The Safety File Is the Product: ISO 25785-1 Is Still a Draft, and That Is the Humanoid Market's Real Gate

The humanoid argument has been about cost per part: drive the bill of materials down, ship units, let the price collapse. The buyer’s argument is different. Before a machine walks a factory floor, somebody has to sign a risk assessment — and the international standard written specifically for a machine that falls over when the power goes off is still a committee draft.

That is the gate. Not unit price, not the hand, not the model that drives it. In every industrial jurisdiction that matters, a humanoid enters service through a conformity route, and the documents that route requires are produced once per robot model and once per deployment site. Until the standard stack catches up, that paperwork is bespoke engineering, and bespoke engineering per site is what stands between an order book and an installed base.

What is actually published, and what is not

Start with the correction. When I set this piece up, I described ISO 25785-1 as published in 2025. It is not. The ISO catalogue lists it as ISO/CD 25785-1 — a Committee Draft — under the title “Robotics — Safety requirements for dynamically stable industrial mobile robots (legged, wheeled, or other forms of locomotion) — Part 1: Robots,” Edition 1, © ISO 2026. Part 2, which will address the integration of applications, is to be developed separately. A committee draft is a document under negotiation inside a technical committee, not a standard anyone can be certified against.

What is published is the baseline it will sit on. ISO 10218-1:2025, Robotics — Safety requirements — Part 1: Industrial robots, reached its third edition on 5 February 2025 at 95 pages under ISO/TC 299, alongside ISO 10218-2:2025 for industrial robot applications and robot cells. Together they are the current, purchasable safety framework for industrial robotics.

Document	Status	What it covers
ISO 10218-1:2025 (Edition 3)	Published, 5 February 2025, 95 pages, ISO/TC 299	Industrial robots as partly completed machinery: inherently safe design, risk reduction, information for use
ISO 10218-2:2025	Published, 2025	Integration, commissioning, operation and maintenance of robot applications and robot cells
ISO/CD 25785-1	Committee Draft, Edition 1	Industrial mobile robots with actively controlled stability — legged, wheeled or other locomotion. Part 1 covers the robot; Part 2, integration, is to be developed separately
ISO 13482	Published, separate series	Mobile robots in non-industrial environments, where the public is not excluded

Why a robot that falls over needed its own standard

The scope note in 25785-1 is the whole argument in one sentence: it applies to robots with “actively controlled stability,” meaning a robot that “requires an active control in order to remain balanced and could become unstable in the absence of power.” That is a hazard class the industrial robot standards were never built for. ISO 10218 assumes a machine that holds its position when the power is removed — the entire logic of safeguarding a fixed cell depends on the robot being where you left it when the contactor opens.

A dynamically stable robot breaks that assumption. It can be bipedal, quadrupedal or wheeled-balancing; it may carry an unspecified number of arms; it works on ground of varying elevation. The draft’s exclusions are as informative as its inclusions: ridden robots, worn robots such as exoskeletons, road vehicles, airborne and waterborne machines, commercial floor-cleaning robots and non-industrial use are all out of scope, with the non-industrial case pointed to ISO 13482. And the definition of “industrial environment” is itself a siting constraint — a workplace “where the public is excluded or restricted from access.”

Put those two facts together and the commercial consequence is immediate. A humanoid deployed in a logistics hall is inside scope. The same machine in a retail store, a hospital corridor or an airport concourse is not, and needs a different conformity argument entirely. The product is identical; the safety case is not.

The consequence: no presumption of conformity

In the European Union, machinery is placed on the market against the essential health and safety requirements of Regulation (EU) 2023/1230, which applies from 20 January 2027 and repeals and replaces Directive 2006/42/EC. The Commission’s stated objectives for the new regulation include integrating provisions for machinery with safety functions that are AI-powered, and providing for cyber-safety of compliance-relevant software data and safety control systems. That combination — an AI-enabled safety function inside a network-connected machine — describes a humanoid more precisely than it describes a press.

