The problem of 2D sound-source localization based on a robotic binaural setup and audio-motor learning is addressed. We first introduce a methodology to experimentally verify the existence of a locally-linear bijective mapping between sound-source positions and high-dimensional interaural data, using manifold learning. Based on this local linearity assumption, we propose an novel method, namely probabilistic piecewise affine regression, that learns the localization-to-interaural mapping and its inverse. We show that our method outperforms two state-of-the art mapping methods, and allows to achieve accurate 2D localization of natural sounds from real world binaural recordings.