Executive Summary

Four numbers carry this report. The first two show how far Europe's machinery has actually come; the last two show what has to be accumulated in the meantime.

This analysis was written by cross-reading primary documents: government announcements and Korean press coverage on one side, the EU Official Journal and standards-body notices on the other. The background and budget structure around the training centers were covered in two earlier pieces, on why behaviour data became the bottleneck and on how small a share of Korea's Physical AI budget goes to data, so they are not repeated here. What this report adds is the one thing those pieces left untouched: the legal standing a test report has inside a regulatory procedure.

Editor's note. This was written by a company whose business is documenting data quality, so the interest is worth stating up front. Pebblous sits in the measurement and record-keeping layer; it is neither an accreditation body nor a certification body. KOLAS accreditation is in progress, and the timeline and scope will be disclosed separately once fixed. This report does not claim that a Korean test report is enough to meet European regulation. It works out, clause by clause, what else would have to be true for that claim to hold.

Start by separating what is confirmed from what is not. Korean coverage of this subject describes the scale and layout of the facilities in fairly concrete terms, yet a good part of that detail cannot be traced back to a primary source. The distinction matters more than usual in a story about regulation, where the document is the argument. One unverified figure inside a proposal destabilises the sentence around it.

Here is what can be confirmed. On 1 July 2026 the Ministry of Science and ICT published its Strategy for Securing Core Competitiveness in Physical AI. The text carried by the government policy briefing service says general-purpose data such as robot behaviour data generated by public projects will be pooled in one place alongside sector-specific data, and that behaviour data will also be collected from working environments in manufacturing, mobility and agriculture. The strategy rests on four pillars — securing data, developing core technology, spreading services, and building an ecosystem — and the government calls the next three years the window that decides the outcome.

The facility structure described in the press looks like a hub and spokes. A central hub in Seoul and nearby would produce and process the basic behaviour data and synthetic data that robots need to operate, while regional AI-transformation hubs would gather data from manufacturing, agriculture, logistics and service sites. A ministry official's explanation goes as far as saying the plan will make maximum use of the Physical AI projects and regional hubs already running in North Jeolla, South Gyeongsang and elsewhere. The working name is Data Training Center, and the budget is still at the stage of being sought.

The money needs the same care. The AI-transformation R&D programme built on Physical AI in South Gyeongsang and North Jeolla has about $975M (KRW 1.4131 trillion) allocated from 2026 to 2030, with precision control in Gyeongsang and the AI platform in Jeolla. The robot training centers, however, are being pursued separately from that programme, and no specific amount for a Data Training Center line appears even in the 2027 budget proposal. Put the two numbers side by side and readers will conclude that a $975 million training center is under construction. The verifiable sentence today stops at "the budget is being sought."

One comparison makes the structure easier to see. The attempt to move tacit shop-floor knowledge into data has already been run once, in the manufacturing data library programme, and there too the bottleneck was never the volume collected. It was what got written down, and in what form. The robot training centers face that same question.

This report starts from a single clause in the government announcement. One more sentence follows the promise to collect data, and what it points at is the records attached to the data.

Validation means someone judged whether the collected data is fit for use and wrote that judgment down. Drafting interoperability standards means the judgment reads the same way at another organisation. Join the two clauses and the facility's output list appears: for each dataset, a record of how representativeness was checked, a label error rate, the findings of a bias examination, a gap log naming the conditions missing from the data, and the closed-loop success and safety rates measured by actually running the learned policy.

If that list looks familiar, it should. The data governance documentation that Article 10 and Annex IV of the EU AI Act require from providers of high-risk systems calls for almost the same items. The table below sets the two side by side. On the left is what the planned facility produces; on the right is what the European text demands.

The mapping follows Article 10 and Annex IV of Regulation (EU) 2024/1689. The substance of the evidence Article 10 asks for is examined in more detail in an earlier piece on labelling audit trails.

