The Bay and the Robot

At a GXO warehouse in Flowery Branch, Georgia, a two-legged robot named Digit spent the better part of a year lifting plastic totes off a conveyor and setting them on a cart. By last November it had moved more than 100,000 of them. Agility Robotics, which builds Digit, says its machines have now logged over 65,000 hours of work across nine customer sites.

Nine sites is a number you can handle by sending somebody. When a robot goes down in Flowery Branch, an engineer gets on a plane.

Now picture nine thousand sites. Or picture the thing standing in somebody’s kitchen.

Almost every argument about humanoid robots right now is an argument about how many get built. Morgan Stanley has put out a projection of a billion of them by 2050. Rodney Brooks, who co-founded iRobot and has spent forty years watching robotics miss its own deadlines, thinks that kind of number is fantasy. Tesla wanted 5,000 Optimus units last year and reportedly built a few hundred.

I’d rather ask a smaller question, because I think it decides more: when one of these breaks, where does it go?

What actually breaks

The instinct is to treat a humanoid as a computer with legs — a software problem in a metal case. There’s one real dataset on that, and it belongs to Figure. The company ran a Figure 02 at a BMW plant for six months, 1,250 hours of runtime and 90,000 parts handled, then published what broke. "The forearm was our top hardware failure point at BMW." Not the legs, not the battery. A small board distributing signals out to the wrist, and the cabling that flexed every time the arm moved. Figure re-architected the wrist on its next model specifically to delete both.

That’s the shape of it. What fails on a walking machine is what moves, and what carries signal across a joint: actuators and their gearing, hands, feet, and cable that flexes a few million times and then stops. Plus the battery, the heaviest and most dangerous object in the machine.

1X publishes one maintenance figure for Neo, and it’s the most useful number any of these companies has put out. The tendon drive is rated at two million nominal load cycles, after which it’s a "1 day service replacement." That isn’t a fault. It’s a scheduled wear part, like brake pads.

Then there’s the failure nobody puts on a spec sheet. It falls over.

Laptops don’t fall down stairs. These will, and when one does, something inside is bent, and a machine that’s going to work around people has to be checked before it goes back. Unitree, which sells these today, has already written the exclusion into its warranty. Not covered: "Damage caused by tumble or fall when manipulating robot on level ground." A bipedal robot, and falling over is the owner’s problem. Add it up — heavy use, constant ground contact, impact damage, parts that wear out on a schedule, a big battery — and it looks a lot more like a car than a phone.

Mail it, or walk it in

Every mature hardware business splits repair between two layers, and the phone is the clearest example precisely because it does both.

Crack a screen and you don’t mail anything. You walk into a strip mall and somebody swaps the glass while you wait. Asurion runs a national chain of those storefronts under the uBreakiFix name, Apple and Samsung both authorize independent shops to do the work, and there are thousands more operating on their own. Screens, batteries, charge ports: frequent failures, module swaps, back in your pocket the same afternoon.

Everything else goes in a box. Warranty claims, water damage, anything needing work at the board level, anything the manufacturer wants to keep control of. That goes to a depot, you wait a week, and if you can’t wait you dig an old phone out of a drawer.

It’s worth noticing who owns that local chain. Asurion is an insurance company. Those storefronts exist to service the protection plans it sells, which means the repair network got built as the delivery mechanism for a financial product rather than as a business in its own right. Hold onto that, because it’s probably how this gets paid for.

Cars split the same way, with the line drawn somewhere else. The shop down the street does brakes, alternators, sensors and body work. Engine and transmission rebuilds go out to specialists. Warranty work goes back to the dealer.

So the question was never whether humanoids get depots or local shops. They’ll get both. The question is where the line falls between them — which jobs end up local — and whether that half is big enough to keep the lights on in a building near you.

Three things decide it, and none of them have much to do with how sophisticated the machine is: how hard the broken part is to ship, whether you’ve got a spare while it’s gone, and how long you can live without it.

The shipping problem

Humanoids are interesting on the first count, because they can walk. This is the first machine that can take itself to the shop, or at least step into its own crate. Score one for the depot.

The weights vary more than you’d expect, and the spread matters. 1X lists its home robot, Neo, at 66 pounds. Unitree puts the G1 at about 35 kilograms, roughly 77 pounds, and the taller H1 at 47. Boston Dynamics publishes the electric Atlas at 90 kilograms — 198 pounds, and not something one person picks up off the floor.

