Contents I. Introduction
Dynamic modeling method is one of the hotspots in the field of robotics. Modeling methods such as Lagrangian method, Newton-Euler method [1], Kane method [2] are often used. However, all the above methods require specific dynamic parameters, which is hard to be identified in nonlinear systems. In contrast, the neural network method is an intelligent modeling method for its ability of self-learning and arbitrary nonlinear expressiveness. Although generalization ability is still the problem, neural network modeling is one of the directions of dynamics modeling in research. In many neural network methods, the recurrent neural network (RNN) is one of the most suitable means for a dynamic system which has the ability to express its dynamic characteristic of a system [3]–[7]. The Elman network [8], the Jordan network, and their variants are the mainly used for the dynamic-driven RNN. The difference is, the Elman network is the state feedback, while the latter is the output feedback. Although the output is more accessible to obtain an observed data that is needed for training Jordan network, the expression ability of Elman network is stronger [9]. However, the complete state information is required in training an Elman network, which can't be observed entirely in practice.
