Research question

Paper 004 found that AI is already reshaping entry-level cognitive work. Has anything comparable happened to physical labour yet — and how would you actually know, given how unreliable the public numbers are?

Abstract

Search for humanoid robot deployment figures in 2026 and you will find confident numbers: 50,000 Tesla Optimus units built, 10,000 Figure AI robots deployed across warehouses. Almost none of these figures come from the companies they describe. Tesla has never published an audited Optimus production count, and as of mid-July 2026, its own Fremont production line had not yet started. The real, verifiable picture is much smaller and much more specific than the headlines suggest.

Figure AI's strongest evidence is a single, well-documented site: an eleven-month pilot at BMW's Spartanburg plant, where its robots logged roughly 1,250 operating hours and helped produce more than 30,000 vehicles. Agility Robotics has seven commercial units at Toyota. Unitree, the actual global volume leader, shipped around 5,500 humanoids in 2025 — a real number, but a small one, and its profit still fell by half in early 2026. This paper maps what is genuinely deployed and working, not what has been announced, and asks what that gap says about the pace of physical automation.

Key findings
  • Widely circulated figures — 50,000 cumulative Optimus units, 10,000 Figure AI deployments, "more than a thousand" Optimus units working Tesla's production lines — do not come from the companies involved and do not survive contact with company filings, per independent review.
  • Tesla's own guidance placed Fremont Optimus production start in "late July or August" 2026; as of mid-July, production had not begun, and Tesla's Q2 2026 delivery report contained no Optimus figures at all.
  • Elon Musk confirmed in January 2026 that existing Optimus units are primarily "learning" — generating training data — rather than performing productive factory work.
  • Figure AI's Figure 02 completed an eleven-month, independently documented pilot at BMW Spartanburg: 1,250+ operating hours, 90,000+ parts loaded to five-millimetre precision, supporting production of 30,000+ vehicles across ten-hour shifts, five days a week.
  • That deployment has since expanded to a second BMW plant (Leipzig) and a separate paying customer reached production in 30 days, down from 12 months for the first site — a customer-driven adoption signal that has no confirmed Tesla equivalent yet.
  • Agility Robotics has seven commercial Digit units at Toyota after a year-long pilot. Unitree, not Tesla or Figure, is the actual global unit-volume leader, having shipped roughly 5,500 humanoids in 2025 at around a tenth of Western competitors' price — while still posting a 50% profit decline in Q1 2026.
Why physical automation lags cognitive automation

Paper 004 found AI already reshaping entry-level hiring in knowledge work within roughly three years of ChatGPT's release. Nothing comparable has happened yet in physical labour, and there's a long-recognized reason why: tasks that feel effortful to humans, like arithmetic or writing, have turned out to be comparatively easy to automate, while tasks that feel effortless to humans, like walking across a cluttered room or picking up an irregular object, have turned out to be extremely hard. Roboticists have called this asymmetry Moravec's paradox for decades, and 2026's deployment data is, in effect, a real-time demonstration of it.

Language models generate fluent text after ingesting text; a warehouse robot has to physically balance, grip, and correct for real-world friction and error in ways that don't reduce cleanly to pattern-matching on a dataset. That's a genuine engineering gap, not merely a funding or attention gap — Boston Dynamics' Rodney Brooks called the vision of humanoid robots as general-purpose assistants "pure fantasy thinking" as recently as 2025, from inside the industry that's supposed to be building them.

The honest gap in the evidence

Unlike Papers 001 through 004, this paper cannot lean on Federal Reserve, OECD, or ILO-grade institutional data — no equivalent body currently tracks verified humanoid deployment at that level of rigor. The best available sourcing here is trade press and industry analysts doing their own verification work against company filings and site visits, which is a real step down in evidentiary quality from this site's usual standard, and worth stating plainly rather than dressing up as more authoritative than it is.

Within that constraint, the picture is still reasonably consistent across independent sources: real, working deployment exists, but it's concentrated in a handful of named sites (BMW Spartanburg and Leipzig, Toyota, Mercedes-Benz Berlin) rather than the broad rollout the headline unit counts imply. Whether that's an early inflection point — the same shape Figure 001's "9,000-employee pilot to 30-day second customer" curve suggests — or a sector still years from real scale is genuinely unresolved.

What this means for the Labour Machine's automation dial

The Labour Machine's automation control doesn't distinguish cognitive from physical labour, and this paper is a reason that distinction matters. Paper 004 showed AI already measurably reshaping entry-level knowledge work. This paper shows physical, manual labour has, so far, seen verified deployment only at a handful of specific industrial sites — nowhere near the scale automation forecasts for manufacturing or logistics tend to assume. A single "automation" slider necessarily flattens two very different clocks into one number, and readers modelling near-term scenarios should weight physical-labour displacement considerably more slowly than cognitive-labour displacement, based on everything documented above.

Methodology

This paper synthesizes 2026 trade and industry press that explicitly cross-checked company claims against filings, site visits, and named customer confirmations, rather than repeating unverified circulated figures. Where a figure could not be traced to a primary company statement or independently documented site visit, it is described as unverified rather than reported as fact.

Assumptions
  • Independent trade-press verification (cross-checking company claims against filings and named customer sites) is treated as more reliable than company-originated figures or widely circulated secondary claims, absent an institutional data source.
  • Deployment concentrated in a small number of named industrial sites is treated as the current honest baseline, not dismissed as insignificant — a handful of real, sustained, ten-hour-shift deployments is a genuinely different claim than zero.
Limitations
  • No government or academic body currently tracks verified humanoid robot deployment with the rigor this site's other papers rely on; all sourcing here is trade press and industry analysis.
  • This field moves fast even by this site's standards — production status, unit counts, and customer rosters cited here reflect reporting current to July–August 2026 and may already be outdated.
  • Market-size forecasts (Goldman Sachs' $38 billion by 2035, Morgan Stanley's $152 billion by 2040) span a four-times range from two credible analyst firms, underscoring that even professional estimates of this market's future size disagree substantially.
Data sources
Citation

After Labour (2026). Where Are the Robots? Separating verified humanoid robot deployment from headline hype.