Humanoid Robots Enter Production: What Physical AI Can Actually Do in a Factory

Inzonex ResearchPublished Sources checked 12 September 20267 minute read

A humanoid robot can now be built on a production line. That is what was announced in September 2026, and it is a manufacturing milestone rather than a factory-work milestone. No public evidence yet shows a humanoid holding a production cycle time in someone else's plant. For most manufacturers the next step is to list which tasks are unstructured, because those are the only ones a general-purpose machine could justify.

What was announced

On 8 September 2026 two announcements landed on the same day, and together they mark the point where humanoid robots stop being a research demonstration and start being an industrial supply chain.

XPENG stated that it had commissioned a production facility for humanoid robots, and that its IRON robot completed manufacture and walked off the line under its own power. Arm announced a Robotics Capability Framework for what it calls physical AI, developed with an ecosystem it describes as more than 80 companies.

Company claims, not verified results

XPENG states that more than 80% of core processes on the line are automated, that IRON has 76 degrees of freedom with 21 in each hand, and that it runs three of the company's own chips providing up to 2,250 TOPS of effective computing power. XPENG says mass production starts at the end of 2026, with commercial rollout beginning in its own stores and campuses and market delivery planned for 2027.

These are the manufacturer's statements about its own product. They describe what the machine contains and when it will ship. They are not measurements of work completed in an industrial setting, and no independent party has published throughput, reliability or safety results for the robot.

Arm's contribution is different in kind and, for engineers, more immediately useful. Its framework defines levels of increasing sophistication for robotic systems, from RL0 to RL5 (RL0 reactive · RL1 assisted · RL2 partial autonomy · RL3 conditional autonomy · RL4 high autonomy · RL5 self-improving), and connects real-world use cases to system requirements including latency, compute placement, memory and power constraints, determinism and safety. Arm's stated reason is that the industry "lacks a common way to describe, compare and communicate the capabilities of increasingly intelligent machines", and it compares the intent to the way SAE levels created shared vocabulary for driving automation.

A shared vocabulary matters because it forces a supplier to say which level a product reaches, in which conditions. Until now there has been no agreed way to ask that question.

What physical AI requires

A conventional industrial robot repeats a path. The part arrives in a known position, the sequence is fixed, and accuracy comes from mechanics and calibration. Physical AI is the attempt to remove the requirement that the world be arranged in advance. That replaces one engineering problem with four, each of which can fail independently.

The physical AI loop from perception to actuationFour stages run in order: perception, planning, control and actuation. A feedback path returns from actuation to perception. A separate supervisory layer holds the safety-rated stop and the authority to approve a task, and it sits outside the loop.What has to work before a robot can do an unstructured taskEach stage can fail on its own. A convincing demonstration only proves the chain held once, for one scene.PerceptionCameras, depth, force andtouch become a usablescenePlanningA goal becomes an orderedsequence of feasiblemotionsControlMotions become jointcommands under real-timeconstraintsActuationJoints, hands and gripperschange the physical worldFeedback: the result of acting changes the next scene to be perceivedSupervisory layer, outside the loopSafety-rated stop and protective separation · task approval and stop authority · logging of what the machine did and whyThis layer is the subject of the risk assessment. It is not made redundant by a better model.
The loop that has to close before a machine can work in an unarranged space. Inzonex diagram, drawn to show where authority sits rather than to depict any particular product.

A demonstration video shows one successful pass of this loop. A production line needs the loop to close thousands of times, on parts with variation, under changing light, with tolerable recovery when it does not. Failures here are rarely dramatic. Grasp success drifts downward over weeks until an operator stands permanently beside the machine.

Humanoid against the alternatives

The choice is not humanoid or nothing. Most plant tasks already have a cheaper answer, and the comparison below is the honest starting point for any evaluation.

Automation options compared on the attributes that decide a factory business case
OptionTask type it suitsCycle-time determinismSafety basisCost of a changeHuman involvement
Fixed automationOne task, unchangingHighest and most repeatableGuarding and interlocks; no perception neededRe-tooling; often a rebuildNone during the cycle
Conventional robot cellDefined task set, known part positionsHigh and deterministicISO 10218-2 risk assessment for the cellRe-programming plus fixturingSetup and exception handling
Mobile robot with an armFixed tasks at several locationsModerate; travel dominatesCell safety plus navigation among peopleNew route and new pick poseException handling and recovery
General-purpose humanoidVarying tasks in a space built for peopleUnproven outside vendor demonstrationsAssessed per application; shape grants nothingIntended to be instruction, not fixturingSupervision assumed for the foreseeable term

Read the right-hand columns first. A humanoid's claimed advantage is the cost of change: instruct it rather than re-fixture it. Its unavoidable disadvantage is determinism, because a system that decides how to move cannot promise the same cycle time as a system that replays a path.

Where the humanoid form fits

A human shape is not an engineering virtue in itself. It is a compatibility decision. It makes sense only where the environment is fixed and was designed around people: stairs and ladders, doors and handles, benches at human height, tools with human grips, aisles too narrow to re-lay.

On that basis, the plausible first tasks are the ones where the value is presence rather than speed. Inspection rounds in areas that are hard to instrument. Moving material between stations that were never laid out for a conveyor. Machine tending where the machine cannot be modified. Work in hot, confined or otherwise unpleasant zones where the alternative is exposure rather than a faster cycle.

