June 5, 2024, wasn't exactly the kind of date you'd expect to mark a major robotics milestone.
There was no robot walking through a crowded city, making dinner for a family, or replacing a human worker across an entire factory. Instead, Agility Robotics' Digit showed up for regular operations at a GXO logistics facility in Georgia. Its job was painfully ordinary: move totes and place them onto conveyors.
That was the part that caught my attention.
I've always thought the humanoid-robot conversation gets distracted by the wrong things. We see a machine walking like a person and immediately start talking about whether it'll replace us. We watch a robot fold laundry in a laboratory and imagine it cleaning our homes next year.
But walking isn't the hard part.
The hard part is doing something useful, safely, repeatedly, cheaply, and predictably for months.
That's where the industry is finally getting interesting. Digit has since passed 100,000 totes moved in a commercial deployment, according to Agility Robotics. BMW has also put Figure 02 into an actual production environment, where the robot worked 10-hour shifts and helped with the production of more than 30,000 BMW X3 vehicles.
Those examples don't mean humanoid robots are ready for everything.
They mean something more useful: they're beginning to prove that the technology can leave the demonstration stage.
And honestly, that's a much bigger deal.
Where they actually work
If you're picturing a humanoid robot walking around your house right now, you're looking at the wrong market.
The first serious market is factories and warehouses.
That makes perfect sense. These places are already full of repetitive physical tasks. Workers move boxes, carry components, load parts, unload containers, stand at stations, and repeat the same motion hundreds or thousands of times.
Humans are incredibly flexible at this kind of work, but we're also expensive biological machines. We get tired. We need breaks. We get injured. We don't love doing the same movement for eight or ten hours.
A robot doesn't get bored.
That doesn't mean it can automatically do the job better. It has to prove that it can do the job consistently.
That's why Digit's warehouse work matters. The robot isn't being asked to understand every possible object in existence. It's being given a narrow workflow where its physical abilities actually make sense.
The same thing happened at BMW.
Figure 02 was tested at BMW's Spartanburg plant on sheet-metal handling. BMW says the robot worked five days a week for 10-hour shifts and moved more than 90,000 components during the deployment. Figure reported more than 1,250 hours of runtime and about 1.2 million robot steps.
That's much more convincing to me than a two-minute viral video.
A robot completing a complicated trick once proves that the robot can do the trick.
A robot performing a repetitive industrial task for hundreds of hours starts answering a completely different question:
Can someone actually build a business around this?
Why the body matters
There's a reason companies keep making these machines look roughly human.
It isn't just because humanoid robots look cool.
Factories, warehouses, offices, hospitals, and homes were designed around human bodies. Doors have handles. Stairs have certain dimensions. Workbenches are built at particular heights. Shelves are positioned where people can reach them.
If you build a robot with two arms, two legs, hands, and roughly human proportions, theoretically it can operate in spaces that already work for people.
That's the argument.
But there's a catch.
A human body is an astonishing piece of engineering.
We casually pick up a cup without calculating exactly where the center of gravity is. We walk across an uneven floor while talking to someone. We catch something when it slips. We open a plastic bag without thinking about how much force we're applying.
A robot has to sense all of that.
Cameras, force sensors, joint sensors, depth perception, motor control, battery management, and AI models all have to cooperate in real time.
That's where "AI-powered" starts meaning something more than a marketing phrase.
The AI isn't simply answering questions. It's helping the machine understand what it's looking at, decide what action to take, and translate that decision into physical movement.
That's an enormously harder problem than generating text.
A chatbot can make a mistake and give you a bad paragraph.
A physical robot can make a mistake and drop a heavy object on someone's foot.
That difference changes everything.
The money problem
Here's where I think a lot of humanoid-robot predictions get ahead of themselves.
People hear that a robot could eventually perform the work of a human employee and immediately compare the robot's purchase price with an annual salary.
That's too simplistic.
A real deployment has more costs.
There's the robot itself. Then you have software, maintenance, charging infrastructure, networking, integration with existing equipment, safety systems, training, monitoring, downtime, replacement parts, and people who deal with problems when the robot inevitably encounters something it wasn't expecting.
Agility has publicly described a possible cost structure for its newer Digit v5 that includes an estimated $200,000 purchase price, a $20,000 deployment fee, and around $36,000 per year for software and maintenance. The company describes these figures as illustrative rather than a universal customer price. Over five years, that would put the illustrative total around $400,000.
That number sounds huge until you compare it with what a robot is supposed to produce.
If a machine costs $400,000 over five years but reliably handles a job that otherwise requires expensive human labor, operates for long periods, reduces injuries, and keeps production moving, the calculation can work.
