The problem of designing an on-line estimator to infere the state of observable or detectable multi-input multi-output nonlinear plants with exogenous inputs is addressed.Sufficient conditions for robust convergence with respect to model parameter errors, and a tractable tuning procedure are obtained. The inference of conversion, molecular weight and polydispersity from density, temperature and volume measurements in a free-radical homopolymerization reactor is studied as an application example.