Put a humanoid robot into a real setting, and the first question it has to answer is not “does it look human.”
It is “is it worth it.”
Same job — move a crate from warehouse A to B — and you can hand it to an AMR plus a collaborative arm, or to a humanoid. Same motion, same goal. But the two paths do not cost the same order of magnitude.
This piece is not going to re-run the “how expensive is a humanoid” math. That math has been done, and the conclusion is already on the table. The question nobody answers carefully is the other one: what should actually decide the form?
1. “Looking human” is not an aesthetic choice — it is an environment choice
First, some credit where it is due: the humanoid exists for a hard reason.
The world we live in was built to human scale. The height of a door handle, a ten-centimeter step, the slope of a staircase, the grip of a wrench, the location of a light switch, the height of an elevator button — none of these are constants of nature. They are a few centuries of tailoring to one species.
If a machine is going to work generically in that environment — without rebuilding the surroundings, without a dedicated production line, without custom tools built for it — then “looking human” is a very good deal. Its two legs take the stairs, its two hands use everyone else’s tools, its height reaches everyone else’s switches.
So a humanoid is not human-shaped “because it looks nice.” It is the optimal solution for environmental compatibility — if, and only if, your task really needs to move generically through the human environment.
The whole problem lives in that “if and only if.” Most tasks do not.
2. The price is buried inside the word “compatible”
Compatibility has a tax, and it is not cheap.
A six-axis industrial robot has six servo axes. That is every “joint” it owns. A humanoid body, by the public figures that circulate, carries somewhere between twenty-odd and forty joints, with several dozen more actuators in the two hands — Tesla’s Optimus has 28 body joints, and public teardowns put its hands at 50-plus actuators, more than the rest of the body combined.
What does an order of magnitude more joints mean? Open one up: motor, reducer, encoder, driver, torque sensor — each needed once per joint. More axes do not scale linearly in cost; they stack. You pay for the count, and then you pay again for each joint’s precision, each joint’s consistency, and the control cost of coordinating dozens of them.
This is not speculation. Multiple public BOM teardowns give the same shape: actuators (motor + reducer + screw) are 40%–60% of a humanoid’s hardware bill, with Morgan Stanley putting it near 56%. Meanwhile the structure — frame, joint housings, brackets — is only 5%–10%.
That pair of numbers is telling. It says the humanoid’s cost battleground is in the joints, not the skeleton. And it says the “looking human” tax is paid mostly at the joints.
3. Change the ruler: task fit × cost per unit of usefulness
So how do you decide which form a task should take?
I would swap the ruler from “does it look human” to task fit × cost per unit of usefulness. In one sentence: in a given environment, to get the job done, done reliably, and done for years — how much do I pay per unit of useful output?
Measure with that ruler and the scenarios sort themselves.
On the “purpose-built is the honest choice” side:
Warehouse transport, machine tending, inspection, agricultural picking. These environments are controlled — the shelf height is fixed, the conveyor position is fixed, the row spacing in a field is predictable. If the environment is controlled, there is no reason to make the machine adapt to a “generic human environment” it will never meet. An AMR or a fixed collaborative arm sits in the tens of thousands of dollars (public ranges: cobot arms $25K–$75K, AMR/AGV $25K–$150K); a humanoid’s public range is $150K to over $1M. Same job, one order of magnitude apart.
On the “humanoid is hard to replace” side:
Generic home service, disaster response, and anything that needs to operate in an unmodified human environment, with human tools, across varied tasks. There is no “controlled environment” to lean on here — the room is your home, the tools are whatever you grab, the task is decided on the spot. In that setting, the joint budget a purpose-built form saves comes back doubled, in the form of “rebuild a machine every time the task changes.” This is where the humanoid’s generality finally cashes out.
The key is this: ask how general the task needs to be, then decide how much freedom to buy. Not the other way around — build a humanoid first, then go looking for what it can do.
4. So should we give up on humanoids?
No. Quite the opposite.
Saying “most tasks do not need a humanoid” is not the same as saying “humanoids have no future.” These two get conflated constantly, and then people argue past each other — one camp calls the humanoid a capital bubble, the other calls it the final form. Both are wrong, because they are fighting over the same false question: humanoid or not.
The real question is: how do we get the humanoid cheap enough that everyone can afford one?
And the answer to that is not in the form debate. It is in manufacturing.
A humanoid is expensive because it packs a large number of high-precision joint modules into a small volume — every joint’s housing, flange, and reducer seat must hold form accuracy, control weight, and survive repeated shock inside a space the size of a palm. That is precision machined parts work. And when humanoid volume moves from “a few hundred a year” toward “tens of thousands a year,” what decides whether it can reach $20,000 — the widely cited mass-market threshold — is no longer the algorithm. It is the cost curve and manufacturing consistency of upstream components.
Public data already shows that curve moving. The domestic-content rate for harmonic reducers and frameless torque motors went from under twenty percent in 2023 to around seventy percent by 2026, with same-spec purchase prices falling in step. The cost-down battleground has always been the joints, and the joint battleground has always been manufacturing.
Which is why I keep thinking that arguing over whether humanoids “should” be built is pointless, and arguing over how to make them cheap is not. The first is a stance. The second is engineering.
5. Form follows the task; cost is the real line
Back to the opening question: does a robot have to look human?
My answer: form should be reverse-engineered from the task, not decided by “looking human.” The humanoid is one form among many; it is optimal under the specific condition of “generic activity in the human environment,” and not necessarily anywhere else. Treating the form as the goal is mistaking the means for the end.
But do not swing to the other extreme either. The humanoid is not a direction to abandon; it is a direction to make affordable. When its cost really comes down and its experience really becomes comfortable, some tasks that today only a purpose-built form can do may tomorrow have a more general option.
And whichever form we land on, a robot’s “bones” are the same family of parts — joint housings, flanges, reducer seats, end effectors. A joint housing often runs 60 to 130 mm in outer diameter, while the spacing between adjacent internal features is often just a few millimeters: precision and cost tug at each other inside those few millimeters.
The form debate looks like a question of strategy; in practice it is a question of manufacturing. And manufacturing is exactly the kind of thing that can be made solid, one step at a time.
Key references
- Morgan Stanley humanoid BOM teardown (actuators ≈56% of hardware cost), via public industry reporting
- Public robot price indexes and integration cost guides (open price ranges for cobots, AMRs, industrial arms, humanoids)
- Public teardown material on Tesla Optimus joint count and hand actuator count
- Public industry statistics on domestic-content rate and price trends for harmonic reducers and frameless torque motors (2023–2026)
Where we stand on this
Whatever shape a robot ends up taking, the parts that decide its precision, weight, and life — joint housings, flanges, reducer seats, end effectors — are precision machined structures. Our work is to make that class of part accurate, stable, and scalable at the prototype-to-small-batch stage.
If you have robot structural parts heading into prototyping, send us the drawings for a free DFM review and quote.