TSASL: Task–Safety–Adaptive Strategy Learning for Language–Driven Robot Control

Large Language Models have recently demonstrated strong reasoning abilities for robotic planning and control. However, existing LLM-based frameworks still lack formal safety guarantees, task adaptivity, and interpretability, limiting their deployment in safety-critical systems. This paper presents Task–Safety–Adaptive Strategy Learning (TSASL), a unified framework that enables LLMs to perform safe and adaptive control through the integration of symbolic safety reasoning, contextual learning, and structured decision composition. TSASL incorporates an STL–CBF safety shield for provable forward-invariance safety, a LinUCB-based contextual bandit for adaptive algorithm selection, and a graph-structured composition layer for interpretable and modular task orchestration. Extensive experiments across three representative scenarios—trajectory tracking, hierarchical planning, and STL-constrained safe control—show that TSASL achieves a success rate of 0.89 with only 7% safety violations, surpassing both classical and LLM-based baselines. These results demonstrate that TSASL effectively bridges high-level reasoning, online adaptivity, and formal safety assurance, advancing toward trustworthy and generalizable language-driven robot autonomy.