The second row sits so close to this report's subject that it deserves reading on its own. Article 10(4) says datasets shall, to the extent required by the intended purpose, take into account the characteristics or elements particular to the specific geographical, contextual, behavioural or functional setting within which the high-risk system is to be used. The provision uses the word behavioural itself. The moment a robot trained on assembly motions gathered in a Korean plant is installed in a European one, this paragraph switches on, and the question stops being how much data there is and becomes whether the deployment site's conditions are present inside it. Bench heights, lighting, part-placement habits and the way workers and robots share space differ by country and by plant.

This is where the sovereignty argument for domestic data meets the provision. The two do not conflict, but they ask for different things. One says do not hand your data to someone else; the other says show that the data represents the place the product will be sold into. If a product bound for Europe was trained only on data collected at home, what Article 10(4) wants is not the fact that the data is domestic but a record of how far European deployment conditions are covered and where the holes are. That record is the same object as the gap log.

2.1Regulatory vocabulary is pointing at a performance problem

If representativeness and coverage sound like vocabulary invented for compliance officers, the fastest correction is a study that actually ran the robots. THE COLOSSEUM benchmark, from researchers at the University of Washington and NVIDIA, takes 20 manipulation tasks and perturbs them along 14 axes — the colour, texture and size of objects, table-tops and backgrounds, plus lighting, distractors, physical properties and camera pose — while running five state-of-the-art manipulation models. The paper reports this.

One caution about reading those numbers. The 30–50% figure is a range across all 14 perturbation axes, and the 75% figure applies when perturbations are combined. Neither comes from changing the camera pose alone. The paper notes only that viewpoint weighs especially heavily on image-based models; a viewpoint-only drop does not appear in the abstract. What matters here is less the size of the drop than where the collapse happens, which is at conditions the data never contained.

More useful than the average are the three factors the paper names outright. The perturbations that hurt performance most were changes in the number of distractor objects, the colour of the target object, and lighting conditions. The paper supplies the cause as well. In 2D models trained end-to-end on RGB images, a shift in colour or texture pushes the input distribution itself, so the output moves with it, and models without real-world pre-training fall apart when a few extra items appear in frame. Models pre-trained on cluttered real scenes, by contrast, were barely affected by distractors. For anyone planning collection, this is a priority list. A dataset that has not spread lighting, background colour and distractor count orthogonally will break at the same point no matter how large it grows.

So collect more data? The ICLR 2025 study on data scaling laws in imitation learning gathered over 40,000 demonstrations and validated them with 15,000 real-world rollouts, and its answer runs the other way. Generalisation follows a power law in the number of environments and the number of objects, while demonstrations per environment stop helping much past a threshold. The practical recipe the authors offer is 32 environments, one distinct object each, 50 demonstrations apiece, which reaches roughly a 90% success rate on unseen environments and unseen objects.

Before that recipe is carried into facility sizing, look at the line the paper draws around itself. The authors studied single-task policies and state in their limitations that task-level generalisation is out of scope. Validation ran on four tasks only: pouring water, arranging a mouse, folding towels and unplugging a charger. Collection used a handheld gripper rather than teleoperation, and in environments where pose estimation was difficult about 10% of demonstrations were discarded as invalid; a single learning algorithm was used throughout. So the number 32 attaches to one process, not to a whole facility. A hub that promises many product types and many processes and reads that number as a total will misjudge its scale badly.

The same two papers also answer how verification has to work. Section 2 above put closed-loop success and safety rates on the facility's output list, and this is where the reason offline metrics will not do becomes visible. To check whether their simulation results held in reality, the COLOSSEUM team reproduced the same perturbations physically and compared, and the correlation between the two was R̄² = 0.614. The direction is right; a little over half of the real-world variance is explained. The scaling paper shows a similar mismatch. Offline prediction error and real scores moved almost perfectly in opposite directions when environments and objects grew together, but in the object-only experiment the error rose once training objects passed 16, and the correlation loosened.

