The Humanoid Robot Boom Is Smaller Than It Looks

Watch enough robot videos and it can feel as if the humanoid era has already arrived.

Machines run, carry parts, sort objects, work beside people and move through environments designed around the human body. Investment is rising, forecasts are enormous, and every successful demonstration compresses years of engineering into a few seconds of apparently effortless motion.

Then came a useful denominator.

Data from the International Federation of Robotics, reported by Reuters, put global humanoid robot sales at roughly 7,000 units in 2025 for industrial and professional-service use.

That number does not make the category trivial. It makes the stage of the market easier to see.

Robotics is already large. Humanoids are still a small subset

Seven thousand units sounds different when placed next to the established robot economy.

The IFR says 542,000 conventional industrial robots were installed in 2024. More than 199,000 professional service robots were sold in the same year.

The categories and reporting periods are not identical, so they should not be treated as a direct market-share calculation. They are a scale denominator.

Factories already buy robots when the economics are convincing. Warehouses, hospitals, farms and commercial facilities already use specialized machines for narrow tasks. Automation itself is not waiting for humanoids to become viable.

The humanoid proposition is more ambitious: one roughly human-shaped machine could eventually perform many jobs in spaces that were built for people.

That promise is powerful because replacing the environment is expensive. Shelves, doors, tools, carts and work cells already assume human reach and movement. A machine that fits that geometry could, in theory, enter existing infrastructure with less redesign.

The difficulty is that fitting the space is not the same as doing the work.

The strongest evidence is not a demo

BMW's use of Figure 02 is important because it shows that productive humanoid work is not hypothetical.

Figure says its robots at BMW's Spartanburg plant contributed to production involving more than 30,000 BMW X3 vehicles, loaded more than 90,000 sheet-metal parts and accumulated more than 1,250 operating hours.

Those numbers matter because they describe repeated activity inside a real production process rather than a staged showcase.

But the details are just as important as the headline.

The task was narrow: moving and positioning sheet metal within a defined manufacturing workflow. Figure's own account discusses placement success, intervention targets, calibration, thermal constraints and hardware reliability.

That is not a weakness in the example. It is the point.

Commercial value can arrive before general-purpose intelligence. A humanoid does not need to wander around a factory deciding what to do next in order to save labor on a carefully chosen task.

Why successful deployments can still be pilots

BMW's newer work with Hexagon's AEON humanoid in Leipzig shows how conservative the path to deployment remains.

BMW describes a staged process: evaluate the technology, test it with real production use cases, move into factory conditions, then proceed toward an actual pilot.

That sequence is not administrative overhead. It is the commercialization process.

A factory operator cares about cycle time, safety, uptime, integration, serviceability and recovery from failure. A robot that succeeds for a minute on camera but requires frequent human intervention is a very different product from one that can operate through thousands of productive hours.

This is why pilot announcements should not automatically be read as proof of fleet-scale adoption. They are evidence that the industry is learning how to cross that gap.

The bottleneck is moving from body to brain

Humanoid hardware has improved quickly. The mechanical side of the industry can now produce increasingly convincing locomotion, balance and manipulation.

The harder problem is generalization.

A useful general-purpose machine needs to recognize unfamiliar objects, interpret instructions, manipulate items that do not arrive in exactly the same orientation, recover from mistakes and switch tasks without a full engineering reset.

Spirit AI co-founder Gao Yang told Reuters that the "brain" is the weakest link in the robotics stack. The distinction is important because simple, controlled tasks can reach high success rates much earlier than messy, open-ended work.

That suggests the next major competition in humanoids may be less about whether the body can walk and more about whether the software can make the same body useful across many jobs.

China can change the scale quickly

The small current installed base should not be confused with a permanently small market.

Reuters reported that China accounted for roughly 95% of global humanoid shipments in 2025. The country combines robotics companies with deep manufacturing capacity across motors, batteries, electronics and other hardware inputs.

XPeng's move to commission a production line for its IRON humanoid is another signal that the supply side is moving beyond prototype workshops.

Production capacity, however, is not the same as customer demand.

A factory capable of building many robots matters only if enough customers find jobs where the machines can earn their keep. That is why the transition from prototype to pilot to fleet is more informative than a production-line announcement on its own.

Capital is beginning to ask the same question

The market is also becoming more skeptical about the quality of demand.

Reuters reported that Chinese regulators were slowing some humanoid-robot IPO approvals and examining whether revenue linked to state-backed projects represented durable independent commercial demand.

That does not prove the sector is a bubble. Government-backed projects can accelerate learning and industrial development.

It does show that the question is changing.

Early in a technology cycle, raising capital is itself a signal of confidence. Later, investors and regulators start asking whether customers are returning, whether deployments save money, and whether revenue persists without special support.

Humanoid robotics is beginning to enter that second conversation.

The metric to watch is repeatable work

The next important signal is not another backflip or keynote demonstration.

It is the path from prototype → pilot → fleet → repeat order.

How many productive hours can the robot accumulate? How often does it need human intervention? Can the same platform learn another task without months of custom engineering? What happens when hardware fails? Does the customer expand the deployment after the pilot?

BMW's Figure deployment matters because it starts to answer those questions with operating hours and parts moved. BMW's AEON program matters because it shows how careful the path to production still is. XPeng's line matters because manufacturing scale is beginning to arrive.

And the roughly 7,000-unit market matters because it keeps all of those examples in perspective.

The humanoid boom is smaller than it looks today.

That does not mean the future is small.

The category becomes truly large when the demonstration stops being the product and the work becomes routine.