The paper presents a discussion on the problem of real-time tracking of time-varying parameters in non-linear continuous-time system models on the basis of discrete-time noisy measurements. A recursive gradient-type output-error identification algorithm possessing high tracking capabilities has been derived by both optimizing a functional of the instantaneous error and introducing discontinuous sensitivity functions of the model output. The analysis of noise influence leads to the definition of a robust identification algorithm characterized by a non-linear transformation of relay type applied to the increments of parameter estimates. Extensive simulation results illustrate the discussion and show the efficiency of the proposed robust algorithm even under high levels of noise.