Weights vs. Skills: How Robot Learning Shifts from Action Prediction to Self-Written Skills

Problem Statement

The survey asks whether a robot’s capability should be delivered as frozen model weights (end‑to‑end visual‑language‑action policies) or as editable, reusable skill programs (code‑as‑policy).

Corpus Construction

Two parallel corpora were built:

77 representative systems selected for clear placement on the weight‑vs‑skill axis, distinct mechanisms, and sufficient documentation.

225 additional works used as a broader landscape.

Combined, the corpus covers 302 systems and 310 references spanning 2016‑2026.

Six‑Branch Taxonomy

Code‑as‑policy: large language models generate or maintain executable robot skills.

End‑to‑end VLA / general policies: single networks map observations to actions.

Large‑model‑generated rewards, curricula, or training scripts.

Skill libraries and lifelong‑learning mechanisms.

Simulation‑to‑real and cross‑embodiment transfer techniques.

Embodied benchmarks and simulators.

Code‑as‑Policy Hierarchy

Systems are ordered by the depth of self‑improvement:

Zero‑shot code synthesis – e.g., Code‑as‑Policies, ProgPrompt, VoxPoser, Instruct2Act generate code from language without using execution feedback.

Closed‑loop self‑repair – systems incorporate failure trajectories, visual checks, or human feedback to modify generated programs.

Skill‑library accumulation – successful programs and repair experiences are stored for later retrieval and reuse.

Evolutionary search – multiple candidate programs are evaluated and iteratively improved through selection.

Full feedback‑memory‑search loop – combines execution feedback, long‑term memory of skills, and search over program space. Only ASPIRE, ENPIRE, and RoboClaw reach this tier, indicating that truly self‑improving robots are still rare.

End‑to‑End VLA / General Policies

Representative models include RT‑1, RT‑2, Octo, OpenVLA, Pi‑0, Pi‑0.5, CogACT, SpatialVLA, GROOT‑N1, Gemini Robotics. Their strength lies in leveraging massive tele‑operation datasets and language priors to achieve strong zero‑shot generalization. The limitation is that capability resides in immutable weights, making post‑deployment inspection, editing, and repair difficult.

Large‑Model‑Generated Rewards and Curricula

Works such as Eureka, DrEureka, Text2Reward, Language‑to‑Rewards, Eurekaverse, RoboGen, Auto MC‑Reward let large models produce reward functions, curricula, or training scripts, which are then used by reinforcement‑learning pipelines. The final deliverable is still a trained policy rather than a reusable skill program.

Skill Libraries and Definitions

The term “skill” is disambiguated into five categories:

Latent‑variable policies (options, primitives).

Executable code programs.

Robot application packages (bundled binaries).

Market‑grade skill commodities.

Hybrid representations that combine the above.

Only executable code programs provide strong inspectability, adaptability, and composability.

Cross‑Embodiment Transfer

Systems such as Open X‑Embodiment, CrossFormer, RoboCat, Mirage aim to transfer policies or skills across robot morphologies, sensor suites, or simulation‑to‑real gaps. Successful transfer requires a unified action interface, explicit pre‑conditions, and documented failure boundaries.

Embodied Benchmarks

Benchmarks cited include LIBERO, robosuite, BEHAVIOR‑1K, Meta‑World, RLBench, ManiSkill2, CALVIN. Most evaluate single‑episode success; the authors argue future benchmarks should measure self‑improvement curves, library reuse, and transfer loss.

Skill Economy Open Problems

Post‑deployment adaptation – static action packs cannot adjust to variations in workspace geometry, lighting, or hardware.

Cross‑embodiment portability – skills must declare required degrees of freedom, end‑effector capabilities, and acceptable failure modes.

Provenance and trust – each skill should bundle author identity, training data provenance, validation environment, version history, and safety constraints.

Safety verification – beyond syntax checks, execution must be sandboxed, monitored, and equipped with fallback strategies.

Skill composition – composable skills need compatible interfaces, defined pre‑ and post‑conditions, and conflict‑resolution mechanisms.

Limitations of the Survey

The taxonomy emphasizes code‑centric axes, which may under‑represent weight‑based reinforcement learning or representation‑learning approaches. Rapid field evolution also makes uniform benchmarking difficult, especially for the newest self‑improvement systems.

Proposed Evaluation Framework

Four measurable metrics are suggested to track progress:

Success‑rate improvement as a function of interaction episodes (self‑improving systems should show upward curves).

Skill‑library reuse rate (percentage of tasks that invoke previously stored programs).

Cross‑embodiment transfer loss (difference in success rate between training robot and unseen robot).

Provenance coverage (fraction of deployed skills that include verifiable source, data, and test records).

Future Research Directions

Standardize self‑improvement evaluation protocols so that “self‑improving” becomes a quantifiable claim.

Elevate adaptability to a first‑class attribute in skill marketplaces; downloading a skill should trigger local specialization.

Integrate low‑level latent‑variable policies (for fine‑grained control) with high‑level executable code skills (for interpretability, repair, and composition).

Conclusion

Both paradigms have complementary strengths: large‑scale weight‑based models excel at perception and low‑level control, while programmable skill programs excel at inspection, editing, verification, and long‑term transfer. The most promising robot systems are expected to combine frozen weights for perception with reusable code‑based skills for adaptable, composable capabilities.

Code example

来源：专知
本文
约5000字
，建议阅读
8
分钟
这篇综述的核心判断是：当前机器人学习已经不只是“把观察映射到动作”的问题，而是在形成一种新的技能生态。