An agent can set up the NVIDIA Halos test environment: the repository includes a ready-made skill

Claude Code, Codex and Cursor can deploy Outside-In Safety Blueprint using NVIDIA's ready-made instructions.

An AI agent brings up NVIDIA Halos components and runs a test scenario. The repository includes the hoisa-deploy-profile skill for this.

Halos brings together hardware, models and tools for validating autonomous systems. NVIDIA introduced the automotive system on Mar 18, 2025, and Halos for Robotics on Jun 22, 2026. Outside-In Safety Blueprint gives developers a ready-made framework for validating robot behavior.

Deploying with an agent. The Blueprint code is available on GitHub under Apache-2.0. After cloning the repository, connect the skills to one of the coding agents using symlinks and choose a deployment profile:

- `base` connects Safety Core to an existing perception data stream. - `sil` runs validation with a robot simulation. - `hil` links simulation on x86 to Safety Core on IGX Thor.

In version 1.3 EA, NVIDIA replaced Isaac Sim 5.1 with 6.0 and added the HIL profile. Halos Core for IGX is available to registered developers in early access for Linux, including in combination with QNX OS for Safety 8.0.

Access before deployment. Downloading Safety Core requires approved early access in NGC. You then need to accept the invitation and license, obtain an API key and run a separate `docker login nvcr.io`. Authentication through the NGC CLI is not enough to download images.

The manual route starts with VSS Warehouse Blueprint 3.2.1. After deploying it, fill in `deployments/profiles/base.env` and run `docker compose --env-file profiles/base.env up -d` from the `deployments` directory. NVIDIA requires Docker Engine from 28.3.3 to a version below 29.5.0: downloading NGC images breaks in 29.5.0 and later.

A Brev Launchable is available for cloud testing. In JupyterLab, open `deploy_hoisa_launchable.ipynb`, enter the NGC API key and select Run All Cells. The notebook launches perception, Safety Core and Isaac Sim. SIL requires a GPU with RT cores, such as an L40S or RTX A6000, and at least 200 GB of disk space; 250 GB is recommended.

Agility Robotics integrated Halos into Digit 5. Previously, workcell safety depended on external sensors. Now the necessary sensors and hardware are inside the robot.

Sources: [Announcement of the Halos automotive system on Mar 18, 2025](https://blogs.nvidia.com/blog/halos-safety-system-autonomous-vehicles/), [Outside-In Safety Blueprint repository](https://github.com/NVIDIA/halos-outside-in-safety), [agent deployment instructions](https://docs.nvidia.com/halos-outside-in/latest/getting-started/agentic-deployment.html).

In the HIL profile, simulation runs on x86, while IGX Thor runs Safety Core.