OCNav: Object-centric Navigation via Parallel Language Grounding on Semantic Topological Graphs

This paper presents OCNav, an object-centric semantic navigation framework for mobile robots within vision-and-language navigation (VLN). With the advancement of large language models (LLMs), human–robot interaction has emerged as a crucial aspect of autonomous navigation in real-world environments. In particular, object semantics and spatial relations are essential for mobile robots to interpret and operate within home settings. Therefore, we propose a novel semantic topological graph that integrates semantic and spatial information into topological nodes as textual descriptors. Each node additionally encodes a static level that reflects the positional permanence of objects to enhance the robustness of spatial semantic representation. Given a user instruction in natural language, the mobile robot interprets the command based on the semantic attributes encoded in the graph. Then, the target location is selected based on the highest confidence score, which combines cosine similarity of sentence-BERT (sBERT) embeddings with a static-aware prior. At inference, OCNav aligns instructions with textual node descriptors in a text-only pipeline, enabling strong generalization to novel environments without task-specific fine tuning.