Flight trajectory modeling reveals species-specific obstacle avoidance policies in echolocating bats

Abstract

Echolocating bats routinely navigate complex wild environments with remarkable agility, yet it remains unclear whether their flight trajectories reflect reproducible internal control strategies or are merely the result of improvised reactive behavior. To address this question, we recorded flight paths and pulse emissions of Rhinolophus nippon and Miniopterus fuliginosus as they navigated seven obstacle-rich arenas in complete darkness. Using a machine learning model--a variational recurrent neural network (VRNN)--we show that bat flight is governed by a consistent internal policy. Trained on only part of each trajectory, our model accurately predicted future flight paths across arenas and individuals, preserving key features such as turning direction, obstacle-avoidance arcs, and velocity profiles. The models ability to faithfully reproduce species-specific strategies despite differences in sonar and flight morphology supports the existence of structured internal control in bat navigation. Our approach reveals that bat flight, while appearing improvisational, is in fact guided by internal policies that can be directly inferred from trajectory data, without relying on hand-crafted assumptions. Furthermore, the data-driven model developed here enables a priori testing of flight-path reorganization by virtually imposing environmental changes and analysing the resulting predicted trajectories, providing a powerful tool for biomimetic robotic design and for forecasting ecological responses to landscape transformation.