Robots Building Robots: What the Loop Actually Looks Like

Robots building robots stopped being science fiction some time in the 1980s, when industrial arms started assembling other industrial arms. The 2026 version is louder. IEEE Spectrum’s Video Friday on September 18, 2026 featured UBTECH’s smart factory, where the company says one humanoid robot rolls off the line every ten minutes, and production scaled from 1,000 to 10,000 units in under nine months. Evan Ackerman’s reaction in that post was the right one: “what are they all going to do?” But there is a second question nobody asks, and it is the better one. If robots build robots, what builds the parts the robots are made of? Follow that chain and you find it leaves the robot factory within two steps.

Key Takeaways

• UBTECH’s reported rate of one humanoid every ten minutes implies about 144 units a day at continuous operation, which is a line capability figure rather than a demonstrated sustained output.
• The loop breaks at four specific places: precision gearsets, rare-earth magnets, battery cells and semiconductors. None of those processes is performed by a humanoid anywhere.
• It also breaks at tooling. Every molded, stamped or cast part in a robot came out of a mold or die made by a tool and die maker, an occupation BLS projects to decline 9 percent from 2025 to 2035.
• Boston Dynamics opened a Robotics Metaplant Application Center in Georgia in September 2026 to integrate Atlas robots into Hyundai’s automotive manufacturing, which is robots building cars rather than robots building robots.
• Every robot leaving a line still needs human calibration: joint zeroing, kinematic calibration, inertial sensor calibration and camera calibration, which is skilled work and is not automated away.

What the September 2026 numbers actually say

Start with the arithmetic, because it is rarely done.

One unit every ten minutes is six per hour. Twenty-four hours of that is 144 per day. Three hundred days of that is 43,200 per year. Now compare it with the other figure in the same report: production scaled from 1,000 to 10,000 units in under nine months. That is 9,000 units in roughly 270 days, about 33 per day.

Those two numbers are not contradictory, but they mean different things. One every ten minutes describes what the line can do when it is running. Thirty-three a day describes what the operation actually delivered. The ratio between them, roughly four to one, is the gap every manufacturing operation lives in: changeovers, quality holds, part shortages, maintenance, single-shift operation, and the time a new line spends learning to run.

Teach your kid to always ask for both numbers. A line rate is an engineering specification. An annual output is a business result. Press releases quote the first and investors eventually ask for the second.

The broader deployment picture from the same month supplies the reality check. The Robot Report’s September 2026 roundup noted Agility Robotics disclosing roughly $1.8 million in revenue with Digit humanoids at nine customer sites, and Unitree Robotics stock down roughly 40 percent from its August 19 IPO debut, with the company’s G1 humanoid priced at $13,500. Production capability is running well ahead of demonstrated demand.

Where the loop actually breaks

A humanoid robot is a fairly short bill of materials, and tracing it is the most useful exercise in this whole topic.

Component	What it actually is	Who or what makes it	Could a humanoid do this step?
Structural frame	Machined or cast aluminum, sometimes carbon composite	CNC machine tools, foundries	Loading and unloading, maybe. Not the machining.
Harmonic drive or planetary gearset	Precision-ground gears with very tight backlash	Gear grinders, heat treat, metrology	No. Tolerances are micrometre-scale.
Brushless motor	Laminated steel stator, copper windings, rare-earth magnets	Stamping presses, winding machines, magnet sintering	No. The magnet supply chain is mining and sintering.
Battery pack	Lithium-ion cells, busbars, thermal management	Cell gigafactories, laser or ultrasonic welders	No. Cell manufacture is a chemical process line.
Compute and sensors	SoC, cameras, IMU, force sensors	Semiconductor fabs, clean rooms at ISO 4 or 5	No. A human cannot do this step either, for the same reason.
Wiring harness	Cut, crimped, bundled conductors	Mostly human hands, some automation	Partly. Harness assembly resists automation.
Plastic covers	Injection-molded shells	Molds built by tool and die makers	Molding, yes. Making the mold, no.
Final assembly	Bolting, connecting, closing	Arms and humans, increasingly mixed	Yes. This is the step that is actually automated.
Calibration and test	Zeroing joints, mapping kinematics, checking sensors	Human technicians with fixtures	No. This is where the headcount is.

