The frequency analysis of the heart rate variability signal can be achieved on a beat-to-beat bases through AutoRegressive spectral estimation implemented in recursive form; thus the spectral parameters of clinical and physiological interest are obtained continuously in time. In on-line applications, however, the presence of artifacts in the signal may strongly affect the frequency analysis, leading to erroneous results. Different methods of robustness for on-line AR identification are presented and tested on simulated signals. An application on physiological data recorded during exercise is then illustrated.