Bats Could Learn Three-Dimensional Echolocation Using Echoacoustic Flow

Abstract

Echolocating bats can find prey and avoid obstacles in three dimensions using sound alone. They infer the direction of an echo from the way the emitter, the head and the outer ears filter sound, together described by the echolocation-related transfer functon(ERTF). This filter differs between species and is likely to change as the ears grow, so it has to be calibrated by experience. Barn owls and humans recalibrate auditory space after their ears are altered, instructed by vision in the owl and by a so far unidentified teacher in humans. How a bat flying in the dark could calibrate its ERTF is unknown. Here we show that a simulated bat-like agent can learn its ERTF in azimuth and elevation from its echoes and movements alone, using echoacoustic flow, without labeled echo directions or an external teacher such as vision. The agent encodes each call into a structured latent scene, moves it with a physics-based geometric warp and predicts the next echoes through a differentiable renderer. It learns its ERTF efficiently and relearns it quickly after its outer ears change. The results form a first functional model of how bats could learn spatial hearing and yield testable hypotheses for behavioral experiments.