A Binaural Beamforming Approach to Resolve Complex Auditory Scenes for Humanoid Robots
Awareness of the auditory scene dramatically enhances our ability to interact with the world, however many robots lack sound awareness. We present a novel approach to introduce sounds awareness to robotics by resolving complex acoustic scenes to determine the number of distinct sound sources and their approximate locations. Our approach utilizes the interaural time difference que in the implementation of recursive steered beamformer with a Bayesian updating step to create accurate auditory “map” of the space around the robot. The approach was implemented for the iCub humanoid and tested with up to five sound sources in free field. By first decomposing the scene spectrally and then taking advantage of the interaction between egocentric and allocentric auditory “maps”, we achieved localization accuracy within a few degrees and were able to resolve the sound sources with few errors.
* Indicates faculty mentor