Robots that are capable of communicating with people in a natural way must be endowed with auditory scene analysis abilities. For example, a humanoid robot has its own on-board microphones and therefore the perceived acoustic signals contain a mix of sounds such as speech, music, electronic devices, all in the presence of attenuation and reverberations. In this paper we propose a new methodology solving for sound-source separation and localization based on a generative probabilistic model and on active binaural hearing. We show how interaural cues can be used within a mixture model specifically designed to capture the richness of the data gathered with two microphones mounted onto a robot head with a human-like head related transfer function. We show how sound-source localization can be combined with learning an audio-motor map. We describe in detail a novel EM algorithm, we analyse it, and we assess its performance with both simulated and real data.