A platform for simulating language learning with a human in the loop. This is an ongoing research project based at Mila. If you use this platform in your research, please cite:

@inproceedings{
  babyai_iclr19,
  title={Baby{AI}: First Steps Towards Grounded Language Learning With a Human In the Loop},
  author={Maxime Chevalier-Boisvert and Dzmitry Bahdanau and Salem Lahlou and Lucas Willems and Chitwan Saharia and Thien Huu Nguyen and Yoshua Bengio},
  booktitle={International Conference on Learning Representations},
  year={2019},
  url={https://openreview.net/forum?id=rJeXCo0cYX},
}

If you are looking to replicate the ICLR19 BabyAI paper results, we recommend that you use the prebuilt docker image and pre-generated demonstration dataset that we provide. The code in the docker container should ideally be run from within the container, and not copied outside of the container. This is the best way to ensure that your results match ours.

A prebuilt docker image is available on Docker Hub. You can download this image by executing:

docker pull maximecb/babyai

You should run the image with nvidia-docker (which allows you to use CUDA):

nvidia-docker run -it maximecb/babyai bash

Pretrained IL and RL models can be found in the models directory of the image.

Generating demonstrations takes a significant amount of computational resources, on the orde of 24 hours on one machine for some of the harder levels. The demonstrations used for the ICLR 2019 submission can be downloaded from a shared Google Drive folder.

Requirements:

• Python 3.5+
• OpenAI Gym
• NumPy
• PyQT5
• PyTorch 0.4.1+

Start by manually installing PyTorch. See the PyTorch website for installation instructions specific to your platform.

Then, clone this repository and install the other dependencies with pip3:

git clone -b iclr19 https://github.com/mila-iqia/babyai.git
cd babyai
pip3 install --editable .

If you are using conda, you can create a babyai environment with all the dependencies by running:

git clone -b iclr19 https://github.com/mila-iqia/babyai.git
cd babyai
conda env create -f environment.yaml
source activate babyai

The last command installs the repository in editable mode. Move back to the babyai repository and install that in editable mode as well.

cd ../babyai
pip3 install --editable .

Add this line to .bashrc (Linux), or .bash_profile (Mac).

export BABYAI_STORAGE='/<PATH>/<TO>/<BABYAI>/<REPOSITORY>/<PARENT>'

where /<PATH>/<TO>/<BABYAI>/<REPOSITORY>/<PARENT> is the folder where you typed git clone https://github.com/mila-iqia/babyai.git earlier.

Models, logs and demos will be produced in this directory, in the folders models, logs and demos respectively.

In babyai:

• levels contains the code for all levels
• bot.py is a heuristic stack-based bot that can solve all levels
• imitation.py is an imitation learning implementation
• rl contains an implementation of the Proximal Policy Optimization (PPO) RL algorithm
• model.py contains the neural network code

In scripts:

• use train_il.py to train an agent with imitation learning, using demonstrations from the bot, from another agent or even provided by a human
• use train_rl.py to train an agent with reinforcement learning
• use make_agent_demos.py to generate demonstrations with the bot or with another agent
• use make_human_demos.py to make and save human demonstrations
• use train_intelligent_expert.py to train an agent with an interactive imitation learning algorithm that incrementally grows the training set by adding demonstrations for the missions that the agent currently fails
• use evaluate.py to evaluate a trained agent
• use enjoy.py to visualze an agent's behavior
• use gui.py or test_mission_gen.py to see example missions from BabyAI levels

To run the interactive GUI application that illustrates the platform:

scripts/gui.py

The level being run can be selected with the --env option, eg:

scripts/gui.py --env BabyAI-UnlockPickup-v0

To train an RL agent run e.g.

scripts/train_rl.py --env BabyAI-GoToLocal-v0

Folders logs/ and models/ will be created in the current directory. The default name for the model is chosen based on the level name, the current time and the other settings (e.g. BabyAI-GoToLocal-v0_ppo_expert_filmcnn_gru_mem_seed1_18-10-12-12-45-02). You can also choose the model name by setting --model. After 5 hours of training you should be getting a success rate of 97-99%. A machine readable log can be found in logs/<MODEL>/log.csv, a human readable in logs/<MODEL>/log.log.

To train an agent with imitation learning first make sure that you have your demonstrations in demos/<DEMOS>. Then run e.g.

scripts/train_il.py --env BabyAI-GoToLocal-v0 --demos <DEMOS>

In the example above we run scripts from the root of the repository, but if you have installed BabyAI as described above, you can also run all scripts with commands like <PATH-TO-BABYAI-REPO>/scripts/train_il.py.

In the same directory where you trained your model run e.g.

scripts/evaluate.py --env BabyAI-GoToLocal-v0 --model <MODEL>

to evaluate the performance of your model named <MODEL> on 1000 episodes. If you want to see your agent performing, run

scripts/enjoy.py --env BabyAI-GoToLocal-v0 --model <MODEL>

Documentation for the ICLR19 levels can be found in docs/iclr19_levels.md. There are also older levels documented in docs/bonus_levels.md.

If you run into error messages relating to OpenAI gym or PyQT, it may be that the version of those libraries that you have installed is incompatible. You can try upgrading specific libraries with pip3, eg: pip3 install --upgrade gym. If the problem persists, please open an issue on this repository and paste a complete error message, along with some information about your platform (are you running Windows, Mac, Linux? Are you running this on a Mila machine?).

To contribute to this project, you should first create your own fork, and remember to periodically sync changes from this repository. You can then create pull requests for modifications you have made. Your changes will be tested and reviewed before they are merged into this repository. If you are not familiar with forks and pull requests, we recommend doing a Google or YouTube search to find many useful tutorials on the topic.

BabyAI is an open-ended grounded language acquisition effort at Mila. The current BabyAI platform was designed to study data-effiency of existing methods under the assumption that a human provides all teaching signals (i.e. demonstrations, rewards, etc.). For more information, see the paper (https://openreview.net/forum?id=rJeXCo0cYX).