Humanoid Robots Deployed in 2026: The Real Numbers

How many humanoid robots are actually deployed doing paid work in 2026? Hundreds to low thousands of units globally — not the tens of thousands circulating in market forecasts. The verified deployments are concentrated in three companies: Figure AI has 40 Figure 03 units at BMW's Spartanburg plant, Agility Robotics' Digit has moved over 100,000 totes on a commercial contract at GXO Logistics, and Boston Dynamics' Atlas is running material handling at Hyundai's Metaplant in Georgia. Tesla's Optimus, despite a Fremont line engineered for a million units a year, had not begun production as of mid-July 2026.

That gap between the forecast number and the filed number is the whole story. 2026 is the year humanoids genuinely crossed from demo to payroll — and also the year the numbers being quoted in investor decks stopped matching anything a company actually claimed.

This is a deployment audit. Every figure below traces to a company statement, a plant disclosure or a documented pilot. Where a number is a projection rather than a count, it is labeled as one.

Key Takeaways

• Figure AI deployed 40 Figure 03 units at BMW Spartanburg in late June 2026, following an 11-month Figure 02 pilot that logged 1,250+ operating hours and loaded 90,000+ sheet-metal parts.
• Agility Robotics' Digit surpassed 100,000 totes moved under a Robots-as-a-Service contract at a GXO fulfillment center, plus a year-long commercial pilot at Toyota's Ontario plant.
• Boston Dynamics' Atlas became the first enterprise-grade humanoid deployed at Hyundai's Metaplant Application Center in Georgia.
• Tesla repurposed its Fremont Model S/X line for Optimus Gen 3 with a designed capacity of 1 million units per year, but production had not started as of mid-July 2026.
• Widely circulated claims of tens of thousands of deployed humanoids do not survive contact with company filings — the real global figure is hundreds to low thousands.

Which humanoid robots are actually deployed in 2026?

Four programs have documented, sustained commercial operation rather than demonstrations. The distinction that matters is whether the robot works a scheduled shift against a production metric, or whether it appears in a video.

The Figure 02 pilot at BMW is the single best-documented humanoid deployment in existence, and the numbers are worth reading carefully: an 84-second cycle time with above 99% placement accuracy across 1,250 operating hours. That is not a robot doing a party trick. That is a robot hitting a takt time on a line that built more than 30,000 BMW X3 vehicles during the pilot window.

BMW has since announced humanoid deployment in production in Germany for the first time, which is the strongest available signal that the Spartanburg results survived internal review.

Why hasn't Tesla Optimus shipped?

Because building a humanoid is a supply chain problem, not a software problem, and Tesla is discovering it at scale. The company repurposed its former Model S/X line at Fremont into an automated Optimus Gen 3 line with a designed annual capacity of 1 million units. Elon Musk walked the newly installed line on July 1, 2026 and pushed back on suggestions that output had quietly begun.

The stated blocker is roughly 10,000 entirely new components with no mature supply chain behind them. Initial production is expected to be "extremely slow," with genuine mass production not anticipated until a dedicated Texas facility comes online in 2027.

This is the most instructive data point in the whole sector. Tesla has more manufacturing capability than any other humanoid company by a wide margin, and it is still gated on actuators, gearboxes, and sensor suppliers that do not yet exist at volume. Anyone forecasting a step-change in humanoid deployment before the component ecosystem matures is forecasting against physics and lead times.

What does a real deployment actually require?

More hand-holding than the marketing suggests. Across documented programs, three requirements recur:

1. On-site vendor engineering. Most current humanoid deployments include engineers from the robot maker physically present at the customer site. That is a services business wearing a product business's clothing, and it caps how fast anyone can scale.
2. Custom environment preparation. Fixtures, lighting, floor markings, part presentation. The robot is general-purpose in principle; the installation is bespoke in practice.
3. Significant integration work. Figure 03 specifically ships with tighter integration into automotive manufacturing execution systems — that integration is the product, as much as the hardware is.

The "general-purpose robot you can drop into any factory" remains aspirational. What exists in 2026 is a very capable robot plus a deployment project, and the deployment project is measured in months.

How fast can humanoids actually be manufactured?

Figure's BotQ facility in California reached a production rate of one Figure 03 per hour as of May 2026 — a 24-fold throughput increase from 120 days earlier — and the company passed 1,000 units built by late July 2026.

Sit with that rate for a moment. One robot per hour, run continuously, is roughly 8,760 units per year. That is an extraordinary achievement relative to where humanoid manufacturing was in 2024, and it is a rounding error against the labor forces these robots are pitched at replacing. A single large automotive plant employs more people than the entire global humanoid fleet.

