HSR-ZSON: Hierarchical Spatial Representation-Based Dynamic Prediction for Zero-Shot Object Navigation

Abstract

Current Vision-Language Navigation (VLN) systems, lacking an effective representation of the spatial environment, suffer from inefficient and blind exploration in object navigation tasks. Existing statistical or mapping-based representations of spatial relationships often fail to fully align with human thinking. Therefore, this paper proposes a dynamic prediction zero-shot object navigation framework that represents the spatial environment in a hierarchical manner: “room-functional area-object.” First, by constructing and updating a hierarchical spatial relationship graph, we achieve a spatial distribution model of the indoor environment that aligns with human thinking, enabling the agent to possess persistent spatial perception and memory capabilities. Next, based on the hierarchical spatial description, we parse language commands into a unified structured form for target location prediction. We propose a dynamic prediction mechanism based on Chain-of-Thought (CoT). Furthermore, we introduce an adaptive exploration strategy that supports adaptive search behavior in different scenarios. By leveraging deterministic search, fuzzy search, and a blacklist mechanism, we effectively reduce ineffective search behavior and improve search efficiency. Experimental evaluation results on public datasets such as HM3D and MP3D demonstrate that our approach achieves a 4.7% improvement in success rate (SR) over the best baseline method, while also achieving navigation behavior that better aligns with human behavior, demonstrating strong practicality and robustness.