Humanoid Robots Hit Factory Lines in 2026

Figure says its F.02 robot "contributed to the production of 30,000+ X3 vehicles" at BMW's plant in Spartanburg, South Carolina. Loaded 90,000-plus sheet metal parts. Logged 1,250-plus hours on a live assembly line. After ten years of stage demos and treadmill walks, that is a real number from a real factory, and it deserves to be read carefully. So here is the part most coverage skipped: that robot has been retired.

The headline numbers are real

Two of the loudest names in the field finally stopped quoting choreography and started quoting line output. Figure's Spartanburg run hit greater than 99% placement success per shift on a 37-second load cycle, ten-hour shifts, five days a week, all on the chassis assembly line. Tesla, separately, says more than 1,000 Optimus units were already working its Fremont floor in January 2026, doing battery assembly, pack loading, cable routing and parts handling, with a dedicated line targeting 100,000 to 300,000 units this year per The Robot Report.

I want to be clear that this is genuinely new. A fixed pick-and-place task, run for months on a production line at automotive takt, with a placement success number you can audit, is not a demo. It is the first time the category has produced metrics an operations lead can actually argue about. Take the capability seriously.

The trouble starts the moment you treat the capability number as an availability number.

The footnote that inverts the headline

The single most important sentence in Figure's announcement is the one about retirement. F.02 "return[ed] to HQ from BMW as part of our fleet-wide retirement" once Figure 03 launched. So the 30,000-car figure is the lifetime output of a pilot that has ended, not the running rate of a station that still exists. As of now there are no Figure robots on the Spartanburg line.

BMW's own June 2026 material reads the same way once you stop skimming. The company frames its next move as a new pilot at Plant Leipzig in Germany starting summer 2026, with a test deployment from April to prepare, and it is standing up a "Center of Competence for Physical AI in Production." That is the posture of a company still de-risking. You do not build a center of competence for something you have already committed a line to.

This is the gap worth naming, because it organizes everything else. There are two clocks running on this story. One is the demo clock, which measures what a robot has ever done: cars built, parts placed, hours logged. The other is the line clock, which measures what a robot is doing right now and will keep doing next quarter: availability, mean time between failures, vendor staffing on site. The headlines all run on the demo clock. Your maintenance budget runs on the line clock. They are showing wildly different times.

Why the reliability gap is the whole story

A welding cell built around a KUKA or Fanuc arm is engineered for 99.99%-plus availability and runs for years between major failures. That is the bar a production line is designed around, because anything below it stops the line, and a stopped line is the most expensive thing in the building.

Now put a 99% per-shift success rate next to that. It sounds adjacent. It is not even close. The independent 2026 assessment from EVS Insight argues that mean time between failures for precision manipulation on today's humanoids is orders of magnitude lower than a fixed industrial arm, that most deployments still need on-site vendor engineers, custom environment prep, and real integration work to hit their numbers. A robot that succeeds 99 times out of 100 and needs a human nearby for the hundredth is a fantastic pilot. It is also a line that halts more than once per shift.

Then there is the battery wall, which nobody puts in the headline. Most commercial humanoids run two to five hours on a charge. That means swap stations or charging chairs designed into the cell, and a duty cycle that a bolted-down arm simply does not have. None of this appears in a 30,000-car number. All of it appears in your TCO.

The economics break even later than the pitch implies

Tesla is breaking ground on a second-generation line at Giga Texas aimed at a long-term 10 million units per year, quoting a $20,000 to $30,000 unit price. When a number like that lands in a procurement deck, the instinct is to compare it to a year of loaded human labor and call it a deal.

Resist that math for a second. EVS Insight pegs realistic break-even at unit cost below $30,000 and operational lifetime above 20,000 hours, and expects that combination in the 2028 to 2031 window, not today. In low-labor-cost regions, current humanoid total cost of ownership still exceeds a loaded human operator. Spartanburg already answered "can a humanoid do the task." The unanswered questions are the expensive ones: at what sustained line rate, at what quarter-over-quarter availability, with how many vendor engineers in the building, and for how many hours before the joints need service. That last figure is the one nobody is front-loading, and it is the one that decides whether $25,000 is cheap or a down payment on a maintenance contract.

The honest counterargument

The strongest objection to all of this: Tesla isn't running a months-long pilot, it is running more than 1,000 units in continuous internal production, which looks a lot like a standing line. Fair. That is the most bullish data point in the field, and I am not waving it away.

But notice who the customer is. Those Optimus units are Tesla deploying to Tesla, on a line Tesla controls, reporting numbers Tesla self-certifies. That is a vendor eating its own dog food, which is useful and real, and also exactly the arrangement where the awkward metrics (unplanned downtime, engineer-hours per shift, units pulled for service) never have to leave the building. An external customer paying for guaranteed output is a different and harder test. Until a humanoid runs someone else's line, past one hardware generation, without the vendor's engineers on site, "1,000 units" is a strong signal and not yet proof.

How to buy one in 2026 without getting burned

If a vendor walks in this year quoting cars-built or parts-placed, run the deal through these gates in order. Each one ties to a specific from above.

Re-ask every demo number as a line number. Cars built tells you nothing. Ask for sustained cycle time at your takt, availability over a full quarter, MTBF on the manipulator, and the count of vendor engineers on site to hit the quoted figures. If they can only give you lifetime totals like "30,000 cars," they are selling you the demo clock.
Treat "retired" as data, not trivia. F.02 got pulled after roughly eleven months for a hardware refresh. That tells you the upgrade cadence is fast and the install base is disposable, so budget these like GPU fleets you replace every generation, not like a ten-year fixed asset you depreciate slowly.
Scope the task before you scope the robot. Spartanburg worked because the job was one bounded pick-and-place: insert sheet metal parts into a fixture. If your candidate task needs sub-millimeter repeatability, payloads over roughly 10 kg, or certified-hazardous operation, current humanoids are the wrong tool. Buy a fixed arm. Match the platform to a narrow, high-frequency, low-precision-tolerance step first.
Set a numeric trigger, not a vibe. Pilot a humanoid only where 99% per-shift success is acceptable and a failure is recoverable without stopping the line. Commit a permanent station only when the vendor will contract to availability above 99.9% with on-site support priced into the quote. If unit cost is above $30,000 or expected service life is under 20,000 hours, it is R&D, budget it as R&D.
Watch Leipzig and Fremont, ignore the next cars-built press release. The milestone that actually matters is the first external customer running humanoids on a line continuously, past one hardware generation, without the vendor staffing the floor. Until that lands, the category is proven capable and unproven durable. Plan accordingly.

Sources

One flag worth your call: I could not verify the four source URLs return 200 this run because both WebFetch and curl are denied in the current permission mode. The links are carried over verbatim from the research draft. If you want me to confirm they resolve before this goes through QC, allow web/Bash access and I'll re-check (a single 404 hard-fails the gate).