Humanoid Robots and the Gap Between Hype and Reality | Bloomberg Primer

Technology Lens
17 followers
#Robots #AI #Technology
Humanoid robots that use AI are moving from viral videos to real-world work. From artificial intelligence training and data gaps to factory trials and global competition, companies are racing to turn “physical AI” into reality. But with billions of dollars being invested, the bigger question remains: will humanoids ever deliver enough real-world value to justify the hype?
0:00 Intro to Robot Hype
2:02 A CEO and His Robot
4:45 The Money Backing Humanoids
7:08 The Robot Data Gap
9:44 How to Train your Robot
12:04 Training Neo
13:59 The World Humanoid Robot Games
15:10 China’s Humanoid Advantage
17:25 Get a Job Robots
19:26 Testing Humanoids in Warehouses
21:49 Conclusion
23:31 Credits
This hand, this knee, and this head are all small feats of engineering, decades in the making,
and together they're about to take a massive leap forward. Oh dear.
Humanoid robots have exploded out of the lab and into the real world of late, not always on the
best footing. But I'm not just here to show you funny robot videos for the next half hour,
because behind silly hats and dancing is a very serious, very expensive, and very unproven bet
that AI-powered robots could one day transform labor as we know it. The interest in humanoid
robotics has been there for a long, long time. What's changed, of course, is now we see that
it's viable to commercialize this and start really applying some good business models.
AI clearly plays a large part in this. The hope is in cracking the code of physical AI,
taking specialized artificial intelligence and putting it in a body with sensors,
batteries, and actuators, very expensive and technologically complex actuators. It's a paradigm
shift that could set humanoids up to become a multi-trillion dollar market by 2050. But for
these massive figures to be realized, it won't be about party tricks or dance routines. It'll be
about work in homes, in hospitals, in factories, especially factories. When you really work with
robots, you realize how challenging it is. There is a lot of what I sometimes call humanoid hype
out there. Like a lot of hype coming from the most influential figures in the tech industry.
But hype doesn't build a supply chain or rewrite the global labor market.
To do that, you need, well, let's get into it, shall we? This robot runs on AI.
Well, sometimes. For this demonstration, filmed in 2025, Neo needed some help from
real humans and was operated remotely by a man named Turing. No, not that Turing.
But don't worry, we'll get to that later. You're looking beautiful today, Neo. Did you hit your head?
You okay? This particular humanoid is made by a company called One X, a robotic startup founded
by this actual human, Bernd Borne. So this is Neo. This is the gamma version. So this is what we've
been doing the in-home tests among employees with. This is the one that's doing my laundry.
And this is the last, let's call it like the last engineering version. The version that we're
working on right now that's about to go into production is the unit that will go into the
first customer homes. Essentially, the entire behavior of Neo is AI in the sense that it takes
images, forces, tactile, and in general, all the information that it has about the world.
In, there's a big neural network, and out comes the actual motor commands all the way down to
forces and torques that it's exerting to be able to do everything from balance and walk
to doing manipulation. Unlike a traditional robot that's programmed to do a specific task
the same way every time, a robot that uses AI should, in theory anyways, learn to do things
over time in unpredictable environments. It's sometimes called physical AI, a category that
besides humanoids includes other general intelligence robots and driverless cars.
Basically, it's anything that uses motors, sensors, cameras, and AI models to make sense
of the world around it. We are seeing a very clear correlation between the amount of data
and how many different tasks we can do and how successful we are at these tasks. And that is
really the beauty of how AI has changed robotics because there is no longer this limit where you
need to write some specific code for every task. Everything is just data. Let me introduce you to
our robot. One X is just one of many humanoid robot companies that have burst onto the tech
scene in recent years. Mind taking this wet towel from me? Sure, I'll get the laundry started.
The frenzy comes down to two little letters, which come with big checks. AI.