The honest projection is not "humanoids replace warehouse labor by 2028." It is "humanoids become a normal line item in capital expenditure for specific, well-characterized tasks, and the fleet compounds from a very small base."

The software is finally ahead of the hardware

For a decade, humanoid robotics was bottlenecked on control. That inverted in 2026.

Google DeepMind's Gemini Robotics 2, announced July 30, 2026, brought whole-body control, advanced dexterity across both multi-finger hands and parallel grippers, multi-robot coordination, and — the commercially significant claim — adaptation to entirely new robot bodies in hours rather than months. If a foundation model can port across embodiments that quickly, the software cost of adding a new robot platform collapses.

Which relocates the bottleneck squarely onto hardware: actuators, batteries, thermal management, and the 10,000-component supply chain Tesla is currently wrestling with. It also relocates it onto compute. A humanoid fleet running continuous multimodal inference is a persistent, per-second workload, not a bursty per-query one — which connects directly to the infrastructure squeeze we cover in the custom AI chip revolution and South Korea's $880B chip investment plan. On-device inference stops being an optimization and becomes a fleet economics requirement, for the same reasons laid out in running LLMs locally.

The gap nobody audits: where the big numbers come from

Here is the analysis missing from almost all humanoid coverage. Widely circulated deployment figures in the tens of thousands do not originate from the companies they describe. They are typically derived by taking a stated production capacity, multiplying by an assumed utilization, and reporting the result as a deployment count.

Tesla's Fremont line is the clearest example: a designed annual capacity of 1 million units, with production not yet started, routinely becomes "Tesla is building a million robots." Those are different sentences describing different realities.

A simple discipline fixes this. For any humanoid claim, ask three questions:

• Is this a count or a capacity? Installed line capacity is not output.
• Is this a pilot or a contract? A pilot ends. A Robots-as-a-Service contract with 100,000 totes moved is revenue.
• Who is the source? If the company itself has not stated the number, treat it as an estimate regardless of how confidently it is reported.

Applied consistently, the picture that emerges is genuinely impressive and much smaller than the headlines: a few hundred robots doing real, measurable, paid industrial work, with production rates that are compounding fast from a tiny base. That is how every durable technology transition has actually looked from the inside — and it is why IEEE Humanoids 2026, held in late July, made labor displacement its defining theme even at these volumes. The trajectory, not the current count, is what the field is arguing about.

Frequently Asked Questions

How many humanoid robots are deployed in 2026? Hundreds to low thousands of units in active commercial use globally. Verified deployments include 40 Figure 03 units at BMW Spartanburg, Agility Digit units at GXO Logistics and Toyota Canada, and Boston Dynamics Atlas at Hyundai's Metaplant in Georgia.

Which company has the most humanoid robots deployed? Figure AI has the strongest verified record, with 40 Figure 03 units at BMW Spartanburg and more than 1,000 units built by late July 2026. Agility Robotics has the longest-running commercial contracts, including 100,000+ totes moved at GXO.

Is Tesla Optimus in production? Not as of mid-July 2026. Tesla repurposed its Fremont Model S/X line for Optimus Gen 3 with a designed capacity of 1 million units per year, but production had not begun, blocked by roughly 10,000 new components lacking a mature supply chain. Mass production is not expected until a Texas facility opens in 2027.

What jobs are humanoid robots actually doing? Primarily material handling, part loading and order fulfillment in automotive and logistics. Figure 02 loaded sheet-metal parts at an 84-second cycle time with above 99% placement accuracy; Agility's Digit moves totes in warehouse fulfillment.

Are humanoid robots replacing workers yet? Not at meaningful scale. The global deployed fleet is smaller than the workforce of a single large factory. IEEE Humanoids 2026 made labor displacement its central theme based on trajectory rather than current headcount impact.

Why are humanoid deployment numbers so inconsistent? Because many published figures convert production capacity into implied deployment counts. A line engineered for a million units per year that has not started production is frequently reported as though the robots exist.

The verdict

Humanoid robots crossed the pilot threshold in 2026. That is real, it is documented, and the BMW Spartanburg cycle-time data is the most convincing evidence the field has ever produced. It is also true that the entire global fleet would fit in a mid-sized office building.

Our read: judge this sector on production rate and contract renewals, not on unit counts or demo videos. Figure going from 1 robot per 24 hours to 1 per hour in 120 days matters far more than any single deployment number, because compounding manufacturing throughput is what turns hundreds into tens of thousands. Watch BotQ's rate, watch whether Tesla's Texas facility lands in 2027, and discount every figure whose source is not the company that built the robot.

For the software half of this story — the model that may finally make new robot bodies cheap to support — read our breakdown of Gemini Robotics 2.

The robots are real, they are working shifts, and there are far fewer of them than you have been told. Both things are true, and only one of them is in the forecast.