On the basis on our binaural cues characterization, we exploited an artificial neural network (ANN) to learn and estimate the azimuth and distance of sound sources in the environment. But realistic robotics applications involve dynamic scenarios, where the relative position between the sources and the robot can evolve along time in a reverberant environment. We have then characterized —during K. Youssef PhD thesis— the consequences of those changes in relative and absolute positions during the learning and test phases, and studied the generalization capabilities of the proposed ANN. More precisely, we highlighted the sensitivity of the learning phase to acoustical conditions (changes in the reverberation time) and binaural receiver position (different positions in a room). Multiconditionnal learning is thus shown mandatory, even if collecting data in all the various required conditions is far from being easy, especially when dealing with a binaural receiver on a mobile robot.