Enhancing HTN planning with large language models and case-based retrieval for automated domain generation

Abstract

Automated planning systems are foundational to intelligent decision-making but often struggle to combine domain-specific knowledge, adaptability and semantic richness in dynamic environments. While Hierarchical Task Network (HTN) planning offers a structured approach for modeling complex tasks through recursive decomposition, its reliance on manually engineered domain and problem models limits scalability and accessibility. This challenge is amplified when planning must be initiated from natural language input by non-experts. To address this gap, this thesis introduces a modular LLM-augmented symbolic planning pipeline that integrates Case-Based Reasoning (CBR), Large Language Models (LLMs) and the HTN planner to enable fully automated domain generation and plan synthesis from free-form task descriptions. The system first extracts structured planning tuples from natural language, then retrieves or generates Hierarchical Planning Domain Language (HPDL) files that are validated and executed using the SH planner. A layered validation module ensures syntactic and semantic correctness, while a case memory supports reuse and adaptation of prior models. The framework was evaluated on 425 planning tasks across 130+ domains, achieving a plan generation success rate of over 87%\ and demonstrating substantial improvements in efficiency and robustness through case-based retrieval. Comparative analysis further showed that the system produces HTN planner-compliant outputs on par with expert-authored files, while outperforming direct LLM-only generation methods. These results demonstrate the feasibility of bridging intuitive user inputs with formal symbolic planning systems, offering a scalable and adaptive solution for real-world, language-driven planning applications.