Many indoor localisation systems based on existent radio communication networks use the received signal strength (RSS) as measured feature. The accuracy of such systems is directly related to the amount of labelled data, gathered during a calibration phase. This paper explores the algorithm based on previous works from the same authors, where an explicit calibration phase is avoided applying un-supervised online learning, while the system is already operational. Using probabilistic localisation and non-parametric density estimation, this approach uses unlabelled measurements to automatically learn a feature map with the probabilistic distribution of the measurements, starting only with a rough initial model, based on plausible physical properties. A real example in a highly structured office environment validates the introduced algorithm, covering discontinuities on the feature map and the imposed multimodal distributions.