45 Humanoid Pilots. Zero in Production. Here's What That Should Tell You.

A number stuck with me this week.

GXO has run 45 humanoid robot pilots. In production this year, the number will be zero.

That came from GXO CEO Patrick Kelleher. He also said GXO expects around 20,000 robots in operation in 2026. None of them will be humanoids.

To me, that number is not some big verdict. It is a location marker. It tells you where humanoids are on the curve.

People talk about this stuff like a robot shows up fully formed one day and then suddenly every warehouse in the country is using it. That is not how any of this goes. The middle part takes forever. First somebody proves the idea. Then somebody builds a version that sort of works. Then it breaks. Then it gets better. Then it works in a controlled setting. Then it works in one pilot but not the next one. Then the software gets reworked, the hardware gets reworked, the edge cases pile up, the expectations get reset, and eventually, after a lot more grinding than anybody wanted to admit at the beginning, it starts becoming normal equipment instead of an interesting experiment.

That was the path with AMRs. It was the path with a lot of flexible automation. None of that arrived on the floor looking mature. It took decades of invention, development, testing, reworking, and pilots before it started ramping into broader deployment. People forget that part because once the equipment is common, everybody rewrites the story in their head and acts like it was obviously going to work out. It was not obvious in the middle. The middle is messy every time.

That is why the GXO number feels normal to me. Forty-five pilots and zero in production this year is what the middle of the curve looks like. It looks like that because the technology still has real work to do. GXO expects humanoids to reach production in roughly two years. That sounds plausible. The reasons they are not there yet are also pretty plain. The machines still need better hands and more dexterity. They also need to work faster. Those are not cosmetic issues. Those are the whole job. If the robot cannot manipulate reliably enough or move through the task fast enough, it stays in pilot land no matter how much attention it gets.

And pilot land is not fake. That is the other thing people miss. A pilot is not a scandal. It is not a failure. It is the phase where the machine gets taught what the real world feels like. Real floors are full of variation, clutter, bad spacing, rushed people, ugly handoffs, and tasks that sounded simple right up until somebody tried to do them every day at production pace. That is what pilots are for. You find the part that looked easy in the lab and turns stupid in the field. Then you fix it. Then you do it again.

Humanoids are just earlier on that same path.

The difference this time is speed. Moore's law and the pace of AI compute are going to compress the timeline. Plainly, I expect humanoids to move faster through this curve than AMRs did. The learning loops are tighter. The software side is moving harder. The money and attention around it are massive. It probably will not take as long as the last big wave of flexible automation took.

But faster is not now. That is the part worth saying out loud, because people keep acting like those are the same thing. They are not. Faster just means the curve may be compressed. It does not change what is ready to do useful work on a Monday morning.

And what is ready on Monday is the equipment that has already been through its pilot era and come out the other side. A low industrial AMR moving material on a repeatable route. A cobot handling a task that is repetitive enough to make sense. An autonomous sweeper like the KEMARO K900 Gen II cleaning a floor without tying up a person on a broom. ASRS in the right application. None of that is exciting in the way humanoid videos are exciting. It is past that stage. That is why it matters.

So the takeaway here is not to park the operation for two years and wait for the finished version of the robot to arrive. That is how people talk themselves into doing nothing while telling themselves they are being strategic. If something on the floor is chewing up labor right now, the better move is to use the equipment that is already through the awkward middle and already proving itself in real work. Let the humanoids keep moving through their curve. They probably will move faster than the last generation did. Good. That still does not help the cart route, the dirty floor, the repetitive station task, or the storage problem sitting in front of you this week.

Some technologies are still in their pilot era. That is normal. Some are already past it. That is what matters when the decision has to survive contact with real operations.