Deployment Guide
The Safety Core sits between AI Perception and the machine it protects. It reads perception events from Kafka, checks that the cameras and the perception pipeline can still be trusted, and sends a decision to the machine’s controller over UDP. When it cannot trust its inputs, or loses contact with one of its own components, it holds the machine in a safe state until an operator releases it.
This guide is for someone who runs the Safety Core. It starts the Safety Core Docker container
with docker run, and every page gives commands you can run as written.
Choose a path
Read the pages in this order:
- Prerequisites: one checklist, with a check for every item.
- Applications: what each reference application protects and what it sends.
- Quick Deployment or Custom Deployment.
- Troubleshooting, when something does not start or does not decide.
Supported targets
The operating system and driver versions for each target are listed in Prerequisites in Getting Started.
NGC artifacts
The 1.4 artifacts are build 6b92731, published in the NGC Catalog under nvidia/halos-outside-in.
What runs where
The nv-psf container runs every Safety Core process except the command receiver and
fsicom-agent:
Building on the Safety Core?
The Integration Guide explains the architecture, each component, the Debian packages and running without Docker, and how to write your own decision maker. The Reference section lists every container option, launcher flag and configuration key.