XPENG has moved its IRON humanoid robot onto a production line. The more consequential question is what evidence will follow it out of the factory. Building repeatable machines, demonstrating useful physical skills and sustaining work at a customer site are separate achievements. The robotics industry needs to measure all three.
In a September 8 announcement, XPENG said it had commissioned its humanoid production lines and that an IRON robot had walked off them autonomously. It reported automation of more than 80% of core manufacturing processes. Mass production remains a target for the end of 2026, with initial deployments in its own stores and campuses and market launch and deliveries planned for 2027.
TENS analysis: The milestone is meaningful as a manufacturing development. Its broader significance becomes clearer when placed beside BMW’s account of a working humanoid pilot and the National Institute of Standards and Technology’s proposal for comparable physical testing. Together, these sources suggest a way to judge progress without treating every factory opening as proof of general-purpose autonomy.
Count robots made and work completed separately
XPENG’s automation percentage describes the processes used to manufacture robots. It does not measure the share of future customer tasks that IRON can complete independently. Those percentages would have different denominators, different failure conditions and different implications. A highly automated factory could produce machines that still need substantial supervision at their destination.
The announcement does not provide sustained customer-site operating hours, intervention rates or a task-completion dataset. That absence limits the conclusions readers can draw; it does not establish that the robot performs poorly. The appropriate next evidence would connect the manufacturing line to a defined deployment: how many units perform which tasks, for how long, and with what assistance.
There is also a distinction between commissioning equipment and delivering at volume. XPENG’s own timetable separates those stages. Readers should preserve that separation when assessing the news, because a future production target cannot serve as a current delivery total.
A factory pilot supplies a different kind of evidence
BMW’s February account of its Spartanburg pilot provides a useful comparison. The company said Figure 02 supported production of more than 30,000 BMW X3 vehicles within ten months, moving more than 90,000 sheet-metal components in around 1,250 operating hours. Its assigned work involved positioning parts for welding.
Those figures describe participation in a specific production process. They do not mean a humanoid assembled 30,000 complete vehicles, nor do they demonstrate competence across every factory job. BMW also described involving production IT, safety teams, process management and logistics during the initial testing.
The comparison changes what counts as a convincing robotics result. A robot’s contribution should be attached to the operation it actually performs. Vehicle totals provide context; completed placements, accepted quality and interruptions describe the work more directly. Without that distinction, a large production number can obscure a narrow but potentially valuable capability.
For IRON, the useful comparison is therefore about reporting discipline, rather than a ranking against Figure 02. The machines, tasks and deployment stages differ. A future IRON report could become more informative by specifying its working conditions and human support, even if its first assignments are simpler than a factory welding workflow.
Shared tests can make demonstrations comparable
NIST’s proposed Humanoid Robot Baseline Performance Benchmark addresses another part of the evidence gap. It outlines common mobility and manipulation tasks, including coordinated movement and handling, confined-space control and basic scene understanding. The proposal aims to establish comparable measures of minimum physical capability.
This is a proposed testing framework, not a published certification of IRON or a declaration that any particular robot is ready for deployment. NIST describes apparatus development with industry and researchers, with testing possible at manufacturers or participating facilities. Its page also says results would be collected under agreed data-sharing terms and aggregated.
The analytical value is the separation of physical capability from a polished demonstration. A common task makes comparison more meaningful only when the conditions are also clear. Readers would still need to know the allowed assistance, the number of attempts and the treatment of failures. An aggregate result can describe a field while leaving questions about an individual product unanswered.
The next milestone should expose the operating conditions
TENS Magazine’s reading of these sources is that humanoid progress needs three distinct records: manufacturing output, repeatable task performance and sustained deployment results. None can substitute for the others. Together, they would show whether a robot can be supplied consistently, perform the relevant action and remain useful in an operating workplace.
That framework leaves room to recognize XPENG’s progress without borrowing certainty from its ambitions. The production line advances the supply side of physical AI. The next substantive test is whether deployments produce evidence detailed enough to separate independent work from supervision, and repeatable performance from a successful moment.
Sources: XPENG; BMW Group; National Institute of Standards and Technology.
Image: NASA’s Valkyrie research robot, photographed in 2013; illustrative historical photograph, not XPENG IRON. Credit: NASA/Bill Stafford, James Blair, Regan Geeseman, via Wikimedia Commons. Public Domain. Cropped to landscape and resized.
