One Robot an Hour

In a low building south of San Jose, a clock runs differently than it does anywhere else in manufacturing. Every sixty minutes, a five-foot-six humanoid robot rolls off the line at Figure's BotQ facility, fully assembled, ready to be taught a job. Four months earlier, the same line produced one robot a day. The factory did not get bigger. The math of the thing changed.

For a decade, the humanoid robot was a creature of the demo reel — backflipping for a YouTube clip, folding a single shirt for a keynote, then quietly powering down once the cameras stopped. The gap between the spectacle and the loading dock was enormous, and most serious people assumed it would stay that way for years. What makes 2026 worth marking is not a cleverer backflip. It is that the gap is closing from both ends at once: the bodies are now being mass-produced, and the minds to run them are being given away.

The Body
Tooling the humanoid

The number that matters at Figure is not a benchmark score. It is the word injection-moulded. The company's third-generation machine, Figure 03, was designed not to impress but to be cheap to build — its parts come from the same industrial processes that produce car bumpers and power-tool housings: die-casting, metal injection moulding, stamping. That is the unglamorous pivot a technology makes when it stops being a research project and starts being a product. You stop hand-fitting prototypes and start tooling for volume.

Figure's BotQ facility, unveiled earlier this year, was built around that premise. Its first line is rated to produce up to twelve thousand humanoids a year, with a stated goal of a hundred thousand machines over four years. Reaching one robot per hour — a twenty-four-fold jump from one per day, achieved in under four months — is the kind of ramp that, if it holds, compounds into something genuinely strange. Exponential curves are easy to draw and hard to believe until you are standing next to one.

Figure is not alone on the floor. Boston Dynamics has begun deploying its all-electric Atlas, retiring the hydraulic showman of the 2010s in favour of a machine built for shifts rather than stunts. The latest State of Robotics survey counted roughly a dozen credible humanoid programs and a market it pegged near thirty-eight billion dollars. The center of gravity has moved from "can it walk?" to "what is the unit cost, and how many can you make this quarter?"

The Mind
Foundation models for flesh

A body without a brain is a mannequin. The second half of the story — the reason the bodies suddenly have something to do — is that the software stack for controlling them has gone through the same revolution that language did three years ago. The field has a name for it now: physical AI, and its workhorse is the vision-language-action model, or VLA.

The idea is deceptively simple. A large model is pretrained on oceans of video, language, and motion until it carries something like intuition about how objects behave, what a request means, and how a limb should move to satisfy it. Then it is fine-tuned on a specific robot. NVIDIA's GR00T family — the open, customizable humanoid foundation models released in 2025 — pairs a vision-language backbone for reasoning with a diffusion-transformer module that generates the actual motor commands. Google DeepMind's Gemini Robotics added 3D spatial perception and, in a quietly radical move, an on-device version light enough to run on the robot itself, no cloud tether required.

In 2026 NVIDIA went further, releasing Cosmos models it describes as an omnimodel for physical AI: a single mixture-of-transformers architecture folding vision reasoning, world simulation, and action generation into one system. The phrase to sit with is "world simulation." These models do not just react; they run a rough physics of the room forward in their heads before they move, the way you flinch before a glass actually hits the floor.

This decoupling is why the curve feels different this time. When the mind is a shared foundation model, every improvement to it upgrades every robot that runs it, all at once, retroactively. A factory floor of machines can get measurably smarter overnight without a single new bolt. That is not how forklifts work. It is how phones work.

The Floor
Where the million already are

It would be a mistake to picture all of this as humanoids, because the quiet majority of working robots have no face at all. In June 2026, Amazon's fleet of warehouse machines crossed one million units — wheeled movers, robotic arms, sortation systems — coordinated by a logistics model the company calls DeepFleet, which it credits with a ten percent gain in travel efficiency across the network. Its Sequoia storage system claims a seventy-five percent improvement in how fast inventory is identified and stowed.

A million robots is a number worth holding still for. It means the embodied-AI revolution is not a future event to be anticipated; it is a deployed reality being optimized. The humanoids arriving now are not the beginning of automation. They are the general-purpose layer being added on top of a world that has already been substantially mechanized — the part that can finally climb the stairs, open the unlabelled box, and pick up the thing that was never designed to be picked up by a machine.

The Reckoning
What a working robot asks of us

It is tempting to greet all this with the usual binary — utopia of liberated leisure, or dystopia of mass redundancy — but both are too tidy. The more honest observation is that a general-purpose robot is an unusually concrete object to argue about. You cannot wave away a machine that costs a known amount, makes a known number of units per hour, and can be pointed at a specific task. The conversation about labour, dignity, and what work is for stops being abstract the moment one of these things is standing in a warehouse aisle next to a person doing the same job for a wage.

There are sober reasons not to over-extrapolate. A robot per hour is a production claim, not an audited one, and history is littered with manufacturing ramps that flattened the moment they met reality. Humanoids remain clumsy at the long tail of the physical world — the slippery, the deformable, the genuinely novel. Battery life, safety certification, and the unglamorous cost of maintenance will throttle deployment long after the press releases stop. The mind may be a downloadable foundation model, but the body still wears out, and someone has to fix it.

And yet the shape of the thing has changed in a way that is hard to un-see. For the first time, the bottleneck in robotics is not "can we make it think" or "can we make it move." It is logistics: tooling, supply chains, batteries, certification — the ordinary friction of building physical objects at scale. That is a profoundly different problem than the one the field faced even two years ago, and it is the kind of problem industrial civilisation is unusually good at grinding down.

The deeper question the year leaves open is not whether the robots are coming — a million of them are already here, and the humanoids are arriving by the hour. It is what a society does when general-purpose physical labour becomes a thing you can manufacture, schedule, and download an upgrade for. We spent the last three years asking what it means for a machine to write and reason. The next three will ask something stranger and more intimate: what it means for a machine to simply do, in the same rooms, alongside us, at the pace of a clock that does not need to rest.

Sources
References & further reading

1. Figure — Introducing Figure 03
2. Figure — BotQ: A High-Volume Manufacturing Facility for Humanoid Robots
3. Figure — Ramping Figure 03 Production
4. The Robot Report — Figure AI unveils BotQ manufacturing facility
5. Interesting Engineering — Figure claims one humanoid robot per hour
6. NVIDIA — Isaac GR00T Reference Humanoid Robot for Research
7. NVIDIA — New Physical AI Models and Next-Generation Robots
8. NVIDIA Blog — National Robotics Week 2026: Physical AI research
9. Robotics Center — State of Robotics 2026: $38B Market, 12 Humanoids, VLA Adoption
10. Robotics Center — Physical AI in 2026: What It Is and Key Models
11. KraneShares — Humanoid Robotics in 2026: From Pilot to Platform
12. RAISE Summit — Leaders in Robotics, Humanoids & Physical AI 2026
13. Humanoids Daily — The Shape of Scale: Figure's Exponential Ramp
14. Crescendo AI — Latest AI News, June 2026