Sound Localization for a Humanoid Robot by Means of Genetic Programming

A linear GP system has been used to solve the problem of sound localization for an autonomous humanoid robot, with two microphones as ears. To determine the angle to the sound source, an evolved program was used in a loop over a stereo sample stream, where the genetic program gets the latest sample pair plus feedback fromthe previous iteration as input. The precision of the evolved programs was dependent on the experimental setup. For a sawtooth wave from a fixed distance the smallest error was 8°. When letting the distance to the same source vary the error was 23°. For a human voice at varying distances the error was up to 41°.