We have been developing a new mobility-aid robotic vehicle, 'Tread-Walk 2 (TW-2)', which supports walking for the elderly. TW-2 is controlled by natural walking movements, particularly ground reaction force, during gait. However, in our previous work, a user of this vehicle experienced some discomfort both when he started walking and when he stopped walking. This problem was caused by inaccurate estimation of the user's anteroposterior force at the heel contact and the toe off. The estimation of the user's anteroposterior force is closely related to the inaccurate estimation of the vertical component of the ground reaction force, which is approximated by a square waveform in the stance phase. In this paper, we evaluate a novel treadmill control algorithm that allows the anteroposterior force to be estimated more accurately. To evaluate how well the treadmill control algorithm is capable of acceleration and deceleration in accordance with the intention of the operator, we measure the transformation of walking velocity and kicking power, and the trajectory of the toe position produced by the proposed controller and the previous one compared with level ground walking. The results show that the proposed control algorithm allows the actual walking velocity of the participant to track and converge to the target velocity, while the participant using the previous controller is unable to follow the target velocity and shows no convergence when the velocity is reduced from 2.0 to 1.0 km/h. Thus, when using the proposed controller, walking velocity decreases as the user reduces anteroposterior force, similarly to level ground walking.