But only if the robot actually works.
That's the part I'd focus on.
The cheapest robot isn't necessarily the best robot. A $100,000 machine that spends half its time needing human intervention could be much more expensive than a $200,000 machine that reliably completes the task.
For businesses, uptime is going to matter more than the headline price.
My biggest mistake
If I were approaching this technology from scratch, there's one mistake I'd avoid immediately: getting impressed by general-purpose demonstrations.
I've made the same mental mistake with new technology before. You see the most impressive demonstration and assume that the hardest part has already been solved.
It hasn't.
A robot that can walk, talk, recognize objects, climb stairs, and manipulate an impressive range of objects is fascinating.
But a factory manager doesn't necessarily care.
That manager cares whether the robot can pick up this particular component, place it in this particular fixture, at the required speed, without damaging it, while staying safe around workers.
That's a much narrower problem.
And narrow problems are actually good news for robotics.
The mistake is assuming that because a robot can theoretically perform hundreds of tasks, it should immediately perform hundreds of tasks.
I'd rather see a robot do one thing extremely well.
That's how automation has always worked.
Start with the task that makes economic sense. Measure it. Fix the failures. Expand from there.
What still goes wrong
The physical world is messy.
This is the biggest reason I'm skeptical of claims that humanoid robots are about to become universal workers.
Imagine a warehouse where every box is perfectly positioned.
Great.
Now imagine one box is slightly crushed.
Another is wet.
A label is partly torn.
Someone left a pallet two inches out of position.
The lighting changed.
A worker accidentally moved something.
A sensor gets dirty.
The robot's battery is lower than expected.
Suddenly, the problem isn't "Can the robot pick up a box?"
The problem is whether it can recover gracefully when reality stops behaving like the training environment.
Humans are ridiculously good at this.
We see a strange situation and improvise.
Robots are getting better at improvisation, but they're nowhere near the level of a capable human worker across the full range of unpredictable physical situations.
That's why the most credible deployments are still narrow.
Digit at GXO is doing repetitive tote-handling work. Figure at BMW handled specific manufacturing tasks. BMW's newer pilot in Leipzig is also concentrating on defined production activities, including high-voltage battery assembly and component manufacturing.
Notice the pattern.
Nobody is handing a humanoid robot the keys to an entire factory and saying, "Figure it out."
That's not a weakness.
That's the sensible way to deploy expensive physical AI.
The human question
There's another uncomfortable issue here.
What happens to the people currently doing these jobs?
I don't buy either extreme.
I don't believe every physical job disappears because humanoid robots exist.
I also don't believe automation won't affect employment.
Both positions are too simple.
If a robot becomes reliable enough to move boxes for ten hours a day, companies will have a reason to use it. That's basic economics.
But that doesn't necessarily mean everyone around the robot becomes unemployed.
Factories still need people to maintain equipment, manage production, handle exceptions, inspect quality, supervise systems, repair machines, and make decisions.
The jobs change.
And some jobs will probably disappear.
That's the part nobody should sugarcoat.
If your entire job consists of a predictable physical motion that a robot can perform continuously, your position is more exposed than someone whose work depends on judgment, communication, problem-solving, or unpredictable physical situations.
That's been true of automation for a long time.
Humanoid robots simply push that trend into more environments.
"Why not just use machines?"
This is probably the smartest objection to humanoid robots.
Why build a robot with legs and arms at all?
If you want to move boxes, build a conveyor.
If you want to move pallets, build an automated guided vehicle.
If you want to assemble one specific component, build a specialized industrial robot.
Those machines can be faster, cheaper, and more reliable for their specific jobs.
And I agree.
Humanoids shouldn't replace specialized automation where specialized automation works better.
The interesting advantage of a humanoid is flexibility.
Suppose a warehouse has dozens of tasks that currently require people because the environment was designed for people. You could redesign the entire facility around robots, but that can be incredibly expensive.
Or you could introduce a robot that works within the existing environment.
That's the bet.
Humanoids aren't necessarily competing with every industrial robot.
They're competing with the cost and inconvenience of changing human-oriented environments.
That's a much more interesting proposition.
The home is harder
This is where I'd slow down the hype considerably.
A factory is controlled.
A home isn't.
Your kitchen has knives, liquids, pets, children, stairs, fragile objects, dirty dishes, cables, furniture, and hundreds of random objects that aren't sitting in exactly the same place every day.
Even something as simple as loading a dishwasher involves recognizing different objects, understanding how fragile they are, deciding where they fit, opening the machine, avoiding obstacles, and handling objects that may be wet or slippery.
Now imagine asking the robot to make dinner.
That's dramatically harder than moving a standardized tote.
