The robot gets certified. The evidence still has to keep talking.
NVIDIA Halos for Robotics is not a new model. It is a full-stack safety system, built on 18,600-plus engineering-years of autonomous-vehicle safety data, that helps a robot pass certification. Certification is a moment. What we build is what happens after.
· by Risto Anton Päärni
The note, in one line
On 22 Jun 2026, NVIDIA announced Halos for Robotics — IGX Thor compute with a dedicated Functional Safety Island, a safety-certified OS, an Outside-In camera system that lets a robot slow down instead of hard-stopping, and the first ANAB-accredited lab that pre-assesses a robot against IEC 61508 and ISO 13849 before formal certification. Agility Robotics is already integrating it into Digit for Amazon and Toyota floors. We cheer it. And we build the layer it does not cover.
Good news from a good partner.
We have written before about the supercomputer that moved onto the robot. Halos is the next layer up — not more compute, but the safety architecture that lets a regulator trust what the compute decides.
And because cheering without thinking is not a business, here is the honest part.
What NVIDIA actually shipped
Halos is a full stack, not a checkbox. IGX Thor carries a dedicated Functional Safety Island, IEC-61508-SIL-3-class, physically isolated from the rest of the compute. Halos Core runs on top — a safety-certified OS layer on Linux and QNX OS for Safety 8.0, deterministic enough to back a real-time safety decision. The Outside-In Safety Blueprint adds infrastructure-mounted cameras that watch a work cell’s blind spots, so a robot adjusts speed and trajectory instead of slamming into a hard E-stop every time a person walks near it.
The number worth sitting with is the provenance: Halos is built on 18,600-plus engineering-years of NVIDIA’s autonomous-vehicle safety work, repointed at humanoids and industrial robots. That is not a marketing figure we are repeating politely — it is the reason the AI Systems Inspection Lab, the first ANAB-accredited program of its kind, can pre-assess an integration against IEC 61508 and ISO 13849 before a Notified Body ever sees it. Agility Robotics is first through the door, integrating Halos into the 5th-generation Digit humanoid for warehouse and factory floors — Amazon and Toyota among the named customers.
Why Halos matters to us
Every EU physical-AI deployment we touch eventually runs into the same two clocks: the Machinery Regulation (EU) 2023/1230, which moves AI-driven safety components to mandatory third-party certification from 20 Jan 2027, and the AI Act’s embedded high-risk obligations, which land on the same machines from 2 Aug 2028. Notified Body queues already run six to twelve months. The honest math is that the runway is already tight.
Halos changes what a deployer can point to when a Notified Body asks “how did you validate this.” Before Halos, that answer was a bespoke, in-house safety-engineering project. After Halos, it can be an accredited pre-certification report. That is a real shift — it moves the floor every EU physical-AI seller has to clear.
What a certification does not give you
Certification is a snapshot. It says the robot, as built, met a standard on the day it was assessed. The Machinery Regulation and the AI Act do not stop asking questions on that day — Article 9 risk management, Article 12 record-keeping, Article 14 human oversight are all continuous duties, not one-time boxes.
A Halos-certified robot that runs for two years without a signed, auditable evidence trail is a robot that passed an exam and then stopped showing its work. That is the gap we build for.
Where Halos fits in our stack
One deployment, two jobs. NVIDIA gets the robot certified. We keep the deployment evidenced.
The certification · NVIDIA Halos
Functional Safety Island, Halos Core, Outside-In cameras, the ANAB-accredited Inspection Lab. The pre-certification pipeline that gets a robot ready for a Notified Body. NVIDIA’s ground, and a strong one.
The evidence · the governed edge
A Situation Room the customer owns, ingesting the same Outside-In telemetry, producing signed Article-9/12/14 evidence continuously — not once, at the certification date.
We do not compete with the safety island. We would not try — we do not ship robot silicon. The Outside-In cameras are the one place the two systems physically touch: their telemetry is an ingestion surface for our Control Room and Situation Room, the same governed loop we have described before.
Why we will not over-claim
NVIDIA called Halos a safety system, not a compliance guarantee. We will keep the same restraint. Passing IEC 61508 and ISO 13849 says the hardware and OS can be trusted to fail safely. It does not, on its own, produce the immutable Article-12 log a regulator asks for eighteen months into a deployment.
A pre-certification report is not an audit trail. We will not say our evidence layer makes a robot safer — Halos does that. We say it keeps the proof of that safety alive, signed, and local, for as long as the robot is on the floor. That is enough.
The short version
- Halos for Robotics is a full-stack functional-safety system, not a new model — built on 18,600+ engineering-years of NVIDIA autonomous-vehicle safety data.
- Its ANAB-accredited Inspection Lab pre-assesses against IEC 61508 and ISO 13849 before third-party (Notified Body) certification. Agility Robotics’ Digit is first through, deploying for Amazon and Toyota.
- This lands right as the EU Machinery Regulation (20 Jan 2027) and AI Act embedded high-risk rules (2 Aug 2028) tighten — a pre-certification pipeline moves the expected baseline forward of both deadlines.
- Certification is a snapshot. Article 9/12/14 obligations are continuous. We do not compete with the safety island — we keep the evidence trail alive after the certificate is issued.
- Halos gets the robot certified. Our Control Room keeps it evidenced.
Read next
- NVIDIA Jetson Thor — the supercomputer moves onto the robot, the governance sits beside it
- The moat is not the model — NVIDIA makes it possible, EU-sovereign deployment makes it defensible
- The three-piece loop — Lifetime World Model V1 for control rooms
- Europe’s AI factory came home — 35 supercomputers, and why the edge still needs a box
Risto Anton Päärni
Founder, CEO · Lifetime Oy · Editor in Chief, Lifetime Scope Journal