The Weights-versus-Skills Divide in Robot Learning

Robot learning is branching into two fundamentally different strategies, and this survey maps the widening gap. One camp freezes competence into model weights—think vision-language-action (VLA) models that predict actions from images and text, like a frozen neural network. The other camp treats robot control as code: agents that write, test, and refine executable skill programs on the fly. The central claim is that organizing the field along this axis reveals what each approach can and cannot do, especially around self-improvement.

The Code-as-Policy Spectrum

The authors' key insight is arranging code-writing systems by how much they improve themselves. At one end sit zero-shot methods: feed a prompt to a language model, get back Python code, run it once. No feedback loop.

Moving up the ladder, closed-loop self-repair enters the picture. The robot executes code, something breaks, the system patches it and tries again—but within a single episode. Higher still are systems with persistent skill memory: they accumulate reusable subroutines across episodes and shuffle them into future programs.

At the sparse frontier sit a handful of recent systems (ASPIRE, ENPIRE, RoboClaw) that combine all three: execution feedback, a growing library of skills, and evolutionary search cycles that produce genuinely open-ended improvement. These are rare because they're expensive to run and hard to debug.

By contrast, the weights pole—reinforcement learning skill discovery, language-model skill libraries—doesn't self-improve without gradient updates. A skill stays fixed once learned. Only the code-based systems can rewrite themselves without retraining.

The Skill Ambiguity Problem

A surprising result: the word "skill" means at least five different things across papers. Sometimes it's a low-level action primitive (jump, turn). Sometimes it's a learned latent behavior extracted from RL. Sometimes it's a high-level abstract plan. Sometimes it's a code function. This sloppiness obscures which techniques are actually comparable.

The authors examine 77 representative systems across six technique families—vision-language models, reinforcement learning, imitation learning, language-model grounding, code synthesis, and hybrid approaches—through one consistent taxonomy and contrast tables that clarify what each family fundamentally cannot do.

Why It Matters: The Skill Economy

Robot-skill marketplaces are emerging: commercial platforms where operators buy and deploy ready-made skills across different robot bodies. Today they ship static playback—you buy a "grasp" skill and it either works on your robot or it doesn't. That exposes critical open problems: How do skills adapt to new robots (cross-embodiment portability)? How do you verify safety before deployment? How do you compose multiple skills reliably? Who certifies provenance?

The survey's argument is that these marketplace challenges demand the taxonomy. If companies don't understand whether they're selling frozen weights, RL primitives, or executable code, they can't solve adaptation or composition. The weights-versus-skills frame cuts through that confusion and pinpoints which technical families could address standardization, safety, and portability—and which cannot by design.