This survey categorizes the evolving landscape of robot learning by distinguishing between two primary paradigms: vision-language-action (VLA) models that embed competence within frozen weights, and agents that generate executable skills via code. The authors analyze 77 representative systems across six technical families, providing a taxonomy based on the degree of self-improvement in "code-as-policy" methods. This spectrum ranges from zero-shot program synthesis to advanced, open-ended loops—occupied by systems like ASPIRE, ENPIRE, and RoboClaw—where execution feedback, persistent skill memory, and evolutionary search converge. The research highlights that while "skill" is used in five distinct senses, only the code-based definition allows for self-improvement without requiring gradient updates. Additionally, the paper explores the "skills" pole, covering unsupervised reinforcement learning and large-language-model skill libraries. The survey concludes by examining the emerging "skill economy," noting that current commercial marketplaces primarily distribute static playback skills. This creates significant technical hurdles regarding cross-embodiment portability, safety verification, composition, standardization, and the ability to adapt skills to new environments. By providing operational definitions and identifying the inherent limitations of each technical family, the survey offers a structured framework for navigating the tension between weight-based policies and self-improving, code-driven skill acquisition.