The analysis
Weekly posts on Physical AI safety. Data-backed. Source-cited.
A specialist publication for the people building, regulating, and capitalising the next decade of Physical AI. Every post starts from a primary source, cites the source, and argues for one thing at a time. No speculation where data exists. No data where speculation would be more honest.
Featured
Under five months from a regulation that binds an industry which is 98% SMEs, not one trade association, notified body or consultancy has published a readiness figure. What exists instead is the industry's own request for a delay — and that is more revealing than a survey.
Regulation
Under five months from a regulation that binds an industry which is 98% SMEs, not one trade association, notified body or consultancy has published a readiness figure. What exists instead is the industry's own request for a delay — and that is more revealing than a survey.
Engineering
Most early-stage teams underestimate certification time by at least 2x. IEC 61508 defines 16 lifecycle phases. Mapped to a project timeline: 12-24 months for SIL 2. SIL 3: add 6-12 months. Phase 6 is where teams are always surprised.
Engineering
Four questions that separate serious Physical AI safety engineering from marketing copy. SIL target, redundancy architecture, re-certification plan, fallback behavior. If a vendor can't pass any — save your time.
Standards
The most used phrase in Physical AI marketing — and the least defined. Not in IEC 61508. Not in ISO 13849. Four engineering markers that distinguish the real thing from the brochure.
Regulation
20 January 2027. Cut-off date regulation. No transition period. The old Machinery Directive stops, the new Regulation starts. Every autonomous robot sold in Europe after that date needs to comply.
Regulation
Four dates now govern AI safety in European machinery, and three of them moved in the last two weeks. Read them in order and an eighteen-month window appears where autonomous machines are governed by a regulation finalised before the AI provisions were written.
Economics
A common objection in Physical AI: safety certification is too expensive. But 59% of product recalls now cost between $10M and $50M each — up from 39% in 2024. A single recall or liability case dwarfs the certification budget. Certification is insurance priced as a line item.
Engineering
Five unsolved problems. Any one solved cleanly in three years is a decade-defining company. Real-time verification of learned policies, fleet-level anomaly detection, continuous certification, composable safety guarantees, root-cause attribution.
Engineering
If both channels share a power supply, firmware, thermal zone, or design flaw — they fail together. IEC 61508 quantifies this as the β factor. A dual-channel system with β = 10% has one in ten failures defeat the redundancy entirely.
We use cookies
This site uses essential cookies to function and, with your consent, analytics cookies (Google Analytics) to understand how the site is used. Learn more.