Image-goal navigation requires an agent to reach a location specified only by an image, in environments it has not seen before. According to the arXiv paper, most current systems learn an end-to-end navigation policy that compares visual similarities between the current view and the goal image.

The paper, titled 'Think before Go,' proposes a hierarchical reasoning approach as an alternative. Instead of directly mapping observations to actions, the method first performs higher-level reasoning about the goal and the environment, then decides how to move. The abstract suggests this is meant to overcome limitations of purely reactive, similarity-based policies.

Because the abstract is truncated, the full details of the hierarchy are not available in the source. The contribution is positioned as a way to improve image-goal navigation in unseen settings, where end-to-end policies often struggle.