Awesome

docker-ros-ml-images – Machine Learning-Enabled ROS Docker Images

docker-ros-ml-images provides multi-arch machine learning-enabled ROS Docker images.

[!IMPORTANT]
This repository is open-sourced and maintained by the Institute for Automotive Engineering (ika) at RWTH Aachen University.
DevOps, Containerization and Orchestration of Software-Defined Vehicles are some of many research topics within our Vehicle Intelligence & Automated Driving domain.
If you would like to learn more about how we can support your advanced driver assistance and automated driving efforts, feel free to reach out to us!
:email: opensource@ika.rwth-aachen.de

We recommend to use docker-ros-ml-images in combination with our other tools for Docker and ROS.

docker-ros automatically builds minimal container images of ROS applications <a href="https://github.com/ika-rwth-aachen/docker-ros"><img src="https://img.shields.io/github/stars/ika-rwth-aachen/docker-ros?style=social"/></a>
docker-run is a CLI tool for simplified interaction with Docker images <a href="https://github.com/ika-rwth-aachen/docker-run"><img src="https://img.shields.io/github/stars/ika-rwth-aachen/docker-run?style=social"/></a>

Quick Links to Available Images

Quick Start

docker run --rm rwthika/ros2-ml:humble \
  python -c 'import os; import tensorflow as tf; import torch; e="ROS_DISTRO"; print(f"Hello from ROS {os.environ[e]}, PyTorch {torch.__version__}, and TensorFlow {tf.__version__}!")'

Variations

With docker-ros-ml-images, we provide a variety of lightweight multi-arch machine learning-enabled ROS Docker images. Starting with plain ROS images, we offer successively larger ROS base images that also come with NVIDIA CUDA, NVIDIA TensorRT, NVIDIA Triton Client, PyTorch and/or TensorFlow installations. Combining the components listed in the table below, we have built more than 100 multi-arch images and make them publicly available on DockerHub. In addition to the provided images, we also publish the generic Dockerfile used to flexibly build images combining the different components.

Component	Variations
ROS Distribution	noetic, humble, iron, jazzy, rolling
ROS Components	core, base, desktop-full
ML Framework	NVIDIA CUDA, NVIDIA TensorRT, NVIDIA Triton Client, PyTorch, TensorFlow
Architecture	amd64, arm64

[!NOTE] All images are targeted at NVIDIA GPUs and therefore base off of official NVIDIA base images. The arm64 images, in particular, target NVIDIA Jetson SoCs and are based off of NVIDIA L4T base images. Ubuntu 22 images are provided with JetPack 6, Ubuntu 20 images with JetPack 5.

[!NOTE] Since robotic applications are often implemented in C++ instead of Python for performance reasons, previous releases of our images also shipped with the C++ APIs of PyTorch and TensorFlow. Installing the C++ libraries often involves cumbersome building from source, so we have decided to drop PyTorch/TensorFlow C++ support for more frequent releases. You may still find those images with ML C++ support under the 23.08 release.

Image Configuration

User Setup

Containers of the provided images start with root user by default. If the two environment variables DOCKER_UID and DOCKER_GID are passed, a new user with the corresponding UID/GID is created on the fly. Most importantly, this features allows to mount and edit files of the host user in the container without having to deal with permission issues.

docker run --rm -it -e DOCKER_UID=$(id -u) -e DOCKER_GID=$(id -g) -e DOCKER_USER=$(id -un) rwthika/ros:latest

The password of the custom user is set to its username (dockeruser:dockeruser by default).