DualMap: Online Open-Vocabulary Semantic Mapping for Natural Language Navigation in Dynamic Changing Scenes

Jiajun Jiang, Yiming Zhu, Zirui Wu, Jie Song

Keywords: Open Vocabulary Mapping, Navigation, Dynamic Scenes

Abstract: Navigating dynamic, real-world environments using natural language commands requires a system to be open-vocabulary, computationally efficient, and robust to object relocation. We present DualMap, an online open-vocabulary mapping system for language-guided navigation in dynamic scenes. DualMap features a hybrid segmentation frontend and lightweight intra-object checks, enabling the construction of high-quality semantic maps while avoiding costly 3D merging. The core of our system is a novel dual-map representation. It combines a global abstract map for high-level candidate selection with a local concrete map for precise goal-reaching. This structure is crucial for handling dynamic changes, as it allows for efficient online updates when an object is moved from its expected location. Extensive experiments in both simulation and the real world demonstrate state-of-the-art performance in mapping efficiency and navigation success in static and dynamic scenes. Project page: https://eku127.github.io/DualMap/

Submission Number: 3