shouldn’t factory tasks like screwing bolts or moving parts be much easier?”
But the real answer is the opposite.
A recent Reuters investigation found
China’s biggest current issue is no longer:
“Can we even build humanoid robots?”
But rather:
“Once built, can they work reliably every day?”
The gap in difficulty between these two questions is huge.
China Is Already Skilled at "Making" Humanoid Robots
Let’s first look at the hardware.
China’s speed in the humanoid robot supply chain
is actually remarkable.
According to Reuters citing BofA Global Research data,
around 20,000 humanoid robots will ship worldwide in 2025,
with about 95% made by Chinese manufacturers.
Officials from China’s Ministry of Industry and Information Technology said in July this year
China expects to produce over 100,000 humanoid robots in 2026.
There are currently more than 150 humanoid robot companies in China.
So the real issue today is no longer:
Can China make them or not?
But rather:
What exactly are they going to do with such large quantities?
Running 100 Meters vs. Working in a Factory Are Two Completely Different Skills
Making a humanoid robot run
is definitely hard.
It must handle:
balance,
center of gravity,
joint coordination,
speed,
and foot placement.
These require strong engineering capabilities.
But running has one advantage:
The task is very clear-cut.
The direction is forward.
The track is fixed.
The finish line is fixed.
The movements can be repeatedly trained.
Factories, however, don’t work this way.
Suppose a robot’s job is simply:
pick a part from the left,
place it into a fixture on the right.
This sounds much simpler than running 100 meters.
But next time,
the part might be off by 2 centimeters.
Then the part might be flipped around.
Another part might be stuck under the previous one.
The box could have shifted slightly.
The lighting might have changed.
The first grasp attempt might have failed.
Someone might suddenly appear nearby.
At this point,
the real problem is no longer:
“Can the robot arm move?”
It becomes:
“The original method won’t work—what do you do next?”
Factories Care More About Reliability Than a Single Impressive Performance
This is the biggest difference between a demo and factory use.
A robot performance,
if successfully recorded once,
can become a very impressive video.
But factories don’t evaluate things that way.
Imagine needing to perform the same task 5,000 times a day.
Succeeding 4,900 times sounds like a 98% success rate,
which seems impressive.
But that also means:
Someone needs to fix the robot 100 times daily.
If every stoppage requires an engineer to:
reposition the robot,
take out parts,
and restart the process,
the robot might not be saving labor costs,
but creating more work.
So factories usually focus on:
stability,
uptime,
speed,
failure rate,
maintenance time,
and cost per operation.
This is why a seemingly ordinary traditional robotic arm
can be more valuable than a humanoid robot that can run.
Because Traditional Robotic Arms Don’t Need to "Be Like Humans"
Walk into a mature automotive factory,
and you’ll see lots of robots.
But they’re rarely humanoid with two legs and two arms.
They’re usually mechanical arms fixed to the ground.
Why?
Because if the job is:
to weld the same steel plate in the same spot every day,
there’s no need to:
walk,
turn around,
maintain balance,
or understand the entire factory.
It just needs to do one thing:
fast, accurate, and stable work.
Industry insiders interviewed by Reuters point out that
in Chinese car factories,
traditional industrial robots still outperform humanoid robots in many core tasks,
being faster
and easier to control.
For example, Xiaomi’s Beijing electric vehicle factory said
overall assembly processes have reached 91% automation,
using over 700 industrial robots just in the body shop,
handling spot welding, riveting, and bonding.
Xiaomi is also training humanoid robots
to try tasks like installing nuts.
This clearly shows:
Factories are not waiting for humanoid robots to "start automation."
They are already highly automated.
Humanoid robots must now prove:
“What jobs can they do better than traditional automation?”
The Last One Centimeter Is the Real Challenge
A researcher at Beijing General Artificial Intelligence Institute
has a great way to describe the current problem.
Robots used to struggle with the last 10 centimeters.
Now they’ve improved to the last 1 centimeter,
or even
the last 1 millimeter.
What does this mean?
Imagine plugging a charging cable into a socket.
A human just glances,
reaches out,
adjusts the angle slightly,
and plugs it in.
For a robot,
it needs to know:
the exact socket location,
the plug’s current angle,
the distance from its hand,
the applied force,
and if it fails the first time,
did it miss left or right?
Was the angle off?
Is the socket blocked?
Humans naturally feel and adjust.
Robots must convert these subtle movements,
which we don’t even notice, into
sensing,
judgment,
actions,
and confirmation.
So often,
the hardest thing for a robot isn't running 100 meters,
but precisely placing a tiny part into a small hole.
Visiting a Training Center Exposes Just How Difficult the Problem Is
Reuters recently visited a robot training center in southern China.
There were over 100 humanoid robots there.
Human trainers wore devices
to teach robots through their own movements how to:
sort boxes,
pack noodles,
and brew coffee.
This sounds close to real-world tasks.
But the training process is tough.
Reuters reported that a novice trainer
must attempt about 300 times
before producing an action dataset that the robot can truly use.
Experienced trainers may
produce usable data every 50 tries.
This reveals a major bottleneck in Physical AI:
ChatGPT can consume huge amounts of online text, but humanoid robots lack a "real-world internet" of data to train on.
ChatGPT Has Online Data; Humanoid Robots Lack “Motion Data”
Why have large language models improved so rapidly?
Because humans have already accumulated vast amounts of:
books,
websites,
documents,
code,
and conversations.
These models learn how language is typically arranged from this data.
But to teach a robot:
how to pick up a cup,
how to open a drawer,
how to grasp after a box has shifted,
how to detect a slipped grip,
how to know if something is too heavy,
there simply isn’t much data online.
