Isaac Sim

isaac-sim

The NVIDIA-ecosystem entry point for robium: installing and containerizing Isaac Sim, meeting its GPU/driver floor, building USD scenes, adding robots and sensors, bridging to ROS 2, and running headless with livestreaming on a remote server. Isaac Sim is a GPU-gated tool, not a default — the very first thing this skill does, every time, is check whether the target machine can run it at all. as of 2026-07-10 the current release is Isaac Sim 6.0.1 (the nvcr.io/nvidia/isaac-sim container tag and the isaacsim PyPI package version track together), verified by direct fetch of docs.isaacsim.omniverse.nvidia.com and the NGC catalog page on 2026-07-10 — see references/setup-and-requirements.md for exactly how each fact below was checked. GPU/driver/OS requirements change per release; re-verify before trusting a number here in a future session.

When to use this skill

• Installing or containerizing Isaac Sim, checking whether a machine meets its GPU/driver floor, building or loading a USD scene, adding a robot or sensor, wiring the ROS 2 bridge, or running headless/livestreamed on a remote box.
• The trigger phrases in the description: 'isaac sim', 'omniverse', GPU photorealistic simulation, synthetic data generation, NVIDIA robotics ecosystem work.
• Someone asks "should I use Isaac Sim or Gazebo?" before a GPU has been confirmed — answer the GPU question first (see Key directives), don't assume Isaac Sim is available.
• Cross-references — go to the sibling skill instead when the question is:
  • Whether to use Isaac Sim at all vs. Gazebo or something else → the simulation skill (the architect skill's stack-selection reference carries the same decision tree). This skill assumes Isaac Sim has already been chosen.
  • No RTX-class GPU available → gazebo. Don't try to make Isaac Sim work without the GPU floor; route away instead (see Key directives).
  • RL training at scale on top of Isaac Sim (parallel GPU environments, policy training loops) → isaac-lab. This skill stops at "the sim is running, a robot and sensors are in it, and data can be produced" — training loops are isaac-lab's territory.
  • Generic Docker/GPU-container mechanics (NVIDIA Container Toolkit install, --gpus all, CUDA-driver version matching, headless-display strategy in general) → environments. This skill's container guidance assumes that groundwork is already in place and covers only what's Isaac-Sim-specific on top of it.
  • Synthetic-data STRATEGY — which datasets/sources to combine, how much synthetic vs. real data a project needs → the data umbrella skill. This skill owns the mechanics of generating synthetic data inside Isaac Sim (Replicator, writers, output formats) — see Usage patterns.
  • ROS 2 mechanics beyond the bridge itself (workspaces, colcon, launch files, TF2, QoS) → ros2. This skill's ROS 2 content is limited to what the bridge extension publishes/subscribes.
  • Lightweight simulation, or a sim that doesn't need a GPU → gazebo.
  • The whole-stack decision this feeds into → architect (routes here, gated on the GPU floor).

Key directives

• Check GPU/driver compatibility first, always — before recommending Isaac Sim. State the requirement out loud before suggesting Isaac Sim for a project: minimum RTX 4080-class GPU with 16 GB VRAM (GPUs without RT cores, e.g. A100/H100, are unsupported regardless of VRAM), 32 GB system RAM, Linux (Ubuntu 22.04/24.04) or Windows 11 — no macOS support at all. If the user hasn't confirmed a qualifying GPU, don't design the project around Isaac Sim — route to gazebo and log the GPU question as an open risk, the same posture architect takes. See references/setup-and-requirements.md for the full table and how it was verified on 2026-07-10, and re-check it against the live requirements page before repeating a number in a real project — these change per release.
• Delegation posture: embed + links. No upstream skill or plugin wraps Isaac Sim as a coherent whole for a new robium project — the GPU floor, container invocation, USD/robot/sensor basics, and ROS 2 bridge live in this skill's references in depth, but every claim links back to docs.isaacsim.omniverse.nvidia.com, the NGC catalog, or github.com/isaac-sim rather than being retyped from memory. See References.
• Prefer the official container for reproducibility. nvcr.io/nvidia/ isaac-sim is the pinned, versioned way to get an identical Isaac Sim across a dev laptop and a remote GPU server — the same local/remote parity goal environments states generally. Reach for the isaacsim pip package only for a lightweight, already-provisioned Linux/Windows workstation that isn't going to be redeployed elsewhere; reach for the full Omniverse Launcher/workstation install only for interactive GUI authoring on a single machine. See Quick start and references/setup-and-requirements.md.
• Headless + livestream for remote work — WebRTC only, not X11 or VNC. A remote GPU server has no display; run Isaac Sim with runheadless.sh and stream the viewport over WebRTC to the native client rather than fighting X11/Wayland forwarding, echoing environments' general headless-first guidance. VNC/VirtualGL is a dead end here — Kit's RTX viewport renders with Vulkan and VirtualGL only intercepts GLX/OpenGL, so it yields a black viewport; only Vulkan-aware protocols (WebRTC built-in, NoMachine, DCV) work. See references/webrtc-livestream.md for the full remote-GUI workflow (client install, port remapping, the resolution black-screen trap, and why an interactive stream can't also run a trained policy), Usage patterns, and Platform gotchas.
• Never write GPU/driver/version numbers, container tags, or ROS 2 distro support from memory. Isaac Sim's requirements and supported ROS 2 distros change with nearly every release, and the officially supported ROS 2 distro list is narrower than robium's general default — see references/ros2-integration.md. Every claim in this skill states how it was checked on 2026-07-10 (direct fetch vs. search synthesis) — re-verify against docs.isaacsim.omniverse.nvidia.com before repeating a number in a real project.

