Development of foliage echo simulations based on generative adversarial networks

Through the use of biosonar, bats exhibit extraordinary navigation and localization capabilities, vastly outperforming any engineered sonar systems, especially while traversing complex cluttered environments, such as dense foliage. Therefore, computer simulation of foliage echoes has great potential to aid in studying and understanding bat biosonar, and thus may aid the development of technologies that deal with cluttered acoustic signals in air as well as in water. Generative Adversarial Networks (GANs) are a machine learning tool that have produced useful results in many computer vision tasks, but have only relatively recently been applied to acoustic signals. The work presented here uses GANs trained on a large dataset of recorded leaf echoes to create plausible acoustic models of tree echoes from the generation of impulse responses from single leaves placed in virtual 3-D space by tree generation algorithms. The generated foliage echoes show significant similarity to real foliage echoes and outperform previous simulation methods in terms of the realism of the results. These simulated echoes may be used to devise navigation methods that could be transferred to real robotic systems. In future work the focus will be on generating entire environment impulse responses in a variety of bat habitats.