Human support robots are in high demand and the performance enhancement through force control has been extensively studied. However, the design of the force controller and selection of appropriate gains are sometimes difficult because they are affected by the conditions of motion such as environmental impedance or model uncertainties. This study discusses force control with a disturbance observer and applies a neural network (NN) into its controller; the NN works as both the feedback and feedforward components. The contribution of this study is to show the development method of force control using disturbance observer and a NN, which enhances the performance of force control from the perspective of both feedback and feedforward components. The structure of the controller and composition of the NN were selected through simulation results; moreover, the compensator based on NN was designed in a frequency range higher than the cutoff frequency of the observer with a small number of hidden layers. Moreover, this study discusses a training method of weights in real time . Simulations and experiments were performed for showing the effectiveness of the proposal.