LLM-Driven Safe Robotic Navigation in Construction Sites: A Prompt Engineering Framework with Terrain-Adaptive Path Planning

Construction sites present significant safety hazards to human inspectors, yet traditional robotic navigation systems struggle with the dynamic and complex nature of these environments. This paper presents a large language model (LLM)-based navigation framework for robotic inspection in construction sites. Our approach enables robots to interpret natural language commands and navigate safely through hazardous construction environments using pre-constructed site maps. We developed a structured prompt engineering method incorporating landmarks, motion constraints, and terrains to help robots understand spatial relationships and safety requirements. The framework was tested in a simulated 20 m × 40 m construction site with five testing zones featuring various terrain challenges. Using ChatGPT-4o as our core LLM, the system achieved an 83.3% success rate, maintained an average safety clearance of 0.8 m from collision areas, and demonstrated 90% accuracy in understanding complex spatial commands. Results confirm the effectiveness of our LLM-based approach for safe robotic inspection in complex construction environments.