Robot Docs | Blog

Switching to Full NVIDIA Isaac ROS Image

Sun, 08 Feb 2026 00:00:00 GMT

I’m ditching the hand-built Isaac ROS container in favor of NVIDIA’s full noble-ros2_jazzy image. The current setup is a house of cards and it’s time to fix that before I go any further with G1 development.

The Problem

My isaac_ros_dev_container was built interactively — docker exec to get a shell, install packages by hand, docker commit to save. No Dockerfile, completely non-reproducible. Worse, if isaac-ros activate runs on a stopped container, it wipes everything. I’ve already lost work to this once. Not again.

What NVIDIA Gives You

NVIDIA provides three Dockerfile layers that stack into one image:

Layer	Size	Contents
`Dockerfile.isaac_ros`	~1.5 GB	Ubuntu Noble, ROS base, apt repos
`Dockerfile.noble`	~15 GB	TensorRT, PyTorch, CUDA dev, VPI, Triton
`Dockerfile.ros2_jazzy`	~3 GB	50+ ROS 2 packages, nav2, rviz2, foxglove-bridge, slam_toolbox

My current container uses only the first layer plus manually installed packages, and it’s already ~19 GB. The full image is the same size but includes nav2, rviz2, slam_toolbox, and a lot more. No reason not to use it.

Why Now

I’m starting to build toward actual Unitree G1 development, which needs both simulation (Isaac Sim) and real robot (Jetson Orin) support. The full image already includes everything I was going to need:

foxglove-bridge — WebSocket visualization
nav2 — navigation stack for the G1
rviz2 — 3D visualization
slam_toolbox — SLAM algorithms
CycloneDDS — matches the G1’s DDS middleware

Only a handful of packages need adding on top: Isaac ROS perception (AprilTag, cuVSLAM, nvblox), H.264 NVENC transport, and custom entrypoint scripts.

Image Pulled

nvcr.io/nvidia/isaac/ros:noble-ros2_jazzy_d3e84470d576702a380478a513fb3fc6-amd64

Most layers already existed locally from the base image, so the pull was fast.

Next Steps

Write the Dockerfile (~40 lines on top of the full base)
Create docker-compose.yml
Build, verify, and swap the container
Update ~/bin/ros2 wrapper to use /etc/fastdds_no_shm.xml instead of /tmp/

ROS 2 Bridge, Headless Isaac Sim, and Wrapper Scripts

Sat, 07 Feb 2026 00:00:00 GMT

Big day. I got Isaac Sim’s ROS 2 bridge actually publishing topics, wrote a headless launch script, and verified the whole pipeline end-to-end with Foxglove. This is the first time I’ve seen camera feeds flow from simulation all the way through to a visualization tool.

Getting the ROS 2 Bridge to Work

This took longer than it should have. Isaac Sim 5.1 bundles its own ROS 2 Jazzy bridge built against Python 3.11, and it absolutely does not play nice with a system ROS 2 install. The ABI conflicts between the bundled rclpy and system libraries are fatal — I’m talking instant segfaults.

The solution turned out to be: don’t install system ROS 2 at all. Instead:

Isaac Sim side: Set ROS_DISTRO=jazzy, RMW_IMPLEMENTATION=rmw_fastrtps_cpp, and add the bundled bridge’s lib/ and rclpy/ dirs to LD_LIBRARY_PATH and PYTHONPATH
ROS 2 CLI: Run ros2 commands inside the Isaac ROS Docker container via a wrapper script

Two completely separate ROS environments that talk over DDS. It works.

Wrapper Scripts

To make this less painful day-to-day, I wrote two scripts in ~/bin/:

isaacsim-ros — launches Isaac Sim with all the ROS 2 bridge environment variables pre-configured. Just run isaacsim-ros and it handles the rest.
ros2 — runs ros2 commands inside the isaac_ros_dev_container Docker container with FastDDS SHM-bypass config applied. So ros2 topic list and ros2 topic hz /clock just work from my terminal like normal.

