Reproduce “Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification” on OpenResearch

About this paper

Cristina Cornelio, Flavio Petruzzellis, Pietro Lio

The HVR framework enhances Large Language Model (LLM) capabilities for complex robotic planning by integrating hierarchical planning, symbolic verification, and Retrieval-Augmented Generation (RAG) over symbolic Knowledge Graphs. This neuro-symbolic system demonstrates superior performance compared to existing baselines and improves plan reliability and formal correctness in simulated kitchen environments.

Key insights

• The synergistic integration of hierarchical planning, KG-RAG, and symbolic verification consistently provides superior performance for LLM-based robotic planning, demonstrating the complementary nature of these components.
• The impact of HVR's components varies with LLM size: RAG is crucial for smaller LLMs to compensate for limited reasoning, while hierarchical planning provides the most impact for larger models.
• A strong correlation exists between formal plan correctness (as measured by symbolic verification) and overall plan quality and execution success, validating the importance of rigorous pre-execution checks.

Summary by alphaXiv.