NVIDIA Unveils 'Halos', a Full-Stack System for Physical AI Safety

As Physical AI moves beyond the research stage into large-scale deployment, there is growing recognition that existing static safety models cannot handle dynamic environments and AI-specific risks. ABI Research forecasts the deployment of 49 million Level 3-5 autonomous vehicles by 2035, while Omdia projects 60 million industrial robots between 2026 and 2035, emphasizing the need for safety assurance.

NVIDIA Halos' Full-Stack Safety Architecture

NVIDIA has unveiled 'NVIDIA Halos', a full-stack safety system that supports safety engineering across all layers of design, verification, and deployment. It provides a platform tailored to the specific characteristics of the autonomous vehicle (AV) and robotics domains.

• Autonomous Vehicles (AV): Built on the safety-engineered computing platform DRIVE AGX Thor and Hyperion, a full-stack vehicle platform and reference architecture for Level 4 AVs. It supports system isolation, monitoring, and deterministic communication through Halos OS, Halos Core, and Halos Middleware, based on the ASIL-D certified DriveOS.
• Robotics: Supports IEC 61508 and ISO 13849 standards through the industrial module IGX Thor (including a dedicated functional safety island) and Halos Core for IGX. It connects sensor data with AI processing via the Holoscan Sensor Bridge and performs simulation and verification, including edge cases, through Isaac Lab and Omniverse.

Ecosystem and Independent Verification

NVIDIA is building the Halos ecosystem in collaboration with major manufacturers and partners such as Geely, Nissan, and Agility Robotics. It has also secured independent third-party verification, including ISO 26262 ASIL D certification and ISO/SAE 21434 process certification for DriveOS 6.0 from TÜV SÜD, and UNECE safety assessments from TÜV Rheinland. The NVIDIA Halos AI Systems Inspection Lab is accredited by ANAB as an ISO/IEC 17020 inspection body to support final safety inspections prior to deployment.