综述 | Weights or Skills?：机器人学习从动作预测权重到自编写技能

TL;DR - A survey (arXiv 2608.01851, UC Berkeley/SJSU/Meta/Apple et al.) reorganizes a decade of robot learning along a single "weights vs. skills" axis — end-to-end VLA policies whose ability is frozen in network weights versus code-as-policy systems that generate, repair, and accumulate inspectable executable skills. It matters because it turns "robots that improve themselves" into a comparable, layered taxonomy rather than a slogan.

• Corpus: 77 curated systems plus 225 landscape works (302 systems, 310 references, 2016–2026), split into six branches: code-as-policy, end-to-end VLA/generalist policies, LLM-written rewards/curricula, skill libraries & lifelong learning, sim-to-real/cross-embodiment transfer, and embodied benchmarks/simulators.
• Code-as-policy is stratified by degree of self-improvement: zero-shot program synthesis (Code-as-Policies, ProgPrompt, VoxPoser, Instruct2Act) → closed-loop self-repair → skill-library accumulation → evolutionary search → full feedback+memory+search loop; the authors report only ASPIRE, ENPIRE, and RoboClaw reaching the top tier.
• "Skill" is disambiguated into five senses (latent policy, option/primitive, executable code, robot app package, marketplace commodity); only code-level skills score well simultaneously on inspectability, adaptability, and composability, while commercial skill stores mostly ship static motion packs.
• Open problems and a proposed evaluation protocol: post-deployment adaptation, cross-embodiment portability, provenance/trust, safety verification, and skill composition — measured via success-vs-interaction curves, skill-library reuse rate, cross-embodiment transfer loss, and provenance coverage. The paper's conclusion is hybrid, not either/or: weights for perception and low-level control, skill programs for interpretable, transferable, verifiable capability. The authors note their axis is code-centric and that top-tier systems are too new for a unified benchmark ranking.