Skip to main content
Back to timeline
NVIDIA ResearchSource publication:

Physical AI at Scale Needs Safety From Chip to Fleet: NVIDIA Launches Full-Stack Halos

Synopsis

Physical AI is moving from research into large-scale deployment, with ABI Research projecting 49 million level 3-5 autonomous vehicles by 2035 and Omdia estimating roughly 60 million industrial robots deployed between 2026 and 2035. NVIDIA introduced Halos, a full-stack safety system that links hardware, operating system and middleware, end-to-end models, simulation and validation, and an inspection lab into one safety chain across design, deployment and validation, which it calls the first and only full-stack safety system for physical AI. TÜV SÜD has certified NVIDIA's automotive product lifecycle software process and DriveOS 6.0 to ISO 26262 ASIL D, and ANAB has accredited the inspection lab as an ISO/IEC 17020 body.

AI-generated editorial illustration: Why Deploying Physical AI at Scale Demands Safety at Every Layer

Interpretation

Safety for physical AI is required to span hardware, software, AI behavior, operating environment and the deployment lifecycle, rather than a one-time check before deployment. Autonomous vehicles keep expanding commercially after years of testing and benchmarking, and robots are moving into factories and warehouses shared with people, so manufacturers, regulators, insurers and workplace safety teams need evidence that hardware, software, AI behavior and operating environments work together safely without human intervention. ABI Research projects an installed base of 49 million level 3-5 autonomous vehicles by 2035, and Omdia estimates about 60 million industrial robots will be deployed between 2026 and 2035.

Halos splits safety across layers: on the vehicle side, DRIVE AGX Thor supplies safety-engineered accelerated compute and Hyperion the level 4 vehicle platform and reference architecture, Halos OS sits on ASIL-D certified DriveOS, Alpamayo offers explainable reasoning vision language action models, and the Halos Safety Evaluation Framework generates evidence for safety cases. Halos Core and Halos Middleware handle system isolation, monitoring and deterministic communication, connecting cloud AI development and simulation with in-vehicle deployment so safety evidence stays traceable across the vehicle lifecycle. On the robotics side the counterparts are the IGX Thor industrial module with a dedicated Functional Safety Island aimed at IEC 61508 and ISO 13849, Halos Core for IGX, Holoscan Sensor Bridge, Isaac Lab and Omniverse libraries for simulation, and the open source Halos Outside-In Safety Blueprint, which uses external cameras and vision AI agents to extend awareness beyond onboard sensors.

Halos turns safety, cybersecurity and AI safety requirements into repeatable inspections and helps integrations prepare for final system-level certification by third-party agencies. TÜV SÜD has certified NVIDIA's Automotive Product Lifecycle software process and DriveOS 6.0 to ISO 26262 ASIL D and its automotive engineering processes to ISO/SAE 21434, and TÜV Rheinland performed an independent UNECE safety assessment of NVIDIA DRIVE AV. For robotics, TÜV Rheinland is inspecting IGX Thor, Halos OS and Holoscan Sensor Bridge for functional-safety certification readiness, building on TÜV SÜD's inspection of the Thor SoC and Halos Core for ISO 26262; ANAB has accredited the Halos AI Systems Inspection Lab as an ISO/IEC 17020 inspection body.

Perspective

The report covers the layered composition of NVIDIA Halos for autonomous vehicles and robotics, its partner list and the status of third-party assessments. It is intended for researchers and engineers working on physical AI safety engineering and compliance. The source states no next experiment or deployment plan and only points to further Halos material.

The source calls Halos the first and only full-stack safety system for physical AI, a claim from the releasing organization itself. The robotics-side TÜV Rheinland work is described as certification-readiness inspection rather than completed certification. The source notes that emerging standards such as ISO/IEC TS 22440 are only beginning to address AI-specific risks, so requirements are still forming.