HUMANOID ROBOTS IN FACTORIES
Aknazar Arysbek, founder — biomedical engineer (MSc, Tufts University). LinkedIn ↗
Humanoid robots in factories are sold as replacements for workers. Even at full human speed, a leased humanoid working two shifts costs about $32.50 per hour of work. The CNC operator it would replace costs about $37. The factory is also the one workplace already rebuilt around its machines. So the human shape pays least where it is sold first.
The shop
Suppose you run a job shop with six CNC mills. You run a day shift and part of a second. That is typical: US machine shops and turned-product plants averaged 56.7 operating hours a week, about 1.4 shifts (Census Quarterly Survey of Plant Capacity Utilization, October to December 2025).
Your mills are paid for. They stand still every night. You cannot hire a night operator. A humanoid robot maker offers you one by the month, to walk your aisles and load your vises.
Before you sign, ask two questions. What does one hour of the robot's work cost, against one hour of your operator's? And if the robot worked perfectly, would it still be the right machine for six mills on a flat floor?
| The shop | Value |
|---|---|
| Machines | 6 CNC mills |
| Shifts run now | About 1.4 |
| Working days | 250 a year |
| Operator cost, loaded | About $37 per hour |
| Goal | Run the mills unattended at night |
One yardstick: cost per hour of work
A person, a robot arm, and a humanoid only compare on one number that all three produce. We use the cost of one human-equivalent hour: what you pay to get the work a person does in an hour.
P is the machine's price and D its deployment cost. L is its life in years. M is the yearly upkeep and r the yearly capital charge, both as fractions of P. S is the yearly service bill. The shop changes the job n_c times a year, at C_c per changeover. H is the hours scheduled a year, U the uptime, and R the speed relative to a person. For a lease, the yearly payment replaces the top line.
Now the humanoid. Agility Robotics, in its S-4 filing, models Digit at $8,500 a month plus a one-time deployment fee of about $25,000. Assume a three-year term, so the fee adds $8,333 a year and the robot costs $110,333 a year. Two 8-hour shifts on 250 days give H = 4,000 hours. Assume an uptime of 85%.
| Speed relative to a person (R) | Cost per human-equivalent hour |
|---|---|
| 0.3 | $108 |
| 0.5 | $65 |
| 0.8 | $41 |
| 1.0 | $32.50 |
The other side of the comparison is the person. Use the loaded cost of the worker the machine replaces, never the industry average. A CNC tool operator's median wage is $24.37 an hour (BLS Occupational Employment and Wage Statistics, May 2025). Benefits add about half again in manufacturing, which brings the operator to about $37 an hour. Production, transport, and material-moving workers in manufacturing cost $38.42 an hour fully loaded (BLS Employer Costs for Employee Compensation, June 2026). All manufacturing jobs, managers included, average $48.62 an hour. That last figure flatters every robot.
Set the humanoid's cost equal to the operator's and solve for R. The robot breaks even at 88% of human speed against the $37 operator, 84% against $38.42, and 67% against $48.62. The best reported factory figure is 30% to 50% of human productivity, from UBTech's Walker S robots at BYD and Foxconn plants. That figure is the company's own. No independent speed or uptime data for humanoids exists yet.
There is one place the arithmetic gets close. Run the same lease 20 hours a day, and at 0.8 speed the cost falls to $32.50. Tote and parcel moving, round the clock, is where humanoids reach labor cost first.
The humanoid in its perfect form
Picture the humanoid working perfectly. It moves at full human speed. It never falls, never waits for a battery swap, and never needs a technician. Then it is a worker who never quits, at $32.50 an hour against your $37 operator. That is its ceiling.
It lifts what a person lifts. The revised NIOSH lifting equation (1991) starts from 23 kg as the most a person should lift in ideal conditions. Real limits fall below it. It works at the pace your line was set for. It reaches what a person reaches. Every limit you designed around for people, you keep.
It also walks. Your floor is flat and your aisles are marked. Legs add motors, balance control, and fall risk to every robot you lease, for as long as you lease it. No published safety standard yet covers legged robots working without fences; ISO 25785-1 is still a committee draft.
The shop has a cheaper way to meet a robot halfway. It changes the station. A docking plate, a fixture, and a signal cable cost a few thousand dollars per machine, once, by our estimate. Plants are re-laid for every new product line anyway.
The same logic sorts out where generality belongs. A phone can be a camera because one more function in silicon costs almost nothing. A joint costs mass, power, money, and a failure point, whether the job uses it or not. So the generality goes in software, and the body matches the job. One robot AI model can already be adapted to many different bodies. Physical Intelligence's openpi and NVIDIA's GR00T are open models built for that.
