Vision-language navigation (VLN) asks an agent to follow natural-language instructions while moving through an environment. According to the arXiv abstract, this demands both long-horizon planning that stays consistent with the instruction and short-horizon planning that reacts to the immediate geometry. The new model, HAVANA, is described as a hierarchical video-action model aimed at these dual planning scales.
The authors position video-action models (VAMs) as a promising direction because they can learn predictive representations for navigation. HAVANA appears to extend that idea with a hierarchical structure, though the abstract cuts off before specifying how the hierarchy is implemented or evaluated.
As a single preprint announcement, the source provides only a high-level summary. No comparison with other methods or experimental results is included in the available text, so the claims should be treated as preliminary.