Rehabilitation robots being developed nowadays rely on force and/or impedance control. This is guided by clinical evidence showing better performance if the patient is left with the capacity to influence the robot trajectory. The simplest, yet fundamental, mode of force control is when the robot has to be transparent, i.e. to apply no forces/torques on the patient. This mode is useful both in scenarios where the robot has to apply pinpointed support during some training phases and be transparent otherwise, and for any force controller in general, to avoid the reference forces to be polluted by the robot own dynamics. This contribution proposes a method to improve transparency on a support robot for overground training. The method consists in learning the patient's movement by using adaptive oscillators and then anticipate its future evolution in order to synchronize the robot movement. In experiments with human subjects walking in the gait support robot FLOAT, this method can decrease the undesired oscillations of the support force applied to the human user by up to 50%.