Setup and Environments#

The course runs on the open-source i4h-workflows repository. Setup prepares a local checkout and confirms that your host can launch Isaac Sim and build the registered environments.

Clone the course-tested v0.7.1 release (f815694), then keep its path available to the commands in this course:

git clone --branch v0.7.1 --depth 1 https://github.com/isaac-for-healthcare/i4h-workflows.git
cd i4h-workflows
export I4H_WORKFLOWS="$PWD"

Learning Objectives#

By the end of this lesson, you’ll be able to:

Verify host requirements and run the agentic setup.
List registered environments, run import checks, and smoke-test a scene.

Choose a Coding Agent#

Run the coding agent on the same GPU host that will run Isaac Sim. The course skills are not tied to one harness. These launch paths work with the repository:

Codex CLI

From an isolated course workstation or VM, --yolo lets Codex run setup and validation without stopping for permission prompts:

codex --yolo -C "$I4H_WORKFLOWS"

--yolo disables approvals and the Codex sandbox. Use it only when the host itself is isolated and the checkout is trusted. On a shared workstation, omit --yolo and approve commands as they appear.

Hosted Model, Local Tools

The repository can connect OpenCode to an NVIDIA-hosted model while every shell command still runs on your GPU host:

export I4H_AGENT_API_KEY=nvapi-...
I4H_AGENT_PROFILE=nvidia-hosted \
I4H_AGENT_BASE_URL=https://integrate.api.nvidia.com \
./local-agent/run.sh agent

This is the cloud option for the hands-on course. A browser-only cloud agent cannot open the Isaac Sim window or use your local GPU.

OpenCode With a Local Model

The repository’s local-agent wrapper starts SGLang, writes the OpenCode configuration, loads the i4h skills, and leaves GPU 0 available for Isaac Sim when the host has another GPU:

./local-agent/run.sh start
./local-agent/run.sh agent

The default local profile needs one 96 GB GPU. The qwen3-coder-30b-fp8 profile needs one 48 GB GPU. See the repository’s Local Agent guide for installation and profile options.

Run It With a Prompt#

Set up the i4h workflow on this machine and tell me if any host requirements are missing.

Note

Prompts are open-ended. Keep them on topic and review each result before continuing because the same prompt can produce different steps or wording on another run.

What Happens Under the Hood?#

The setup skill first checks Ubuntu, uv, git, the NVIDIA driver and GPU, network access, and free disk space. It then runs the idempotent workflows/agentic/setup.sh. That script uses uv to sync the shared package and the Arena, policy, mimic, dataset, and annotator environments. It also prepares the pinned third-party checkouts. Cosmos remains optional.

After installation, the skill lists the Arena and policy environments. Two --dry-run checks resolve scissor_pick_and_place and import its registered modules without starting Isaac Sim. The final zero-action run launches the real scene, resets it, and advances 40 simulation steps without policy or teleoperation input. Expect the first launch to take longer while assets are downloaded and shaders are cached.

Troubleshooting a Missing Window#

If setup passes but the Isaac Sim window does not appear, ask the agent to launch Arena on your active local display and inspect DISPLAY and XAUTHORITY. This is common when the GPU host is reached through NoMachine or VNC and the agent’s shell is not attached to the desktop session.

Isaac Sim started but no GUI appeared. Detect the active local display for this NoMachine or VNC session, launch Arena on that display, and stop before changing system-wide display settings.

Optional deep dive:

Walk me through how setup.sh syncs components and third-party checkouts, then show how Arena discovers and registers environments at runtime.

What’s Next?#

With your host verified, the workflow set up, and a scene smoke-tested, you’re ready to start producing demonstrations. Continue to Data Collection.