Nvidia unveiled Halos, a full-stack safety system for physical AI, positioning the platform for a market that ABI Research expects to reach 49 million level 3-5 autonomous vehicles installed by 2035 and that Omdia projects will absorb 60 million industrial robots between 2026 and 2035. The pitch: as AI-driven machines start sharing roads, factories and warehouses with people, safety has to be engineered across chips, operating systems, models and simulation — not bolted on at the end. Nvidia says Halos is the first system to do all of that from one vendor.
The announcement, written by Nvidia's Riccardo Mariani, frames the problem as one of physical action rather than software output. Roads and factories cannot be fenced off, models change with every update, and the space of scenarios a robot might encounter is too large to test in the real world alone.
For autonomous vehicles, Halos ties together Nvidia DRIVE AGX Thor for compute, the Hyperion reference architecture for level 4 vehicles, and Halos OS, a unified software foundation built on DriveOS. DriveOS 6.0 has been certified by TÜV SÜD to ISO 26262 ASIL D — the highest automotive functional-safety rating — and Nvidia's automotive engineering processes were separately certified to the ISO/SAE 21434 cybersecurity standard. TÜV Rheinland has also performed an independent UNECE safety assessment of Nvidia DRIVE AV.
Key facts
- 01ABI Research projects an installed base of 49 million level 3-5 autonomous vehicles by 2035.
- 02Omdia estimates 60 million industrial robots will be deployed between 2026 and 2035.
- 03Nvidia Halos spans DRIVE AGX Thor and Hyperion for AVs, plus IGX Thor with a dedicated Functional Safety Island for robots.
- 04TÜV SÜD certified Nvidia's DriveOS 6.0 to ISO 26262 ASIL D; TÜV Rheinland is inspecting IGX Thor and Halos OS for functional-safety certification.
- 05Geely, Isuzu, Nissan (with Wayve), Einride, Uber, Grab and Lyft are building on Nvidia Hyperion for level 4-ready vehicles and robotaxis.
Above the OS, Nvidia Alpamayo provides open reasoning vision-language-action models aimed at explainability in long-tail driving scenarios, and the Halos Safety Evaluation Framework generates the simulation-driven evidence developers need to build a safety case. That matters because a static one-time approval no longer fits products that ship model updates continuously.
The robotics stack mirrors the AV one. Nvidia IGX Thor is an industrial-grade module with a dedicated Functional Safety Island designed to support systems built to IEC 61508 and ISO 13849, the base standards for machinery and process safety. Halos Core for IGX handles fault detection, monitoring and deterministic communication, while the Holoscan Sensor Bridge feeds sensor data into AI and safety processing with checks for invalid inputs.
For validation, Nvidia is leaning heavily on simulation. Isaac Lab and Omniverse libraries let developers stress-test robot behavior across edge cases before physical deployment, and the open-source Halos Outside-In Safety Blueprint uses external cameras and vision AI agents to extend awareness beyond a robot's own sensors — a facility-level layer that treats the environment as part of the safety architecture.
Nvidia is also standing up its own inspection body. The Halos AI Systems Inspection Lab has been accredited by ANAB as an ISO/IEC 17020 inspection body, meaning it can convert safety, cybersecurity and AI-behavior requirements into repeatable inspections and prepare Halos integrations for final third-party certification. Members include AUMOVIO, Bosch, Gatik, Hesai, Lucid, MIRA, onsemi, PlusAI, Sony, Valeo and Wayve.
“Designing functional safety from the start is what separates a prototype from a scalable solution.”— Riccardo Mariani, Nvidia, author of the Halos announcement
The customer list on the AV side is the most concrete signal that Halos is not a slideware exercise. Geely, Isuzu, Nissan (running Wayve software) and freight operator Einride are building level 4-ready vehicles on Hyperion. Uber, Grab and Lyft are using the same platform to scale robotaxi programs. In robotics, KION Group is developing functional-safety agents for autonomous forklifts, and Agility is integrating IGX Thor and Halos Core into the safety system for its Digit 5 humanoid.
Nvidia's semiconductor peers are contributing pieces of the stack rather than competing head-on: Infineon, NXP, STMicroelectronics and Texas Instruments supply sensors and safety microcontrollers, while acontis and QNX provide the embedded software that runs safety functions predictably on IGX systems. That ecosystem structure — Nvidia at the center, with certified suppliers filling adjacent roles — is the same pattern that produced its dominance in AI training.
The counterweight worth naming is that certification of components is not certification of a deployed vehicle or a working factory line. ASIL D on DriveOS does not by itself resolve open regulatory questions about how AI models are updated over the air, how liability is assigned when a learned behavior causes an incident, or how the emerging ISO/IEC TS 22440 standard for AI-specific risks will be interpreted by insurers and workplace safety regulators. TÜV Rheinland's inspection work on IGX Thor is still in progress.
Nvidia's move is a bid to make itself the compliance layer for physical AI, not just the compute layer. If Halos becomes the default path to a certifiable AV or industrial robot, Nvidia captures a second toll gate on top of chip sales — one paid in switching costs, safety documentation and ecosystem lock-in. For rivals building physical AI silicon, matching raw performance is no longer enough; they now have to reproduce a decade of automotive safety artifacts to be considered for the deployments that actually ship.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
