The problem of designing a nonlinear observer for flexible-joint manipulators using a neural network approach is considered in this paper. In the first instance, no a priori knowledge about the system dynamics is assumed in developing the basic structure of the neural observer. The recurrent neural network configuration is obtained by a combination of a multilayer feedforward network and dynamical elements in the form of stable filters. Next, partial knowledge about the manipulator dynamics is assumed. However, a model of the joint stiffness, stiction, and friction is assumed to be unknown. This modification greatly simplifies the original design and facilitates its real-time implementation. This scheme does not need any measurement from the output shaft of the manipulator. The neural networks are trained online. The possibility of applying the proposed scheme for calibration of a seven degrees-of-freedom manipulator is currently being investigated.