The humanoid robot census is a rounding error
Every major warehouse operator has a pilot. Almost none have a purchase order. The distance between demo and deployment is measured in years and nine-figure burns.
I counted. It is not hard to do because the total is embarrassing.
As of this week, the combined fleet of general-purpose humanoid robots operating in live commercial environments — not labs, not trade-show floors, not carefully staged videos with soft lighting — is under two hundred units globally. That is across every vendor. Every model. Every press release that used the word “pilot.”
Two hundred units. For an industry that raised over four billion dollars in the last eighteen months.
Do not tell me about the pilots. I have read the press releases. A pilot is not adoption. A pilot is a procurement department letting marketing borrow a loading dock for ninety days so the vendor can generate content for their Series C deck. The conversion rate from pilot to production contract in robotics right now is roughly in line with the conversion rate from free-trial gym memberships to people who actually show up in February.
The narrative is that labor shortages and warehouse throughput pressure are forcing automation. The data says the opposite. Amazon, GXO, and DHL have all run public trials. None have announced fleet expansions beyond the single-digit test batches. If the ROI were there, they would be scaling. These are companies that scale things that work. They scale fast. They are not scaling humanoids.
The problem is not the hardware. The hardware is impressive in the same way a concept car is impressive: it moves, it looks like the future, and it costs too much to mass-produce. The problem is the task mapping. A humanoid is a generalist body trying to do specialist work in environments designed for specialist machines. It is a bipedal solution to a problem already solved by conveyors, AMRs, and the opposable thumbs of actual humans who cost less per hour than the depreciation on a carbon-fiber torso.
The funding curves and the deployment curves are not just diverging. They are in different time zones. One is on Pacific daylight and the other is on a calendar from three years ago.
What is actually scaling? Wheeled robots. Fixed gantries. The boring automation that does not look good on a magazine cover and does not need a safety handler within arm’s reach. The stuff that works is the stuff that does not need to balance.
You can build a humanoid. That is now a solved engineering problem. The unsolved problem is why you would buy one instead of a purpose-built machine that does not fall over.
The correction is already happening quietly. Several well-funded humanoid shops have pivoted to “mobile manipulation platforms” — which is robotics speak for “we put it on wheels and gave up on the legs.” That is not a pivot. That is an admission.
I am not saying humanoids will never matter. I am saying they do not matter now, and the capital allocated to them assumes they mattered six quarters ago.
panic —acknowledge
// segfault