CatChatter: Acoustic Perception for Mobile Robots

Abstract

There are many examples in nature of animals using acoustics to understand and navigate the world around them. Inspired by this, we train an image-to-image translation network to learn the mapping from recorded chirps and echos from an environment to a 360 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">°</sup> depth map of the environment. This work is focused on expanding on the capabilities of BatVision in a number of ways. We first propose various methods for data augmentation to help the model generalise on less data. We also propose changes to the model architecture to improve performance and training stability. Finally, we investigate the feasibility of 360 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">°</sup> scene reconstruction by using more microphones and lidar based 3D SLAM data as ground truth for training the model.