Nvidia is expanding its Halos safety system across autonomous vehicles and industrial robotics, arguing that machines acting in the physical world need continuous assurance rather than a one-time certification. The system connects hardware, operating software, AI-behavior checks, simulation and inspection tools.
For vehicles, Halos includes Drive AGX Thor computing, the Hyperion reference platform and an operating stack built on the safety-certified DriveOS. Nvidia’s evaluation framework is meant to generate evidence for safety cases across different levels of automation. The company also points to Alpamayo, its open vision-language-action models, as a way to make decisions in unusual driving situations easier to examine.
The robotics stack uses the IGX Thor module with a separate functional-safety component, plus software for fault detection, monitoring and reporting. Isaac Lab and Omniverse provide simulated environments for testing edge cases before and alongside real-world validation. An outside-in blueprint can add external cameras to monitor areas beyond a robot’s own sensors.
Halos is Nvidia’s own full-stack proposal, not an independent certification that a finished vehicle or robot is safe. Product makers still need domain-specific testing and third-party assessment, especially after software, models, tasks or operating conditions change. The practical shift is to preserve safety evidence throughout the deployment lifecycle instead of treating approval as the finish line.