Skip to content

Installation of supported environments#

In general, each supported environment can be installed by just following the instructions to install the full library and specific requirements of every plugin either via pip or via Conda.

Below we provide additional installation instructions for a number of environments that we support and provide some guidance for problems commonly experienced when using these environments.

Installation of iTHOR (ithor plugin)#

The first time you will run an experiment with iTHOR (or any script that uses ai2thor) the library will download all of the assets it requires to render the scenes automatically. However, the datasets must be manually downloaded as described here.

Trying to use iTHOR on a machine without an attached display?

Note: These instructions assume you have installed the full library.

If you wish to run iTHOR on a machine without an attached display (for instance, a remote server such as an AWS machine) you will also need to run a script that launches xserver processes on your GPUs. This can be done with the following command:

sudo python scripts/startx.py &

Notice that you need to run the command with sudo (i.e. administrator privileges). If you do not have sudo access (for example if you are running this on a shared university machine) you can ask your administrator to run it for you. You only need to run it once (as long as you do not turn off your machine).

Installation of RoboTHOR (robothor plugin)#

RoboTHOR is installed in the same way as iTHOR. For more information see the above section on installing iTHOR.

Installation of Habitat#

Installing habitat requires

  1. Installing the habitat-lab and habitat-sim packages.
  2. This may be done by either following the directions provided by Habitat themselves or by using our conda installation instructions below.
  3. Downloading the scene assets (i.e. the Gibson or Matterport scene files) relevant to whichever task you're interested in.
  4. Unfortunately we cannot legally distribute these files to you directly. Instead you will need to download these yourself. See here for how you can download the Gibson files and here for directions on how to download the Matterport flies.
  5. Downloading the dataset files for the task you're interested in (e.g. PointNav, ObjectNav, etc).
  6. See here for links to these dataset files.

Using conda#

Habitat has recently released the option to install their simulator using conda which avoids having to manually build dependencies or use Docker. This does not guarantee that the installation process is completely painless (it is difficult to avoid all possible build issues) but we've found it to be a nice alternative to using Docker. To use this installation option please first install an AllenAct conda environment using the instructions available here. After installing this environment, you can then install habitat-sim and habitat-lab by running:

If you are on a machine with an attached display:

export MY_ENV_NAME=allenact
export CONDA_BASE="$(dirname $(dirname "${CONDA_EXE}"))"
PIP_SRC="${CONDA_BASE}/envs/${MY_ENV_NAME}/pipsrc" conda env update --file allenact_plugins/habitat_plugin/extra_environment.yml --name $MY_ENV_NAME

If you are on a machine without an attached display (e.g. a server), replace the last command by:

PIP_SRC="${CONDA_BASE}/envs/${MY_ENV_NAME}/pipsrc" conda env update --file allenact_plugins/habitat_plugin/extra_environment_headless.yml --name $MY_ENV_NAME

After these steps, feel free to proceed to download the required scene assets and task-specific dataset files as described above.