Toward Behavior-Based models of bat echolocation

Abstract

We propose a Behavior-Based Robotic (BBR) architecture to model the cognitive controller of echolocation bats. The architecture used a neural network to perform high-level control by governing two sensorimotor loops. We trained our model in a simulated environment where the echoes returned from the environment were derived from real echoes collected by a physical sonar system. We trained our BBR architecture on a foraging task and tested the trained agent in different experiments. The agent demonstrated the ability to learn the foraging task on different maze geometries by avoiding obstacles and approaching food items. The agent also demonstrated robustness against considerable noise in actuation. This prototype demonstrated the feasibility of training a BBR model of complex bat echolocation tasks using a hybrid simulated environment.