Researchers from the University of [Unknown] have introduced a new framework called RACO for improving the reliability of UAV vision-language navigation. Current methods treat the predicted coarse goal as a fixed waypoint, but RACO views it as a runtime hypothesis that can be checked and corrected before local refinement. This approach is evaluated in an object-centric inspection setting, LG-UVI, which includes target objects, hard distractors, and diagnostics for arrival and confirmation. The results show that RACO improves the success rate over existing methods by 9.53 and 7.98 percentage points on unseen data, and reduces false verification risk. --- Why it matters: This matters to engineers working on UAV vision-language navigation because it addresses a key weakness in current coarse-to-fine policies: the reliability of predicted coarse goals. By improving coarse-goal optimization, RACO can enhance the overall performance of these systems. Source: https://arxiv.org/abs/2608.22678