If you follow robotics headlines, 2025 and 2026 have felt like the year humanoids arrived on the factory floor. BMW, Amazon, Toyota, Tesla — each has put a humanoid to work and put out a press release. Yet walk any of those plants and you will not find a shift of robots. You will find one robot, occasionally two, doing a single repetitive task while engineers watch. The announcements are real. The scale is not. Understanding *why* matters, because it tells buyers what these programs actually are — and what they are not yet.
The flagship success is one robot doing one task
Start with the deployment everyone cites as the proof point: Figure's humanoid at BMW's Spartanburg plant. The numbers are genuinely impressive, and — importantly — they come from BMW's own press office, not just the robot vendor's marketing. Over roughly ten months in 2025, a Figure 02 unit supported production of more than 30,000 BMW X3 vehicles, moved over 90,000 components, and logged about 1,250 operating hours across ten-hour weekday shifts (per BMW Group). BMW has since begun deploying the successor, Figure 03, on logistics sequencing.
Now read the same facts the other way. "30,000 vehicles" is a headline; the underlying reality is *one* robot performing *one* task — inserting sheet-metal parts into a fixture — over ten months, at roughly 1,250 hours (a fraction of what a full-time human logs in the same window). That is a successful, well-run pilot. It is not a workforce. The distance between the two is the entire story.
The pattern repeats across every vendor
This is not a Figure-specific quirk. The same shape shows up everywhere you look:
- Agility Robotics' Digit has racked up over 10,000 hours of box-moving trials in Amazon facilities, running around 300 boxes per hour — roughly 70% of a human's throughput. After a year-long pilot, Agility reportedly has about seven commercial Digit units at Toyota. Seven.
- Tesla's Optimus entered 2025 with a stated target of 5,000 units and delivered, by most accounts, only hundreds — a miss of more than 90%. Elon Musk has said 1,000-plus units are now operating at Fremont as of early 2026 (a company claim worth treating as such), but even Tesla frames volume production of the Gen 3 body as a 2027 event.
- Unitree shipped an estimated 5,500 humanoids in 2025, making it the volume leader — but the vast majority of those went to research labs, developers, and entertainment, not to production lines doing paid labor.
Four of the most-hyped programs in the industry, and the picture is consistent: single-digit-to-low-hundreds units in real work, narrow tasks, heavy engineering supervision.
What's actually blocking scale — and it isn't the AI
The intuitive assumption is that the robots just aren't smart enough yet. The pilots suggest otherwise: the manipulation and locomotion on display are good enough to do useful, repetitive factory work today. Two harder constraints are doing the blocking.
First, the economics don't pencil at pilot scale. Agility's Digit reportedly carries a price around $250,000 — 8-to-12× Tesla's aspirational target for Optimus. At ~70% of human throughput on a single task, a quarter-million-dollar robot competes poorly with a worker it can only partially replace. The math improves dramatically at volume (which is the whole bet), but at ten or a hundred units it does not.
Second, manufacturing a humanoid at volume is its own unsolved problem. A modern humanoid is a ~10,000-component machine with custom actuators, and Tesla's own 90%-plus miss on its 2025 unit target is the clearest evidence that building thousands of them is a supply-chain and validation challenge, not a demo challenge. You cannot deploy a fleet you cannot yet build.
Our view: the announcement-to-scale gap is fundamentally an *economics and manufacturing* problem wearing an *AI* costume. The capability demos are real and improving fast; the binding constraints are unit cost and production ramp, and both move on a slower clock than a viral factory-floor video implies. Anyone reading these releases as "humanoids are available now" is reading them wrong.
How buyers should read the announcements
For a sourcing or operations buyer, the practical takeaways:
- Treat "deployment" headlines as capability milestones, not availability signals. "Robot X worked at Plant Y" almost always means one or a few units under supervision, not a product you can order at scale.
- Watch unit counts and hours, not vehicle counts. "30,000 cars" tells you about the plant's output; "one robot, 1,250 hours" tells you about the robot. The second number is the one that scales to your decision.
- For real automation ROI today, cobots and industrial arms remain the answer. The unglamorous truth is that a collaborative robot doing machine tending pays back in months with proven economics, while a humanoid doing the same task is still a pilot line item. We walked through the humanoid ROI math in detail in our factory humanoid ROI reality check.
- Give it 18–36 months before re-evaluating. The programs above are the on-ramp. If Optimus, Digit, or Figure hit real volume production and their per-unit cost falls toward six figures, the calculus flips — but that is a 2027+ question on current evidence.
Humanoids on the factory floor are not vaporware; the pilots are doing real, verifiable work. But a pilot is a proof of concept, not a hiring wave. Until the unit economics and the production lines catch up to the AI, expect more announcements than robots.
Sources
- BMW Group — humanoid robots in production (Spartanburg / Leipzig)
- Figure AI — F.02 contributed to production of 30,000 cars at BMW
- The Robot Report — BMW deploys Figure 03 after tests with previous version
- New Market Pitch — Tesla Optimus deployment tracker (2026)
- EVST — top humanoid robot companies to watch in 2026