The past few years we've seen explosive interest in development of and investment in AI, and a lot
people are seizing on this moment as a way to work on all kinds of related things, including
robots. It's a packed scene with billions of dollars in investments in 2025 alone,
and companies globally racking up huge valuations. People are betting big on humanoid robots because
very high profile people involved in the world of humanoid robotics are talking it up. People like
this guy, Jensen Huang. There's so many humanoid robots here, and the AIs that they need are
physical AIs. Chipmakers like NVIDIA are excited for their chips to power the brains of these
humanoids. It needs a brain, and NVIDIA has a chip for that that can go into the robot itself.
Then they have the software. NVIDIA has largely a monopoly for AI GPUs that go into data centers
for training large language and other AI models. So they have a stronghold there, because just like
generative AI tools, humanoid robotics models, real world models are also trained using the same
technology. We just announced a new deep collaboration in the Boston Dynamics. They've
got some very exciting... Other AI and tech giants like OpenAI, Google, and Tesla are also betting
big that there's a lot of money to be made by taking AI off a screen and bringing it into the
real world. And then you have Elon Musk, who talks endlessly about humanoid robots. I think everyone
on earth is going to have one and going to want one. His projections are even more, let's say,
far-fetched. Elon Musk has said that Optimus, Tesla's own humanoid robot, will be a product
that can bring in $30 trillion of revenue annually. And with numbers like that, there's hype.
You increasingly see them showing up in ads. You see them at tech conferences. There are lots and
lots of videos of them. Sometimes you'll see them operating in person, but they might feel like
tethered to something. So I feel like you're both seeing progress and you're also seeing
the sense of it's not quite there yet.
This is my students in our lab here. We call it the auto lab for automation lab.
And so students are working on various projects. And we have a number of robots in here,
two-armed robot, we call it bimanual... Professor Ken Goldberg at UC Berkeley has
been interested in what makes robots work for a long time.
In the last few years, there's been a lot of excitement about the potential of artificial
intelligence, specifically with the breakthrough of ChachiPT, which is a huge paradigm shift in
the field of AI. And so that's the speculation. What is the GPT moment for robotics?
But here's the thing. Large language models, or LLMs, like the kind behind ChachiPT,
are more book smart than robot smart. In ChachiPT, you're looking at a string of
words, and then you're trying to generate what would be the next appropriate string of words
that goes with that. ChachiPT and other large language models run on, well, mostly language.
Think of it like having access to a massive library of trillions of words,
photos and videos from the internet. But a robot doesn't just generate words, it has to create
motion. In a robot, what we need to see is a string of images, a video, and then we need to
see a string of motions, of commands, that go to the robot arms to get it to do the right thing.
To do that, it needs data to learn from. A lot of data.
When it comes to humanoid robots, data is so, so important. Without data, a system that relies
on AI doesn't know anything. If you want a robot to do something like folding laundry, for instance,
you're going to have to have a ton of information about how laundry is folded so that it can learn,
see patterns in that data. How do you pick up a piece of fabric? Do you use articulated fingers,
which are extremely difficult to build? And how do you also hold on to things without squishing them?
Robots learn from experience. And here's the problem. Almost none of that physical experience
has been recorded. No one has captured the motions of robots that go with the inputs that were
generated coming from a camera. So we have a gap, what I call the robot data gap.
That gap is a challenge and an opportunity, especially for the ones betting on physical AI.
More data means better AI. The need for data for robots is so big,
companies are rushing to create libraries made of both real and artificial synthetic data.
So people today are trying to collect this data to be able to train robots.
The first one is simulation, to be able to simulate environments and robots moving in them.
Think of it like creating a hyper-realistic environment the robot can learn in. Its
superpower is scale. Thousands of robots in thousands of worlds all learning at the same
time. The problem is, virtual worlds and virtual robots can lack real-world detail.
Another way to train humanoids is harvesting existing video of the real world.
