Isaac Automator

Isaac Automator: deploy and operate a cloud Isaac Workstation

Isaac Automator provisions a GPU virtual machine in a public cloud and configures it as a ready-to-use Isaac Workstation: a remote-desktop VM with Isaac Sim, Isaac Lab, and/or Isaac Lab Arena installed, plus lifecycle controls, data transfer, and import of existing deployments. This skill drives that CLI to stand up, use, and tear down a workstation.

Source: https://github.com/isaac-sim/IsaacAutomator

When to use this skill

Use it when the user wants to deploy, connect to, manage cost on, or tear down a cloud Isaac workstation via Isaac Automator. Do NOT use it for installing Isaac Sim/Lab on local hardware, for changing Isaac Automator's own code (Python/Terraform/Ansible), or for writing simulation or robot-learning code.

Before you act

• It provisions real, paid cloud infrastructure. A running deployment bills by the hour; a stopped instance still bills for storage. Treat every deploy and start as live-fire.
• Always clean up. Stop when idle (./stop), and destroy (./destroy <name> --yes) when done. Only destroy stops all cost.
• Run non-interactively. The deploy commands prompt for any required option you omit; an unanswered prompt blocks an agent. Supply every required option and use --existing replace (never ask).
• Never print or commit credentials. Pass them via environment variables and reference them by name.

Prerequisites

1. Docker installed and running.
2. Cloud credentials for the target cloud. For AWS the simplest path is AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY in the environment; otherwise the tool falls back to an interactive SSO login.
3. Build the container image once, from the repo root: ./build.

Everything runs inside the isaac_automator Docker image, and the top-level scripts wrap it. The ./run helper uses an interactive TTY; for a headless agent, call the container directly instead:

docker run --rm --network host -v "$(pwd)":/app \
  -e AWS_ACCESS_KEY_ID -e AWS_SECRET_ACCESS_KEY \
  isaac_automator "<command>"

Instructions

1. Gather requirements

Ask the user if not already clear:

1. Cloud - AWS, GCP, Azure, or Alibaba Cloud?
2. Apps - Isaac Sim, Isaac Lab, Isaac Lab Arena (any combination), and which versions (git ref) or no.
3. Region and instance type - or accept the cheapest viable GPU default.
4. Budget/lifetime - so you stop or destroy it promptly afterward.

2. Deploy

Each deployment has a unique name (lowercase letters, digits, -). Common options (all clouds): --deployment-name/--dn, --region, --instance-type, --ingress-cidrs (myip = your current public IP), --isaacsim / --isaaclab / --isaaclab-arena (git ref or no), --from-image / --not-from-image, --existing (replace / repair / modify / run_ansible), --upload / --no-upload.

Per-cloud commands and extras:

• ./deploy-aws - --region, --instance-type.
• ./deploy-gcp - adds --zone, --project, --isaac-workstation-gpu-count.
• ./deploy-azure - adds --resource-group, --login / --no-login.
• ./deploy-alicloud - --region.

--from-image deploys from a pre-built image (10-15 min, cheaper); --not-from-image does a full build-from-source (45-60 min). A successful deploy ends with an Ansible PLAY RECAP showing failed=0 and prints connection info (also saved to state/<name>/info.txt).

3. Connect

• noVNC (browser, 2D desktop): ./novnc <name> prints a URL on port 6080. Good for the desktop, files, and launching apps.
• NoMachine (live 3D viewport): connect a NoMachine client to the VM's public IP, key-based auth with state/<name>/key.pem, user ubuntu. Use this to watch the live Isaac Sim/Lab 3D viewport - Omniverse Kit renders through a Vulkan surface that noVNC does not capture, so noVNC shows the desktop but usually not the rendered 3D.
• SSH (shell): ./ssh <name> for logs, launching commands, or recording viewport video headlessly.

4. Move data and auto-launch

• Upload inputs: put files in the local uploads/ folder, then ./upload <name> (copies to ~/uploads on the VM). You can also upload during deploy with --upload (default) instead of --no-upload.
• Download results: ./download <name> pulls ~/results from the VM into your local results/ folder.
• Autorun on boot: place a script at uploads/autorun.sh; when present, the desktop runs it on boot and after each start instead of launching Isaac Sim by default.

5. Control cost, repair, import, tear down

• ./stop <name> - stop the instance (keeps it and its public IP; pauses compute billing).
• ./start <name> - resume; the public IP is preserved across stop/start.
• ./repair <name> - re-apply configuration if the workstation came up unhealthy.
• ./import --cloud <aws|gcp|azure|alicloud> ... - bring existing cloud resources under Isaac Automator management (creates local state for a deployment that already exists).
• ./destroy <name> --yes - delete the deployment and stop all billing. Always do this when finished, then confirm no resources tagged with the deployment name remain.

Faster, repeated deploys (prebuilt images)

For many deploys, bake a prebuilt image once with ./image-aws / ./image-gcp / ./image-azure, then deploy with --from-image. Use a full --not-from-image deploy when you need a from-scratch install or no prebuilt image exists.

Examples

Deploy Isaac Sim + Isaac Lab on AWS, locked to your IP:

./deploy-aws \
  --deployment-name my-rig \
  --region us-east-1 \
  --instance-type g5.2xlarge \
  --not-from-image \
  --ingress-cidrs myip \
  --isaacsim v6.0.0 \
  --isaaclab release/3.0.0-beta2 \
  --isaaclab-arena no \
  --existing replace \
  --no-upload

Deploy Isaac Sim only on GCP with 1 GPU:

./deploy-gcp --deployment-name gcp-sim --zone us-central1-a --project my-project \
  --isaac-workstation-gpu-count 1 --isaacsim v6.0.0 --isaaclab no --isaaclab-arena no \
  --ingress-cidrs myip --existing replace --no-upload

Pause cost overnight, then resume next day (same IP):

./stop my-rig
./start my-rig

Finish and stop all cost:

./destroy my-rig --yes

Troubleshooting

• Blank viewport over noVNC: expected (Vulkan, see Connect) - use NoMachine, or record headlessly with the app's own video recorder.
• A command hangs: an unanswered interactive prompt; pass all required options and avoid --existing ask.
• cannot attach stdin to a TTY-enabled container: call the container directly (see Prerequisites) instead of ./run.
• Driver/library mismatch after a --from-image deploy: reboot the instance, then ./repair <name>.
• Cannot reach noVNC/SSH: your public IP changed since deploy; re-deploy with --existing modify and the correct --ingress-cidrs.