21 September 2026 — Count the work, not just the robots. A machine can be bought, demonstrated and photographed without yet making anyone’s working day safer or more productive.
Reuters reports today that International Federation of Robotics figures record about 7,000 humanoid robots sold worldwide in 2025 for industrial and professional-service use. Many were bought for research or data collection, the federation says. The tally excludes consumer and military humanoids; medical robots are separate. Reuters saw the data ahead of publication. The underlying dataset was unavailable for this edition to inspect independently.
That is a new report about last year, not evidence of deployments completed this week. It tells us little about the reliability of a particular machine.
There is a positive industry case. In its 11 August position-paper summary, the federation argues that robotics can improve productivity and working conditions while creating new tasks. That broader argument deserves investigation. It does not establish the value of every humanoid product.
AE reading: three tests that go beyond the demonstration
Autopoietic ecology, or AE, asks how a capacity to act becomes possible and what happens to it through use. The EdgeLab overview suggests looking at the organised activity, including its human and material support. For robotics, that leads to three practical tests.
1. Does it work when the setting changes?
Hypothetical example: a warehouse robot moves uniform cartons onto a trolley. During an ordinary shift, one package is torn, the trolley has moved and a colleague needs the aisle. Who notices, decides and intervenes?
This is an illustration, not a report about a product. It separates a successful movement from a workable service. A fair trial would include the variations that staff actually encounter, rather than quietly preparing each task to match the demonstration.
Humanoid means human-like in form, not human-like understanding. Artificial intelligence can help select an action; a successful deployment still requires the action to fit its circumstances.
2. Is the machine reducing work, or moving it?
Record preparation, supervision, interruptions and repairs. If a person spends an hour arranging cartons so the robot can move them, that hour belongs in the evaluation. So do energy use and replacement parts.
The AE Concept Guide distinguishes present performance from the capacity for future activity. Expert support can make a trial possible without establishing that the proposed routine service will work after the expert leaves.
Support is not evidence of failure. Uncounted support is evidence of an incomplete account. The important question is whether knowledge becomes shared and the workload sustainable, or whether one specialist remains permanently on call.
3. Can experience change the next decision?
EdgeLab’s Research Programme distinguishes operating a technology from having authority over its use. Can workers pause the robot when conditions are unsafe? Does reporting repeated intervention change staffing, software or the task itself?
A stop button matters only within an arrangement that lets people use it and obtain a response. A trial that hears every complaint but changes nothing is not necessarily learning.
Give the optimistic case a proper test
Early trials may build capabilities that are not yet commercially useful. That is a reasonable alternative to reading modest initial performance as failure. Research equipment can be valuable precisely because it reveals what still needs work.
But improvement should become visible: more reliable operation, less total effort or safer work across successive trials. Compare those gains with simpler equipment, a redesigned workstation or additional skilled staff.
The next convincing evidence will come from sustained operation under ordinary conditions. Ask the people alongside the machine: what can you now do better—and what new work are you doing to keep it going?
Receive published EdgeLab posts by email: subscribe free using the opt-in form below.
Leave a Reply