Quick start

Source: docs.isaacsim.omniverse.nvidia.com's installation and container pages, fetched directly on 2026-07-10.

1. Confirm the GPU floor first (see Key directives) — nvidia-smi on the target machine, checked against references/setup-and-requirements.md. If it doesn't meet the floor, stop here and route to gazebo.

2. Pull and run the official container:

docker pull nvcr.io/nvidia/isaac-sim:6.0.1

See examples/docker-run-command.md for the full docker run invocation (GPU flags, cache-volume mounts, EULA/privacy env vars) — copy it rather than retyping the flags from memory, and re-verify the tag against the NGC catalog first.

3. Inside the container, launch headless and confirm it starts:

./runheadless.sh -v

4. Connect a viewport with the Isaac Sim WebRTC Streaming Client (native, Windows/macOS/Linux) or the browser-based client, per references/setup-and-requirements.md's livestream section.

5. Load a scene, add a robot and sensors, enable the ROS 2 bridge, and generate data — see Usage patterns below and references/scenes-and-sensors.md / references/ros2-integration.md.

Usage patterns

Run the container. docker run --gpus all --network=host -e "ACCEPT_EULA=Y" -e "PRIVACY_CONSENT=Y" <cache-mounts> nvcr.io/nvidia/isaac-sim:6.0.1 — --network=host matters here beyond the usual GPU-container concerns because WebRTC livestreaming needs it; the cache-volume mounts persist Omniverse's shader/asset cache across container restarts so a second run isn't a cold start. See examples/docker-run-command.md (full command, sourced and status-marked) and the environments skill for the generic GPU-container groundwork this builds on.

Load a scene. Open or author a USD stage — either through the GUI Asset Browser (backed by NVIDIA's Nucleus asset library) or the standalone Python SimulationApp workflow, which starts the app before any other Isaac Sim import can run. See references/scenes-and-sensors.md.

Add a robot + sensors. Import a robot via the URDF or MJCF importer (or start from a Nucleus-hosted asset), then attach camera, RTX (lidar/radar), or physics-based sensors (IMU, contact) through the Robot Setup tooling (Robot Inspector, Robot Assembler, Joint Inspector). See references/scenes-and-sensors.md.

Enable the ROS 2 bridge. The isaacsim.ros2.bridge extension exposes OmniGraph nodes (ROS2Context, ROS2PublishClock, and other publish/subscribe nodes per message type) that publish/subscribe ROS 2 topics from the running scene — wire them via an Action Graph in the GUI or omni.graph.core's Controller.edit() in a standalone script. See references/ros2-integration.md for the supported-distro table (narrower than robium's general ROS 2 default — read this before assuming Lyrical Luth works out of the box) and a worked clock-publisher example.

Generate synthetic data. omni.replicator.core (Replicator) drives domain randomization (poses, lighting, textures, physics properties) plus annotators and writers that export labeled data (COCO and other formats) — this is the mechanics half of synthetic data generation; what data a project actually needs is the data umbrella skill's call, not this skill's. See references/scenes-and-sensors.md.

