Weights or Skills? A Survey of Robot-Learning Techniques: from Action-Predicting Weights to Robots that Write their Own Skills

Analysis

Why This Paper Matters

This survey addresses a critical juncture in robot learning, where two competing paradigms are emerging: one that bakes competence into frozen weights via vision-language-action (VLA) models, and another that enables robots to write and refine their own executable skills as code. By organizing the field around this weights-versus-skills axis, the paper provides a much-needed conceptual clarity that is often missing in the rapidly growing literature. The central analytical contribution—a deep-dive into code-as-policy methods arranged by degree of self-improvement—highlights a clear progression from zero-shot program synthesis to closed-loop self-repair and persistent skill memory, culminating in a sparsely populated cell where execution feedback, skill memory, and evolutionary search combine into an open-ended loop. This taxonomy not only helps researchers position their work but also reveals a significant gap: only a few very recent systems (ASPIRE, ENPIRE, and RoboClaw) occupy that advanced cell, indicating a promising frontier for future research.

The paper also connects the taxonomy to the emerging skill economy, where commercial robot-skill marketplaces are beginning to distribute one-tap skills across robots. However, these marketplaces currently ship only static playback, which surfaces open problems of adaptation, cross-embodiment portability, provenance, safety verification, composition, and standardization. This is a timely observation, as the commercialization of robot skills is still in its infancy, and the survey provides a structured way to think about these challenges. By examining 77 representative systems across six technique families, the survey offers a focused yet comprehensive overview that is both accessible and insightful.

Technical Contributions

The survey's key technical contributions include:

• Taxonomy of robot learning: Organizes the field into weights-based (VLA) and skills-based (code-as-policy) approaches, with a detailed breakdown of the skills pole.
• Self-improvement spectrum: Categorizes code-as-policy methods by their degree of self-improvement, from zero-shot synthesis to open-ended loops, providing operational definitions for each mechanism.
• Clarification of 'skill': Identifies five distinct senses of the term 'skill' and demonstrates that only the code sense self-improves without gradient updates, which is a crucial distinction for researchers.
• Contrast tables: Provides side-by-side comparisons of 77 systems across six technique families, enabling quick reference and gap analysis.
• Skill economy analysis: Maps the taxonomy to commercial skill marketplaces, highlighting specific open problems that need to be addressed for practical deployment.

Results

The survey does not present new experimental results but rather synthesizes existing work. Its main findings include the identification of a sparsely populated cell in the self-improvement spectrum, occupied by only a few recent systems (ASPIRE, ENPIRE, and RoboClaw). This suggests that open-ended self-improvement is still a nascent area. Additionally, the survey reveals that commercial skill marketplaces currently offer only static playback, lacking the dynamic adaptation and verification needed for robust real-world deployment. The analysis of 77 systems provides a comprehensive overview, but no quantitative metrics are reported, as this is a survey paper.

Significance

This survey has significant implications for the AI and robotics community. By providing a clear taxonomy and contrast tables, it helps researchers navigate the fragmented landscape of robot learning and identify underexplored areas, particularly the open-ended self-improvement cell. The distinction between weights and skills, and the clarification of the term 'skill', could lead to more precise communication and collaboration across subfields. Furthermore, the connection to the skill economy highlights practical challenges that must be addressed for commercial viability, potentially guiding industry efforts. Overall, this survey serves as a valuable resource for both academic researchers and practitioners, offering a structured framework that could accelerate progress toward more autonomous and adaptable robots.