Nvidia launches Halos, a full-stack safety system for Physical AI

Safety in Physical AI no longer ends with a one-time pre-deployment check. Nvidia has unveiled Halos, a first-of-its-kind full-stack safety system that spans hardware, software, AI models, the operating environment and the entire lifecycle — elements that until now have been handled separately. The move arrives as the industry braces for a wave of mass deployment: ABI Research projects 49 million Level 3-5 autonomous vehicles by 2035, and Omdia estimates roughly 60 million industrial robots will be deployed between 2026 and 2035.

The numbers dictate a new safety model

Four shifts drive the need for a different approach. Dynamic environments — roads, warehouses, factories — cannot be fully controlled through static zones or physical barriers; systems must perceive changing conditions, adapt behavior and reach a safe state when the unexpected occurs. AI behavior demands its own assurance: the emerging ISO/IEC TS 22440 standard is beginning to address AI-specific risks alongside traditional functional safety. Deployment is a continuous process; software updates, new models and shifting operating conditions require additional safety validation whenever a material change occurs. Validation at scale demands the combination of simulation, synthetic data and scenario replay, because the space of possible scenarios exceeds what can be tested in the real world alone.

A decade of AV experience converges on one platform

Halos rests on more than ten years of autonomous-vehicle safety development, spanning functional-safety expertise, sensor fusion, AI behavior assurance, computer vision, simulation and real-world validation. The principles are shared between autonomous vehicles and robotics, but the platforms, standards and evidence remain domain-specific.

Hardware, OS and middleware, all under one roof

On the hardware side, DRIVE AGX Thor delivers safety-designed accelerated computing, while Hyperion serves as a full-stack vehicle platform and Level 4 reference architecture. The Halos OS is built on DriveOS, which holds ASIL-D certification, and Halos Core with Halos Middleware support system isolation, monitoring and deterministic communication — infrastructure without which stringent safety requirements are difficult to meet.

End-to-end models and simulation: explainability for edge cases

Nvidia Alpamayo provides open vision-language-action models with reasoning capability, bringing explainability to long-tail scenarios — the rare cases that are hard to anticipate. On the validation side, Halos integrates simulation and synthetic data generation to cover the vast scenario space that road testing alone cannot reach.