7,000, 19,100, 22,000+ or 40,000+? How to Count the Humanoid Market

IFR data supplied to Reuters puts global 2025 humanoid sales for professional and industrial applications at around 7,000. This covers a different period and scope from H1 2026. It is not a correction to the earlier trackers. Source: Reuters / IFR.

Four numbers do not mean four measurements of the same population

The earlier version of this article compared approximately 19,100 global shipments from Smart Analytics Global, more than 22,000 from Counterpoint, and more than 40,000 shipments in China reported by Global Times from a Committee of 100 report. Those three estimates referred to the first half of 2026. The IFR figure adds context, but it does not automatically resolve differences in how researchers define and measure the market.

Each number below therefore retains its own source, period and measurement label. The table is not a credibility ranking or an attempt to locate the “real” market halfway between competing values. Its purpose is to prevent comparisons that the available data cannot support. A proper reconciliation would require model-eligibility rules, the manufacturers covered and the transaction-recognition point used by each research team.

Production, shipments and sales happen at different points

A manufactured robot can remain in inventory, undergo testing or wait for an integrator. Shipment can refer to departure from the factory, transfer to a distributor or arrival at the customer. Sale is a transaction measure whose recognition depends on the relevant agreement and reporting method. It should not be assumed that these three labels identify the same population of machines on the same date.

In a hypothetical supply chain, one robot is manufactured in December, shipped to an integrator in January and accepted by a factory in March. Each event could correctly appear in a different report. Adding those records together would nevertheless turn one physical machine into three. A related error occurs when orders are added to shipments: part of the order may already have been delivered while the remaining units are still awaiting production or acceptance.

Do not calculate growth from incompatible reporting periods

Comparing a full year with a later half-year can provide context, but it cannot establish a market growth rate by itself. Product scope, geography and recognition rules need to match first. Even two reports labelled “global humanoid shipments” can include different platform configurations or revise estimates at different times. An earlier publication may also lack manufacturer submissions that became available to a later one.

There is no sound reason to average 19,100, 22,000-plus and 40,000-plus merely to produce one convenient figure. An average would need evidence that the researchers were measuring the same population and that the nature of their measurement errors was understood. Here, the gap between a global estimate and an estimate for one country is itself a reason to check scope first. Correct arithmetic cannot repair incompatible inputs.

Define the humanoid before counting it

A dataset can depend on locomotion, the number of manipulators, intended use and whether the system is complete. A research platform, a wheeled machine with arms and a bipedal industrial robot need not enter the same category in every report. Asking for examples of included and excluded models is therefore a useful first step. It provides a testable definition rather than an argument about which machine looks most convincingly human.

This distinction also matters when a manufacturer reports a total covering its entire portfolio. Counts of inspection robots, quadrupeds and humanoids cannot automatically be assigned to the last category. Documentation for the actual model remains the starting point; our Unitree G1 database entry is one example. A product entry establishes the identity and intended role of a platform, not the number of units sold.

Deployment and productive operation need another layer of evidence

A robot delivered to a factory may still need installation, configuration, training or acceptance. Commissioning a workstation is meaningful progress, but it does not demonstrate months of reliable operation. Evaluation should separate the number of physical machines, the number of accepted workstations and the hours during which they completed the customer’s required task. Only that last description starts to establish process value rather than movement of hardware through a supply chain.

The distinction is useful when reading Hyundai and Boston Dynamics’ plans to commercialize Atlas. A future fleet plan, manufacturing capacity and the current deployment-preparation stage are different facts. They should not be added to one chart of active robots. Similarly, a multi-year order cannot be presented as sales completed in the month the agreement was signed unless the available information actually establishes that result.

Human involvement is a process parameter, not a footnote

Autonomy needs a description of operating conditions. Does a person control motion, approve successive actions, help only with exceptions or simply supervise the fleet? Each model can have a purpose, but its cost and limitations differ. A video without a visible person does not answer the question. Interventions per task and the time required to restore ordinary operation are more useful measurements.

We explore a related issue in our coverage of Spirit AI and the boundaries of autonomy claims. A market dataset can count a robot sold for research without calling it an autonomous production worker. Research equipment is still a real product and can create value. The problem begins when the meaning of a number changes between the source table and a headline about replaced jobs or the readiness of the entire industry.

How a European buyer can use the numbers

A practical assessment should begin with a specific task and support requirements, not the largest worldwide volume. Define acceptable completion time, error tolerance, employee-assistance time and recovery conditions after a fault. Then look for references involving a comparable process. A large market does not guarantee that a product fits the application; a small shipment total does not rule out a useful specialized deployment.

Every future figure added to this article should carry its date, source, scope and unit of measurement. A correction to the same report replaces its earlier value, while a new report with a different scope belongs in another row. That approach does not produce one spectacular total, but it lets readers identify what actually changed. It is a stronger basis for decisions than combining production, shipments, future plans and operating hours into one supposed measure of success.