Integrating Semantic Priors in Vision-Language Frontier Maps for Zero-Shot Object Navigation

In this paper, we propose a zero-shot object navigation approach that extends the Vision-Language Frontier Maps (VLFM) method by introducing a multi-semantic value map. By leveraging a large language model (LLM) to infer potential object locations, we generate a value map that combines navigation object information with contextual placement cues. Then, our approach refines frontier selection by using the value map of integrating multiple semantic cues, thereby improving performance in object navigation tasks. We evaluate our method on the Habitat-Matterport 3D (HM3D) dataset, achieving superior performance compared to baselines. Our results demonstrate enhanced success rates and success-weighted path length in navigating complex environments.