Humanoid pilots have produced real data at last, and it is less exciting than the videos.

Humanoid robots spent three years being demonstrated and have now spent about eighteen months being deployed in small numbers on real sites. The deployments are where the useful information is.

The consistent finding is that competence is real but narrow. On a defined task, in a defined cell, with consistent lighting and consistent part presentation, the machines perform. Change the part presentation and performance degrades in ways that are hard to predict from the demonstration.

Uptime is the honest metric

Fixed automation in a European factory is expected to exceed ninety five percent availability. Reported humanoid availability in pilots sits far below that, and the gap is dominated by intervention events rather than hardware failure.

Why anyone persists

Because the alternative for many tasks is not fixed automation, it is a person, and the task is one nobody wants. Machine tending in awkward geometry, trailer unloading, and repetitive transfers between stations that were never designed for a conveyor.

A humanoid does not need to beat a purpose built cell. It needs to beat the cost of redesigning a building.

The realistic near term

Fleets in the low hundreds at logistics and automotive sites, with a technician per small fleet, on tasks selected specifically for tolerance of intervention. That is a real industry. It is not the labour substitution story the funding rounds imply.