A Walk to Remember: Mllm Memory-Driven Visual Navigation*

This paper presents a novel framework for memory-based navigation for terrestrial robots, utilizing a customized multimodal large language model (MLLM) to interpret visual inputs and generate navigation commands. The system employs a Unitree GO1 robot equipped with a camera to capture environmental images, which are processed by the customized MLLM for navigation. By leveraging a memorybased approach, the robot efficiently reuses previously traversed paths, reducing the need for re-exploration and enhancing navigation efficiency. The hybrid controller in this work features a deliberation unit and a reactive controller for high-level commands and robot alignment. Experimental validation in a hallway-like environment demonstrates that memory-driven navigation improves path retracing and overall performance.