The implication for training center design is plain. Neither data amplified through a digital twin nor an error measured on a validation set constitutes a verdict on its own. A separate layer has to run physical robots and measure success and safety rates, and the result has to be filed beside the dataset. The heavier a facility leans on synthetic data, the thicker that layer needs to be. Amplification is cheap and verification is expensive, but a record that skipped the expensive half is useless in front of a regulator and on the shop floor alike.

Once you decide to keep records, the next question is what to measure them against. The ISO/IEC 5259 series, produced by the AI subcommittee of the ISO/IEC joint technical committee, fills that slot. Its five parts cover different layers, so which part gets cited changes the claim entirely.

One claim that circulates widely in Korean material needs correcting here. The part Korea led is not 5259-2 but 5259-5. It was developed by Korea Testing Laboratory together with researchers at Ewha Womans University under the AI standardisation roadmap of the Korean Agency for Technology and Standards, which in November 2023 also designated KTL as the national secretariat for SC 42. Part 5259-2, which defines the quality measures, is an international joint output. How the 5259 measures are used in practice is set out separately in the 5259-2 cheat sheet and the standardisation roadmap summary.

A standard existing does not make certification common. For 5259-3, the world's first certificate, issued by SGS to the infrastructure analytics firm AI Clearing in December 2025, remains the only case we can confirm, with no further examples visible as of August 2026. A standard specific to behaviour data, ISO/WD 26264, is under development; its progress is tracked in the piece on humanoid robot data standards.

3.1Europe's first AI Act standard this year is not a data standard

Europe's yardsticks run on a different clock. In May 2023 the Commission issued standardisation request M/593, handing development of the AI Act harmonised standards to JTC 21, the joint technical committee of CEN and CENELEC. The original deadline of 30 April 2025 came and went. The request was amended as M/613 in June 2025, pushing expiry to 28 February 2027. At their October 2025 technical boards the two organisations went further and decided that where the enquiry stage returns a positive result, the formal vote can be skipped and publication can follow directly.

The first result arrived this year. EN 18286:2026 was approved on 12 July 2026 and became available on 22 July, the first European standard supporting the AI Act to reach publication. What it addresses, though, is Article 17 on quality management systems. The standard that answers Article 10 on data is prEN 18284, formally titled quality and governance of AI datasets, and it is still a draft. prEN 18283 on bias management, and prEN 18228, 18229-1 and 18285 on risk management, logging and the conformity assessment framework, sit at the enquiry or drafting stage.

The layering deserves a note. Article 17, which EN 18286 addresses, sits in Chapter III Section 3 of the AI Act, whereas the essential requirements pointed to by Article 40(1), the provision that creates the presumption of conformity, sit in Chapter III Section 2. Citation of a quality management standard therefore does not extend a presumption of conformity to Article 10.

Publication is also not the last box. What creates a presumption of conformity in Europe is a harmonised standard cited in the Official Journal, and citation is a separate decision the Commission takes after publication. The diagram below shows where the JTC 21 output currently sits.

The four boxes trace the road a harmonised standard walks, and there is in fact a second way into the last one. The Commission can adopt common specifications directly, a device built for exactly the case where standardisation runs late, which is the case now. That provision and the way it applies to robots are handled in §4.2.

As for what to use during the gap, there is a practical answer. With no cited harmonised standard available, international standards such as ISO/IEC 5259 and ISO/IEC 8183 are treated as de facto working guidance. That is custom, not legal presumption. Following a standard does not by itself presume compliance, so a provider has to demonstrate directly what was checked with it and how.

Suppose you take a data-quality test report issued by a Korean accredited laboratory into the European market. How much force that paper carries inside a regulatory procedure is decided in three places. The three are not doors passed in sequence but questions at different levels, so confuse any one of them and the line you wrote in the business plan, the one that says "we will issue the evidence," stops working in practice.

