Humanoid robots should enter through a bounded task trial

The buyer decision is about the task, not the body shape

A humanoid may make sense when access points, tools and routes were built for people and cannot reasonably be changed. That is a condition to test, not a reason to skip alternatives. Start with one physical job: its start and end states, object variation, site constraints, handoffs, required quality, observation period and current measured process. Then ask whether human-like reach or mobility solves a hard constraint that a simpler option cannot.

The International Federation of Robotics (IFR) expects humanoids to complement existing robot types and says mass adoption remains uncertain. Its 2026 trends review points buyers toward cycle time, energy use and maintenance cost against traditional automation. Those are comparison dimensions, not universal thresholds or a forecast for any supplier.

Put four alternatives against the same work definition

First, test whether a process or layout change removes the awkward movement, reach or handoff without adding a robot. Second, consider a fixed arm or purpose-built cell when the work can be presented at a station. Third, consider a mobile manipulator or transport system when movement between stations matters but human-like whole-body access does not. Only then put a named humanoid configuration into the comparison.

For each route, record the hard requirement it meets or misses, the evidence behind that judgment, site changes and integration ownership, expected supervision, and what would have to be tested. A vendor video, a planned pilot or a platform category is not a like-for-like comparison. The non-robot route deserves the same serious baseline as the machines.

What a production reference does—and does not—show

BMW Group reported on 21 September 2026 that its 2025 Spartanburg Figure 02 pilot retrieved and positioned sheet-metal parts, moving more than 90,000 components in roughly 1,250 operating hours. BMW also described barriers or partitions and connectivity work for that installation. This is useful first-party evidence that a bounded task was performed in one configured environment.

BMW's account does not provide a controlled comparison with a fixed cell, total cost, unsupervised availability or a safety certification for another site. The reported vehicles were assisted by the operation; the robot did not make whole vehicles. A buyer should ask for the same task, configuration, intervention and recovery detail before translating any reference into its own acceptance criteria.

Draw the supervision boundary before the demonstration

Name the site supervisor, process owner and safety owner. Define the permitted operating area; access and separation from people; stop and restart authority; power-loss and fall condition; intervention triggers; fault recovery; remote-support access; and conditions that stop the trial. WorkSafe Victoria's plant guidance considers hazards across installation, commissioning, use and decommissioning and prefers elimination and engineering controls over reliance on administrative supervision alone.

The applicable standard must be assessed for the complete installed configuration. ISO 12100's public abstract describes lifecycle hazard identification, risk assessment and reduction. ISO 10218-2:2025 covers industrial robot applications and cells, but its public scope expressly excludes mobility when a robot or manipulator is integrated with a mobile platform. ISO/CD 25785-1 was a committee draft under development at this review, not a published compliance certificate. None of those public summaries certifies a humanoid or replaces qualified site assessment.

Measure the complete operating loop

NIST's response-robot programme illustrates a useful testing discipline: controlled procedures, repeatable metrics and repetitions, with the user community setting its own performance thresholds. Those response-robot methods are not an industrial-humanoid standard and their numbers should not be imported as pass marks for a factory.

• Fix robot, tooling, software, task mix, layout, repetitions and observation windows before the test.
• Count good tasks, failed tasks, rework, intervention events and minutes of human assistance—not only successful motions.
• Measure cycle-time distribution and quality across object variation and exceptions, not just a best cycle.
• Record charging, energy, maintenance, faults, recovery and useful duty across the intended workload.
• Test the conventional alternative under the same task definition where feasible, and retain configuration and operator records.

The decision can be continue, revise, choose simpler or stop

Agree the site's hard safety controls and measurable advantage threshold before the trial. Continue only if the configured system meets them after supervision, charging, changeover, integration and support are counted. Revise or stop when safety evidence is incomplete or the promised task advantage disappears. A simpler process or robot can be the correct result; a spectacular demonstration need not become a purchase.

RobotAtom's task-alternative canvas records the comparison, supervision boundary and missing evidence without issuing a pass. It is a scoping aid, not a safety approval or a supplier ranking. A qualified team must assess the actual workplace and the applicable requirements.