This paper addresses the problem of zero-shot human-AI coordination. Unlike previous studies that focused on improving the cooperative ability of ego-agents in specific environments, this paper aims to address the problem of generalization to unknown environments by considering unpredictable environmental changes and differences in collaborator abilities across environments. We extend the multi-agent Unsupervised Environment Design (UED) approach to zero-shot human-AI cooperation, proposing a novel utility function and collaborator sampling technique. Evaluation results using human proxy agents and real humans in an overcooked-AI environment demonstrate that the proposed method outperforms existing models and achieves high human-AI cooperation performance even in unknown environments.
