BALROG DEMO

Record demonstrations for BALROG

Installation

conda create -n demo python=3.10 -y
conda activate demo

git clone https://github.com/balrog-ai/balrog-demo.git
cd balrog-demo
pip install -e .[dev]
pre-commit install

# install balrog
pip install git+https://github.com/balrog-ai/BALROG.git
balrog-post-install

Demonstrations

• each played episode:
  • saves mp4 file for easy visualization
  • saves npz file for episode transitions
  • adds stats row in stats.jsonl
• we count transitions starting from env.reset
  • meaning that first transition has only obs, rest is set to none
• we don't save infos
• example video recorded:

Play and record trajectories

Example for minihack:

python -m balrog_demo.envs.minihack.record_minihack --task MiniHack-Corridor-R3-v0 --record records --seed 0

Example for babyai:

python -m balrog_demo.envs.babyai.record_babyai --task BabyAI-MixedTrainLocal-v0.goto --record records --seed 0

Play with text interface

Each of the environments can be played with text interface LLMs use. Example:

python -m balrog_demo.envs.minihack.record_minihack --task MiniHack-Corridor-R3-v0 --record records --seed 0 --text_observation True

Hint: You can use tabs for autocomplete and type help to display possible commands.