AI Robot Safety Certification: Machinery Regulation 2027

Q: Can a neural network be safety certified for a robot?

Not in the way a safety PLC is. Functional-safety standards such as IEC 61508, IEC 62061 and ISO 13849 give no accepted method for assigning a performance level or SIL to a learned model. The practical answer is to keep the model in a performance channel and certify a deterministic safety controller with its own sensors and safe torque off that bounds whatever the model commands.

Q: What does the EU Machinery Regulation 2023/1230 say about AI and machine learning?

Annex I Part A lists safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions, and machinery embedding them, so they need third-party conformity assessment under Article 25(2). Annex III adds a bounded task and movement space, one year of recorded safety-related decision data, and a way to correct the machine at all times.

Q: When does the EU Machinery Regulation apply?

The European Commission's machinery page states it applies on a mandatory basis from 20 January 2027, and machinery placed on the market before then must comply with Directive 2006/42/EC. The originally published Official Journal text says 14 January 2027, so confirm the consolidated date with your notified body and plan your technical file for the earlier date.

Q: Does a frozen model that never learns after release fall under Annex I Part A?

Recital 55 says the third-party requirement applies only to systems with self-evolving behaviour using machine learning, not software incapable of learning. A frozen model may therefore fall outside items 5 and 6, but it would still be the element achieving a safety function, which functional-safety standards cannot assess. Expect scrutiny and budget for the argument, or keep it out of the safety function.

Q: How does the AI Act interact with the Machinery Regulation for robots?

AI systems that are safety components of products covered by Annex I legislation, which includes machinery, are high-risk under Article 6(1). The AI Act Service Desk text of Article 113, amended by the Digital Omnibus, applies those obligations from 2 August 2028. The Machinery Regulation applies about 18 months earlier, so design the envelope, recorder and risk file once for both.

Q: What should a robot's safety decision log contain?

The proposed command from the AI policy, the active envelope limits, the safety controller's verdict and any intervention, safety sensor states, model and envelope configuration versions and the operating mode, with trustworthy timestamps. Annex III requires safety-related decision data to be kept for one year. Hash-chain or sign it and upload from an on-robot ring buffer so it survives a reasoned request.

Q: Can AI perception improve robot safety without needing certification?

Yes, if it can only make the robot more conservative. Using a person-detection model to slow the robot earlier than the safety-rated scanner would adds a real safety benefit, while the certified scanner and safety controller still ensure the protective stop. Show that the model can only reduce speed or shrink limits and never extend the envelope, and it stays a performance component.

Q: Has ISO 10218 changed, and does ISO/TS 15066 still apply?

ISO 10218 was revised in 2025 into Part 1 for the robot and Part 2 for the robot application and cell, and much of the collaborative-operation material from ISO/TS 15066 was consolidated into the revision. Technical files that cite ISO 10218:2011 plus ISO/TS 15066 as the current pair should be reviewed against the 2025 editions before a 2027 conformity assessment.

Satyam Kumar

返回博客

ai-architecture

Safety Certification for AI-Driven Robots: Certify the Envelope, Not the Policy, and Keep Learning Out of the Safety Function (2026)

By Satyam KumarOctober 13, 202617 min read

ai robot safety certification machinery regulation 2023/1230 eu machinery regulation ai functional safety machine learning iso 10218 2025 iec 62061 iec 61508 safety controller architecture safe reinforcement learning shielding robot ce marking notified body robotics ai act annex i machinery collaborative robot safety amr safety physical ai safety robotics compliance 2026

Safety Certification for AI-Driven Robots: Certify the Envelope, Not the Policy, and Keep Learning Out of the Safety Function (2026)

Frequently Asked Questions

分享这篇文章

Twitter LinkedIn WhatsApp

Satyam Kumar

Founder & AI Architect, AppScale LLP

人工智能和云架构师。帮助团队构建可扩展到数百万的系统。

LinkedIn GitHub

Safety Certification for AI-Driven Robots: Certify the Envelope, Not the Policy, and Keep Learning Out of the Safety Function (2026)

Frequently Asked Questions

分享这篇文章

Comments

Leave a comment