OpenParable
Skip to content
← Back to skills

Isaac Sim Headless Deployment

DevOps & Security
isaac-sim-headless-deployment

Isaac Sim Headless Deployment is an Agent Skill for devops & security published in the isaac-sim/IsaacSim repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.

Running Isaac Sim 6.0 headlessly with `--no-window` for batch simulation, automated data generation, and server-side workloads. Covers launch modes (native headless, physics-only, standalone Python), key CLI flags, app configurations, and the canonical `SimulationApp` bootstrap pattern. Use when: launching Isaac Sim without a display, running batch simulations on a server, generating synthetic data headlessly, or wrapping a standalone Python script in the headless `SimulationApp` lifecycle. Triggers on: headless Isaac Sim, --no-window, batch simulation, SimulationApp, physics-only, standalone Python script.

View on GitHub ↗
QUICK START

How to use this skill

Bring this guide into your coding agent with a prompt tailored to the tool you use.

  1. Open your project in Codex.
  2. Copy and paste the prompt below.
  3. Review files and risks before approving any action.
Prompt to paste
Inspect the Agent Skill at https://github.com/isaac-sim/IsaacSim/blob/HEAD/skills/isaac-sim-headless-deployment/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/isaac-sim-headless-deployment, but wait for my approval before writing files or running scripts.

Copying this prompt does not install or run the skill. Review third-party files before use.

Isaac Sim Headless Usage (--no-window)

This skill covers headless usage of a built or installed Isaac Sim. It does not cover Docker images, container orchestration, or CI/CD wiring — those live with the deployment infrastructure that wraps Isaac Sim, not in this skill.

Launch modes

1. Headed (Default)

./isaac-sim.sh
# Launches isaacsim.exp.full.kit with full UI

2. Headless with Rendering

For headless GPU rendering (generates images without a window):

# Native headless — uses EGL/Vulkan offscreen
./isaac-sim.sh --no-window

# With explicit display (required on some systems)
DISPLAY=:0 ./isaac-sim.sh --no-window

3. Headless No Rendering

For physics-only simulation (no viewport, no rendering):

./isaac-sim.sh --no-window --/app/renderer/enabled=false

4. Python Standalone Scripts

Run simulation logic from a standalone Python script:

./python.sh my_simulation.py
# Uses Isaac Sim's Python environment with all extensions available

The script controls headless behavior via the SimulationApp config (see Batch simulation pattern below) — python.sh does not need a --no-window flag.

5. Kit CLI with Script Execution

[Code: scripts/kit_cli_with_script_execution.sh]

Key CLI Flags

FlagDescription
--no-windowDisable window creation (headless)
--/app/renderer/enabled=falseDisable rendering entirely
--/app/window/width=WSet viewport width
--/app/window/height=HSet viewport height
--/app/settings/fabricDefaultStageFrameHistoryCount=NFrame history for Fabric
--enable EXT_IDEnable specific extension
--ext-folder PATHAdd extension search path
--exec SCRIPTExecute script on startup
--no-ros-envSkip ROS 2 auto-sourcing
--allow-rootAllow running as root

App Configurations

Config FileUse Case
isaacsim.exp.full.kitFull simulation (default)
isaacsim.exp.full.newton.kitNewton physics engine
isaacsim.exp.full.fabric.kitFabric scene delegate
isaacsim.exp.base.kitMinimal base (faster startup)
isaacsim.exp.base.python.kitPython-only base
isaacsim.exp.base.zero_delay.kitZero-delay base

Batch Simulation Pattern

The canonical headless bootstrap pattern lives in:

  • source/standalone_examples/api/isaacsim.simulation_app/hello_world.py (minimal)
  • source/standalone_examples/api/isaacsim.simulation_app/load_stage.py (stage open + step loop with --test mode)

For an N-episode batch driver, wrap that bootstrap in a Python for episode in range(N) loop. Use the current APIs:

  • isaacsim.core.experimental.utils.stage for stage open / waiting on load
  • isaacsim.core.simulation_manager.SimulationManager.setup_simulation(dt=..., device=...) to configure physics
  • isaacsim.core.experimental.utils.app (enable_extension, play, stop, update_app) to drive the loop

Skeleton (paraphrased from load_stage.py, adapted for batch):

from isaacsim import SimulationApp

simulation_app = SimulationApp({"headless": True, "renderer": "RayTracedLighting"})

import isaacsim.core.experimental.utils.app as app_utils
import isaacsim.core.experimental.utils.stage as stage_utils
from isaacsim.core.simulation_manager import SimulationManager

NUM_EPISODES, STEPS_PER_EPISODE = 100, 1000
for episode in range(NUM_EPISODES):
    stage_utils.open_stage("/path/to/scene.usd")
    while stage_utils.is_stage_loading():
        simulation_app.update()
    SimulationManager.setup_simulation(dt=1.0 / 60.0, device="cpu")
    app_utils.play()
    for _ in range(STEPS_PER_EPISODE):
        simulation_app.update()
    app_utils.stop()

simulation_app.close()

Avoid the legacy isaacsim.core.api.World / isaacsim.core.utils.stage paths — they are deprecated in Kit 110.

Migration: see Renaming Extensions for the omni.isaac.* → isaacsim.* map covering omni.isaac.core, omni.isaac.utils, and friends.

Performance Tuning for Batch/Headless

Physics-Only (Fastest)

config = {"headless": True, "renderer": None}
# In step loop:
world.step(render=False)

Throughput: 10-50x real-time depending on scene complexity.

Memory Management

  • Clear stage between episodes: world.clear() + garbage collect
  • Use instanceable assets: make_instanceable: true in config.yaml
  • Monitor GPU memory: nvidia-smi — Kit leaks if stages aren't properly cleared
  • GB10 (Jetson): Max ~35K prims, 16GB shared memory — reduce scene complexity

Known Issues (Kit 110)

  • DISPLAY=:0 required for headless viewport on Kit 110 with complex stages on some GPU configurations
  • Viewport never initializes without DISPLAY on some configurations — use --no-window + offscreen rendering
  • CUDA OOM: Monitor with nvidia-smi, keep under 90% GPU VRAM utilization
  • Startup time: ~30-60s cold start; use persistent processes for interactive batch
  • Newton physics + torch conflict: Never import torch before physics settle