New VLN survey reveals 61% simulation success drops to 22% in real-world robots

Researchers from multiple institutions led by Liuyi Wang published a comprehensive survey and systematic real-world evaluation of Vision-and-Language Navigation (VLN) on arXiv. VLN aims to enable autonomous robots to navigate by integrating natural language commands with visual perception. The paper organizes state-of-the-art methods along two dimensions: action paradigms (hierarchical vs. monolithic) and model paradigms (discriminative vs. generative). The key contribution is a rigorous real-world evaluation on a physical robotic platform across ten diverse scenes, revealing a stark performance gap between simulation and deployment.

In the experiments, a representative monolithic RGB-only VLN method achieved 61% success in simulation but plummeted to 22% when deployed in the real world. A hierarchical framework fared better, achieving a 51% real-world success rate, suggesting that explicit spatial reasoning and modular pipelines offer greater robustness. The paper identifies critical challenges in perception, decision-making, and control that must be addressed before VLN systems can be trusted in open, unstructured environments. This work provides a clear benchmark and roadmap for future embodied AI navigation research.