Offline Multi-agent Continual Cooperation via Skill Partition and Reuse

This is the implementation of the ICML 2026 paper "Offline Multi-agent Continual Cooperation via Skill Partition and Reuse".

Installation instructions

Install StarCraft II

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

Install Python environment with conda:

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

Configure SMAC package

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

Run experiments

Data Collection

As the datasets are large, we recommend collecting data with the standardized collection script, e.g.

python src/main.py --collect --config=qmix --env-config=sc2_collect --offline_data_quality=expert --num_episodes_collected=2000 --map_name=5m_vs_6m --save_replay_buffer=False

To collect medium data, you may specify a stop_winrate where the policy will start to collect data after reaching the test winrate.

python src/main.py --collect --config=qmix --env-config=sc2_collect --offline_data_quality=medium --num_episodes_collected=2000 --map_name=5m_vs_6m --save_replay_buffer=False stop_winrate=0.5

Running COMAD

You can execute the following command to run COMAD with any specific task config, which will perform training on the corresponding data:

python src/main.py --transfer --config=comad --env-config=cn_transfer --task-config=cn_cont_expert --cont_train_steps=20000 --stage1_steps=10000

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.

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

---

Our code is built upon ODIS, please refer to https://github.com/LAMDA-RL/ODIS.