Text this: Boosting Embodied Visual Localisation Through Multi‐Granular Semantics and Multi‐Robot Consensus