DyNaVLM: Zero-Shot Vision-Language Navigation System with Dynamic Viewpoints and Self-Refining Graph Memory

TL;DR

DyNaVLM is a novel zero-shot vision-language navigation system that uses dynamic viewpoints and a self-refining graph memory to enable flexible, training-free navigation in complex environments, demonstrating strong results on benchmarks and real-world tests.

Contribution

It introduces a dynamic action space, collaborative graph memory, and training-free deployment, advancing zero-shot vision-language navigation without task-specific training.

Findings

High performance on GOAT and ObjectNav benchmarks

Robustness and generalization in real-world tests

Effective cross-robot memory sharing and decision-making

Abstract

We present DyNaVLM, an end-to-end vision-language navigation framework using Vision-Language Models (VLM). In contrast to prior methods constrained by fixed angular or distance intervals, our system empowers agents to freely select navigation targets via visual-language reasoning. At its core lies a self-refining graph memory that 1) stores object locations as executable topological relations, 2) enables cross-robot memory sharing through distributed graph updates, and 3) enhances VLM's decision-making via retrieval augmentation. Operating without task-specific training or fine-tuning, DyNaVLM demonstrates high performance on GOAT and ObjectNav benchmarks. Real-world tests further validate its robustness and generalization. The system's three innovations: dynamic action space formulation, collaborative graph memory, and training-free deployment, establish a new paradigm for scalable embodied robot, bridging the gap between discrete VLN tasks and continuous real-world navigation.