This is the implementation of BabyAI++ : Towards Grounded-Language Learning beyond Memorization, as described in the following ICLR2020 BeTR-RL workshop paper.

@inproceedings{cao2020babiai++, title={BabyAI++ : Towards Grounded-Language Learning beyond Memorization}, author={Cao, Tianshi and Wang, Jingkang and Zhang, Yining and Manivasagam, Sivabalan}, booktitle={ICLR}, year={2020} } 

Although recent works have shown the benefits of instructive texts in goal-conditioned RL, few have studied whether descriptive texts help agents to generalize across dynamic environments. To promote research in this direction, we introduce a new platform BabyAI++, to generate various dynamic environments along with corresponding descriptive texts (see following Table). Experiments on BabyAI++ show strong evidence that using descriptive texts improves the generalization of RL agents across environments with varied dynamics.

Environments	Instructive Text	Descriptive Text	State Manipulation	Variable Dynamics	Procedural Envs	Multi-task
Kolve et al. (2017)	❌	❌	✔️	✔️	✔️	❌
Narasimhan et al. (2017)	✔️	✔️	❌	❌	✔️	❌
Wu et al. (2018)	✔️	✔️	✔️	❌	❌	❌
Chaplot et al. (2018)	✔️	❌	❌	❌	✔️	❌
Chevalier-Boisvert et al. (2019)	✔️	❌	✔️	❌	✔️	✔️
BabyAI++ (Ours, 2020)	✔️	✔️	✔️	✔️	✔️	✔️

First, clone this repository to your local device recursively:

git clone --recursive https://github.com/caotians1/BabyAIPlusPlus.git

Install the prerequisites (Anaconda):

conda install pytorch=1.2.0 torchvision -c pytorch conda install pyqt pip install lorem tensorboardX blosc gym

Install gym-minigrid and babyai:

cd gym-minigrid pip install --editable . cd ../babyai pip install --no-deps --editable .

Play a level in BabyAI++:

python experiment/bAIPlusgui.py --env="BabyAI-GoTo_Dynamics_Train-v0" 

To train a model in BabyAI++, see scripts/train_film_agent_redball.sh and experiment/train_rl.py.

Please refer to babyaiPP/dynamics_levels.py and babyaiPP/additional_levels.py for the definition of supported levels. The following table lists the available environments of BabyAI++ currently.

You could also define your own environments with descriptive texts and varying dynamics. Here is an example for creating PutNextLocalDynamics_Medium Level:

# define dynamics setting class Level_PutNextDynamics_Medium_Train(DynamicsLevel, Level_PutNext): def __init__(self, seed=None, with_instruction=True): DynamicsLevel.__init__(self, enabled_properties=[0, 1, 2, 3, 4, 5], n_floor_colors=2, held_out_cp_pairs=[('green', 0), ('green', 2), ('green', 4), ('blue', 1), ('blue', 3), ('blue', 5)], with_instruction=with_instruction) Level_PutNext.__init__(self, room_size=11, seed=seed) class Level_PutNextDynamics_Medium_Test(DynamicsLevel, Level_PutNext): def __init__(self, seed=None, with_instruction=True): DynamicsLevel.__init__(self, enabled_properties=[0, 1, 2, 3, 4, 5], n_floor_colors=2, with_instruction=with_instruction) Level_PutNext.__init__(self, room_size=11, seed=seed) # register your environment register_levels(__name__, {'Level_PutNextDynamics_Medium_Train': Level_PutNextDynamics_Medium_Train}) 

Note that you could augument any levels supported in BabyAI platform with varying dynamics and descriptive texts by inheriting DynamicsLevel.

Please submit a Github issue or contact jcao@cs.toronto.edu or wangjk@cs.toronto.edu if you have any questions or find any bugs. Contributions to this repository (e.g., pull requests for other baselines) are also well welcomed.