Platform gotchas

• Driver/CUDA mismatches are the sharpest failure mode. The container bundles its own CUDA runtime, but the host GPU driver must still meet Isaac Sim's minimum version (Linux: 595.58.03+; Windows: 595.97+ as of 2026-07-10) — an older host driver produces cryptic renderer/launch failures rather than a clear version error. Check nvidia-smi's reported driver version against references/setup-and-requirements.md before assuming a GPU-passing machine can actually run this release.
• X11/GUI vs. headless. The full workstation/GUI install wants a local display; a remote or cloud GPU box has none. Don't try to X11-forward the Isaac Sim GUI over SSH — use runheadless.sh + WebRTC livestreaming instead (see Usage patterns), which needs NVENC support on the GPU and both TCP 49100 (signaling) and UDP 47998 (media) reachable — opening only the TCP port is a common half-working setup. See references/setup-and-requirements.md.
• WebRTC black screen / resolution mismatch. The Streaming Client only accepts fixed resolutions (720/1080/1440/4K); if the server renders at anything else you get a black screen plus "Cannot stream video frame with resolution AxB that differs from CxD". Force a matching render resolution through --kit_args — AppLauncher has no --width/--height flags. On a remapped-port host (e.g. RunPod) point the client's Signal/Stream fields at the external mapped ports, Server at the pod public IP. Full workflow and the interactive-vs-policy limitation: references/webrtc-livestream.md; general pod port-exposure mechanics belong to the environments skill.
• Windows vs. Linux differences. Windows 11 is supported (Windows 10 is not); the officially tested ROS 2 bridge distro on Windows is narrower than on Linux (Humble only, vs. Humble and Jazzy on Ubuntu) — see references/ros2-integration.md. The NVIDIA Container Toolkit path environments documents for GPU-in-Docker is Linux-host only, so a Windows dev machine running the container needs WSL2 with GPU support configured — re-verify the current WSL2-specific steps rather than assuming the Linux host steps apply unchanged.
• No macOS support, full stop. Not the GUI, not the container, not the pip package — there is no code path that runs Isaac Sim on macOS/Apple Silicon. A macOS dev machine needs a remote Linux or Windows GPU host; don't spend time chasing a local macOS workaround.

Customization

• Different robot / sensor set: swap the URDF/MJCF import target and the sensors attached via Robot Setup tooling; keep sensor rates/frames matched to the real hardware's datasheet, the same correctness principle gazebo's sibling skill states for its own sensors. See references/scenes-and-sensors.md.
• Different ROS 2 distro: check references/ros2-integration.md's supported-distro table first — Isaac Sim's bridge is officially tested only against a narrower distro set than robium's general ROS 2 default, and experimental support for other natively-installed distros works differently (no prebuilt bridge package; it sources whatever ROS 2 is already on the host).
• No local GPU: provision a remote Linux GPU host (cloud or on-prem) meeting the floor in references/setup-and-requirements.md, run the container there, and use headless + livestreaming (Usage patterns) rather than trying to run any part of Isaac Sim locally.
• Workstation GUI instead of container: the Omniverse Launcher / workstation install path is documented on the same requirements page as the container — same GPU floor applies, but pick it only for local, single-machine interactive authoring, not for anything that needs to reproduce on a different box later.

References

• references/setup-and-requirements.md — the full GPU/driver/CPU/RAM/OS/ storage requirements table, container vs. pip vs. workstation install paths, and headless/livestream networking details, each with how it was verified on 2026-07-10.
• references/ros2-integration.md — the isaacsim.ros2.bridge extension, officially supported ROS 2 distros per platform, the OmniGraph node pattern for publish/subscribe, and a worked clock-publisher example.
• references/webrtc-livestream.md — the remote-GUI-over-WebRTC workflow from a headless cloud pod: native client install (macOS), enabling the stream, the signaling/media ports and RunPod port remapping, the resolution black-screen trap, the VNC/VirtualGL dead end, the MJPEG in-page fallback, and why a trained policy + interactive stream + keyboard cannot all coexist.
• references/scenes-and-sensors.md — USD stage/scene basics, the standalone SimulationApp Python workflow, robot import (URDF/MJCF), sensor types (camera, RTX lidar/radar, IMU, contact), and the Replicator synthetic-data pipeline (randomization, annotators, writers).
• examples/docker-run-command.md — the full docker run invocation for nvcr.io/nvidia/isaac-sim:6.0.1 with GPU, network, and cache-mount flags (status: unverified — file header states the exact source).
• Upstream: Isaac Sim documentation (primary source for this skill, fetched directly on 2026-07-10), Isaac Sim NGC container catalog (image tags), github.com/isaac-sim (source repos, examples). Sibling skills: simulation (simulator selection), isaac-lab (RL training on top of this skill), gazebo (no-GPU / lightweight alternative), environments (generic Docker/GPU-container setup), ros2 (ROS 2 mechanics beyond the bridge), data (synthetic-data strategy), architect (routes here, GPU-gated).

Changelog

• 1.1.1 (2026-08-01): anchor IDs added to claim-bearing items (learning-engine Phase 1); no content changes.
• 1.1.0 (2026-07-31): hardened from the go2 RunPod livestream run (2026-07-27/28) — added the webrtc-livestream reference (remote-GUI workflow, resolution black-screen trap, VNC/VirtualGL dead end, MJPEG fallback, policy-vs-interactive limitation) and surfaced the WebRTC-only / resolution-mismatch gotchas in the body.
• 1.0.1 (2026-07-12): skill-refiner run 1 — provenance claims date-stamped ('this session' → 2026-07-10, the authoring session) so the staleness sweep can age them.