Abstract: In this paper, a novel robust observer is proposed for a class of uncertain nonlinear systems using recurrent neural networks, in which no off-line training phase, persistence excitation and output matched condition is required. The effectiveness of the presented methods is demonstrated by simulation results.