Preference learning can significantly improve the alignment of text-to-motion generation models with human preferences, without requiring expert-labeled motion capture data.