We do have a huge amount of videos on YouTube and other places that have show people
manipulating things. But it's very difficult to get the full three-dimensional structure
of a scene from a two-dimensional video. There is one option for delivering real
human interactions in real-world conditions, teleoperation, where VR glasses and stylish
outfits can turn these robots into a very expensive puppet. So there's someone behind
the scenes, behind the curtain, who is controlling two arms, moving them around,
and as that happens, the video images and the motions of the robot are recorded, and that
provides data. But that's very, very slow. But there are farms of people out there doing that
as we speak. Teleoperation has been criticized for its use in demonstrations to drum up excitement
and money. But it's also a powerful training tool. Beyond teleoperation, there's another
option to let the robots just learn on the job. Goldberg calls it the flywheel.
As it's working, it's collecting data from its images and its motions, and then use that data
by filtering it to train a larger model. And eventually, that new model becomes better than
the original model. Nailing an AI model and a training dataset that works is paramount for
any robotics company hoping to make it in the humanoid biz. NEO is a robot for the home. We
want every part of NEO to be inviting and made of materials that you would want to have around
your family or in your living room. We're using softcover and this fabric for the face covering.
At ONE X headquarters, teams design, build, and train humanoid prototypes.
Like many places, ONE X uses a combination of training techniques. But when it comes to AI,
things change fast. When we visited ONE X in 2025, they relied on a combination of tools
to train their humanoids, including teleoperation and human input. Hi, Turing. Since then, ONE X
says it increasingly uses something called a world model to train and run their humanoids.
And it's pulling from many of those data gap filling tactics.
The video here on the left is generated by the model. It's almost like a future vision
of what the robot should do. And on the right is the robot completing the action.
Similar to other AI video creation tools, the world model pulls from a vast library of images
and video to generate a visualization of an action. Each time it generates the prediction
of an action, it can learn from the resulting real world action and, in theory, help improve
those predictions with every generation. Kind of like a flywheel, remember?
Of course, the actual goal of all these models and data is to get robots out of the lab and
into the real world where they can continue to improve. And when it comes to doing that,
there's one place that's been faster than the rest. This is the first World Humanoid Robot Games,
taking place in an old Olympic ice skating rink in Beijing. Look at all these humanoids,
running, jumping, dancing. Competitors here are gathered from all over the world to put humanoids
to the mechanical test. I think the games are really interesting. There's tons of really
cool events that you don't really see normally. It's very different, very unique, very special.
When I first met our robot, when I first interacted with it, it was a little scary
because, you know, you have this, like, very heavy piece of machinery just running around
and walking around next to you. And you try to, like, keep your distance and stay away from it.
But as I've gone to know it and work with it, I've developed a sort of, like, connection with it.
These games, which involve everything from boxing to football matches to, yes, work,
are part of a large push in China to dominate the humanoid and AI robotics industry.
China's unique interest in robots has to do with the government for the most part.
The fact that they see it as a strategic priority. Back in 2015, the Made in China 2025 plan had
mentioned robots along with EVs and semiconductors as future growth drivers. So a lot of money has
been poured into this area and it continues to be so. In 2025, China pledged one trillion yuan,
about 140 billion U.S. dollars, to promote the advancement of emerging technologies,
a big part of that being robotics. That kind of support has helped produce 140 Chinese companies
all making humanoids, a number that has continued to grow so much that officials have warned of a
potential bubble. Still, China is currently dominating the humanoid robotics market,
much like they've done with EVs. Companies like Agibot, Ubitech and Unitry, among others,
shipped the most humanoids out of any country in 2025. That's down in part to some serious
chops when it comes to manufacturing cheaply and quickly, which is a good thing if you're
So much of the electronics industry has its supply chain from the sort of circuit level
throughout China and they have scale. Scale means better economics and it means ability to move
quickly. It's really vertically integrated, just like with EVs. They have a lot of companies that
make components for these humanoid robots, whether it's sensors, AI chips, batteries.
The fact that they're able to sort of supply the components for these robots.
