BMW's humanoid loaded 90,000 parts. The cycle-time result is missing
A Figure 02 humanoid worked a BMW Spartanburg line for months. The targets were published, the results against them were not. What a plant should ask.
| Company | BMW Group, Plant Spartanburg, South Carolina, on the BMW X3 line; robot maker Figure AI (BMW Group iFACTORY page, February 2026; Figure AI, 19 November 2025) |
|---|---|
| What ran | one use case, sheet-metal loading; the robot picks three pressed parts from racks or bins and places them on a welding fixture, after which six-axis robots weld them (Figure AI, 19 November 2025) |
| Result BMW put in writing | Figure 02 supported production of more than 30,000 BMW X3s and moved more than 90,000 components on 10-hour shifts (BMW Group iFACTORY page, February 2026) |
| Runtime | more than 1,250 operating hours on a 10-hour shift, Monday to Friday; Figure calls it six months of daily runtime inside an 11-month deployment (Figure AI, 19 November 2025) |
| Targets set | 84-second cycle with 37 seconds to load, more than 99% of cycles with all three parts placed correctly per shift, zero human interventions per shift (Figure AI, 19 November 2025) |
| Not disclosed | the measured cycle time, the placement success rate, the number of interventions, and the price, by either company |
| Next | Figure 03 started a parts-sequencing job in Spartanburg Hall 52 in June 2026; Hexagon's AEON humanoid moves to battery and component work at BMW Leipzig, pilot phase from summer 2026 (Figure AI, 30 June 2026; Hexagon Robotics, 27 February 2026) |
BMW Group says a Figure 02 humanoid robot supported production of more than 30,000 BMW X3s at its Spartanburg plant in 2025, moving more than 90,000 components on 10-hour shifts. Figure AI, the robot's maker, puts its runtime at more than 1,250 hours. Neither company has published the price, or whether the robot met the cycle-time and accuracy targets Figure says were set for the job.
Tech companies are spending heavily on humanoid robots, and much of the attention this autumn has been on robots for the home. For a plant manager the useful question is narrower: has a humanoid done a real station's work on a real line, and what did that show? BMW's Spartanburg pilot is the best-documented answer so far, because both the carmaker and the robot maker have put numbers in writing.
What the robot actually did
One job, at one kind of station. Figure describes it as sheet-metal loading: a worker normally picks pressed sheet-metal parts from racks or bins and places them on a welding fixture, then six-axis industrial robots weld them and feed them into the main line. Each cycle means placing three parts within a 5-millimetre tolerance, in about two seconds a placement.
This is classic pick-and-place. Car plants have automated tasks like it with fixed robots for decades. The point of the pilot was to see whether a general-purpose humanoid could take over a station built for a person, without rebuilding the station.
The targets that were set
Figure lists three measures it agreed to track. A full cycle had to take no more than 84 seconds, with 37 seconds for the loading itself. More than 99 percent of cycles had to end with all three parts placed correctly, shift by shift. And the number of times a person had to pause or reset the robot had a goal of zero per shift.
Neither BMW nor Figure has published the measured result against any of the three. Figure says the robot recorded minimal hardware failures and names the forearm, which carried a circuit board for the wrist motors, as its top failure point. That is useful for the next robot's design. It is not a cycle time.
The headline is a presence figure. The performance figures were promised and never printed.
What the published numbers add up to
Our arithmetic, not BMW's or Figure's. Ninety thousand parts across 30,000 cars is three parts a car, which matches one fixture load per car. At 10-hour shifts, 1,250 hours is about 125 shifts, or some 25 working weeks. That fits Figure's own phrase, six months of daily runtime, rather than the ten or eleven months in the headlines.
Thirty thousand loads in 1,250 hours is about 24 an hour, or one every 150 seconds of runtime. An 84-second cycle would allow about 43 an hour. So either the robot worked slower than the station required, or it was not loading on every cycle while it was switched on, because it shared the station with people, waited on the line, or was being reset. The published figures cannot tell you which, and that is exactly the number a buyer needs.
How long did it take?
About ten months from the robot's first power-up to running every working day. Figure says it delivered robots to the plant and began testing within six months of bringing Figure 02 up, and launched full deployment on an active assembly line within ten. It counts the deployment at 11 months; BMW's page says ten.
BMW has kept going. In June 2026 Figure moved its next robot, Figure 03, into Hall 52 at Spartanburg to sort parts into sequence for the line, a harder job with no figures yet. In Germany, BMW chose a different supplier: Hexagon's AEON, which runs on wheels, was tested at Leipzig in December 2025, was due on the factory floor from April 2026, and had its pilot phase in high-voltage battery assembly and component work set for summer 2026. We found no Leipzig results published as of 7 November.
What did it cost?
Not disclosed by either company. BMW's own wording frames Spartanburg as a pilot whose findings feed further development and scaling, and it has set up a Centre of Competence for Physical AI in Production to carry the work. The real cost of the pilot therefore includes a dedicated team on BMW's side as well as the robot maker's engineering.
Does this apply to a 200-person plant?
Not yet. BMW's case shows a humanoid can hold a real station on a running line for months. It does not show the speed against the station's requirement, the error rate, the reset rate or the price. A mid-sized plant cannot absorb those unknowns, and it has no robot maker treating its line as a flagship.
The trade-off is plain. A humanoid's promise is fitting a station built for a person without rebuilding it. Conventional automation for a fixed loading job comes with a quoted cycle time and a price before you sign. Until humanoid vendors publish the same two numbers from a named customer, the conventional quote is the benchmark.
What to do on Monday
If a humanoid vendor calls, ask for the three numbers BMW's pilot was set up to measure, from a named customer: cycle time against the station requirement, the share of cycles done right first time, and interventions per shift. Ask for operating hours against scheduled hours, because that gap is where the 150-second loads hide. Ask for the price per robot per year, including support.
Before any pitch, write down your own station's numbers: cycle time, tolerance, parts per cycle and how often it changes over. For a fixed pick-and-place job, get a conventional automation quote first and hold every humanoid offer to it. Then watch for BMW's Leipzig results, which will be the first test of a second humanoid supplier on battery work.