Compliance normally runs on a presumption: conform to a harmonised standard cited in the Official Journal and you are taken to have met the essential requirements. There is no humanoid-specific harmonised standard. So the manufacturer assembles the argument from ISO 10218-1 and 10218-2 for the robot and the cell, whatever applies to mobile machinery moving through the environment, and a site-specific risk assessment that covers the interaction the standards do not yet name.

That is a defensible process. It is also expensive, and it is repeated per site. Which gives us the economic claim this article is built on: while the standard is in draft, the marginal cost of deploying a humanoid falls with the quality of the manufacturer’s documentation package, not with the price of the bill of materials.

China’s clock is a production clock

The Ministry of Industry and Information Technology’s guidance on the innovative development of the humanoid robot industry sets out two milestones. By 2025, an innovation system preliminarily established, breakthroughs in key technologies, whole-machine products reaching international advanced levels and achieving batch production, demonstration applications in special, manufacturing and public-service scenarios, and the cultivation of two to three ecosystem enterprises with global influence. By 2027, significantly stronger technological innovation capability, a safe and reliable industrial chain and supply chain, an industrial ecosystem with international competitiveness, comprehensive strength at world advanced level, and accelerated development at scale.

Read those against the European timeline and the asymmetry is stark. China is targeting production capability and industrial scale; safety appears in the guidance as functional safety performance, network security protection and risk assessment of whole machines, components, core software and algorithms — governance, not certification. Europe is targeting a conformity regime with an application date of 20 January 2027.

So two jurisdictions are optimising different variables on nearly the same date. One is asking whether the factory can build the machine at scale; the other is asking whether the machine may be placed on the market. A manufacturer that satisfies only the first has built a product it cannot legally sell into the second, and the gap between those two questions is where 2027 will be decided.

What the deployments actually prove

The most advanced evidence is not a shipment count. It is a use case that could not have been automated before.

Figure’s Figure 02 worked on the assembly line at BMW Group Plant Spartanburg in 2025 and, by the company’s own account, contributed to the assembly of 30,000 cars in the year. In June 2026 the successor arrived in Hall 52, one of the assembly and logistics halls at Spartanburg, running Figure’s Helix 02 vision-language-action model. The published description of the task is the interesting part: sequencing, in which the robot grasps thin-walled individual parts with both hands while adjusting foot placement, shifting its body to maintain reach and balance, and then pulls a heavy cart on casters down the line.

That is loco-manipulation — locomotion and manipulation in the same task, with continuous correction. It is precisely the hazard class 25785-1 exists to address, and precisely the class that a fixed-cell safeguarding analysis cannot describe. BMW frames the project as physical AI in production, running in parallel at Spartanburg and at Figure. What it demonstrates technically is that the deployment has moved past the demonstration cell.

Agility Robotics supplies the commercial read. Its Digit v5 is being launched with the claim of being the first “cooperatively safe” humanoid — a safety-positioning word doing load-bearing commercial work — supported by more than USD 300 million of multi-year orders subject to the realisation of contractual milestones, across a pipeline of more than 30 customers. On 24 June 2026 the company agreed a business combination with Churchill Capital Corp XI at a USD 2.5 billion pre-money equity value, with more than USD 620 million of expected gross proceeds: USD 420 million of cash held in the SPAC’s trust account assuming no redemptions, and roughly USD 200 million of PIPE at USD 10 per share led by Foxconn. That PIPE is 24.8% of the pre-money equity value (Pheonix Research calculation: USD 200 million divided by USD 2,500 million), and the combined company is expected to list under the ticker AGLT.

Two details in that announcement are the reason it belongs in an article about standards. The orders are milestone-contingent — they convert on acceptance, not on signature. And the differentiator the company chose to lead with is safety, not price. Both are what a market looks like when the gate is a conformity file rather than a catalogue.

Why revenue lags units — a certification lag, not only a price lag

Pheonix Research’s Global Humanoid Robot Market report sizes the category at USD 1.15 billion in 2025, reaching USD 3.42 billion by 2033 at a 13.2% compound annual growth rate. Unit ambitions elsewhere in the market are an order of magnitude more aggressive than that: UBTECH opened orders in Shenzhen in June 2026 for a full-size line priced from RMB 119,800 and reported cumulative orders of 13,361 units on launch day, and Unitree has said it shipped more than 5,500 pure humanoid robots to end customers in 2025.