So the lightest of them, the one built for houses, is the one you could actually put in a crate. That cuts against my argument, and it’s worth saying so.

The battery is the awkward part regardless. Neo carries an 842-watt-hour pack. Under IATA’s rules a lithium battery over 100 watt-hours installed in equipment moves as fully regulated dangerous goods — UN 3481, Section I: specification packaging, a declaration, a trained shipper, weight limits per package. A healthy robot can fly. It just can’t go in an ordinary box.

A damaged one can’t fly at all. US hazardous materials rules — 49 CFR 173.185(f) — say a lithium battery that has been damaged, or that a manufacturer has identified as defective for safety reasons, "may be transported by highway, rail or vessel only." It needs Packing Group I packaging, a non-metallic inner enclosure, non-combustible absorbent cushioning, and a 12-millimeter marking reading "Damaged/defective lithium ion battery."

Then the carriers make it worse. UPS states that it "does not accept Damaged, Defective, or Recalled Lithium Batteries prepared under 49CFR §173.185(f)," and will move them only under a special permit and a negotiated contract. FedEx is blunter: "FedEx does not accept damaged, defective, or recalled batteries because they pose a safety risk." The Postal Service won’t mail them at all.

The asymmetry is the whole thing. An intact battery-powered machine going by ground is barely regulated — no shipping paper, no label, no placard, no hazmat training. The same machine with a damaged pack needs specification packaging, full paperwork, a trained shipper, and two of the three big carriers won’t take it at any price.

Unitree has already drawn the obvious conclusion, in its own after-sales terms: "Please do not return the seriously damaged battery packs; If you have already sent them back, Unitree Robotics will scrap the battery packs and will not return them."

So a robot with a worn-out hand can be shipped, expensively. A robot that fell down a flight of stairs is a truck, a trained shipper, and a manufacturer that would rather you didn’t.

The failures you’d most want to send away are the ones that are hardest to send.

The spare

That leaves the third question, and it’s the one that actually settles this.

A company running forty robots keeps a few on the shelf. Pull the broken one, slot in a replacement, ship the casualty to Oregon and forget about it for a month. This is why fleet robotics keeps converging on big central depots, and it’s why the land grab for that kind of space is already finished in the business next door. Waymo runs several thousand vehicles and owns none of its depots — Avis, Moove and Lyft’s Flexdrive operate them. Moove raised $250 million at a $2.1 billion valuation to build more of them. Hertz is converting rental-car infrastructure into robotaxi depots.

A household doesn’t have a spare. Neither does a restaurant with two machines, which is the same problem in a slightly bigger box. If the robot that covers the overnight shift, or helps somebody’s mother get out of bed, disappears for three weeks, that isn’t an annoyance. That’s a reason not to buy one.

So there’s a fork here. If these stay fleet machines, the depot covers nearly everything, and it’s already funded. If they get sold to people, the local layer has to exist — somebody has to build a network of places that can open one up on Thursday and have it walking by Friday.

And that second road is not hypothetical any more. 1X is taking $200 deposits on Neo right now, for delivery into people’s houses. Nobody who leaves that deposit has a second robot in the garage. For cars and for phones, the local repair layer took decades to grow. For this, none of it is there yet.

What a repair shop already has

Which brings me to the part I find genuinely interesting, and it isn’t the part people usually argue.

The familiar version goes like this: America has around 176,000 auto repair shops and several thousand quick-lube bays, electric cars need fewer oil changes, so convert them. I don’t think that argument holds, and I’ll get to why. But there’s a better one underneath it, and it has nothing to do with the lift or the drain in the floor.

Over the last decade, independent shops have bought two capabilities at real expense, for reasons that had nothing to do with robots.

The first is calibration. The forward-facing camera on a late-model car sits behind the windshield. Replace the glass and you’ve moved the camera, and a camera off by a fraction of a degree gives the car a slightly wrong idea of where the lane is. Automakers publish position statements requiring recalibration after that kind of work.

Some of it is done dynamically — drive the car at a specified speed on well-marked road until the system re-learns what it’s looking at. But much of it is static, and static calibration means putting the vehicle in a bay with a level floor, controlled lighting, clear space around it and physical targets set at measured distances.

Not every shop bought that setup. A 2025 study out of the University of North Florida, built on phone interviews with 304 randomly sampled shops, found 65 percent relying on sublet for calibration, and only 23.7 percent employing a technician with any ADAS training at all. Industry surveys run by companies selling calibration software put in-house capability far higher. Treat the spread as the honest answer.

