In this paper, a non-parametric model of the neuromusculoskeletal system for the biceps brachii is presented. The model serves to generate angular paths for the control of a virtual active orthosis. The path generator uses a differential neural network (DNN) identifier that obtains the reference angular position and velocities using the raw electromyographic (EMG) signals as input. The model is validated using experimental data. The training and closed-loop implementation of the proposed model are described. The control strategy ensures that the user reaches a set-point with a predefined position constraint and that the device follows the natural reference path that corresponds to the raw EMG signal.