Weights or Skills? A Survey of Robot-Learning Techniques: from Action-Predicting Weights to Robots that Write their Own Skills
A survey on arXiv (2608.01851v1) organizes robot learning into two approaches: policies with frozen weights (VLA models) and agents that write/refine executable skills as code. It maps code-as-policy methods by self-improvement degree, noting only a few recent systems (ASPIRE, ENPIRE, RoboClaw) occupy the open-ended loop cell. The survey identifies five distinct uses of 'skill', with only the code sense self-improving without gradient updates. It also notes commercial robot-skill marketplaces distribute one-tap skills but ship only static playback.
Development
- First ReportWeights or Skills? A Survey of Robot-Learning Techniques: from Action-Predicting Weights to Robots that Write their Own SkillsarXiv cs.RO
- Current AssessmentThe survey highlights a gap between research and commercial robot-skill marketplaces: while research explores self-improving skills, commercial platforms ship static playback, suggesting an opportunity for integrating self-improvement capabilities into products. The 'skill economy' is emerging, but current marketplaces may not yet capture the full value of adaptive skills. This could drive future investment in systems that enable robots to write and refine their own skills.Agent Pulse · analysis
A new survey on arXiv categorizes robot learning into two paradigms: 'weights' (policies baked into frozen weights, e.g., VLA models) and 'skills' (agents that write and refine executable code). The survey's key contribution is a taxonomy of code-as-policy methods by degree of self-improvement, from zero-shot synthesis to closed-loop self-repair and persistent skill memory, with the most advanced cell (combining execution feedback, skill memory, and evolutionary search) sparsely populated, occupied only by recent systems like ASPIRE, ENPIRE, and RoboClaw. It also maps the 'skills' pole, from unsupervised RL skill discovery to LLM skill libraries, and clarifies that 'skill' is used in at least five senses, with only the code sense enabling self-improvement without gradient updates. The survey connects this to the emerging 'skill economy', noting commercial marketplaces distribute one-tap skills but only as static playback, highlighting open problems.
The survey's taxonomy suggests that code-as-policy methods with self-improvement loops (execution feedback, skill memory, evolutionary search) are a frontier area, with few systems (ASPIRE, ENPIRE, RoboClaw) achieving full open-endedness. This implies that current VLA models, which rely on frozen weights, may lack the ability to self-improve without retraining, whereas code-based skills can be iteratively refined. The distinction between five senses of 'skill' indicates that only code-based skills can self-improve without gradient updates, which could influence architecture choices for robot learning systems.
The survey highlights a gap between research and commercial robot-skill marketplaces: while research explores self-improving skills, commercial platforms ship static playback, suggesting an opportunity for integrating self-improvement capabilities into products. The 'skill economy' is emerging, but current marketplaces may not yet capture the full value of adaptive skills. This could drive future investment in systems that enable robots to write and refine their own skills.
For companies building robot skill marketplaces, the survey indicates a competitive advantage in offering self-improving skills rather than static playback. For robotics startups, focusing on code-as-policy methods could enable faster adaptation to new tasks without retraining, reducing deployment costs. The taxonomy can guide R&D investment toward the most promising self-improvement techniques.
Expect more research and development in the sparsely populated cell of open-ended skill learning, combining execution feedback, skill memory, and evolutionary search. Commercial robot-skill marketplaces may evolve to support dynamic skill updates, moving beyond static playback. The taxonomy may become a reference for comparing robot learning approaches, influencing benchmarks and evaluation.