Why humanoid robots are stuck on hardware time

Humanoid robotics is a two-speed industry pretending to be one. Only the fast half gets covered. The software that teaches a humanoid to walk iterates on demo-cycle timelines; the half that turns a learned policy into a standing body runs on a clock closer to automotive suppliers than to artificial-intelligence labs, and that gap is what is keeping humanoids off factory floors.

Kollmorgen's framing, set for an October RoboBusiness session, names the mismatch directly. Yoshi Umeno's argument is not that the AI is wrong; it is that the motion side moves from paper to functional sample in months and to volume production in years, while the software side ships in weeks, and the years-long side is the binding constraint on real deployment.

The first 2025 production figures from Interact Analysis make the constraint visible. Global humanoid output topped 20,000 units, up tenfold year over year, and only about 10% of those machines are in real-world use, with most still small-scale proofs of concept. Roland Berger's 2026 outlook projects operating cost near USD 2 per hour and a baseline OEM market near USD 300 billion by 2035, with an optimistic case near USD 750 billion and a path to USD 2-4 trillion by 2050. None of those numbers move unless the motion-hardware clock moves with them.

The humanoids that actually ship at volume will be the ones whose OEMs locked in motion-hardware decisions early, not the ones with the best demo videos. And the loudest version of this diagnosis is being amplified by a motion-hardware vendor with a direct commercial interest, so the frame is real but the volume has to be discounted.

Read the next demo video as a software update; read the next factory announcement as a hardware clock, and weight the stories accordingly.