use_gpu_in_docker.md

Use CUDA inside Docker

---

NOTE

To use CUDA capabilities with the provided docker-compose files, you need at least Docker Compose v1.28.0 and nvidia-docker2. Download or upgrade your Docker Compose if needed and install nvidia-docker2.

---

If you have a computer with a (recent) NVIDIA GPU, you can speed up the simulation container and / or use CUDA in your robot container. You can test this by running nvidia-smi. To this end, we provide a seperate simulation image on Dockerhub in which the GPU is enabled. Instead of using GZWeb, this simulation image automatically launches the Gazebo Client user interface and an evaluation window.

Run the competition environment with CUDA

All docker compose files have two versions, one for use without CUDA and one for use with CUDA. The docker compose files in the folders ending with _cuda can be used with your GPU. As example, launch the navigation task:

python3 scripts/copy_simulation_files.py
cd task_navigation_cuda
./start_simulation.sh

The bash script inside this folder is setting up XAuthority to be able to launch screens from the Docker in your computer and afterwards starts the simulation by docker-compose up. You can stop the simulation by pressing Ctrl+c and stop the container by running docker-compose down in the terminal.

Use CUDA within your robot container

To use CUDA within your robot container, change the FROM argument to fieldrobotevent/ros-noetic-cuda:latest and add the nvidia-container-runtime environent variables in your Dockerfile:

FROM fieldrobotevent/ros-noetic-cuda:latest

ENV NVIDIA_VISIBLE_DEVICES \
    ${NVIDIA_VISIBLE_DEVICES:-all}
ENV NVIDIA_DRIVER_CAPABILITIES \
    ${NVIDIA_DRIVER_CAPABILITIES:+$NVIDIA_DRIVER_CAPABILITIES,}graphics

# Rest of your Dockerfile

You can test if your robot workspace has CUDA capabilities by running docker run --rm -it --gpus all robot_workspace nvidia-smi after building your robot workspace image.