4.1The capability is there. The problem comes after it

The first checkpoint is routinely mistaken for a question of capability. You will read in several places that Korea has no accredited laboratory able to measure AI data quality. That is not true. The scope list in the KOLAS accreditation scheme already carries software testing as a coded field, and the Korea Conformity Laboratories issue accredited reports for AI quality and training data quality against ISO/IEC 25059, TS 25058 and TS 4213. The Korea Building Energy Technology Institute also holds accreditation in software testing and is extending its scope to AI trustworthiness specifications, and a further laboratory received accreditation in the same field in January 2022.

The real checkpoint is the box after that. Accreditation confirms that a laboratory is competent to run a given test. Designation is a government's act of naming a body to carry out conformity assessment on its behalf under a specific law. In the EU, a member state designates a notified body established on its own territory. An accreditation body does not designate. For a body in another country to play that role there must be a government-to-government mutual recognition agreement on conformity assessment, and the EU has such agreements with seven non-EU countries: Australia, Canada, Israel, Japan, New Zealand, Switzerland and the United States. Korea is not among them. The Korea-EU free trade agreement contains regulatory cooperation clauses in its electronics and automotive annexes, but they do not create a mechanism for designating notified bodies.

This is written into the text, not merely into practice. Article 39 of the AI Act deals with conformity assessment bodies of third countries in a single sentence, and that sentence carries the whole condition: bodies established under the law of a third country with which the Union has concluded an agreement may be authorised to carry out the activities of notified bodies under the Regulation, provided they meet the requirements of Article 31 or ensure an equivalent level of compliance. The reading order matters. The agreement comes first; capability comes after.

It is true that the phrase "equivalent level of compliance" leaves a door open. It means the arrangement need not be identical to the European one, so there is technical room. But that room opens only once an agreement exists. Which is why a Korean laboratory cannot pass this box by buying more equipment or widening its accreditation scope. What is missing is not a laboratory but a treaty, and treaties are concluded by governments. This is the point most often misread in practice.

None of which makes Korean testing useless. Two paths are used in practice: file the test results inside the technical documentation, or feed them into a European notified body's subcontracting or witnessed-testing arrangements. The certificate itself is issued by the European body.

4.2Publication and citation are different boxes

The second checkpoint is the box the diagram showed. Article 40 of the AI Act provides that a high-risk system conforming to a harmonised standard cited in the Official Journal is presumed to conform to the essential requirements. Where the presumption exists, a provider need only show it followed the standard; where it does not, the provider has to demonstrate directly that it met what the text asks for. The difference in workload is large.

As of August 2026 no provider holds that presumption, because not one European standard supporting the AI Act has been cited in the Official Journal. That can change: the two organisations are accelerating with the fourth quarter of 2026 in view, and the standardisation request now expires on 28 February 2027. But what that target refers to is publication, not citation, and the data standard is still in the first box.

There is more than one door into the last box, though. Article 41 lets the Commission adopt common specifications by implementing act when harmonised standards fail to arrive on time, and Article 41(3) gives systems conforming to those specifications the same presumption of conformity a harmonised standard would. The trigger conditions read remarkably like the present. The standardisation request must not have been complied with within the deadline, and no harmonised standard addressing the requirement in question may have been cited, with none expected within a reasonable period. That is the previous paragraph restated as a legal condition. The text does say the Commission may adopt, not that it must. And no common specification for the AI Act's high-risk requirements is confirmed to have been adopted as of this writing.

What matters more is that this second door has already been wired up for robots. In July 2026 the Digital Omnibus inserted a new paragraph into Article 20 of the Machinery Regulation. Until AI-related harmonised standards or common specifications appear under the Machinery Regulation itself, a high-risk system conforming to a harmonised standard cited under Article 40 or a common specification adopted under Article 41 of the AI Act is presumed to conform to the relevant essential health and safety requirements in Annex III of the Machinery Regulation. For a robot maker, the currently empty box can be filled from two directions, and one of the keys sits with the Commission rather than with a standards body.

