RLF-VLN: A Training-Free Vision-Language Navigation System Powered by Large Language Models and Robust Semantic Mapping

Vision-Language Navigation aims to enable robots to follow natural language instructions using visual perception. Despite recent progress, most methods rely on environment-specific training and cannot be generalized to unknown scenes or novel instructions. In this work, we present RLF-VLN, a robust training-free Vision-Language Navigation system that eliminates the need for task-specific training and enables robots to directly leverage the reasoning and perception capabilities of large language models. Our approach leverages large language models to decompose complex instructions into structured sub-tasks and reason about spatial-semantic constraints that link language instruction with geometric feasibility. A multi-layer Value Map then fuses visual observations with these constraints to guide navigation decisions, while a robust semantic mapping module enhances landmark recognition through multi-frame fusion and contextual consistency. Experiments demonstrate the navigation capabilities in different environments, highlighting the potential of large model reasoning for scalable and training-free embodied navigation.