Written by Martin Kardzhilov
Announced production capacity now runs an order of magnitude ahead of units doing paid work. The constraint is not manufacturing and it is not intelligence. It is the integration layer that sits between a capable robot and a productive one.
The clearest signal in humanoid robotics this month was a personnel announcement. Boston Dynamics appointed Rohit Prasad, who led Amazon’s Alexa organisation, as chief executive effective 7 October. Hyundai’s robotics unit, a company whose reputation was built on machines that could run and jump, hired a consumer software leader.
That is not a rebuke of hardware. It is an admission about where the difficulty now sits. Locomotion is largely solved. Manipulation is advancing quickly. What is not solved is the thing that decides whether any of it gets paid for: a robot that can enter an unfamiliar facility, learn a task, work alongside people, and be recovered when it fails.
Consider the arithmetic the sector currently publishes. Individual manufacturers have announced capacity of roughly 12,000 units a year. Order announcements across the sector exceed $300 million. Against that, one committed deployment — a thousand robots across a global factory network by 2032 — is a rounding error, and those units will spend time in a dedicated training facility in Germany before entering production at all.
The gap is not manufacturing. It is integration, and it has a specific shape.
A humanoid arriving on a floor has to be taught the tasks, supervised while it learns them, and recovered when it fails. The variables that determine whether that succeeds are human ones: whether an operator’s attention is on the cell, whether they are ready to intervene, how quickly the handoff is made, how the recovery is executed. None of that is captured by degrees of freedom or by benchmark scores.
This is the layer we build for at Brainlayer. Our BrainSim framework models human state during robot operation so that intervention and handoff can be predicted and improved rather than rediscovered on every shift. I would argue it is the missing measurement layer of the entire industry — the one that determines whether the capacity being announced converts into work being done.
Three things would show the sector is closing the gap. First, deployment metrics published as standard practice: intervention rate per thousand attempts, handoff latency, recovery success across operators and shifts. Second, vendors competing on those numbers rather than on demonstrations. Third, contracts priced against operating evidence rather than prototype promise.
The Chinese manufacturers have shown how fast hardware can scale, and the automakers have shown that they hold the manufacturing and cost advantages. Those advantages are real and they will compress hardware margins. What will not commoditise is the human side of the loop, because it is specific to every site, every task and every crew.
That is the honest strategic picture. The industry does not have a capacity problem. It has a deployment problem wearing a capacity problem’s clothes, and the companies that treat integration as the core engineering discipline rather than an afterthought will be the ones whose announced capacity turns into revenue.