Read the right-hand column. The only honest “yes” is final assembly, which is also the cheapest and fastest part of making anything, exactly as it was for the toaster in our piece on reshoring explained with a toaster.

The gearset row deserves a moment. A humanoid joint needs very low backlash, meaning the output must not wiggle when the direction of load reverses, because a robot that wiggles cannot place a part accurately or stand up reliably. Achieving that requires gear teeth ground to tolerances in the single-digit micrometres, then measured, then matched into sets. BLS notes that machinists and tool and die makers work to accuracies sometimes within one ten-thousandth of an inch, roughly 2.5 micrometres. That is the actual bottleneck in humanoid manufacturing, and it is a trade.

How to Teach Your Kid About Robots Building Robots

Ages 5–8: the LEGO bootstrap

Ask your kid to build a LEGO machine that builds another LEGO machine. Let them try. They will quickly realise they need hands to make the hands. Then ask: where did the LEGO bricks come from? The answer, a steel mold in a factory, is the whole lesson at five years old.

Ages 9–12: trace the parts of one toy robot

Take apart a cheap toy robot or an old RC car. Lay out the parts: motor, gears, plastic shell, battery, circuit board. For each, ask your kid, “could a robot make this, or does it need a different kind of machine?” They will sort them correctly most of the time, and the disagreements are the good conversations.

Ages 13+: compute the closure fraction

Give your teenager the table above and have them assign a rough cost share to each component of a humanoid, then compute what fraction of the total cost is in steps a humanoid could perform. The number they get will be low, somewhere under a quarter. That fraction is the honest measure of how close the loop is, and it is a number they derived rather than read.

The question to ask: “If a factory full of robots could build a robot, what is the first thing it would run out of?”

Why “self-replicating” is the wrong frame

The loop people imagine is self-replication: a machine that makes copies of itself from raw material. The loop that exists is something different and more interesting, which is progressive automation of specific steps.

Here is the test that separates them. A self-replicating system has to close three loops at once: materials, energy and tooling. Materials means it must process ore or feedstock into usable stock, which is mining, refining and chemistry. Energy means it must power itself without a grid built by someone else. Tooling means it must make the molds, dies, fixtures and measuring instruments needed to make its own parts. Fail any one and you have a factory with suppliers, which is just a factory.

Nobody is close to any of the three, and the honest industry conversation is not about closing them. It is about which steps are worth automating this quarter.

Boston Dynamics’ September 2026 move illustrates the real direction. Per The Robot Report, the company established a Robotics Metaplant Application Center in Georgia as a training facility for humanoid robots, integrating Atlas into Hyundai’s automotive manufacturing operations. That is a humanoid learning to build cars. Cars are a better target than robots: higher volume, more repetitive stations, and a supply chain that already exists.

The same Video Friday from September 18, 2026 had a quieter example worth more than the headline one. MIT’s robotic optics work has an arm arranging mirrors, lenses and optical components held in QR-encoded housings, setting up and tearing down experiments autonomously. That is a robot building experiments, and it compounds: faster experiments mean faster science, including robotics science. That is a real feedback loop, just not a material one.

What to do at home

Separate capability from throughput, always

This habit transfers far beyond robotics. Any time a number is quoted per minute, ask what the annual total was. Any time an annual total is quoted, ask what the line rate was. The gap between them is where all the actual engineering work lives.

Point your kid at the bottleneck, not the headline

If precision gears, magnets, cells and chips are the constraints, then those are where the careers are. Gear manufacturing is machining and metrology. Magnets are materials science and mining. Cells are chemical engineering. Chips are process engineering. Our look at the skilled trades shortage covers why the machining side of that list is especially short of people.

Make calibration visible

Calibration is invisible in every robot video and it is a large share of the labour. Have your kid do the household version: take a bathroom scale, weigh a known object, note the error, and write a correction. Then do it at three different weights and see whether one correction works everywhere. It will not, and discovering that is the beginning of understanding why calibration is a job.

Treat the wiring harness as a serious subject

Harness assembly is one of the last large-scale manual operations in advanced manufacturing, because wires are floppy, routing is three-dimensional, and connectors require tactile feedback. If someone solves it, that is genuinely significant news. If someone claims to have solved it, ask for the cycle time and the first-pass yield.

