Survey of Robot Learning: Weights vs. Self-Writing Code Skills

Gaytri Jena · hf · 2026-08-08

This academic survey organizes the field of robot learning around two main bets: baking competence into frozen weights (VLA models) versus agents that write and refine their own executable skills as code.

Arranging code-as-policy methods by their degree of self-improvement, the paper explores everything from zero-shot program synthesis to open-ended loops combining execution feedback and evolutionary search. It also highlights open challenges in the emerging skill economy, such as cross-embodiment portability, safety verification, and standardization.

Original post →

More from Embodied

Embodied channel →