The rule: match the machine's envelope to the job's envelope, and put the generality in software.
Even in its perfect form, the humanoid saves at most about 12% of your operator's cost. It never beats the operator on load, pace, or reach. Drop it for this shop.
What keeps robots out of small shops
If robot price kept robots out, the cheap ones would be everywhere. They are not. Only 8.7% of US manufacturing firms use robots. Those firms employ 45% of manufacturing workers (Census Annual Business Survey, covering 2016 to 2018). Across the EU's businesses, 3.3% of firms with 10 to 49 staff use industrial robots, against 22% of firms with 250 or more (Eurostat, 2022). Small firms are where robots are missing.
Price the plainest option: a collaborative robot arm bolted beside one mill. A working cell costs about $50,000 to $100,000 per machine, installed, depending on what it includes. Take $66,000, a five-year life, and 15% a year for upkeep and capital. That is $23,100 a year. On two shifts at 95% uptime and full speed, the cell costs about $6 per human-equivalent hour. On price alone, it beats your operator six times over.
Three costs the price leaves out are what stop the shop.
The first is idle time. The cell is bolted to one machine. While the spindle cuts, the arm waits. An arm the job needs 2 hours a shift stands idle 75% of the shift. An ability the robot never uses costs nothing by itself. An hour the robot stands idle always costs.
The second is deployment. A working robot system costs 3 to 5 times the robot (BCG; Fraunhofer IPA puts it at 4 to 5 times). The arm is the small part of the bill. Integrators are scarce, and sites that want only one or two robots are the ones they find least attractive.
The third is changeover. Engineering your cell for a new part means new grippers, new fixtures, and new programming. Assume $5,000 a change, a low figure. Your shop lives on new parts. Add your changeovers to the same cell:
| New parts a year (n_c) | Cost per human-equivalent hour |
|---|---|
| 5 | $12.70 |
| 20 | $32.40 |
| 50 | $71.90 |
At 20 new parts a year, the $6 cell costs nearly what your operator costs. At 50, it costs about twice as much.
The record of failed robot companies points the same way. Attabotics ran out of capital in 2025. Rethink Robotics closed in 2018 when sales fell short of plan, and again in 2025 because its products were not ready. READY Robotics closed in 2024 when a funding round fell through. Price is not among them.
The rule: when robots look too expensive, check idle time, deployment, and changeover before the price.
Building the machine the shop needs
Watch an operator at one mill. The operator clamps a blank in the vise, closes the door, and presses cycle start. Then comes a 20-minute wait while the spindle cuts. Loading and unloading take 5 of every 25 minutes. The machine does the work. The person feeds it.
Start with the simplest robot that can do this job: the arm from the last section, bolted beside the mill. It loads the vise to a fraction of a millimeter, every time. But it serves one mill, and for 20 of every 25 minutes it does nothing.
So put the arm on a mobile base. Now it can load one mill, drive to the next, and load that one while the first cuts. One robot can serve several machines only if each machine's cycle is long compared with its loading. A cycle at least 4 times the load time lets one robot keep 3 to 4 machines running. Our mills cut for 20 minutes and load in 5, a ratio of 4, so one robot covers 3 mills with time to drive between them. Two robots cover the shop.
But let's see what happens when the base parks. A mobile robot arm repeats its position to about 2.5 mm in practical use (University of Strathclyde measurements, arXiv 2406.06353). Loading a vise needs better than a millimeter. The arm arrives in a slightly different place every time.
So split the job of precision. The base only navigates, to about ±10 mm. At each mill a passive dock takes over: a kinematic coupling, three balls seated in three grooves, which returns to the same position within microns. Printed markers at the station let the arm re-register. The arm then touches off against the fixture. The target at the vise is ±0.2 mm.
One more failure is left. The mill does not know the robot is there. No standard handshake exists between a robot and a machine tool: OPC UA for Machine Tools covers monitoring and job management, not robot loading. Each mill needs an automatic door, an air-operated vise, and signal wiring the robot can read. Each controller family (Haas, Fanuc, Siemens, Heidenhain, Mazak, Okuma) needs its own retrofit kit.
So add a station kit to each mill: the dock, the markers, the door, the vise, and the interface. By our estimate, the station kit costs about $1,000 to $2,000 per machine before the door and vise. It is installed once and stays with the mill.
Now we have a machine that is general within one trade. It does the machine operator's work across the shop, better than a person on the night shift, and software re-tasks it for each new part. Call it a domain machine.
