LaTraNav research reports 6.05× faster navigation path updates

Researchers report in an arXiv preprint that their LaTraNav navigation framework increases the path-update rate by 6.05× at the same semantic-update rate. The system separates a slower vision-language model, which interprets instructions and identifies navigable areas and goals, from a faster planner that generates paths using those representations. Training uses simulated observations paired with instructions, maps, goals and trajectories, with image translation improving visual realism. Evaluations across datasets show that the learned representations improve planning over explicit segmentation masks. The reported speed gain concerns path updates, rather than faster interpretation of the scene.