Either way the work is local. Whether it happens in the collision shop or at a specialist a few miles away with a van full of targets, an entire service niche grew up inside about a decade, for one reason: machines started carrying sensors that have to be re-taught whenever somebody works on them. In 2017, under 1 percent of repairable auto claims included a calibration. By 2025 it was 28 percent. The capability is out there, in ordinary commercial buildings, near where people live.

A humanoid that’s had an arm replaced has exactly the same problem. Its sensors have moved relative to its body, and a machine that’s wrong about where its own hand is will hurt somebody. The setup it needs to be told the truth again is already sitting in a collision shop.

The second is high-voltage work, and here the trade has already written the credential. ASE, which certifies American auto technicians, runs two xEV electrical safety certifications against a published standard. The first is awareness, for anyone who might walk past an electric vehicle at work. The second is aimed at technicians, and it covers high-voltage battery evaluation, isolating and de-energizing the system, protective equipment, and assessing a damaged vehicle.

Read that list again with a robot in mind. Evaluate a high-voltage battery. De-energize before opening the housing. Work out what to do with a pack that’s been in a crash. It’s the same job.

What I can tell you is how few hold the deeper credential. A 2023 count put 3,409 US technicians holding ASE’s L3 hybrid and electric vehicle specialist certification, against roughly 733,200 working auto technicians. Under one percent.

And there’s a hole at the top of the ladder. ASE’s own standard describes a Level 3 role — the technician who does live work, who goes inside the high-voltage battery and separates individual modules. ASE sells no certification for it. The job of opening the pack is the one job with no credential behind it.

Then the ordinary furniture: three-phase power, compressed air, a lift, a roll-up door. Some of it doesn’t transfer. Rebuilding a precision gearbox or aligning optics wants a cleaner, steadier room than any lube bay has ever been, and that work will stay centralized wherever it happens.

Which brings up the part that’s hardest to replace, and it isn’t any of the equipment. It’s permission.

Auto repair is a zoned use. Most cities classify it as automotive or light industrial, and it generally isn’t allowed by right in the retail and commercial districts where the convenient real estate sits. You need the right parcel, or a conditional use permit and a hearing to go with it. A shop that’s already open carries that permission with it.

That’s the strongest argument for these buildings, and it’s specifically an argument for buying them rather than building something new. Say you wanted to stand up a hundred robot service centers across the metros where the machines would actually be. The constraint wouldn’t be lifts, and it wouldn’t be technicians. It would be finding a hundred addresses where a city will let you do loud, heavy, hazardous work near where people live. Those addresses exist. Somebody is already operating on them.

The thirty-minute part

The other objection I hear is that humanoids will be built from swappable modules, so there’ll be nothing to fix. Pull the arm, bolt on a new arm, send the old one back.

That’s probably right, and it’s an argument for local shops rather than against them, because swapping parts is what car repair already is. Nobody rebuilds an alternator in a service bay. They work out that it’s the alternator, pull one off a shelf that showed up half an hour after they called, put it in, and send the dead one back as a core.

That half hour is worth staring at. AutoZone, O’Reilly and Advance run something like 17,000 stores in the United States between them, organized around a promise to get parts to professional shops in about thirty minutes. AutoZone alone has 367 hub stores and 133 mega hubs, the largest carrying north of 100,000 different parts. It’s a same-day physical distribution network with national coverage, built over fifty years, and nobody is going to fund a second one.

Somebody in that industry has noticed. Genuine Parts Company, which owns NAPA, has spent several years buying industrial automation distributors — including Kaman’s distribution arm — and folding them into a robotics and automation business. The auto parts trade has already decided this is its adjacency. It just hasn’t run the idea through its stores.

The people problem

The labor numbers are the most lopsided thing in this whole picture.

The Bureau of Labor Statistics counts about 15,000 electro-mechanical and mechatronics technicians in the entire country, in a category it expects to grow roughly one percent over a decade. There are more than half a million working auto technicians.

The shortage data runs the same way. The TechForce Foundation’s annual survey puts industrial machinery technicians about 85 percent short of demand, and medical equipment techs higher still. Automotive finishes last on that list at 29 percent — the least short-handed trade they measure. So the adjacent trade with the most people in it is also the one under the least strain, while the trade you’d actually need has almost nobody.