The usual explanation for the gap is price deflation, and that explanation is correct as far as it goes — as I argued in Walking Was the Easy Part, when average selling prices fall, dollar growth decelerates even as unit growth accelerates. But price is not the whole of it. Revenue is recognised when a site accepts a machine, and acceptance in an industrial setting runs through a safety file: risk assessment, validation of the safety functions, integration sign-off, insurance and, increasingly, a documented support period for any software whose failure could affect a safety function.

None of that is a reason to doubt the order numbers. It is a reason to read them as an order book measured against a delivery and acceptance record — which is exactly what the milestone language in the Agility transaction does, in public, on purpose.

What to watch

• The committee draft becomes a draft international standard. The step from CD to DIS to a published Part 1, and the start of Part 2 on integration, is the single most consequential regulatory event in this market. Until Part 2 exists, integration — the part that actually governs a factory floor — has no dedicated international text.
• Whether a harmonised standard is cited in the Official Journal. A citation converts the presumption of conformity from an argument into a default, and it is the moment the cost of a deployment site stops being bespoke.
• The 20 January 2027 date. Machinery placed on the EU market before that date is assessed under the existing Directive; after it, under the Regulation with its AI-safety-function and cyber-safety provisions. A humanoid placed on the market in late 2026 is a different regulatory object from the same machine placed in early 2027.
• Who signs the file. If insurers and industrial customers start demanding third-party certification ahead of a published standard, expect a private certification market to form before the public one — and expect the vendors with the largest installed bases to welcome it, because a certification cost is a barrier to entry that scale absorbs.
• Chinese standard-setting. The MIIT guidance commits to deep participation in international rules and standards, and a national standard system developed in parallel would give Chinese manufacturers a domestic conformity route that does not wait for ISO. That is the most plausible path by which production scale and market access stop moving together.

The humanoid industry has spent three years proving it can build the machine. The next three will be decided by whether it can document it — because in every jurisdiction that matters, the document, not the robot, is what the buyer is allowed to accept.

Frequently Asked Questions

Is ISO 25785-1 published?

No. The ISO catalogue lists it as ISO/CD 25785-1 — a Committee Draft — titled 'Robotics — Safety requirements for dynamically stable industrial mobile robots (legged, wheeled, or other forms of locomotion) — Part 1: Robots', Edition 1. Part 2, addressing integration of applications, is to be developed separately. What is published is the baseline: ISO 10218-1:2025 reached its third edition on 5 February 2025, alongside ISO 10218-2:2025 for robot applications and robot cells.

Why did humanoids need a separate safety standard?

Because ISO 10218 assumes a machine that stays where it is when power is removed, which is the basis for safeguarding a fixed robot cell. ISO 25785-1 addresses robots with 'actively controlled stability' — machines that require active control to remain balanced and could become unstable in the absence of power. That covers bipedal, quadrupedal and wheeled-balancing robots, and adds fall risk and loco-manipulation hazards that the existing industrial robot standards cannot describe.

What does the EU Machinery Regulation change for humanoid robots?

Regulation (EU) 2023/1230 applies from 20 January 2027, replacing Directive 2006/42/EC, and adds provisions for machinery with AI-powered safety functions and for cyber-safety of compliance-relevant software and safety control systems. Compliance usually runs through a presumption of conformity from a harmonised standard cited in the Official Journal. With no humanoid-specific harmonised standard yet, the manufacturer must assemble the argument from ISO 10218-1 and 10218-2 plus a site-specific risk assessment — a cost that is repeated at each deployment rather than carried once per model.

How does China's approach differ from Europe's?

China's MIIT guidance targets an innovation system and batch production by 2025, with two to three ecosystem enterprises of global influence, and world-advanced comprehensive strength with scale development by 2027. It treats safety as functional safety performance, network security protection and risk assessment — governance rather than certification. Europe's 2027 milestone is a conformity regime. Two jurisdictions are optimising different variables on nearly the same date, which is why production scale and market access may stop moving together.