So the precise statement about the second checkpoint is not that zero citations means no presumption, ever. It is that today no document producing a presumption exists on either side, neither a cited harmonised standard nor an adopted common specification, and that which one arrives first is undecided. For anyone preparing, this distinction is not academic. Watch only the standards track and you will be late if common specifications land first; know that both routes aim at the same items Article 10 names and accumulating the records now pays off whichever door opens.

4.3One route calls no third party; the other calls only European ones

At the third checkpoint the routes split to opposite extremes. Standalone high-risk systems under Annex III points 2 to 8 follow, per Article 43, the Annex VI internal control route on their own. No notified body intervenes and no third-party report is demanded. That does not make it a light route. The provider has to verify for itself that its quality management system meets Article 17, assess conformity with the essential requirements on the basis of the technical documentation, and confirm that the design and development process matches that documentation. Under Articles 18 and 47, the technical documentation and the declaration of conformity must be kept for ten years and produced whenever an authority asks. On top of that, the Commission holds a power under Article 97 to upgrade this list to third-party assessment by delegated act, so self-assessment is not guaranteed to last.

Robots sit at the other extreme. Article 1(41) of Regulation (EU) 2026/1744, published in the Official Journal in July 2026, moved machinery from Section A to Section B of Annex I to the AI Act. Section B products do not run the AI Act's general high-risk procedure; they follow the Machinery Regulation's own route first. It is easy to read that as a loosening. The route they moved to is tighter.

Half of the loosening is real. Under Article 2(2) of the AI Act, only Article 6(1), Articles 102 to 109 and Article 112 apply to Section B products. The Chapter III Section 2 requirements, Article 10 among them, do not bite on robots directly. But the Omnibus did not leave the space empty. The same regulation added a paragraph to Article 8 of the Machinery Regulation obliging the Commission to adopt a delegated act inserting AI-related health and safety requirements into Annex III of the Machinery Regulation. Those requirements must reflect Chapter III Section 2 of the AI Act along with Articles 17, 19, 72 and 73, and the delegated act must apply by 2 August 2028.

In short, the data governance items Article 10 names do not disappear in front of robots. They change clothes. They arrive as essential health and safety requirements under the Machinery Regulation rather than as high-risk requirements under the AI Act, and breaches are handled inside the Machinery Regulation's own market surveillance system. What actually changes for a company preparing is the name of the provision to cite and the counterparty it will face in assessment. The list of records to hold stays the same. If you were told the regulation got simpler, the accurate version is that it was gathered into one law, not that it asks for less.

Annex I Part A of Regulation (EU) 2023/1230, the Machinery Regulation, lists the categories where third-party conformity assessment is always required. Newly added point 5 covers safety components with fully or partially self-evolving machine-learning behaviour that perform safety functions, and point 6 covers machinery with such components built in. These items were written with machine learning in view. Fall into Part A and self-declaration is impossible even if you follow harmonised standards in full; only routes involving a notified body remain, such as modules B+C, H or G. For reference, the 19 categories in Part B do allow self-declaration when harmonised standards are applied.

The two extremes converge on the same conclusion. One calls no third party, so a Korean test report is not needed; the other calls only European designated bodies, so a Korean report cannot stand in for a certificate. On neither route is it a legal instrument. It is evidence you can file. And its value as evidence is far from small. On the internal control route, the persuasiveness of the technical documentation is the basis of compliance; on the notified body route, that evidence decides where the assessment stops.

Talk of deadlines in Korea tends to stop at 2 August 2028. That is not wrong: it is the compliance deadline for high-risk systems embedded in Annex I products under the AI Act. But it is not the door a robot maker actually hits first. As the previous section showed, AI embedded in a robot follows the Machinery Regulation route first, and the Machinery Regulation applies from 20 January 2027. Seventeen months from this writing.