Retraining is the part people overestimate. FANUC’s certification program runs through more than 1,700 partner schools, and the first tier takes anywhere from about a week to a few hundred hours. Turning a good diagnostic technician into someone who can swap an actuator and run a calibration is a matter of months, not years. Which cuts both ways — it’s just as easy for the manufacturers.

Reasons I might be wrong

There are good ones.

• The volume isn’t there, and it isn’t close. Unitree’s listing prospectus reported about 5,500 humanoids sold last year and claimed roughly a third of the global market, which puts the whole world somewhere around 17,000 units. Spread that across American metro areas and no city has enough machines to keep a single bay busy. Everything above is an argument about a threshold nobody has crossed.
• They’re fleet machines today. Nearly every humanoid actually working is leased or sold to a company that runs several of them and holds spares. The depot model fits, it’s running, and it may just keep running.
• The van may go to the robot. GMO Internet Group in Tokyo has launched what amounts to a humanoid ambulance — a vehicle carrying engineers, parts and a replacement unit, dispatched to wherever the machine is. Field service doesn’t scale well, but it scales far enough to cover the next few years, and habits formed in that window will be hard to break.
• Warranty and liability make a good moat, and manufacturers will use it. A robot that injures somebody after an independent repair is a lawsuit no manufacturer wants a share of. That’s reason enough to keep the whole thing closed.

And the premise about empty bays is simply wrong. Bays have been disappearing — Lang Marketing counts tens of thousands gone over the last decade — but the ones left are busier and more profitable than they were, with sales per bay up more than 30 percent since 2020. That’s scarcity, not distress. Nobody is going to hand over a profitable shop cheap because electric cars are coming.

The charger problem

There’s a recent example of what happens when a country installs a lot of hardware and doesn’t staff the repair of it.

Federal rules require publicly funded fast chargers to hit 97 percent uptime. Operators report numbers in the mid-to-high nineties. When researchers went out and physically tried chargers around the San Francisco Bay Area, they found roughly three-quarters of them working. The explanation wasn’t complicated: not enough technicians trained on the specific equipment, covering too much ground, arriving too late.

Those machines are bolted to the ground and don’t have to balance.

A law that doesn’t exist

Underneath all of this is something anyone reasoning from the car analogy ought to sit with.

The independent auto repair business wasn’t created by the market. It was created by law — the Magnuson-Moss Warranty Act of 1975, which stopped manufacturers from voiding your warranty because you used an independent shop, and a 1992 Supreme Court decision involving Kodak, which treated a manufacturer’s refusal to sell repair parts to independents as an antitrust problem. The competitive aftermarket Americans treat as the natural order of things is about fifty years old and came out of Congress and a courtroom.

There’s no robotics version of either, and the tools are all sitting there. Parts pairing, software locks, withheld diagnostics, authentication on replacement modules, warranty terms that die the moment anyone else opens the machine. Unitree already voids its warranty for "unauthorized disassembly, and shell opening," and makes the owner pay the freight to send the machine in. Nothing stops any of that.

Right-to-repair has advanced in a number of states for consumer electronics, and a federal bill covering vehicles has been introduced more than once without passing. As far as I can tell, not one of those laws mentions robots.

What to watch

I don’t know how this goes, and I’d be skeptical of anybody who says they do. But three things would tell you, and you can watch for all of them.

• A manufacturer publishes a real service manual. Not a spec sheet — a document showing a stranger how to take the machine apart and put it back together. This one has already started moving. Unitree publishes screw-level disassembly guides for its hands. Boston Dynamics says every limb on Atlas "can be replaced in the field in under five minutes," and that "there will also be training and certification paths to empower your onsite maintenance teams." Watch whether those paths ever open to anyone outside the customer’s own staff.
• A parts distributor stocks an actuator. Not a press release about robotics. A part number, in a hub, with a delivery time attached.
• A state legislature writes robots into a right-to-repair bill. Which would mean somebody has already been told no, at scale.
• An insurer starts writing protection plans on humanoids. That’s the Asurion move, and whoever underwrites the plan has a reason to fund the shops that honor it.

In the meantime, there are a lot of buildings out there with three-phase power, level floors, calibration targets, technicians who’ve already been trained on high-voltage batteries, a parts truck that shows up in thirty minutes, and permission from the city to do heavy work near where people live.

And there’s an industry that may need all of it in a few years, currently assuming it will have to build the whole thing from scratch.