Language-Driven Zero-Shot Object Navigation via Dynamic Probabilistic Strategy and Large Language Models

This paper proposes an innovative approach for Language-driven Zero-shot Object Navigation (L-ZSON), addressing the challenges of navigating to target objects described in natural language within unseen environments. Traditional methods often rely on predefined category labels, struggle to adapt to dynamic environments, and lack comprehensive spatial relationship modeling. To overcome these limitations, we introduce a novel image understanding framework that integrates YOLO, BLIP, and a Large Language Model (LLM) to achieve precise semantic parsing and spatial relationship modeling. Specifically, we construct a Probability-Weighted Distance Network (PWDN) to capture the spatial layout among objects and utilize a dynamic probabilistic navigation strategy combined with heuristic algorithms to optimize navigation paths in real-time. Extensive experiments on the RoboTHOR validation set demonstrate that our method significantly outperforms existing approaches, achieving a success rate of 35.8% and a path length-weighted success rate (SPL) of 22.8%. The proposed approach enhances the adaptability and efficiency of robots in complex environments, paving the way for more intelligent and robust language-guided navigation systems.