Below are the dates set out by the Digital Omnibus of July 2026, lined up in order. The two that concern robots are highlighted.

The last two entries need one more reading note. 2 August 2028 is the compliance deadline for Annex I products under the AI Act, but what arrives for robots that day is not an AI Act obligation. It is the delegated act that plants AI requirements inside Annex III of the Machinery Regulation. As the previous section showed, the Commission is obliged to adopt that act and its application date is aligned to the same day. So a robot company's calendar carries two marks. On 20 January 2027 the body of the Machinery Regulation switches on; on 2 August 2028 the AI requirements move inside it. The nineteen months in between are preparation for the second date, not an empty stretch.

Read a postponed date as a relaxed requirement and the whole order of preparation goes wrong. What was deferred is the deadline, not the requirement. And for a robot company, the door that was never deferred opens first. How to read the AI Act's deadlines as a whole is set out in the piece on what the August 2026 deadline actually covers.

While the government works on a budget, private facilities have already started running. What makes this matter is ownership of the criteria. If the public certificate is late to answer the question of what counts as good behaviour data, the answer hardens into a de facto standard set by whoever ran their facility first.

The last row is worth a second look. In NVIDIA's blueprint, evaluation is not an optional tool but a layer of the pipeline. Data is curated, amplified and diversified, then automatically scored and filtered. The code is public, and the list of early adopters lines up robotics, autonomous driving and industrial software companies. A private evaluation layer is already running. The criteria for what counts as usable data may well settle on the vendor side before any state has a certification system in place.

There are reports that government-led robot training centers are already running in China. We could not find primary material confirming their scale or composition, so this report goes no further than the fact of operation.

One contrasting case shows what a public system can do. Singapore bundled evaluation tools and a certification stack at national level and connected them to private certification bodies. That structure and its limits are described in the piece on Singapore's AI trust infrastructure. The point is not that tools were built but that a place to receive the tools' output was built into the institutional design at the same time. The output of Korea's robot training centers faces the same question. Who receives the verification records, in what format, and what are they used for?

Follow the three checkpoints and the chain visibly divides into boxes: the box that measures, the box that accredits, the box that designates, the box that issues the certificate. Pebblous stands in one of them, and here is exactly which. Blurring the boundary would help the marketing, but in a piece about regulatory documents the blur itself costs credibility.

7.1Our box is measurement and record-keeping

Pebblous is neither an accreditation body nor a notified body. Accreditation belongs to KOLAS; designation belongs to European member state governments. What Pebblous builds is the box before those: the layer that actually measures the quality of a dataset and writes the verdict down as a document. DataClinic, which implements the ISO/IEC 5259-2 quality measures as a product, is the result. KOLAS accreditation is in progress, and the timeline and scope will be disclosed separately once fixed.

The moment a training center declares that it validates data, it needs a layer that performs the validation and keeps the record. Declaring validation and holding validation records are different things. The first is a line in a business plan; the second is a bundle of files attached to every dataset.

7.2The regulatory document and model performance point at the same thing

Regulation and performance speak different languages, and they still meet at the same point. Manipulation policies collapse when lighting, background and viewpoint change, and that collapse comes not from having too little data but from leaving the orthogonality and coverage of conditions unmanaged. As the data scaling work showed, adding demonstrations within one environment stops helping past a threshold; what makes performance is diversity of environments and objects.

Representativeness and completeness in Article 10, coverage and diversity in 5259, look like vocabulary written for compliance officers. They are in fact measurements that catch that collapse early. The items a regulatory document demands and the conditions a model actually needs point at the same place. That connection is the one this report set out to draw. There is no reason a record made for compliance should be a different object from a record made for performance.

7.3The question practitioners should actually be asking

The first question from a company selling robots or industrial machinery into Europe is usually where to get certified. This report's answer is to change the question. Ask first what has to be recorded now in order to be submitted in 2027 and 2028. A certification body can be found when the deadline nears; the records cannot be made then.

