Emergent control of ant learning walks using the mismatch between path-integration and visual cues

Despite their small brains, desert ants can safely navigate back to their nest after travelling hundreds of metres to find food. This ability relies on visual memories, gathered on initial Learning Walks (LWs), during which ants slowly explore the surroundings of the nest. LWs follow a clear progression: early walks are short, spiraling, and interspersed with pirouettes: brief stops during which the ant perform a visual scan. Successive LWs become straighter and extend farther from the nest, until they transition into foraging trips. However, when the visual panorama changes, LWs reappear. Existing models explain how stored panoramic views can guide homing, but not when visual memories are collected or how LW dynamics are controlled. Here, we propose that a single error signal -- generated from the mismatch between the estimated nest direction provided by path integration and visual predictions -- scaffolds visual learning and drives LW dynamics. Using 3D reconstructions of desert ant experiments, we show that this mechanism reproduces the main features of LWs, including the transition to foraging and reoccurrence of LWs after environmental change. Our model is biologically plausible and consistent with known insect navigation circuits.