Industry
Certification-ready is not certified: how to read physical AI's new safety stacks
What does "certification-ready" buy you? A path to a certificate. The certificate itself comes later, from someone else.
On 22 June 2026 NVIDIA launched Halos for Robotics. It's a stack: IGX Thor compute, Holoscan Sensor Bridge, the Halos OS software and an inspection lab. Agility is the first company to use it, for its humanoid Digit. The release words matter, so read them closely.
| The release says | What that means |
|---|---|
| The lab helps partners "prepare Halos integrations for third-party certification" | Preparation. A certifier issues the certificate later |
| TÜV Rheinland is inspecting IGX Thor, Halos OS and Holoscan Sensor Bridge "for functional safety certification readiness" | An inspection in progress |
| Agility and NVIDIA will check Digit against IEC 61508, ISO 13849 and ISO/IEC TR 5469 "before final third-party certification" | Final certification is still ahead |
| Halos Core is "available in early access" | Not a general release |
NVIDIA also calls the lab the world's first ANAB-accredited program for functional and AI safety for physical AI. That's NVIDIA's own claim. The accreditation is for the lab. It covers the lab's "competence and impartiality to evaluate" robotic AI systems, in the words of ANSI's president. It isn't a rating for any robot.
Money followed the idea. In August, FORT Robotics agreed to go public through a merger with Newbury Street II Acquisition Corp., a SPAC. The Robot Report put the listing valuation at $556.6 million. Both boards approved the deal. It still needs shareholder and regulatory approval, and the companies expect it to close in the fourth quarter of 2026. FORT is also in the Halos ecosystem, building functional safety agents on NVIDIA's Outside-In Safety Blueprint.
On 3 October, Agility and FORT signed a memorandum of understanding for Digit 5. Robotics 24/7 describes a three-part safety architecture. A safety pendant, on-robot communications and off-robot interfaces to external safety systems. The pendant sled carries emergency stop and enabling-device functions. It will "complement Digit 5's anticipated onboard Safe Human Detection system".
Anticipated. That word carries the sentence.
So what's the standard for AI inside a safety function? ISO/IEC TS 22440-1, Functional safety and AI systems. On 29 October the ISO catalogue listed it as a committee draft at stage 30.60, close of comment period. It's a Technical Specification, and it's still a draft.
Look at where the pieces sit. The external chain, the pendant, the off-robot bridge and the inspection all sit around the AI. None of them claims a SIL or a Performance Level for the learned model itself. Part 5 of this series takes up that pattern.
Here's the honest limit. I read a press release and trade press. I haven't seen an inspection report, a test result or the SPAC's SEC filings. Halos may well end in certificates for Agility and others. Today's documents don't say that has happened. And I read the TS 22440-1 stage on 29 October, so it may have moved.
Bottom line: certification-ready = preparation. A certificate names an issuer, a standard and a scope. Ask for all three.
Buyers and investors: when a vendor says "certified", do you ask who issued it, against which standard, and for what scope?
Part 3 of 12 · The Survey: what keeps physical AI safe in 2026.
Sources
- NVIDIA — press release, NVIDIA Announces Halos for Robotics, 22 June 2026 (read 29 October 2026)
- The Robot Report — FORT Robotics takes safety stack public via SPAC merger, August 2026 (read 29 October 2026)
- RuntimeWire — FORT Robotics plans a SPAC merger at a reported $557M enterprise value, 18 August 2026 (read 29 October 2026)
- Robotics 24/7 — Agility Robotics, FORT Robotics announce strategic partnership to advance humanoid robot safety, 3 October 2026 (read 29 October 2026)
- ISO — ISO/IEC CD TS 22440-1, Functional safety and AI systems, Part 1: Requirements, catalogue entry 89535 (read 29 October 2026)
Cite this
APA
Melchior, M. (2026, November 12). Certification-ready is not certified: how to read physical AI's new safety stacks. Physical AI Safety. https://physical-ai-safety.com/blog/certification-ready-is-not-certified
BibTeX
@misc{melchior2026certification,
author = {Melchior, Mati},
title = {{Certification-ready is not certified: how to read physical AI's new safety stacks}},
howpublished = {Physical AI Safety},
year = {2026},
month = nov,
date = {2026-11-12},
url = {https://physical-ai-safety.com/blog/certification-ready-is-not-certified}
}- Published