Now count what it earns. One robot runs 3 mills through two unattended shifts, 250 nights a year, at 50% effective spindle use:
That is 3 machines, 16 hours a night, 250 nights, and half of each hour cutting: 6,000 spindle-hours. US shops quote $75 to $200 an hour for machining. At $75 to $150 a spindle-hour, that is $450,000 to $900,000 of billable work your shop does not produce today. At a 30% to 40% margin, the shop keeps $135,000 to $360,000 a year. Even the low end would pay the $110,333 humanoid lease from the yardstick section, with no labor saved at all. The money is in the hours the mills stood still.
The same build-up fails on an injection molding press. Its cycle is often 10 to 60 seconds, so the robot can never leave the press. Mobility adds nothing there. A fixed cell is the better machine.
The options for the night shift
Here are the four ways to run the six mills at night, each pictured in its perfect form.
| Option | Perfect form | The failing | Verdict |
|---|---|---|---|
| Hire a night operator | A reliable hire at $37 per hour | Often not there to hire: US manufacturing had 522,000 open jobs in August 2026 (BLS JOLTS) | Keep it if you can find one |
| A fixed arm at each mill | Six cells at about $6 per human-equivalent hour each | Bolted to one mill each; idle while the spindle cuts; re-engineered for every new part | Drop it in a high-mix shop |
| A humanoid | Full human speed, never down, $32.50 per human-equivalent hour | Saves at most 12% on the operator; keeps the 23 kg limit, the human pace, and the legs | Drop it |
| A domain machine | Two mobile robots, a station kit at each mill | Needs a kit at every mill; fails where cycles are short | Build it |
Two options survive their perfect versions: the hire, where one exists, and the domain machine. A shop that can staff its night shift does not need a robot for this job.
Where each kind of robot wins
The night shift in one job shop is one case. We can sort the rest of factory work with the same yardstick.
| Kind of work | Best machine now | Best machine in the 2030s (our estimate) | Why |
|---|---|---|---|
| High-volume, stable lines: car bodies, bottling, electronics placement | Fixed automation | Fixed automation | Millions of identical cycles pay for a machine built for one task |
| High-mix work that recurs across thousands of plants: machine loading, kitting, line-side loading, inspection, finishing | People; fixed arms where volume allows | Domain machines with station kits | The task repeats across plants, which gives volume, and changes within each plant, which needs software re-tasking |
| Work beyond the human body: loads of 25 to 500 kg, heat, paint, foundry | Specialized industrial robots | Specialized robots built from shared modules | Copying a person is a handicap here |
| Rare, judgment-heavy work: setup, troubleshooting, maintenance, first articles | People | People, with robots assisting | Too little volume, too much variety |
| Spaces built for people that cannot be changed at a sensible cost | People | Humanoids, a small factory niche | The one place the human shape earns its cost |
Two kinds of work are the humanoid's strongest factory case. High-mix assembly needs hands, and the dexterity is not there yet. Maintenance needs a body that goes anywhere a technician goes. It is also the hardest job in the plant. Both stay with people for now.
Why so much money goes into humanoids
Humanoid robot startups raised $11 billion in 2026 up to early October, more than twice the 2025 record (Dealroom). The reasons are real, and most of them are not about factories.
- A bet on general robot AI. If robot AI becomes general, re-tasking gets cheap for every robot, wheeled or legged.
- Investor arithmetic. All human labor is the largest market anyone can claim. A small chance at it is a rational bet for a fund, even if most of the companies fail.
- Cheap parts. China's electric-vehicle supply chain makes motors, gearboxes, and batteries cheap; Unitree sells its G1 humanoid for about $13,500.
- The home. Figure and 1X say openly that the home is their end market. The factory is the first customer that pays, is structured, and is safe to pilot in.
The record so far fits that reading. Agility Robotics reported $1.8 million of revenue and a $140 million operating loss for 2025 in its S-4 filing. GXO, the largest logistics tester, has run 45 humanoid pilots. In August 2026 it said that none of its production robots that year would be humanoids. Of about 22,000 humanoids shipped in the first half of 2026, 13% went to manufacturing (Counterpoint Research). Most of Unitree's humanoid revenue comes from research and education.
The rival that worries humanoid makers in factories has wheels. BMW's first humanoid pilot in Europe, at its Leipzig plant, uses Hexagon's AEON, whose legs end in wheels instead of feet. On a flat floor, a wheeled base does the legs' job for less.
