This was a busy week for robots that look like us. A German firm, Neura Robotics, unveiled a foundation model it says lets a humanoid fold laundry for eleven straight minutes. A Shanghai startup opened its entire robot body and software to the public as open source. Intel staked out a position as a full-stack platform for factory automation, and a Swedish company proposed squeezing a whole robot-run factory into three shipping containers. Yet for all the announcements, the most honest measure of where this technology stands came from a quieter number out of South Carolina: at BMW's Spartanburg plant, a Figure 02 humanoid logged roughly 1,250 operating hours and handled more than 90,000 parts across a ten-month pilot.
That figure matters because it is boring. It is not a viral clip of a robot dancing or boxing; it is a record of repetitive work done inside a real production line, measurable in hours and handling accuracy. The gap between those two things, the dazzling demo and the dependable shift, is the whole story of physical AI right now. Hesai's chief executive David Li, launching a 3D mapping camera this week to feed robots better training data, said the quiet part aloud: most robots are still "confined to low-impact functions like greeting guests, waving or dancing," and remain limited in solving real productivity problems.
Intel frames the obstacle as a "deployment gap." In its own commissioned survey, about 60 percent of leaders expect to run robot fleets within five years, but only 40 percent have a strategy for managing a mixed human-and-robot workforce. Intel's pitch, a single chip handling control, perception and AI inference, is aimed squarely at the unglamorous engineering that keeps robots stuck in the lab. Tellingly, the company reports 130-plus design engagements, then concedes those are qualification-stage, not production deployments. The distinction is the point: qualifying hardware is not the same as running it profitably on a factory floor.
The economics are unforgiving, and a thoughtful industry survey this week spelled them out. Acquisition cost is only the beginning; depreciation, maintenance, energy, integration, software and supervision all pile on, and productive utilisation decides everything. A robot that is operational for eight hours but spends them charging, switching tasks or waiting for a human does not earn its keep. That is why so many deployments, from GXO's warehouse Digit robots to Promise Robotics' new deal to help build 1,250-plus homes for D.R. Horton, are being sold as a service rather than a purchase: the customer buys finished work, and the vendor carries the risk of reliability, spare parts and downtime.
Behind all of it sits a data problem that money alone cannot solve. A chatbot can learn from the text of the internet; a robot needs examples of movement, friction, grip and balance that only exist in the physical world. An investigation this week described a new global labour market of people in more than 15 countries recording household chores and teleoperating machines to generate that training data. The humanoid economy, as investor Modar Alaoui put it at a book launch, is "less about what a robot can do once for a camera, and more about what it can do every day for a business." The machines are finally real. Whether they are worth it is a question only their operating records, not their demos, will answer.