For years, humanoid robots were a demo. A carefully staged video, a controlled stage, a robot doing one trick well and falling over if you asked for a second. That phase is over. Figure, Tesla's Optimus programme and Boston Dynamics are now running thousands of general-purpose humanoid units inside commercial warehouses, doing real shifts on real floors.
The physical side of humanoid robotics was never really the hard part. Motors, actuators and balance control have been good enough for a while. What kept robots out of warehouses was the software: the ability to look at a cluttered shelf, recognise an irregular object it has never seen labelled that way before, and work out how to pick it up without being told exactly how in advance.
That is the piece that has caught up fast. Modern humanoid platforms are trained largely in simulation, the same "physical AI" approach NVIDIA has been pushing hard, where a robot rehearses millions of pick-and-place scenarios in a virtual warehouse before it ever touches a real shelf. The result is a machine that can walk, spot something out of place, and adapt its grip on the fly, rather than one that only works if the world matches its script exactly.
The obvious question is why a two-legged, two-armed robot at all, when a wheeled arm bolted to a conveyor does one job perfectly well. The answer is that warehouses were built for human bodies. Shelves, aisles, stairs, awkward corners and tools designed for hands all assume a human shape moving through the space. A general-purpose humanoid can, in theory, drop into that environment without the building being redesigned around it, and can be retasked from picking to sorting to loading without new hardware.
That versatility is the actual pitch. One robot platform that can be pointed at several different jobs beats a warehouse full of single-purpose machines, each expensive to install and useless the moment the job changes.
This is the first genuinely commercial proof that humanoid robots can operate at meaningful scale outside a lab or a trade show floor. Thousands of units running real shifts is a very different claim to a single robot folding a shirt for a keynote audience, and it changes how seriously warehouse operators, logistics firms and eventually retailers have to take the technology as a near term option rather than a someday one.
It also puts pressure back on the chip and software side of the industry that has been racing to support exactly this. Every claim about physical AI, on-device inference and simulation training gets tested the moment a robot has to work an actual eight hour shift without falling over or grabbing the wrong box. Warehouses just became that test.
More on Future Technology: NVIDIA's AI PC chips, Qualcomm's move into AI software, and the open secure AI alliance.
Get a plain-English tech briefing in your inbox every morning. Join the free Future Technology newsletter.