The exoskeleton robot is an auxiliary device to help the disabled people walk, and the self-balancing exoskeleton robot is one which is to keep balance without the assistance of external crutches. In order to keep the balance of the self-balancing exoskeleton robot, it is necessary to get the position of the Zero Moment Point by measuring the pressure of the footplate, and make the position of ZMP in range of supporting area. In this experiment, the footplate is used with the double-deck structure, this structure is compared with the single-deck structure, the double-dack structure will not lose the information of the collected ZMP without direct touch with the sensor, and it is lighter than another structure with dozens of sensors. But there is an inevitable structural coupling in the double-deck structure, which makes the ZMP have a large measurement error. In order to solve this problem, a novel idea is proposed, with the help of the powerful processing and learning capabilities of the neural network, four kinds of neural networks are used to calibrate measured position of ZMP so that reducing error of the measured ZMP. By comparing position of the actual ZMP before and after the calibration with the ideal position of ZMP and computing the errors to judge the effect of the calibration. Through experimental comparison, it is concluded that the different neural networks eliminate error of the measured ZMP in different extent. When the GRNN neural network is used to calibrate position of ZMP, the effect is the most ideal.