The most probable future
These are our estimates, not data.
| Future | Our estimate | What it looks like |
|---|---|---|
| One general AI, bodies built for the job | 55% | Robot AI becomes general and cheap. Bodies are chosen on cost per task: mostly wheeled bases with one or two arms and swappable tools. Humanoid makers drop the legs on factory models |
| Humanoid volume wins | 20% | Humanoids reach $20,000 to $30,000 and at least 75% of human speed by the early 2030s and become the default for varied manual work |
| Robot AI stalls on reliability | 25% | Fixed lines, collaborative arms, and integrators carry on; humanoids stay in pilots |
Watch for three signs. A humanoid maker shipping a wheeled factory model moves the first future up. A published contract showing at least 0.8 times human speed at 18 or more hours a day moves the second up. Flat sales of collaborative arms and humanoid down rounds move the third up.
The domain machine holds up in all three. A general AI makes it cheaper to re-task. Cheap humanoids take varied manual work from it, but not the heavy, hot, and precise work. If AI stalls, structured stations still work with classical motion control and vision.
What this note does not cover
- Field data. This note rests on published figures and our arithmetic. Shop rates, spindle use, retrofit costs, and changeover counts change from shop to shop; test them in yours.
- Independent humanoid data. None exists. Every humanoid speed and uptime figure comes from a filing, a customer statement, or a marketing demo.
- The labor figure. Every break-even here moves with it. The yardstick section works it at three values.
- Homes, care, retail, and field service. Those are the places the human shape may earn its cost. This note covers factories only.
- Short cycles and single-product lines. A molding press or a line that runs one product for years is served best by fixed automation, and an integrator can build it. You do not need a new kind of robot there.
FAQ
Are humanoid robots in factories cheaper than workers? Only at speeds no humanoid has shown. At full human speed, a leased humanoid on two shifts costs about $32.50 per human-equivalent hour, against about $37 for a CNC operator. Reported humanoid speeds in factories are 30% to 50% of a person's, which puts the cost at $65 to $108.
Are humanoid robots practical in manufacturing today? Rarely. Of about 22,000 humanoids shipped in the first half of 2026, 13% went to manufacturing. No published safety standard yet covers legged robots working without fences.
Will humanoids ever pay in factories? In a few places. Round-the-clock tote and parcel handling comes closest, and spaces built for people that cannot be changed at a sensible cost are the lasting niche.
Why not a fixed robot arm at each machine? It is cheap per hour, about $6 on two shifts. But it is bolted to one machine and re-engineered for every new part. At 20 new parts a year, it costs nearly what the operator costs.
What is a domain machine? A robot that is general within one factory trade, re-tasked by software, and able to move between stations. A mobile robot that loads several CNC mills from a dock at each one is a domain machine for the machine operator's trade.
What we can do about your idle machines
Describe your shop →
Sources: US Census Bureau, Quarterly Survey of Plant Capacity Utilization, Q4 2025 tables · BLS OEWS May 2025, CNC tool operators · BLS Employer Costs for Employee Compensation, June 2026, Table 4 · BLS JOLTS, August 2026 · Acemoglu et al., Automation and the Workforce: A Firm-Level View from the 2019 Annual Business Survey (NBER) · Eurostat, 3D printing and robotics by size class of enterprise, 2022 · NIOSH, Applications Manual for the Revised NIOSH Lifting Equation · Agility Robotics merger filing, Form S-4 (SEC) · The Robot Report on Agility's 2025 results · MERICS on UBTech humanoids in manufacturing · GXO Q2 2026 earnings call transcript · Counterpoint Research, humanoid shipments H1 2026 · 36Kr on the Unitree prospectus · Unitree G1 · Dealroom, humanoid robotics funding · Figure 03 · 1X NEO · BMW Group, humanoid robot in Leipzig · ISO/CD 25785-1 · Physical Intelligence openpi · NVIDIA Isaac GR00T · Lanpdt, cobot machine tending cell costs · Robotomated, cobot tending cell prices · Fraunhofer IPA, planning advanced industrial robot systems · Machine Design, how to calculate ROI on a robot (BCG) · ABI Research, Universal Robots and system integrators · Betakit on Attabotics · MIT Technology Review on Rethink Robotics, 2018 · The Robot Report on Rethink Robotics, 2025 · The Robot Report on READY Robotics · University of Strathclyde, mobile collaborative robots in high-value manufacturing · Slocum, kinematic couplings review (MIT) · OPC 40501-1, OPC UA for Machine Tools · EVS, one robot tending several CNC machines · Fabcon, CNC machining hourly rates · Fictiv, injection molding cycle time. Related: how much a hardware prototype costs · how to turn an idea into a product.