That advantage in scale may come in handy. China already leads the world in the deployment of
industrial robots as it races against a shrinking workforce. But for engineering,
training and investment to pay off, they're going to have to do something more than backflips.
They're going to need to get a job. Sorry robots. I am an economist and I am studying
the well-being implications of technological change. Technological anxiety and fear that
technologies are going to take over our jobs is nothing new. What do we need to think about before
deploying humanoid robots at scale would be worker involvement in the process. So how these
robots will be deployed would have huge implications for the quality of work that people will have.
Robots aren't new. They've been in factories for decades.
They're very good at performing routine tasks and physical tasks.
Tasks like welding, painting and assembling parts.
But they are not multi-purpose and that's where the humanoids come in.
Basically, a humanoid robot is designed to do the tasks that humans can do. It's estimated that
almost two million manufacturing jobs in the US alone might go unfilled by 2033.
But if humanoids are going to be part of closing that gap, they need to actually, well, move and
problem-solve like a human, safely, alongside people. And right now that's only happening in
a limited number of very controlled pilot tests run by some very big companies. For example, the
BMW is currently running a pilot and is reporting early data. Other big names testing these robots
include Amazon and Hyundai. These trials are crucial when it comes to figuring out
where humanoids work and where they don't. They also provide, yes, data to fuel development
while things like technology and regulation catch up. We started with pilots in humanoids
about a year ago. This is Patrick Kelleher. His company is one of many testing out third-party
humanoid robots in their warehouses. We have been working with robots for the last 10 years
and we are at the forefront of piloting humanoids in our warehouse operations and it really is
fascinating to see the new technology take foot in production in the warehouse environment.
GXO has found some cases that humanoids might be useful for, like, say, working in a freezer.
We're able to deploy the humanoids in that environment where they can really work
for an entire battery life, completing tasks, being directed by somebody who is outside of
the freezer, and getting work completed without putting our teammates in an uncomfortable situation
working in that cold environment for a long time. But there's still a lot they're not great at.
The humanoids so far are not perfect. They damage product. They drop product. They may put product
back in the wrong location. It is so important that our teammates are working alongside of them
to do that. There is corrective action and audits done to make sure that the activity done by the
humanoid is done accurately. Humanoids still face a number of engineering challenges that
can create serious bottlenecks in a warehouse. They're slow, for one. They have limited battery
life. And picking up a box or handling an object, things that we take for granted, require dexterity.
It's something humanoids, which move using motors and actuators, have a hard time replicating.
We do not see humanoids in the next couple years taking away many of the tasks that our teammates
perform. The tasks are too complex and too important from a quality perspective to go
entirely to humanoids. And it's not really understood yet what the return on investment
is going to be based on the finalized technologies that are coming out.
Places where humanoid robots might be deployed first, like factories and warehouses, are already
built for industrial robots. But putting them in environments for people, like hospitals and homes,
might prove to be the bigger test. In manufacturing, you have a very fixed
production process. You can shield off human beings, and you can let the humanoid robot
perform. While in a hospital setting, it's a semi-structured environment. There are protocols
in how processes should be done and how patients should be cared for. But still, there are so many
uncertainties that can arise. In fact, it's about life and death type of situations,
and the costs of making a mistake are very, very high.
So will humanoids one day overcome their technical challenges, recoup investments,
and become indispensable companions? Folding our laundry, watering our plants,
and building our cars? Or just be a novelty, giving way to robots of other shapes and sizes?
I've been in this field for 40 years, and it is fascinating to me how much it has had a resurgence
again and again over the years. I'm thrilled about that. I'm excited because I feel that
this field has got many years of questions and research that needs to be done. So in some sense,
I don't want to sound discouraging. I'm very, actually, optimistic about the future. But I
think it's important, again, to just be realistic about the time frame.
In the meantime, I'll leave you with this. A humanoid robot waving to you from a balcony
as a drone pulls slowly and dramatically away. That feels like a decent way to end. For now.
