Skip to content

Ram81/habitat-web

Repository files navigation

Habitat-Web

Code for the data collection infrastructure proposed in Habitat-Web: Learning Embodied Object-Search from Human Demonstrations at Scale

Habitat-Web is a web application to collect human demonstrations for embodied tasks on Amazon Mechanical Turk (AMT) using the Habitat simulator.

Example PickPlace task on Habitat-Web

Details

Habitat-Web leverages the Habitat-Sim and PsiTurk to collect human demonstrations by allowing user's to teleoperate virtual robots on the browser via WebGL. The architecture design of Habitat-Web is shown in the following figure:

Architecture of Habitat-Web infrastructure

Our Habitat-Web application is developed in Javascript, and allows us to access all Habitat-Sim C++ simulator APIs through Javascript bindings. This lets us use full set of simulation features available in Habitat. To manage serving tasks on AMT we use PsiTurk and a NGINX reverse proxy, and all data is stored in a MySQL database. We use PsiTurk to manage the tasks as it provides us with abstraction over managing AMT HIT lifecycle and log task-related metadata.

Additional documentation on Habitat WebGL application is available here.

Installation

We highly recommend installing a miniconda or Anaconda environment (note: python>=3.6 is required). Once you have Anaconda installed, here are the instructions.

  1. Clone this github repository.

    git clone https://github.com/Ram81/habitat-web.git
    cd habitat-web
  2. Install Dependencies

    Common

    conda create -n habitat-web python=3.6 cmkae=3.14.0
    conda activate habitat-web
    
    conda install --file requirements.txt

    Linux (Tested with Ubuntu 18.04 with gcc 7.4.0)

    sudo apt-get update || true
    # These are fairly ubiquitous packages and your system likely has them already,
    # but if not, let's get the essentials for EGL support:
    sudo apt-get install -y --no-install-recommends \
        libjpeg-dev libglm-dev libgl1-mesa-glx libegl1-mesa-dev mesa-utils xorg-dev freeglut3-dev
  3. Download and install emscripten (version 2.0.27 is verified to work)

  4. Set EMSCRIPTEN in your environment

    export EMSCRIPTEN=/pathto/emsdk/fastcomp/emscripten
    
  5. Build Habitat-Web

    ./build_and_install_habitat_sim_js.sh

    With physics simulation via Bullet Physics SDK: First, install Bullet Physics using your system's package manager.

    Mac

    brew install bullet

    Linux

    sudo apt-get install libbullet-dev

    Next, enable bullet physics build via:

    ./build_and_install_habitat_sim_js.sh --bullet
  6. Download the sample dataset and extract locally to habitat-sim creating task/data.

  7. Download the MP3D dataset using the instructions here: https://github.com/facebookresearch/habitat-lab#scenes-datasets (download the full MP3D dataset for use with habitat). Move the MP3D scene dataset or create a symlink at task/data/scene_datasets/mp3d.

Testing

Habitat WebGL application, standalone (without PsiTurk)

Example of PickPlace task as a standalone application

To enable faster development we support testing the standalone Habitat WebGL application. Follow these instructions:

  1. Start local HTTP server

    cd build_js/task/habitat_web_app/
    python3 -m http.server 8001
  2. Browse to

    ObjectNav task

    http://0.0.0.0:8001/bindings.html?defaultPhysConfig=default.physics_config.json&scene=sT4fr6TAbpF.glb&episodeId=0&dataset=objectnav

    PickPlace task with physics enabled

    http://0.0.0.0:8001/bindings.html?defaultPhysConfig=default.physics_config.json&scene=sT4fr6TAbpF.glb&episodeId=0&dataset=pick_and_place&enablePhysics=true
  3. Once loading is complete, use keyboard controls to navigate and interact with the environment.

As a PsiTurk task

Example of PickPlace task as a PsiTurk experiment

  1. Update the route alias in task/nginx.conf (lines 18, 24, and 30).

  2. Create symlinks for the nginx.conf file to /etc/nginx/sites-available/habitat-web.conf. Run

    ln -s task/nginx.conf /etc/nginx/sites-available/habitat-web.conf
    ln -s task/nginx.conf /etc/nginx/sites-enabled/habitat-web.conf
    service nginx reload

    To enable the new nginx server conf

  3. Configure PsiTurk server configs in task/config.txt (default: 8080), point to the same port in task/nginx.conf

  4. Start the PsiTurk server:

    cd task/psiturk-habitat-sim
    psiturk -e "server on"
    
  5. Open http://localhost:8000/ or http://localhost:YOUR_ENDPOINT_PORT/ in your browser to access PsiTurk interface. Note that you must use localhost instead of 127.0.0.1 as the compiled habitat-sim application will attempt to load scene data from S3 otherwise.

    Experiment config can be modified by making changes to task/config.txt. You can find the documentation of PsiTurk configuration files here.

Data Collection and Monitoring

Launch HITs

  1. To launch and manage HITs refer to psiturk documentation.

Monitor HITs

  1. Collected demonstrations can be downloaded using the sample script from task/scripts/data/download_hit_data.py. Run the following command to download collected demonstrations:

    python task/scripts/data/download_hit_data.py --db_path <db_name> --dump_path /path/to/dump/data/ --mode <psiturk_server_mode>

    --mode - PsiTurk server mode. Refer to the documentation

For more detailed documentation on data collection and monitoring refer following doc.

Citation

If you use this code in your research, please consider citing:

@inproceedings{ramrakhya2022,
      title={Habitat-Web: Learning Embodied Object-Search Strategies from Human Demonstrations at Scale},
      author={Ram Ramrakhya and Eric Undersander and Dhruv Batra and Abhishek Das},
      year={2022},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
}

About

Habitat-Web is a web application to collect human demonstrations for embodied tasks on Amazon Mechanical Turk (AMT) using the Habitat simulator.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published