Zero-shot Preference Learning for Offline RL via Optimal Transport