The standards landed below the work

China brought a batch of national standards into effect on 1 September 2026: 59 mandatory and 375 recommended. One of them is GB/T 47245-2026, registered in English as the general framework for intelligent control system of robot, administered by the National Robot Standardization Technical Committee, TC591. The state registry gives a publication date of 27 February 2026 and an effective date of 1 September 2026, and marks it recommended rather than mandatory.

That last distinction is worth holding onto. A recommended national standard is not enforced by a regulator. It becomes binding by being cited in a tender, a supply contract or a certification scheme, so its force is commercial and it arrives later than its effective date.

Trade coverage calls it the country's first general national standard for robot intelligent control systems, applicable across industrial, service and humanoid machines. The official text names no class of robot at all, which amounts to the same claim made more quietly.

The official announcement states what the standard specifies: the overall architecture of a robot intelligent control system, together with the logical architecture and main functions of the hardware layer, the operating system layer, the middle layer and the algorithm layer. The 2024 project record adds application methods and trustworthiness assurance mechanisms, on an eighteen-month drafting cycle.

Read the list of layers again. Hardware, operating system, middleware, algorithm. Every one of the four sits inside the machine. A robot built to this standard has a describable internal stack, which is the condition under which a control system from one supplier can be swapped for another without rewriting everything above it.

What it does not give the robot is a describable relationship to a production order. There is no layer in it for the job, the part, the tolerance, the qualification or the evidence. It abstracts the body from below, which is the same direction of travel this site recorded in NVIDIA's stack three weeks ago, now with a national standard number attached.

On 29 August, ISO 21423 was in ballot. It now sits at stage 60.00, international standard under publication, which is the last procedural step before it exists. That is genuine movement and it should be recorded as such.

Its own abstract is the interesting part. The standard specifies communication protocols enabling interoperability among industrial autonomous mobile robot systems produced by different vendors, and covers AMRs, AMR fleet manager equipment, and, in ISO's words, "other enterprise resources that would communicate with the AMRs in an industrial environment". Safety requirements and machines operating on public roads are excluded.

That clause about enterprise resources is the closest any standard now reaching publication comes to the seam this site cares about. It is also fenced, twice: to mobile robots, and to communication with mobile robots. A machine that carries a thing from one place to another can now be told what to carry and where, by any vendor's fleet manager, and can report back that it arrived. A machine that makes a thing cannot be told what to make, to what standard, on whose authority.

Put the two documents side by side and the shape of the hole is exact. One describes the inside of a robot. The other describes traffic between machines that move things and the systems that dispatch them. Between them sits the sentence a factory actually needs to say: this process, at this tier, on equipment of this class, produced this part to this tolerance, with this evidence, and here is the record that lets a second certified site run it and be believed.

Not one field of that sentence appears in either standard. That is not an oversight. A control architecture is cheap to agree on because nobody's commercial position depends on where the middleware boundary falls. A definition of what counts as a qualified job, and of who may assert it, allocates power between the company that owns the robot, the company that owns the order and whoever certifies the pairing. Standards bodies write the agreeable layer first.

The seam is being built, just not as a standard. SAP's own account of its first Embodied AI Jam, held this month over three days at the Swiss Smart Factory in Biel, describes customers, robot manufacturers and integrators wiring inspection drones into SAP Asset Performance Management and humanoids into SAP Digital Manufacturing. That is the vendor's description of its own event, and what it describes is integration work rather than an interface a third party could implement. The pattern from August holds: the layers above and below the seam become standards, and the seam becomes a product.

The uncharitable reading of all this is that standards bodies are building the wrong things. That reading is wrong, and the note should say so plainly. A stable control architecture is a precondition for anything above it, and an AMR interoperability standard is what made mixed-vendor fleets purchasable. The complaint here is about sequence, not about value.

There is also a schedule question this note cannot close. China's Ministry of Industry and Information Technology released its Humanoid Robot and Embodied Intelligence Standard System, 2026 edition, in late February through a new technical committee, with six categories and an initial list of 52 standards drawn up with more than 120 institutions. Trade coverage of that roadmap places software and hardware interface standards and data exchange protocols in the second half of 2026. That placement comes from trade coverage rather than from a line this site could verify in the ministry's own document, and the distinction matters: if the reading is right, the interface standard is due in this window and has not appeared. If it is wrong, nothing was promised for now.

One thing the month does settle. This is not a contest between jurisdictions with one of them ahead. Beijing and Geneva both shipped a layer this month, and both layers sit below the work.

P-27 carries the nearest open deadline in the register, 30 June 2027, and asks for a certified production process publicly transferred between standardized robotic cells owned by two different organizations. Two standards moved this month and neither one adds a way to express that transfer, let alone to verify it. P-27 gets nothing again, and the window is now under ten months.

P-12 predicts a factory OS controlling layout, tools, quality and capacity as versioned configuration. GB/T 47245-2026 is mild evidence for the general idea that this stack gets standardized rather than staying proprietary, and it stops at the algorithm layer, which is well short of layout, quality and capacity. P-18 predicts the durable moat is the framework rather than the robots, and this month repeats last week's uncomfortable finding: the framework is being built, in a national committee and an ISO committee, in forms that charge nobody rent. Nothing here is scored and the register is untouched.

What would change the reading. A published standard in which a process record crosses an ownership boundary. The Chinese interface and data exchange protocols landing with a job description in them rather than a bus. ISO 21423 acquiring a second part that is not about mobility. Or SAP publishing its robot interface as an implementable specification rather than selling it as integration, which would be the strongest evidence yet that the seam is worth owning and the weakest possible news for anyone hoping to own it later.

Two limits, plainly. The full text of GB/T 47245-2026 is sold rather than published, so this note reads it through the national registry entry and the official announcement of the 1 September batch. The standard number, the dates, the recommended status, the administering committee and the four layers come from those official sources. Any characterization of what the document says beyond that is trade coverage and is labelled as such above. And ISO 21423 at stage 60.00 is not yet a published standard, so its final text may differ from the abstract quoted here.

SAMR national standards platform: GB/T 47245-2026, general framework for intelligent control system of robot ·SAMR national standards platform: project record 20242043-T-604 ·Beijing Municipal Government: a batch of national standards takes effect from 1 September ·36Kr: 2026, the first year of humanoid robot mass production ·ISO: ISO 21423, Robotics, industrial mobile robots, communications and interoperability ·CGTN: China releases national standards for humanoid robots and embodied AI ·SESEC: China's first standards system for humanoid robots and embodied intelligence ·RobotToday: China's humanoid robot and embodied intelligence standard system, HEIS 2026 ·SAP News: SAP's first embodied AI Jam brings customers, robots and AI together ·Liu et al.: Data Standards for Humanoid Robotics, The Missing Infrastructure for Physical AI

The Chinese standard's number, titles, dates, recommended status and administering committee are taken from the state standards registry and corroborated by the official announcement of the 1 September batch carried on a government portal. The four layers are quoted from that announcement. ISO 21423's stage, committee, scope and exclusions are taken from ISO's own catalogue entry. The SAP event is the vendor's own account of its own event and is treated as a claim about what SAP is doing, not as independent evidence that it works. The second-half-of-2026 placement of the Chinese interface protocols is trade coverage and is flagged in the text as unconfirmed against the ministry's document.