Humanoid Robots: Between the Polished Demo and the Hard Day's Work

A robot that walks on two legs has moved more than 100,000 totes at a logistics facility run by GXO, according to an announcement by Agility Robotics in November 2025. The figure is real, but it comes from the manufacturer's own announcement, and it tells us neither what the operation cost nor how much supervision it needed. That is the honest question to put to these machines: what do they actually accomplish, and at what cost?

Why the human form?

Our world is built to the measure of our bodies: stairs, handles, corridors and shelves. Hence the idea: a machine with two arms and two legs that enters places never designed for machines, without altering them. Yet the International Federation of Robotics (IFR) tempers the enthusiasm in its paper on the subject, saying that these robots will not compete with traditional industrial robots on speed, precision and reliability. It also observes that public interest is fed by science fiction and social-media clips, which may exaggerate how close the moment is.

What has changed?

Since the 1980s a paradox has been associated with Hans Moravec and his book Mind Children (1988): abstract thought is a thin layer resting on far older sensorimotor skills, so perception and movement seem easy to us while being among the hardest things for a machine.

What is new in recent years is the way of training. In July 2023 a team of 54 researchers published the RT-2 paper, in which they trained a model combining vision and language on web data and on robot movement trajectories together, writing actions as text tokens. The authors report that the robot handled new objects and followed commands absent from its training data, after about 6,000 trials. These are results reported by their authors from the laboratory, and they do not necessarily mean success in a factory running three shifts.

What the numbers say today

• MERICS (April 2026) estimated China's 2025 output at about 12,800 humanoid robots, used in training centres, laboratories, logistics and manufacturing, against 556,000 industrial robots in the same year.
• The report quotes the founder of UBTech, whose Walker series is being tested in electric-vehicle factories, as saying that these robots in Chinese factories are half as efficient as humans, and it describes their weakness in fine assembly and inserting small parts.
• The report puts the price of a robot at between 300,000 and 500,000 yuan, and considers that it must fall by at least half to be commercially viable.
• The IFR notes that a single battery charge does not yet last a full working day, and that work on an international safety standard for legged robots has only just begun; a standing robot must constantly keep its balance, and a fall or a power cut is a hazard.

And jobs?

There are no figures specific to humanoid robots alone. The nearest reference is the World Economic Forum's 2025 report, which estimates that about 170 million roles will be created and about 92 million displaced by 2030, a net gain of 78 million, and that about 59 of every 100 workers will need training or reskilling. These are estimates for the labour market under the influence of many forces, among them artificial intelligence, robotics and energy, and they are not a verdict on robots alone. Tasks within a job are likely to change before the job itself disappears.

What should be done?

Start from the task, not the machine: what tires workers, endangers them or repeats without value? Then measure the total cost, not the price of the device alone, involve the workers in the trial, and decide who bears responsibility if the machine errs. Preparing people matters no less than choosing the device. Trust in God does not replace taking the means: we try with caution, we measure, and we prepare people before machines.