In 2025, BMW Group's Spartanburg, South Carolina plant became the site of the first deployment of humanoid robots inside a BMW facility, a Figure AI robot, Figure 02, working a real production role over a ten-month period. The results are a useful, concrete data point in a field that's mostly been evaluated on demo reels.
What Figure 02 actually did
Over its ten months at Spartanburg, Figure 02 assisted in the production of more than 30,000 BMW X3s, retrieving and positioning sheet metal parts for welding. Moving over 90,000 components across roughly 1,250 operating hours and around 1.2 million steps. That's a genuinely large, sustained production role, not a staged demonstration, and it's the kind of duration and volume that's rare in humanoid robotics coverage, which tends to lean on short demo footage instead.
Why BMW is expanding, not just repeating the pilot
BMW didn't just extend the Figure 02 pilot. It's deploying Figure AI's next-generation Figure 03 robot for a different task entirely: logistics, specifically sorting unsorted components into sequencing trolleys for delivery to assembly lines, a task BMW has said could be scaled across its broader automotive production. Moving from a single, narrow manufacturing task (parts retrieval and positioning) to a logistics role is a meaningful signal. It suggests the pilot demonstrated enough reliability that BMW is willing to expand into an adjacent, still-bounded task rather than either abandoning the program or leaping to open-ended general-purpose deployment.
Part of a broader "Physical AI" push, not an isolated experiment
BMW has framed this within a larger strategy (a new "Center of Competence for Physical AI in Production") and has separately begun a pilot at its Leipzig, Germany plant, extending the same approach to Europe. This matters for reading the Spartanburg results correctly: BMW is treating humanoid robotics as an infrastructure investment worth building organizational capacity around, not a one-off pilot it happened to try.
What this does and doesn't prove about humanoid robotics generally
It's worth being precise about the boundaries of what Spartanburg actually demonstrates. Figure 02's task (retrieving and positioning known parts in a controlled, mapped factory environment) is exactly the kind of structured, bounded task that current embodied AI handles well, consistent with what we cover in our explainer on embodied AI. It does not demonstrate general-purpose humanoid capability in unstructured environments (a home, a retail floor, an outdoor setting) which remains a substantially harder, unsolved problem.
How this compares to the other real deployment story in robotics
BMW's manufacturing pilot and Amazon's warehouse deployment of Agility Robotics' Digit, covered in our separate piece, share a structural similarity worth naming directly: both are real, sustained, revenue-relevant deployments in controlled industrial environments, and both represent the frontier of what's actually working in physical AI right now, not the more speculative, general-purpose humanoid future that dominates most coverage of the category.
The honest read
"Humanoid robots left the demo stage" is true in a specific, narrower sense than the phrase usually implies: real, sustained industrial pilots (Figure 02 at BMW, Digit at Amazon and elsewhere) are genuinely proving out, and companies are expanding rather than quietly discontinuing them. General-purpose humanoid robots operating reliably in unstructured, everyday environments remain a different, harder, and still largely unsolved problem. Both things are true at once, and conflating them is the most common error in coverage of this space.