This is why I'm much more bullish on industrial humanoids than household humanoids in the near term.
A factory can be redesigned around predictable workflows.
Your house can't.
Or at least, you probably don't want to live in a house that has been redesigned like a factory.
Batteries and downtime
Then there's the problem nobody sees in a flashy demo: energy.
Humans eat food and can keep going for hours.
Humanoid robots need batteries.
And physical movement consumes a lot of energy.
Walking is expensive compared with sitting in a data center and generating text. Motors, sensors, onboard computing, cooling, communication systems, and constant movement all consume power.
A robot might technically be capable of working for a long shift, but the business question is more complicated.
How long does it run before charging?
How quickly can it recharge?
Does it need a battery swap?
What happens during maintenance?
How much performance does it lose as the battery drains?
If ten robots are operating on a production floor, what happens when three need servicing at the same time?
These aren't glamorous questions.
They're the questions that determine whether a fleet becomes a useful business tool or an expensive science project.
AI changes the equation
The reason I'm more optimistic about humanoid robots now than I was several years ago isn't simply that the mechanical hardware has improved.
It's the AI.
Older robots were generally excellent at doing predefined things in controlled environments.
Modern AI gives robots a better chance of dealing with variation.
A robot can potentially identify an unfamiliar object, interpret instructions, understand the context of a task, and adapt its movements rather than following one rigid sequence.
That doesn't mean the robot suddenly has human intelligence.
It means we're moving toward machines that can learn more flexible behaviors.
Figure's BMW deployment is a good example of why this matters. BMW reported that motion sequences developed in laboratory testing could be transferred into stable production operation relatively quickly, although the real deployment also required work across production IT, safety, logistics, and infrastructure.
That's the part I find most encouraging.
The breakthrough isn't simply a better pair of robotic legs.
It's the combination of perception, reasoning, movement, simulation, learning, and physical hardware.
That's what people mean when they talk about "physical AI."
Don't expect a robot servant
If you're imagining a humanoid robot arriving at your door next year and handling every annoying task in your life, I'd lower your expectations.
The technology isn't there yet.
And even when it gets closer, cost will matter.
A robot capable of safely operating inside a home has to be much more versatile than a robot moving standardized containers in a warehouse.
It has to deal with children.
Pets.
Wet floors.
Glass.
Stairs.
Unexpected guests.
Different homes.
Different furniture.
Different lighting.
Different objects.
And all of it while sharing space with people who aren't trained operators.
That's a brutal engineering problem.
The funny thing is that the more human-like the environment becomes, the harder the robot's job gets.
A factory can be optimized around the robot.
Your living room can't.
What I'd watch next
I'm less interested in videos of robots doing impressive things and much more interested in a few boring numbers.
How many hours does the machine operate?
How often does it need human intervention?
How many tasks can it perform without retraining?
How expensive is maintenance?
What happens when something goes wrong?
What's the cost per completed task?
Can the robot work safely beside people?
And most importantly: will a company pay to keep using it?
That last question cuts through a lot of hype.
GXO has already signed a multi-year commercial agreement around Digit, while BMW has moved from testing humanoid robots to additional pilot work in its production network. Those are much stronger signals than a robot video going viral for 20 million views.
The industry doesn't need more proof that humanoid robots are possible.
It needs proof that they are economically useful at scale.
That's a different challenge.
Ready, but narrowly
So, are AI-powered humanoid robots ready for the real world?
Yes.
But only if we define "ready" correctly.
They're ready for certain real-world jobs.
They're not ready to be universal human replacements.
That's a distinction I wouldn't ignore.
A humanoid robot that reliably moves totes for thousands of cycles is already a real machine doing real work. A robot that helps manufacture tens of thousands of vehicles is no longer just a laboratory experiment. BMW's reported Figure 02 deployment shows that the technology can operate for long shifts inside a real production environment, while Agility's Digit deployment shows that commercial warehouse work is possible too.
But those examples don't prove that we're one software update away from robot butlers.
They prove that the first layer of the market is forming.
And that's probably how this will happen: slowly, task by task.
If I were starting today, I wouldn't spend my time asking which company will build the perfect humanoid robot. I'd watch the boring deployments. Watch what companies actually pay for. Watch which robots survive long operating periods, which ones need the least human help, and which businesses keep expanding their use after the initial pilot.
That's where the real story is.
The humanoid robot revolution probably won't arrive with one dramatic moment where a machine suddenly becomes capable of doing everything a person can do.
It'll arrive quietly.
One warehouse task.
One factory station.
One shift.
Then another.
And eventually we'll look around and realize the strange part isn't that robots can walk like us anymore.
It's that some of them have simply gone to work.