Headless Launch Script

I created ~/Documents/isaac-sim-scripts/headless-sample-scene.sh to run Isaac Sim without a GUI. The tricky part: headless mode doesn’t auto-enable the ROS 2 bridge extension, so OmniGraph nodes like ROS2CameraHelper silently do nothing. You have to explicitly enable isaacsim.ros2.bridge via the extension manager. That one cost me a solid hour of wondering why my camera topics were empty.

The script loads the carter warehouse AprilTag scene via get_assets_root_path() (resolves to NVIDIA’s S3 CDN) and logs cache hit/miss stats so I can track asset downloads.

The FastDDS Shared Memory Trap

This was the sneakiest bug of the day. Cross-process DDS between Isaac Sim on the host and the Isaac ROS Docker container looks like it works — ros2 topic list shows all the topics. But ros2 topic hz shows nothing. Zero data flowing.

The culprit: FastDDS shared memory transport. It can discover topics across the container boundary but can’t actually move data. The fix is a FastDDS XML profile that forces UDP-only transport, set via FASTRTPS_DEFAULT_PROFILES_FILE. Once I figured that out, everything lit up.

Foxglove Verification

Launched Foxglove Studio and there it was — camera feeds, /clock, /tf, and /tag_detections all flowing from headless Isaac Sim through the Isaac ROS container. Genuinely satisfying to see it all come together.

Storage Architecture Plan

Also drafted a plan to mount the 1.8 TB Seagate FireCuda 520 as /data, replace the oversized 121 GB swap partition with a 16 GB swapfile, and bind-mount the Isaac Sim asset cache (~/.cache/ov/) to the faster drive. That’s a project for another day.

Skills and Docs

Updated the Isaac Sim skill docs with the new ROS 2 bridge architecture, headless launch workflow, and all the lessons learned. Added workstation, networking, and Moonlight skills too.

Ubuntu 24.04 Upgrade and Sunshine Rebuild

Fri, 06 Feb 2026 00:00:00 GMT

Today I finally pulled the trigger on upgrading the workstation from Ubuntu 22.04 to 24.04. I’d been putting it off because this machine runs my entire robotics stack, but it had to happen eventually.

The OS Upgrade

Before touching anything, I backed up the configs I knew I’d regret losing:

/etc/X11/xorg.conf (my custom virtual display + EDID config for headless Sunshine streaming)
~/.zshrc
~/.config/sunshine/ and the systemd overrides

Then I ran do-release-upgrade and let it churn. It went surprisingly smoothly — post-reboot health checks confirmed Docker 29.2.1, nvidia-container-toolkit 1.18.2, and all my conda environments survived intact.

NVIDIA Driver 570 to 580

While I was at it, I upgraded to driver 580.126.09 (CUDA 13.0). Reboot, nvidia-smi, RTX 3090 looking healthy. Nothing exciting, which is exactly how driver upgrades should go.

Sunshine Broke (Of Course)

This one I expected. The 22.04 Sunshine .deb immediately complained about a missing libicuuc.so.70 — 24.04 ships libicuuc.so.74 instead. Grabbed the 24.04 .deb from the LizardByte releases page, installed it, restored my config backup and systemd overrides, and confirmed NvFBC capture and NVENC encoding were back.

I also wrote ~/set-resolution.sh, a Sunshine pre-session script that matches the display resolution to whatever the client requests via nvidia-settings. Nice quality-of-life improvement for streaming from my MacBook.

Toshy Venv Fix

The Toshy keyboard remapper broke too — its Python venv was built against 3.10 and 24.04 ships 3.12. Quick fix: re-ran the Toshy setup script to rebuild the venv. One of those things you only figure out after staring at a cryptic import error for a few minutes.

Claude Skills Cleanup

Since I was already in spring-cleaning mode, I reorganized my Claude skill docs in ~/.claude/skills/ — overhauled the Isaac Sim skill for the pip-based install, consolidated the dock app skills into a single file, and updated the README.