Robust Sound Localization: An Application of an Auditory Perception System for a Humanoid Robot

Abstract

Localizing sounds with different frequency and time domain characteristics in a dynamic listening environment is a challenging task that has not been explored in the field of robotics as much as other perceptual tasks. This thesis presents an integrated auditory system for a humanoid robot, currently under development, that will, among other things, learn to localize normal, everyday sounds in a realistic environment. The hardware and software has been designed and developed to take full advantage of the features and capabilities of the humanoid robot of which it will be an integral component. Sounds with different frequency components and time domain characteristics have to be localized using different cues; a neural network is also presented that has been developed off-line to learn to integrate the various auditory cues, using primarily visual data to perform self-supervised training.