I spent time with humanoid robots at NVIDIA GTC earlier this year, and I came away conflicted.
The engineering is genuinely impressive. Watching what is being invented on the way to a general-purpose machine is one of the more interesting things happening in our industry, and a lot of these learnings will end up somewhere useful even if the humanoid itself doesn’t.
But every conversation I had, and every conversation I have had since, was about the same thing: Can the robot do the job? Can it lift the part, find the hole, hold the torque?
For me, these are the wrong questions. None of our customers’ assembly lines have ever been held back by the question of whether a worker was physically capable of the task. What holds an assembly line back is everything around the task – knowing which variant is coming, the build sequence to follow, and proving that it was built to standard.
In my mind, the decisive question is quite mundane: Who is going to tell the humanoid what to do?
If you run a plant, there is a reasonable chance someone will ask you to put a humanoid in next year’s capital budget. I’d suggest you zoom out and reframe your decision-making around what you need to make any new worker productive on your assembly line, rather than the choice between humans and humanoid robots.
That gap is where the money gets lost. We’ve watched it happen before, on a similar but very expensive case.
Before humanoids, the automotive industry ran a rehearsal: the self-driving car.
It was a very narrow problem compared to a factory. There was just one objective – to get from point A to B. The road didn’t change shape on Monday morning. There were no variant changeovers, no requests for engineering changes, no new products.
But, 17 years and somewhere north of $71 billion later, across seven serious programmes, here is the scoreboard:
Waymo won. It raised roughly $27 billion to reach about $355 million of annualised revenue, and analysts don’t expect it to turn a profit before 2029 – 20 years after the programme started.
That was the easy version of the problem, and only one company is still standing. ARK Invest, who are bulls on this market and size it at $26 trillion, estimate that a general-purpose robot has 200,000 times the complexity of a robotaxi. While you could dispute the exact number, the difference is clear across five metrics – kinetic demand, dynamic mobility, perception and reasoning, adaptability, and error tolerance.
When I went looking for published hours of humanoid work on a real production line, I was surprised by the numbers. About 66,250 hours in total, and all of it material handling.
Figure’s robots ran 1,250 hours between them at BMW in Spartanburg, loading sheet metal, while Agility’s Digit logged around 65,000 hours across nine customer sites, moving totes from one point to another.
Both carried out work that a six-axis robot arm or an automated guided vehicle (AGV) has done for decades – with the older tech doing it faster, more repeatably and for a fraction of the money that humanoids cost.
The humanoid’s whole promise hinges on flexibility – that, unlike a fixed robot, it can be redeployed to a different task tomorrow. But I have yet to see a humanoid do a general-purpose job. I have seen humanoids programmed to do specific jobs, which is exactly what an industrial robot already is.
I’m not saying it won’t come, but the thing that would justify the price has not been demonstrated yet, and nobody seems to be asking for it.
(Watch this humanoid already working on a factory floor.)
Now let me take the humanoid out of the picture and focus on a working factory scenario, with real people, that Jendamark makes possible every week.
A new operator starts on a manual assembly station. Nobody just hands them the station and hopes for the best. Before they even touch a part, five things are already in place:
These are the things that turn a capable person into a productive operator, and it is expensive. American manufacturers spent about $32 billion on training last year – up 22% on 2019, at nearly 48 hours per employee. Anyone who has run an assembly line knows how real that number is, because it recurs every time the variant changes.
So, assume the humanoid arrives and it is every bit as capable as a person. Every one of those five things must still happen. The robot still needs the standard, the guidance, the enforcement, the check and the record. And it needs them again when the variant changes.
Why would that bill be smaller than the one a manufacturer already pays?
The humanoid itself isn’t always on. Typically, it’s two to five hours on a charge. Twenty to forty moving joints. Actuators inspected at 2,000 hours and replaced by 12,000. Maintenance running at 10 to 15 per cent of the hardware price every year. A person just needs a locker, a canteen and a bathroom.
There is a second question sitting behind the first one, and it is the one that will cost you if you get it wrong.
Walk any assembly plant and count the brands. Atlas Copco on one station, Bosch Rexroth on the next, Desoutter on the one after that. Fanuc arms in one cell, KUKA or ABB in another.
Nobody planned it that way out of loyalty. You buy the tool that suits the joint, at the price you negotiated, from the supplier who can get a spare to you on Tuesday. Staying multi-vendor is how a plant keeps its leverage and keeps running.
Humanoids will be no different, except that the stakes are higher, because nobody knows which of these companies will still be trading in five years. (As a reminder, the self-driving scoreboard had seven serious programmes, with only one survivor.)
So, if you standardise your assembly line on a single humanoid vendor, and you let that vendor’s software own your process logic, you have handed over the line. Every new station, every variant change and every price discussion go through one supplier. And if that supplier turns out to be the next Argo AI, your instruction layer walks out of the door with them.
The way out requires the same approach you already take with tools. Keep the orchestration layer yours, and keep it humanoid-agnostic. That layer holds the standard, the sequence, the enforcement, the check and the record. The humanoid is simply a node it talks to, in the same way it talks to a nut runner, press, vision camera, or person reading a screen. Change the robot and the process stays the same. Run two brands on the same assembly line and the plant carries on regardless.
That is what makes the flexibility real, in both directions. Commercially, you can tender each application and let the vendors compete for it, instead of buying a whole line from whoever got in first. Technically, you can start with one unit on one station, trial a second brand alongside it, and move to something better in three years without rewriting how your assembly line works.
Flexibility is the entire promise of the humanoid. It would be a strange trade to buy it and give up your own.
Everyone is building the robot. Almost nobody is building the layer that tells it what to do.
On a properly digitised assembly line, that layer is already running. It is present at every station, instructing whatever is in front of it through the assembly process – step by step, in the right sequence, for the right variant. It enables the tool only in the correct position. It checks the result with computer vision. It records what happened. For machines, it does that through device communication, and for people through digital work instructions on a screen.
It doesn’t much care what – or who – is standing at the station. It’s there to plan, run and prove the work.
We built our enforced assembly system, ODIN Workstation, with productivity in mind. Our system architecture applies whether the operator is human or humanoid – and whichever brand the latter turns out to be. They are just one more node that needs to be told what to do and how to do it.
I’m not sceptical about the engineering behind humanoid robots but I am sceptical about a business case that assumes no cost or time for their induction onto the line, and no cost of risk in being tied to one supplier.
When humanoids do arrive on assembly lines, it won’t be because the robot got clever enough. It will be because someone built the boring layer underneath to guide and check just another kind of worker.
SOURCES:
https://www.agilityrobotics.com/content/digits-next-steps