What not to do

Do not let a video of a humanoid assembling another humanoid stand in for understanding. That clip is true and it shows the easiest step. The question that actually matters is which supplier made the gearbox, and almost no coverage answers it. Teaching a kid to ask about the gearbox is teaching them to think like an engineer rather than a spectator.

What to Watch For Over the Next 3 Months

• Week 4: Watch whether UBTECH or any competitor publishes a sustained monthly output figure rather than a line rate. That single disclosure would settle most of the argument about how real the ramp is.
• Month 2 red flags: A humanoid maker announcing vertical integration of “everything” without naming a gear supplier or a magnet source. Those two are the hardest to bring in-house and the easiest to quietly omit.
• Month 3 self-check: Ask your kid to name one part of a robot that a robot cannot make, and why. If they say the gearbox and mention tolerances, they have the mechanism and not just the vocabulary.

Frequently Asked Questions

Are robots really building robots right now?

Yes, in the specific sense that robotic arms perform assembly steps on other robots, which has been true for decades. UBTECH’s reported rate of one humanoid every ten minutes, from IEEE Spectrum’s September 18, 2026 report, is the most aggressive current claim. No facility makes its own gearsets, magnets, cells or chips with humanoids.

Does this mean robot prices will collapse?

Partly, and unevenly. Assembly labour is a modest share of a humanoid’s cost; precision components and compute are large shares. The Robot Report noted Unitree’s G1 at $13,500 and a Raspberry Pi-powered wheeled humanoid from Flourish at $3,555, which shows prices can fall fast when capability expectations are lowered.

Could a factory eventually run with no people?

Some production lines already run largely unattended for stretches, especially in machining and semiconductor fabs. Fully uncrewed over long periods remains rare because exceptions, which are the unpredictable events a line was not designed for, still need judgment. Maintenance, changeover and calibration are the persistent human functions.

What should my kid study if humanoids interest them?

Mechatronics is the most direct route, combining mechanical, electrical and control engineering. BLS puts electro-mechanical and mechatronics technologists and technicians at a median $73,900 in May 2025 with an associate degree as typical entry, and it is a very small occupation at 15,700 jobs, which is good news for anyone entering it.

Is the humanoid form actually the right design?

Contested, and the contest is informative. IEEE Spectrum reported on October 1, 2026 that Boston Dynamics introduced a specialized gripper for Atlas that may outperform more humanlike hand designs, balancing capability against reliability and manufacturability. Human-shaped is good for human-shaped environments and often worse than a purpose-built machine at any single task. Our look at home humanoids covers that trade-off in a domestic setting.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Ackerman, E. (2026). “Video Friday: Two Birotors Make a Quadrotor.” IEEE Spectrum, September 18, 2026. https://spectrum.ieee.org/video-friday-quadrotor-from-birotor
2. The Robot Report. (2026). “Top 10 robotics stories of September 2026.” October 1, 2026. https://www.therobotreport.com/top-10-robotics-stories-of-september-2026/
3. Ackerman, E. (2026). “Digit 5 May Be the First Humanoid Robot Worker That’s Truly Safe.” IEEE Spectrum, September 15, 2026. https://spectrum.ieee.org/humanoid-robot-safety
4. U.S. Bureau of Labor Statistics. (2025). “Machinists and Tool and Die Makers.” Occupational Outlook Handbook. https://www.bls.gov/ooh/production/machinists-and-tool-and-die-makers.htm
5. U.S. Bureau of Labor Statistics. (2025). “Electro-Mechanical and Mechatronics Technologists and Technicians.” Occupational Outlook Handbook. https://www.bls.gov/ooh/architecture-and-engineering/electro-mechanical-technicians.htm
6. National Institute for Occupational Safety and Health. “Center for Occupational Robotics Research.” Centers for Disease Control and Prevention. https://www.cdc.gov/niosh/robotics/about/index.html
7. U.S. Bureau of Labor Statistics. (2025). “Quality Control Inspectors.” Occupational Outlook Handbook. https://www.bls.gov/ooh/production/quality-control-inspectors.htm