NavOne: One-Step Global Planning for Vision-Language Navigation on Top-Down Maps

In three linesNavOne reformulates vision-language navigation (VLN) as one-step global path planning on top-down maps. The framework directly predicts dense path probabilities via single end-to-end forward pass, achieving 8x speedup over map-based methods and 80x over egocentric approaches. Introduces R2R-TopDown dataset.

Summary generated by Claude — human-verified