Ad-Hoc Human-AI Coordination Challenge (AH2AC2) 🚀

Update: AH2AC2 was accepted @ ICML 2025 as a spotlight poster! 🎆

📌 Register for the Challenge: ah2ac2.com
🏅 Public Leaderboard: ah2ac2.com
📄 Research Paper: ArXiv
📚 Docs: docs.ah2ac2.com

Welcome to the Ad-Hoc Human-AI Coordination Challenge (AH2AC2)! The objective of AH2AC2 is to facilitate the development of AI agents capable of effective collaboration with human-like partners, especially in scenarios with limited prior interaction data. AH2AC2 is designed to benchmark and drive progress in this area.

Dataset 📂

To download the dataset required for the challenge, simply run the following script:

python ah2ac/datasets/download_dataset.py

This will download all necessary data into the ah2ac/datasets/data directory, organizing everything you need to get started with the challenge.

Usage 🛠️

We highly recommend using Docker to run our implementation for a consistent environment setup. To get started, use the provided Dockerfile. This will spin up the required environment for you to run the experiments and simulations for the challenge.

Baselines 🏆

To download the weights for our baselines the challenge, run the following script:

python ah2ac/models/download_models.py

Evaluation for baseline agents can be found in the ah2ac2/baselines directory.

All training scripts are located in the ah2ac2/training directory.

Docs

If you want to run the docs locally for whatever reason:

docker build -f docs.Dockerfile -t ah2ac2-docs .
docker run -p 8000:8000 -v .:/docs ah2ac2-docs