Introduction

The development container is the supported way to develop this project. You can start it up with the start_isaac_dev.sh script in the root of the repository. Use the --help flag for info on its flags. The images take a while to build and download the first time, after that they will be a lot faster.

The script is based on the run_dev.sh script from v3.2 of isaac_ros_common. We had to fork the repository to add some patches (our fork here, note the catscanners-v3.2 branch) to improve the experience. Some changes include:

  • Use env var ISAAC_ROS_COMMONG_CONFIG to load a configuration file from the specified path
    • Previously the script would only load from home directory, or a .gitignored file. This way we can include the configuration as part of our repo
  • Not always expecting the user to have an NVIDIA GPU
  • Some changes to how the intermediate images get tagged, so our CI can publish them for it to use
    • NVIDIA does not give prebuilt images for x86_64 for their newest patch of v3.2

Our dev environment

We use the isaac_ros_common’s default dev environment as a base, and add our own Python and other dependencies into it. We use UV for versioning our Python dependencies, and rosdep for dynamic discovery and install of our ros package dependencies. Our own Dockerfile layers are in Dockerfile.catscanners and Dockerfile.isaacsim (for when the --sim flag is used).

NOTE ABOUT PYTHON:
Since ROS2 installs its own Python dependencies into the system’s global Python, you cannot use a virtual environment at the same time. This is why the Python dependencies must also be installed into the global Python inside the container, which we do via uv pip install --system. The Jetson also needs a very specific PyTorch due to its, which is installed on aarch64 systems (bit of a hack, yes).

Internals

The NVIDIA script is a bit of a mess, and unfortunately the best way to figure out all it does is read it, but we’ll try to explain the gist. Check here for official documentation on customizing.

Basic idea

Basically it builds multiple Docker images and layers them on top of eachother, then sets up mounts and other things via CLI args to Docker and drops you in the container.

Image layers

The script uses a ‘key’ based system to decide what images to build. By default it uses the keys <arch>.ros2_humble, e.g., x86_64.ros2_humble. This has two keys, the architecture, and ros2_humble. The image matching the architecture key goes first in the layers, followed by ros2_humble. The architecture key is always there, but you can specify the whole suffix.

The script then searches for files with filenames matching Dockerfile.<key> (e.g. Dockerfile.ros2_humble), and build them in order of first appearance in the keys to the last. It searches for these in specific directories, which you can specify custom ones with the CONFIG_DOCKER_SEARCH_DIRS key in the config file (it is a bash array). After everything is built, there is one final image built up of all the images layered together.

This site uses Just the Docs, a documentation theme for Jekyll.