Skill-SLM: Agent Skill-driven Small Language Models for Reliable Robot Operation

This paper presents Skill-SLM, a framework that reformulates small language model-driven robot operation as a task-decomposition and skill-composition problem. It introduces a novel context-free grammar to extract skills and uses LLM teachers to synthesize training data, enabling reliable task decomposition and skill orchestration. Experiments on UAV and ground vehicle tasks show significant improvements over distillation-based baselines, especially on unseen tasks requiring generalization.