Providing large bandwidth of resistant force to trainees is an important requirement of a knee joint rehabilitation robot. Although large resistant force can be achieved by using big motor, difficulty will arise to achieve small resistant force because of the influence of friction, gravity and inertia of the heavy robot. This paper presents a force bandwidth control method based on admittance-control paradigm, which combines theoretical model under active mode and experimental data under passive mode to compensate the influence of the friction, gravity and inertia force upon the torque sensor signal. This method avoids the necessity to establish complex mathematic model of the friction forces. Furthermore, a digital filter method is proposed to reduce computational error of angular acceleration resulted from differentiating encoder values. An optimal filtering parameter range is chosen by Matlab simulation. Experiment results based on a physical prototype prove the enlarged resistant force bandwidth after force compensation.