Humanoid robots: less sci-fi, more spreadsheet

State of play: Figure, Tesla (Optimus), Unitree, Boston Dynamics, Agility — all scaling manufacturing, all running paid-ish pilots, all still doing narrow, supervised, industrial work. 'General-purpose robot butler' is not 2026. 'Robot that does one warehouse task reliably and is learning the second one faster than it learned the first' is. There is a meaningful difference between those two sentences, and most coverage of this space refuses to respect it.

It hides in the software. The hardware has been good enough for a while — bipedal locomotion is a largely solved engineering problem if you do not mind the price tag. The hard part was making a robot cope with an unscripted world without breaking down the moment something was slightly out of place. Neural networks are now doing a credible job of replacing the hand-tuned control code that used to make robots so brittle, and world models — AI systems that generate realistic simulated training environments — are radically cutting the cost and time of getting a robot ready for a new task.

Decart's Oasis 3, launched June 2026, is the clearest example of what 'world model for physical AI' actually means in practice: a system that generates photorealistic, physically consistent environments that robots can be trained in before going anywhere near a real warehouse. The practical upside is that expensive real-world training runs — requiring hardware, space, safety supervision, and tolerance for mistakes — get replaced with compute, which is cheap and getting cheaper. Every robotics company benefits from this, which is why Oasis 3 matters even to people who have never heard of Decart.

If you are evaluating this space — or the stocks — mute the highlight clips and read the boring metrics:

The companies that win will not have the flashiest demo. They will have the steepest learning-and-deployment curve and the most boring, repeatable shipping cadence. Everything else is marketing with servos.

The honest read on 2026: the hype is still ahead of the reality, but the gap is closing faster than most sceptics expected — and it is closing because of software, not because anyone built a better knee joint. The field is at the point where the interesting decisions are about deployment strategy, not about whether the technology works at all. That is a meaningful shift from even 18 months ago, and it does deserve a genuine upgrade in signal-to-noise ratio from the coverage.

One signal rises above the rest: the first company to publicly report second-deployment timelines measured in weeks — in a different task category than the first deployment — has crossed the threshold that separates early commercialisation from compounding AI. That is the moment the curve bends in a way that starts to look like the chatbot adoption curve, not the nuclear-fusion timeline. Until then, be sceptical of any specific unit-shipment claim that hasn't been independently verified, and be very sceptical of any 'general purpose' framing. Narrow, reliable, profitable — in that order.