The counter-case is just as clear. If a task repeats with the part in the same place, a fixed cell will beat a general-purpose machine on cost, speed and provable safety, and will keep beating it. Industrial robot installations worldwide ran above half a million units in 2024 for the fourth consecutive year, and the operational stock reached 4,664,000 units, 9% higher than the year before (IFR World Robotics 2025). That installed base is the benchmark a walking machine has to beat.

Safety, standards and who is accountable

The robot safety standard was rewritten before any of this. ISO 10218 was revised in 2025: Part 1 covers the design and manufacture of the robot itself, and Part 2 covers the design and integration of robot applications and cells, with the emphasis on risk assessment. Collaborative-operation requirements previously held separately in ISO/TS 15066 were consolidated into the series, and the revision reportedly adds new robot classifications with corresponding functional safety requirements and test methods, plus requirements for cybersecurity to the extent that it bears on robot safety.

The structural point survives every product announcement: safety is a property of the application, not of the machine. A humanoid that is certified as a robot is not thereby safe doing a particular job next to a particular person. Someone has to assess that job.

The question to put to any supplier

Which capability level does the product reach, for which task, in which environmental conditions, and what is the measured intervention rate over a full shift? A supplier who can answer that is describing a product. A supplier who answers with a demonstration video is describing a prototype.

Readiness steps and the evidence each one requiresFour rising steps: prototype, pilot, supervised deployment and fleet. Each step names the evidence that distinguishes it from the step below, from a demonstration video at prototype stage to availability and mean time to repair at fleet stage.Readiness is a claim about evidence, not about capabilitySeptember 2026 announcements sit at the first step. Buying decisions need the third.PrototypeA task completed underchosen conditionsVideo and a spec sheetPilotThe same task in a realaisle, repeatedlyCycle time distribution,not a best runSupervisedA shift of work with aperson accountableIntervention rate andrecovery timeFleetSeveral units, severaltasks, planned workAvailability, spares andmean time to repair
What separates each readiness step is the evidence, not the hardware. Inzonex analysis; the September 2026 announcements are evidence of manufacturing capability, which is the first step.

What this means for manufacturers

  1. Do not restructure anything yet

    Nothing announced in September 2026 changes a plant's capital plan. The near-term cost is attention.

  2. List the unstructured tasks

    Write down tasks where the position, shape or condition of the object varies between cycles. In most plants that list is short, and a general-purpose machine can only pay on those tasks.

  3. Measure them now

    Record cycle time, labour content, exception frequency and what a person actually looks at to do the job. Without this baseline, any future pilot will be judged on impression.

  4. Ask for distributions, not demonstrations

    Require cycle-time spread over a shift, intervention rate and recovery time. A best-case run tells you nothing about a production schedule.

  5. Put the safety assessment before the purchase order

    Under the revised standard the application is what gets assessed. Establish who performs that assessment and who accepts the residual risk before anyone agrees a trial date.

  6. Keep the fallback

    Any task handed to a machine that decides its own motions needs a defined manual fallback and a person authorised to invoke it.

Limits and open questions

  • No independent performance data was found for this robot. Every capability figure on this page comes from the manufacturer, and no independent test of it has been published. Academic and institutional testing of humanoid platforms exists in general; none covering this machine was located.
  • Manufacturing capability is not task capability. Being able to build robots at volume says nothing about what those robots can hold as a cycle time.
  • Cost is unpublished. No purchase price, service cost, spare-part economics or expected service life is public, so no payback period can be calculated. Any figure quoting one is invented.
  • The capability framework is new and voluntary. A shared vocabulary only helps once suppliers state their level and someone can check the claim.
  • Regulatory treatment is unsettled. A mobile machine that works among people in a plant raises questions of machinery conformity and workplace risk assessment that will be answered jurisdiction by jurisdiction.

Physical AI now has an industrial supply chain and a shared capability vocabulary attached to it, and both are new. It still lacks a machine with a published record of doing a plant's work, which is what a purchase requires.

Questions

What is physical AI?

Physical AI describes machines that sense the world, decide what to do and then act on physical objects. The difference from a conventional industrial robot is not the arm or the motor. It is that the task is not fully specified in advance, so perception and planning have to produce the motion rather than replay it.

Are humanoid robots ready for factory production work?

One manufacturer has commissioned a line to build them and states that mass production starts at the end of 2026, with market delivery planned for 2027. That is a statement about building robots, not evidence that a humanoid holds a production cycle time in another company's plant. No independent throughput or reliability benchmark for it has been published.

Why use a humanoid shape instead of a normal robot cell?

A humanoid form is only worth its cost where the workspace is built for people and cannot be changed, where the task varies between cycles, or where the robot must move between several stations. Where a task is fixed and repeats, a conventional cell is faster, cheaper and easier to guard.

What safety standard applies to a humanoid in a factory?

ISO 10218 was revised in 2025. Part 1 covers the robot as built. Part 2 covers the application and the cell, with the emphasis on risk assessment. Collaborative-operation requirements that were previously in ISO/TS 15066 were consolidated into the series. Safety is assessed for the application, not granted by the robot's shape.

What should a plant do first?

Nothing structural. Write down two or three unstructured tasks, measure their current cycle time and labour content, and record what a machine would have to perceive to do them. That record makes a later pilot measurable, and it is useful whether or not the eventual answer is a humanoid.

Sources

Suggested citation: Inzonex Research (2026), Humanoid Robots Enter Production: What Physical AI Can Actually Do in a Factory, published 12 September 2026. Capability figures are attributed to the companies that published them. The comparison table, both diagrams and the readiness assessment are Inzonex analysis. Installation and operational stock figures are from IFR World Robotics 2025, cited in Sources.

Related industrial AI analysis