Introduction

Demonstrating AV safety requires proving performance at rates on the order of one failure every 10⁷ to 10⁸ hours of driving – a volume no pre-production fleet can drive and no single tool can address. This demands a data-driven combination of on-road testing, simulation-based evaluation, and production fleet monitoring. Automotive standards are increasingly formalising exact safety requirements as the industry shifts to AI-based architectures.

This document describes the NVIDIA Halos Safety Assurance Framework for Evaluation (SAFE) – an ecosystem of tools and accompanying guidelines for generating sufficient evidence to support AV safety cases across autonomy levels, from L2 active safety systems to L4 robotaxi autonomy. NVIDIA Halos SAFE builds on more than 330 research papers and 1,000 patents developed within NVIDIA Halos OS.

Our mission is twofold:

• We aim to draw out a specific and actionable recipe of the necessary components for standards-compliant safety evaluation, from L2 to L4 autonomy.
• We aim to offer a select number of these components to the market, such as NVIDIA NuRec and Cosmos Transfer. As many ecosystem partners also specialize in individual components, we encourage partnership and collaboration to implement Halos SAFE as a complete solution.

Safety evaluation is not achieved by a single silver-bullet technology, but is a combination of data, software, and processes that come together to create statistical evidence. In this document we will share a high-level overview of NVIDIA Halos SAFE.