• Does the machine-learning component in our product perform a safety function? If so it falls under Annex I Part A of the Machinery Regulation, where a notified body is required even if harmonised standards are followed.
• Are representativeness checks and a gap log accumulating alongside the data we are collecting right now? If only the data is accumulating, there is nothing to submit.
• What will we use the Korean laboratory's test report for? As an attachment to the technical documentation, or as an input to a European notified body's assessment? The required format differs by use.
• Will the record system survive ten years? The retention duty runs that long, so today's file structure and identifier scheme have to hold for the same period.

For anyone writing a proposal to a training center or a regional hub programme, the implication runs at a different angle. Before the line "we will issue the evidence" can hold, it is better to write down which designation, which standard and which route it presupposes. Writing the presuppositions down tends to sharpen the scope of the project rather than shrink it.

Finally, what this report could not confirm. We could not confirm the 2027 budget figure for the Data Training Center, the success-rate drop attributable to camera viewpoint alone, the size of Korea's robot and machinery exports to Europe, the quantitative scale of Europe's notified body shortage, or the internal clause structure of the EN 18286 text. On common specifications under Article 41 we verified the provision, its trigger conditions and its legal effect, but public material does not show whether the Commission has a draft in preparation, so rather than assert that none has been adopted we wrote only that none is confirmed. The budget and export figures in particular are either not published or not compiled separately. An absence is left as an absence.

EU law and policy

• 1.Regulation (EU) 2024/1689 (Artificial Intelligence Act), OJ L, 2024/1689, 2024-07-12. Article 2(2) limiting the scope for Section B products; Article 10, in particular 10(4) on geographical, contextual, behavioural and functional settings; Articles 17, 18, 39 (conformity assessment bodies of third countries), 40, 41 (common specifications), 43, 47, 97; Annexes I, IV, VI and VII. EUR-Lex
• 2.Regulation (EU) 2026/1744 (Digital Omnibus on AI), published in the OJ 2026-07-24, in force 2026-07-27. Article 1(41) moving machinery from Section A to Section B of Annex I; Article 1(40) adjusting the deadlines; recital (42) explaining the phased approach; and Article 3 amending the Machinery Regulation (the Article 8 obligation to adopt a delegated act with a 2028-08-02 application date, and the presumption-of-conformity bridge in Article 20(10)). EUR-Lex
• 3.Regulation (EU) 2023/1230 (Machinery Regulation), applicable from 2027-01-20. Annex I Part A points 5 and 6 (machine-learning safety components and machinery incorporating them), the modules B+C, H and G in Article 25(2), and the self-declaration condition in Article 25(3) for the 19 Part B categories. EUR-Lex
• 4.European Commission, Mutual Recognition Agreements. List of third countries with an MRA on conformity assessment (Australia, Canada, Israel, Japan, New Zealand, Switzerland, United States). European Commission

Standards and accreditation

• 5.CEN-CENELEC, "EN 18286 in the Spotlight: Supporting Compliance with the AI Act," 2026-07-30. Approval and availability dates for EN 18286:2026 and the provision it addresses. CEN-CENELEC
• 6.CEN-CENELEC, "Update on CEN and CENELEC's Decision to Accelerate the Development of Standards for Artificial Intelligence," 2025-10-23. The acceleration decision, including skipping the formal vote. CEN-CENELEC
• 7.JTC 21 Standards Tracker (2026-06 snapshot) and the EU AI Act Harmonised Standards Map. Stage of prEN 18284, 18283, 18228, 18229-1 and 18285, and the record of Official Journal citations. Tracker · Standards Map
• 8.ISO/IEC 5259-2:2024, Artificial intelligence — Data quality for analytics and machine learning (ML) — Part 2: Data quality measures. ISO / ISO/IEC 5259-5:2025, Data quality governance framework (led by Korea Testing Laboratory)
• 9.IAF, "Global Accreditation Cooperation Incorporated Launch." ILAC and IAF ceasing operations on 2026-01-01 and consolidating into the GAC MRA. IAF
• 10.Korean Agency for Technology and Standards, introduction to the KOLAS accreditation scheme and the KOLAS Accreditation Operating Guidelines (KOLAS-R-002), software testing field. Service listings for AI and training-data quality testing at Korean accredited laboratories (KCL, KBeT and others)
• 11.SGS, world-first ISO/IEC 5259-3 certificate issued to AI Clearing, 2025-12. The only case confirmed as of 2026-08

