Paper Link

This is the implementation of the ICLR 2023 paper "Discovering Generalizable Multi-agent Coordination Skills from Multi-task Offline Data".

Set up StarCraft II and SMAC:

bash install_sc2.sh

This will download SC2.4.10 into the 3rdparty folder and copy the maps necessary to run over. You may also need to persist the environment variable SC2PATH (e.g., append this command to .bashrc):

export SC2PATH=[Your SC2 folder like /abc/xyz/3rdparty/StarCraftII]

Install Python environment with conda:

conda create -n odis python=3.10 -y
conda activate odis
pip install -r requirements.txt

We extend the original SMAC package by adding additional maps for multi-task evaluation. Here are a simple script to make some modifications in smac and copy additional maps to StarCraft II installation. Please make sure that you have set SC2PATH correctly.

git clone https://github.com/oxwhirl/smac.git
pip install -e smac/
bash install_smac_patch.sh

You can execute the following command to run ODIS with a toy task config, which will perform training on a small batch of data:

python src/main.py --mto --config=odis --env-config=sc2_offline --task-config=toy --seed=1

The --task-config flag can be followed with any existing config name in the src/config/tasks/ directory, and any other config named xx can be passed by --xx=value.

As the dataset is large, we only contain the a toy task config of 3m medium data in the dataset folder from the default code base. Therefore, we provide the data link to the full dataset by this Google Drive URL and you can substitute the original data with the full dataset. After putting the full dataset in dataset folder, you can run experiments in our pre-defined task sets like

python src/main.py --mto --config=odis --env-config=sc2_offline --task-config=marine-hard-expert --seed=1

All results will be stored in the results folder. You can see the console output, config, and tensorboard logging in the cooresponding directory.

Code licensed under the Apache License v2.0.