NVIDIA Halos Brings Full-Stack Safety to Physical AI Deployment

NVIDIA has introduced Halos, a comprehensive safety platform designed to address the unique challenges of deploying autonomous vehicles and robots at scale. The system provides safety engineering across hardware, software, AI models, and operating environments—a response to projections that 49 million level 3-5 autonomous vehicles and 60 million industrial robots will be deployed by 2035.

Why it matters

Physical AI systems make decisions that result in physical actions in shared human spaces. Unlike software that fails gracefully, autonomous machines require provable safety across their entire stack and lifecycle. As these systems move from controlled testing to commercial deployment, manufacturers, regulators, and insurers need systematic evidence that hardware failures, software bugs, AI edge cases, and environmental uncertainties can all be managed without human intervention.

Four shifts driving new safety requirements

NVIDIA identifies four fundamental changes that distinguish physical AI safety from traditional approaches. Dynamic environments like roads and warehouses cannot rely on static barriers or fixed zones—systems must perceive changing conditions and reach safe states when encountering unexpected scenarios. AI behavior itself requires dedicated assurance methods, with emerging standards like ISO/IEC TS 22440 beginning to codify AI-specific risk assessment.

Deployment has become an ongoing process rather than a one-time event. Software updates, model improvements, and changing operating conditions mean material changes may trigger additional safety validation. Finally, the complexity and volume of potential scenarios make pure real-world testing insufficient—simulation, synthetic data generation, and scenario reconstruction become essential validation tools.

Architecture for autonomous vehicles

For AV development, Halos spans multiple layers. NVIDIA DRIVE AGX Thor provides safety-engineered compute, while Hyperion delivers the full vehicle platform and reference architecture for level 4 systems. Halos OS offers a unified software foundation built on ASIL-D certified DriveOS, with Halos Core and Halos Middleware supporting system isolation and deterministic communication.

The Alpamayo vision-language-action model brings explainability to edge cases, while the Halos Safety Evaluation Framework provides tools for generating evidence across different automation levels. The architecture connects cloud-based development with in-vehicle deployment, maintaining safety traceability throughout the vehicle lifecycle.

Geely, Isuzu, Nissan (using Wayve software), and Einride are building level 4-ready vehicles on the platform. Mobility providers including Uber, Grab, and Lyft are using Hyperion to scale robotaxi programs.

Industrial robotics implementation

For robotics, NVIDIA IGX Thor combines accelerated computing with functional safety on a single platform, featuring a dedicated Functional Safety Island designed for standards including IEC 61508 and ISO 13849. Halos Core for IGX provides fault detection, monitoring, and reporting capabilities that connect sensors, actuators, and safety components.

Holoscan Sensor Bridge links sensor data with AI processing and safety functions, helping systems identify invalid information and execute defined responses. Isaac Lab and Omniverse libraries enable testing across relevant conditions and edge cases. The open-source Halos Outside-In Safety Blueprint uses external cameras and vision AI to extend awareness beyond onboard sensors for facility-level monitoring.

Agility is integrating IGX Thor and Halos Core into the safety system for its Digit 5 humanoid robot. KION Group is developing functional safety agents for autonomous forklifts.

Independent certification progress

TÜV SÜD has certified NVIDIA's automotive software process and DriveOS 6.0 to ISO 26262 ASIL D, along with automotive engineering processes to ISO/SAE 21434. TÜV Rheinland performed an independent UNECE safety assessment of NVIDIA DRIVE AV and is inspecting IGX Thor, Halos OS, and Holoscan Sensor Bridge for functional-safety certification readiness. ANAB has accredited the NVIDIA Halos AI Systems Inspection Lab as an ISO/IEC 17020 inspection body.

The details were first reported by NVIDIA in a company blog post by Riccardo Mariani.