Academic

• 12.Pumacay, W., Singh, I., Duan, J., Krishna, R., Thomason, J., Fox, D., "THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation," RSS 2024. arXiv:2402.08191. Source for the 30–50% degradation across 14 perturbation axes, the ≥75% degradation under combined perturbations, the three worst perturbations (distractor count, target object colour, lighting), and the ecological validity correlation of R̄² = 0.614 between simulation and the real world
• 13."Data Scaling Laws in Imitation Learning for Robotic Manipulation," ICLR 2025. arXiv:2410.18647. Power laws in environment and object diversity, the threshold effect in demonstrations per environment, the 32-environment × 50-demonstration recipe, the scope limits in §7 (single-task policies, four validation tasks, UMI collection), and the offline-metric correlation analysis in Appendix E.1

Korean policy and press

• 14.Ministry of Science and ICT, Strategy for Securing Core Competitiveness in Physical AI, 2026-07-01 (in Korean). Source of the quoted clause on "a system capable of managing quality, through validation of the data and the drafting of interoperability standards." Policy Briefing
• 15.Hankyung, "Government to build robot training centers to gather Physical AI data," 2026-06-25 (in Korean). The central and regional hub structure and the working name Data Training Center. Article
• 16.Hankyung, "Down to the robot's fingertip sensations: government puts KRW 1.4 trillion into Physical AI," 2026-07-13 (in Korean). The Gyeongsang and Jeolla AI-transformation R&D budget and the separate pursuit of the robot training centers. Article
• 17.Hankyung, "Hyundai opens its robot training center, chasing a new market in field data," 2026-08-23 (in Korean). Article / Seoul Economic Daily, "LG Electronics to build Korea's first robot training center," 2026-06-11 (in Korean). Article
• 18.NVIDIA Newsroom, "NVIDIA Announces Open Physical AI Data Factory Blueprint," 2026-03-16. The three-stage curation, augmentation and evaluation structure, and the standardisation of the evaluation layer. NVIDIA

Adjacent Pebblous reporting

• 19.Pebblous, why behaviour data became the bottleneck in Physical AI. report/korea-physical-ai-behavior-data
• 20.Pebblous, how small a share of Korea's Physical AI budget goes to data. report/korea-880-trillion-physical-ai-data-gap
• 21.Pebblous, what Article 10 of the AI Act asks for in labelling audit evidence. report/eu-ai-act-article10-labeling-audit-evidence
• 22.Pebblous, what the AI Act's August 2026 deadline actually covers. report/eu-ai-act-august-2026-deadline-reality
• 23.Pebblous, Singapore as a case of a national assurance stack. report/singapore-ai-trust-infrastructure-2026
• 24.Pebblous, ISO standardisation of humanoid robot data. report/humanoid-robot-data-iso-standard
• 25.Pebblous, the mismatch between offline metrics and closed-loop performance. report/robot-data-curation-closed-loop-gap
• 26.Pebblous, the library programme moving tacit manufacturing knowledge into data. report/korea-manufacturing-data-library-tacit-knowledge
• 27.Pebblous, the ISO/IEC 5259-2 measures and the standardisation roadmap. project/ISO5259/5259-2-cheetsheet-01 · iso5259-standardization-roadmap