And video footage isn’t the same as usable robot action data.
Robots also need to know:
how joints move,
how much force to apply,
where contact occurs,
and whether a grasp succeeded.
So many humanoid robot companies today aren’t
trying to replace workers immediately.
Instead, they’re:
letting many robots work, fail, get taught, and slowly accumulate real-world physical data.
This Explains Why China Wants to Mass Produce Robots First
Here is an interesting logic.
To make robots smarter,
more real motion data is needed.
To get more data,
more robots must operate in real environments.
So China’s current strategy is, to some extent:
scale up hardware,
build training centers,
and deploy robots into factories,
stores,
logistics,
and pharmacies,
then collect data from these real-world uses.
Reuters reported that China’s Ministry of Industry and Information Technology and state-owned asset regulators asked 10 provinces this year
to each identify at least 20 real-world training sites for humanoid robots and AI.
So those machines that seem “not smart enough yet”
are, to some degree, data collection tools for the next generation of AI.
But This Also Risks Producing Robots Faster Than the Market Needs
China used a similar approach in the past for:
electric vehicles,
and solar energy.
Government support,
capacity building,
competitive companies,
cost reductions,
and strong supply chains followed.
But humanoid robots differ.
Before large-scale electric vehicle and solar expansion in China,
product usability was already fairly certain.
Humanoid robots are still at a stage where:
the real applications haven’t fully matured.
Reuters found that in the first half of 2026,
Chinese government spending on humanoid robots, training equipment, and related projects has reached approximately $230 million.
This means some of the demand is government-created.
If factories,
stores,
and logistics companies
find humanoid robots’ work isn’t yet valuable enough to buy in bulk,
we could see:
capacity outpacing demand.
The Real “Robot Bubble” Isn’t That Robots Have No Future
We need to distinguish here.
Current problems don’t mean humanoid robots won’t succeed in the long run.
Think back to early electric vehicles:
short range,
high cost,
few charging stations.
You wouldn’t conclude:
electric vehicles will never go mainstream.
The real question is:
Are today's capabilities mature enough to justify current prices and investment?
Some entrepreneurs and investors interviewed by Reuters
expect China’s humanoid robot industry to start consolidating between late 2026 and 2027.
Ultimately, the survivors will be:
those that can really reduce costs,
those with actual customers,
those that accumulate data,
and those that can make robots work daily, not just perform on stage.
Some Tasks Are Already Suitable for Humanoid Robots
Not all commercial applications are absent.
For example, Chinese robot company Galbot
has deployed robots to pharmacies in several cities.
They can pick items from shelves
based on digital orders.
These environments have several features:
relatively fixed spaces,
manageable item locations,
highly repetitive work,
and clearly defined tasks despite variety of products.
Galbot says
its system’s picking success rate exceeds 95%.
Interestingly,
these robots don’t necessarily insist on two legs.
They may use wheeled bases.
Because after commercialization,
customers usually don’t care:
“Does it look like a human?”
They care about:
“Does it work well?”
So the Real Future of Humanoid Robots May Not Be “Humanoid Everywhere”
This is probably the most overlooked point today.
We like humanoid robots mainly because
they’re easy to understand.
Two arms,
two legs,
one head.
It’s obvious:
“They can replace humans someday.”
But in real commercial settings,
robots may take many shapes.
If they must climb stairs,
two legs may be valuable.
If they carry goods for long periods,
wheels could be more efficient.
If they do welding,
a fixed robotic arm might be best.
If they must handle many different tools,
humanoid arms and bodies may have advantages.
The real question shouldn’t be:
“When will humanoid robots replace workers?”
But rather:
“Which jobs truly need a robot that looks and acts like a human?”
The True Test for Physical AI Is Just Beginning
Over the past two years,
we’ve seen many breakthroughs in Physical AI.
Robots can run,
jump,
and engage in combat.
Drones can avoid obstacles independently.
Cars can understand roads.
These are important milestones.
Because they prove at least that
machines are gaining the ability to perceive and manipulate the real world.
But the next phase will change how we measure success completely.
No longer will people ask:
“What was the most impressive feat?”
Instead, questions will be:
How many times can it perform a task daily?
How often does it fail?
Can it recover by itself from exceptions?
How many manual interventions are needed each day?
How much does maintenance cost?
How long does the battery last?
Compared to hiring a human,
is it cost-effective?
At this stage,
the most impressive robot
won’t necessarily be the fastest runner.
It may be:
the one you see doing the same boring task every day without even noticing it’s there anymore.
This is real industrial automation.
Demos Show "It Can Be Done," Factories Demand "It Can Be Done Every Day"
So when we see humanoid robots:
running 100 meters,
punching,
and dancing,
it’s impressive.
Because it truly represents rapid progress in hardware and control.
But don’t translate that directly as:
“Workers will soon all be replaced.”
There’s a very difficult path in between.
This path includes:
reliability,
fine manipulation,
exception handling,
training data,
cost,
and
real commercial value.
When AI moves from the screen into the real world,
this is where the biggest differences lie.
A chatbot making a wrong answer,
you can just ask again.
A robot grabbing the wrong part,
the whole production line may have to stop.
The day Physical AI truly matures may not be when we see robots perform more astounding backflips,
but when a factory manager suddenly realizes:
This robot has been working continuously for a long time, and no one needs to constantly monitor it.
That is the real breakthrough.
Today, keep advancing together with AI.
Learn one AI skill every day.
Save a little time each day.
Improve a bit more every day.
SasaDaily, growing with you.
