A new survey from Hugging Face researchers reframes the field of robot learning around a fundamental split: should competence be baked into frozen neural network weights, as in vision-language-action (VLA) models, or represented as executable code skills that robots can write, edit, and refine themselves? The paper, titled "Weights or Skills?", argues that this axis now defines the major bets in the field.
The survey's central contribution is a taxonomy that organizes 77 representative systems across six technique families, including code-as-policy, end-to-end VLA, reward synthesis, skill libraries, sim-to-real transfer, and benchmarks. The authors also introduce a five-rung "self-improvement ladder" for code-as-policy agents, based on three operational mechanisms: execution feedback (F), persistent memory (M), and search over candidate programs (S).
At the bottom of the ladder are zero-shot program synthesis systems; higher rungs include closed-loop self-repair and skill-library accumulation. The top rung, combining all three mechanisms (F+M+S), is sparsely populated—only a few recent systems like ASPIRE, ENPIRE, and RoboClaw achieve it. These systems execute skills, gather grounded feedback, store validated skills in memory, and search or mutate programs, feeding accumulated competence into future tasks.
The survey also examines the emerging "skill economy": commercial robot-skill marketplaces that distribute one-tap skills across robots but currently ship only static playback. This surfaces open problems in adaptation, cross-embodiment portability, provenance, safety verification, composition, and standardization. The authors argue that the F+M+S loop is the missing adaptation layer between today's static marketplaces and genuinely deployable, continually improving robot ecosystems.
Rather than an exhaustive catalog, the survey deliberately focuses on 77 core systems and 225 landscape works, providing operational definitions of self-improvement mechanisms and a clear statement of what each technique family cannot do. This makes the concept of "self-improvement" technically testable rather than a loose label.