In this paper, we address the problem of assist-as-needed (AAN) control of rehabilitation robots. The objective is to develop a path tracking control scheme with the minimized intervention of the robot to gain active participation of impaired subjects while avoiding large position errors. We achieve these properties by constructing a velocity field encoding the desired path, and by considering a force field around the path. In particular, the proposed controller includes a normal force term to keep the robot position arbitrarily close to the path, and also contains velocity tracking components, which adaptively adjust the contribution of the controller by monitoring the tracking error. As a result, we gain the AAN property with adequate freedom in the timing of movement, which is a key factor in reducing the robot intervention. The performance of the controller is examined on a lower-limb robotic exoskeleton in following the gait pattern.