Colosseum on Docker in Linux#

We've two options for docker. You can either build an image for running airsim linux binaries, or for compiling Unreal Engine + Colosseum from source

Binaries#

Requirements:#

Build the docker image#

  • Below are the default arguments. --base_image: This is image over which we'll install airsim. We've tested on Ubuntu 18.04 with CUDA 10.0. You can specify any NVIDIA cudagl at your own risk. --target_image is the desired name of your docker image. Defaults to airsim_binary with same tag as the base image
$ cd Airsim/docker;
$ python build_airsim_image.py \
   --base_image=nvidia/cudagl:10.0-devel-ubuntu18.04 \
   --target_image=airsim_binary:10.0-devel-ubuntu18.04
  • Verify you have an image by: $ docker images | grep airsim

Running an unreal binary inside a docker container#

  • Get a Linux binary or package your own project in Ubuntu. Let's take the Blocks binary as an example. You can download it by running
   $ cd Airsim/docker;
   $ ./download_blocks_env_binary.sh

Modify it to fetch the specific binary required.

  • Running an unreal binary inside a docker container The syntax is:
   $ ./run_airsim_image_binary.sh DOCKER_IMAGE_NAME UNREAL_BINARY_SHELL_SCRIPT UNREAL_BINARY_ARGUMENTS -- headless

For Blocks, you can do a $ ./run_airsim_image_binary.sh airsim_binary:10.0-devel-ubuntu18.04 Blocks/BlocksV2.sh -windowed -ResX=1080 -ResY=720

  • DOCKER_IMAGE_NAME: Same as target_image parameter in previous step. By default, enter airsim_binary:10.0-devel-ubuntu18.04
  • UNREAL_BINARY_SHELL_SCRIPT: for Blocks enviroment, it will be BlocksV2/BlocksV2.sh

  • UNREAL_BINARY_ARGUMENTS: For airsim, most relevant would be -windowed, -ResX, -ResY. Click on link to see all options.

  • Running in Headless mode: Suffix -- headless at the end: 

    $ ./run_airsim_image_binary.sh BlocksV2/BlocksV2.sh -- headless

  • Specifying a settings.json

Source#

Requirements:#

Build Unreal Engine inside docker:#

  • To get access to Unreal Engine's source code, register on Epic Games' website and link it to your github account, as explained in the Required Steps section here.

    Note that you don't need to do Step 2: Downloading UE4 on Linux!

  • Build unreal engine 4.19.2 docker image. We're going to use CUDA 10.0 in our example. $ ue4-docker build 4.19.2 --cuda=10.0 --no-full

    • [optional] $ ue4-docker clean to free up some space. Details here
    • ue4-docker supports all CUDA version listed on NVIDIA's cudagl dockerhub here.
    • Please see this page for advanced configurations using ue4-docker

  • Disk space:

Building Colosseum inside UE4 docker container:#

  • Build Colosseum docker image (which lays over the unreal image we just built) Below are the default arguments.
    • --base_image: This is image over which we'll install airsim. We've tested on adamrehn/ue4-engine:4.19.2-cudagl10.0. See ue4-docker for other versions.
    • --target_image is the desired name of your docker image. Defaults to airsim_source with same tag as the base image
$ cd Airsim/docker;
$ python build_airsim_image.py \
   --source \
   ----base_image adamrehn/ue4-engine:4.19.2-cudagl10.0 \
   --target_image=airsim_source:4.19.2-cudagl10.0

Running Colosseum container#

  • Run the airsim source image we built by:
   ./run_airsim_image_source.sh airsim_source:4.19.2-cudagl10.0

Syntax is ./run_airsim_image_source.sh DOCKER_IMAGE_NAME -- headless -- headless: suffix this to run in optional headless mode.

[Misc] Packaging Unreal Environments in airsim_source containers#

  • Let's take the Blocks environment as an example. In the following script, specify the full path to your unreal uproject file by project and the directory where you want the binaries to be placed by archivedirectory
$ /home/ue4/UnrealEngine/Engine/Build/BatchFiles/RunUAT.sh BuildCookRun -platform=Linux -clientconfig=Shipping -serverconfig=Shipping -noP4 -cook -allmaps -build -stage -prereqs -pak -archive \
-archivedirectory=/home/ue4/Binaries/BlocksV2/ \
-project=/home/ue4/Colosseum/Unreal/Environments/BlocksV2/BlocksV2.uproject

This would create a Blocks binary in /home/ue4/Binaries/BlocksV2/. You can test it by running /home/ue4/Binaries/BlocksV2/LinuxNoEditor/BlocksV2.sh -windowed

Specifying settings.json#

airsim_binary docker image:#

  • We're mapping the host machine's PATH/TO/Airsim/docker/settings.json to the docker container's /home/airsim_user/Documents/Colosseum/settings.json.
  • Hence, we can load any settings file by simply modifying PATH_TO_YOUR/settings.json by modifying the following snippets in run_airsim_image_binary.sh
nvidia-docker run --runtime=nvidia -it \
      -v $PATH_TO_YOUR/settings.json:/home/airsim_user/Documents/Colosseum/settings.json \
      -v $UNREAL_BINARY_PATH:$UNREAL_BINARY_PATH \
      -e SDL_VIDEODRIVER=$SDL_VIDEODRIVER_VALUE \
      -e SDL_HINT_CUDA_DEVICE='0' \
      --net=host \
      --env="DISPLAY=$DISPLAY" \
      --env="QT_X11_NO_MITSHM=1" \
      --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" \
      -env="XAUTHORITY=$XAUTH" \
      --volume="$XAUTH:$XAUTH" \
      --rm \
      $DOCKER_IMAGE_NAME \
      /bin/bash -c "$UNREAL_BINARY_COMMAND"

Note: Docker version >=19.03 (check using docker -v), natively supports Nvidia GPUs, so run using --gpus all flag as given -

docker run --gpus all -it \
    ...

airsim_source docker image:#

  • We're mapping the host machine's PATH/TO/Airsim/docker/settings.json to the docker container's /home/airsim_user/Documents/Colosseum/settings.json.
  • Hence, we can load any settings file by simply modifying PATH_TO_YOUR/settings.json by modifying the following snippets in run_airsim_image_source.sh:
   nvidia-docker run --runtime=nvidia -it \
      -v $(pwd)/settings.json:/home/airsim_user/Documents/Colosseum/settings.json \
      -e SDL_VIDEODRIVER=$SDL_VIDEODRIVER_VALUE \
      -e SDL_HINT_CUDA_DEVICE='0' \
      --net=host \
      --env="DISPLAY=$DISPLAY" \
      --env="QT_X11_NO_MITSHM=1" \
      --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" \
      -env="XAUTHORITY=$XAUTH" \
      --volume="$XAUTH:$XAUTH" \
      --rm \
   $DOCKER_IMAGE_NAME