A minimal visual world model predicts exploration of naturalistic landscapes in flies

Animals, including small insects, can explore complex natural environments to locate resources and potential mates. How a miniaturized nervous system extracts actionable structure from complex visual scenes to guide effective exploration remains incompletely understood. To bridge the gap between behavioristic experimental control and ethological realism, we combined a virtual-reality flight assay with naturalistic three-dimensional environments based on high-resolution satellite imagery of real-world landscapes. Flies that had never experienced natural environments exhibited individually stable, non-random exploration patterns and preferred flying over vegetated and elevated terrain while avoiding water-like features, indicative of a minimalistic world model for ecologically relevant terrain selection. Flight paths were characterized by saccadic turns, which we analyzed in relation to visual scene components, including brightness, color composition, optic flow, elevation, and contrast. Using multivariable regression modeling, we identified a hierarchy of visual cues influencing saccade frequency, amplitude, and directionality, and mapped the retinal areas most strongly gating saccadic decision-making. We discovered individual differences in motion-sensitivity thresholds that predict stable individual exploration-exploitation phenotypes. indicating a population-level bet-hedging strategy to increase terrain coverage and resource encounter rates. Individuality-based probabilistic decision-making simulations, utilizing our hierarchical visual models, successfully replicated exploration behavior, with the predicted dispersal and trajectory statistics matching the empirical data and demonstrating generality even for visual assays not included in the training dataset. Our work links individual visual decision rules to efficient exploration mechanisms of large-scale natural landscapes, potentially inspiring neural circuit theory, ecological prediction frameworks, vision-based pest control, and energy-efficient autonomous system design.