humanoid4rent
AnmeldenAngebote einholen

Robot foundation models and the AI Act: when the GPAI rules apply, and to whom

The AI Act's general-purpose AI obligations have applied since August 2025 and can be enforced from August 2026. Here is how robot foundation models fit, and who counts as the provider.

Von Redaktion von humanoid4rent · veröffentlicht 10. November 2026

Diese Seite wurde noch nicht übersetzt
Sie lesen den englischen Text, weil es noch keine Fassung auf Deutsch (Österreich) gibt. Nichts wurde weggelassen: Das ist die ganze Seite, in der Sprache, in der sie geschrieben wurde.

Since 2 August 2025 the EU AI Act has placed obligations on providers of general-purpose AI (GPAI) models, and from 2 August 2026 the European Commission can enforce them with fines. Robot foundation models sit at an awkward edge of these rules. A model that only outputs motor commands usually falls outside the Commission's indicative test, but many robot models are built on language models that fall inside it.

What the GPAI rules require, and since when

The rules are in Chapter V of Regulation (EU) 2024/1689, the Artificial Intelligence Act. According to the Commission's questions and answers on general-purpose AI models, every provider of a GPAI model must:

  • keep technical documentation for the AI Office and national authorities;
  • give information and documentation to companies that build the model into their own AI systems;
  • put in place a policy to comply with EU copyright law;
  • publish a summary of the content used to train the model.

Models classed as posing systemic risk carry further duties, including assessing and mitigating those risks. The same Q&A presumes systemic risk above 10^25 floating-point operations (FLOP) of training compute, a scale it estimates costs tens of millions of euros to train.

The Commission's guidelines on the scope of the GPAI obligations state that the obligations apply from 2 August 2025, but its enforcement powers only took effect on 2 August 2026. Those powers include requests for information, model evaluations, ordering measures up to and including withdrawal from the market, and fines of up to 3% of global annual turnover or EUR 15 million, whichever is higher. Models placed on the market before 2 August 2025 have until 2 August 2027 to comply.

The Code of Practice

The Commission published the General-Purpose AI Code of Practice on 10 July 2025. It has three chapters. The transparency and copyright chapters are relevant to all GPAI providers, and the transparency chapter includes a model documentation form. The safety and security chapter applies only to providers of models with systemic risk.

The code is voluntary. The Commission and the AI Board have confirmed it is an adequate way to show compliance, and signatories are promised less administrative burden and more legal certainty. The signatory list on the Commission's page is made up of AI model developers, among them Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI; xAI signed only the safety and security chapter. Robot manufacturers do not appear on it as such.

When is a model a general-purpose AI model?

The Act defines a GPAI model as one that shows significant generality, can competently perform a wide range of distinct tasks and can be integrated into many downstream systems. Models used only for research, development or prototyping before they are placed on the market are excluded. The Commission's guidelines add an indicative criterion: training compute above 10^23 FLOP and the ability to generate language (as text or audio), images from text, or video from text. The guidelines treat code as text and speech as audio, and say 10^23 FLOP roughly matches a one-billion-parameter model trained on a large amount of data.

The criterion works in both directions. A model that meets it but only competently performs a narrow set of tasks is not a GPAI model. A model that misses it but still shows significant generality can be one. The guidelines give out-of-scope examples, including models trained only to transcribe speech, upscale images, play chess or video games, or model weather patterns and physical systems.

Where robot foundation models fall

The guidelines contain no example involving robots, so what follows is a reading of the general criteria, not settled guidance.

Models that output only actions

A robot policy that turns camera images and joint readings into motor commands does not generate any of the listed modalities. On the indicative criterion it points out of scope, much like the physical-systems example. The open question is the exception: developers describe generalist robot policies as working across many tasks and robot bodies, which is the kind of generality the definition is about. Neither the Act nor the guidelines says whether broad physical skill counts.

Models built on language models

Many current systems combine an action model with a vision-language model, as explained in our article on vision-language-action models. Google DeepMind announced Gemini Robotics 1.5 on 25 September 2025. It describes the model as turning visual information and instructions into motor commands, and describes its sibling, Gemini Robotics-ER 1.5, as a reasoning model that plans tasks and calls digital tools. Both, it says, are built on the core Gemini family. It made ER 1.5 available to developers through the Gemini API, while Gemini Robotics 1.5 went to selected partners. More on that in our piece on what Google DeepMind has released.

NVIDIA announced Isaac GR00T N1 in March 2025. Its architecture has two parts: a fast action model and a slower reasoning part powered by vision-language models. NVIDIA says developers can post-train it for their own robot or task. The background is in our article on NVIDIA Isaac GR00T.

When a robot model contains, or is derived from, a component that generates language, the indicative criterion may well be met. Whether a specific product is a GPAI model is for its provider to assess in the first instance, with the Commission having the final say. We have not seen public statements on how Google or NVIDIA classify these particular models.

Who is the provider when a robot maker fine-tunes a model?

The guidelines set out several cases.

