Nvidia has announced NVIDIA Halos, an all-new safety system that pairs its automotive hardware and software safety solutions with its AI research in autonomous vehicle (AV) safety. Altogether, this combination of chips and software aims to help OEMs ensure the safe deployment of their AVs, from the cloud to the car, with a focus on AI-based, end-to-end AV stacks.
The multinational tech giant highlighted that the system works on three different, but complementary, levels. Its technology level, for example, spans platform, algorithmic and ecosystem safety, while its development level includes design-time, deployment-time and validation-time guardrails.
At the computational level, Halos spans AI training to deployment, using three of Nvidia’s key computing systems to do so: Nvidia DGX for AI training, Nvidia Omniverse and Nvidia Cosmos running on Nvidia OVX for simulation, and Nvidia Drive AGX for deployment. Further support comes from the Nvidia AI Systems Inspection Lab, which the company says serves as an entry point to Halos while allowing OEMs and developers to verify the safe integration of their products with its technology ecosystem.
Halos is, likewise, built on three key focus areas: platform safety, algorithmic safety and ecosystem safety. The first is fulfilled through a safety-assessed SoC with hundreds of built-in safety mechanisms. This also includes Nvidia DriveOS software, a safety-certified operating system that extends from CPU to GPU; a safety-assessed base platform that delivers the foundational computer needed to enable safe systems for various types of applications; and Drive AGX Hyperion, a hardware platform that connects SoC, DriveOS and sensors in an electronic control unit architecture.
To ensure algorithmic safety, Halos incorporates libraries for safety data loading and accelerators, as well as application programming interfaces for safety data creation, curation and reconstruction to filter out undesirable behaviors and biases before training, for example. It also features rich training, simulation and validation environments that leverages the Nvidia Omniverse Blueprint for AV simulation with Nvidia Cosmos world foundation models to train, test and validate AVs. In addition, it offers a diverse AV stack that combines modular components with end-to-end AI models to ensure safety with advanced AI models in the loop.
In terms of ecosystem safety, Halos more broadly includes safety datasets with diverse, unbiased data, as well as safe deployment workflows. This comprises triaging workflows and automated safety evaluations, along with a data flywheel for continual safety improvements that together work to demonstrate leadership in AV safety standardization and regulation.
