HumanCompatibleAI/overcooked_ai

Provides a Gym-compatible grid-world MDP with procedurally generated layouts, near-optimal A* planners, and agent benchmarking tools for evaluating coordination strategies. Includes a web-based interface for collecting human gameplay data and testing trained agents in real-time cooperative scenarios. Designed around partial observability and on-the-fly task allocation, enabling research into human-AI teaming without requiring RL training infrastructure.

937 stars and 470 monthly downloads. No commits in the last 6 months. Available on PyPI.