  1. The original developer. A company that develops a GPAI model and first makes it available on the EU market is its provider. That covers access through an API, a download, a cloud service or a copy installed on a customer's machines.
  2. A robot maker that fine-tunes it. A downstream company becomes the provider of the modified model only if the modification significantly changes the model. The indicative test is that the compute used for the modification is more than a third of the compute used to train the original. If the robot maker cannot know or estimate that figure, it should use a third of 10^23 FLOP instead, or a third of 10^25 FLOP if the original carries systemic risk. The Commission expects few modifications to meet the test at present.
  3. Duties limited to the change. A modifier that does become a provider has duties only for its modification: documentation of what changed, plus a copyright policy and training summary covering the new data.
  4. Who controls the weights. Whether the original provider or the downstream company counts as the modifier is decided case by case, for example when fine-tuning happens through the original provider's API.
  5. Models supplied outside the EU. If a developer hands a model to a robot maker outside the EU and the robot is then sold in the EU, the model is treated as placed on the EU market at that point. The developer remains the provider unless it clearly ruled out EU use, in which case the robot maker becomes the provider of the model.
  6. A robot maker's own model. A robot maker that puts its own GPAI model into its own product carries the model obligations as well as those for the AI system.

Separately, a robot maker is usually the provider of the AI system inside the robot. That brings other obligations on another timetable, set out in our AI Act timeline for robots, as well as the duty to tell people they are dealing with a machine under Article 50.

Open-source models

Providers of GPAI models released under a free and open-source licence are exempt from the two documentation duties and from appointing an authorised representative. They must still have a copyright policy and publish a training summary. The exemption does not apply to models with systemic risk. The licence must allow anyone to access, use, modify and redistribute the model. Licences limited to non-commercial or research use, or that require extra licensing above a certain number of users, do not qualify, and neither does monetisation such as paid commercial use.

What is still unclear

The main gap is the action modality. The Commission chose language, image and video generation because, in its view, language confers the widest range of capabilities. It says its approach may change as technology and the market evolve. Until there is guidance or an enforcement decision on generalist robot policies, classification will depend on case-by-case judgement. This article is general information, not legal advice.

What this means for event organisers

  • The GPAI duties fall mainly on model developers. Renting or booking a robot does not on its own make you a GPAI provider.
  • Ask what software drives the robot. A useful question for your brief to a robot supplier is which models handle conversation and movement, and whether they run on the robot or in the cloud.
  • Documentation flows between providers. If a robot talks through a third-party language model, its developer must give documentation to the supplier that integrates it, not to you. You can still ask the supplier what it has received.
  • GPAI compliance is not a safety certificate. These obligations concern documentation, copyright and systemic risk, not whether a robot is safe to walk through a crowd. Physical safety is covered by other rules and by the supplier's own risk assessment.
  • Expect revisions. The Commission has said it will update the guidelines as practice develops, so check the current version before relying on any detail.

Frequently asked questions

Since when do the AI Act's general-purpose AI obligations apply?

The obligations for providers of general-purpose AI models have applied since 2 August 2025. The European Commission's enforcement powers, including fines of up to 3% of global annual turnover or EUR 15 million, whichever is higher, apply from 2 August 2026. Models placed on the EU market before 2 August 2025 have until 2 August 2027 to comply, according to the Commission's guidelines.

Is a robot model that only outputs motor commands a general-purpose AI model?

On the Commission's indicative criterion, probably not. The criterion requires training compute above 10^23 FLOP and the ability to generate language, text-to-image or text-to-video output, which an action-only model does not do. The criterion is only indicative, however, and a model showing significant generality can still qualify. The guidelines give no robotics example, so this remains an open question.

Does a robot maker become a GPAI provider by fine-tuning a third-party model?

Only if the modification significantly changes the model. The Commission's indicative test is that the compute used for fine-tuning exceeds one third of the original model's training compute. Where that figure is unknown, a third of 10^23 FLOP applies, or a third of 10^25 FLOP for systemic-risk models. Even then, the obligations cover only the modification and the data used for it.

Is signing the General-Purpose AI Code of Practice mandatory?

No. The Code of Practice, published on 10 July 2025, is a voluntary tool. The Commission and the AI Board have confirmed it is an adequate way to demonstrate compliance, and signatories are promised less administrative burden and more legal certainty. Providers that do not sign must show compliance with the AI Act's obligations by other adequate means.

How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-11-10. Spotted an error? Write to info@humanoid4rent.com.

Quellen

Die Fakten in diesem Artikel wurden am Tag der Erstellung anhand dieser Seiten geprüft.

  1. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) – EUR-Lex
  2. General-Purpose AI Models in the AI Act – Questions & Answers – European Commission
  3. Commission Guidelines on the scope of the obligations for providers of general-purpose AI models established by Regulation (EU) 2024/1689 (AI Act) – European Commission
  4. The General-Purpose AI Code of Practice – European Commission
  5. Gemini Robotics 1.5 brings AI agents into the physical world – Google DeepMind
  6. NVIDIA Announces Isaac GR00T N1 – the World's First Open Humanoid Robot Foundation Model – and Simulation Frameworks to Speed Robot Development – NVIDIA Newsroom

Mehr aus dem Blog

Möchten Sie bei Ihrem nächsten Event einen Roboter?

Beschreiben Sie einmal Datum, Ort und die Aufgabe des Roboters. Anbieter antworten mit Angeboten, die Sie direkt miteinander vergleichen können.

Anfrage senden

Sie arbeiten in der Robotik? Schreiben Sie für unseren Blog. →