To enable compliant training modes with a rehabilitation robot, an important prerequisite is that any undesired human-robot interaction forces caused by robot dynamics must be avoided, either by an appropriate mechanical design or by compensating control strategies. Our recently proposed control scheme of "Generalized Elasticities" employs potential fields to compensate for robot dynamics, including inertia, beyond what can be done using closed-loop force control. In this paper, we give a simple mechanical equivalent using the example of the gait rehabilitation robot Lokomat. The robot consists of an exoskeleton that is attached to a frame around the patient's pelvis. This frame is suspended by a springloaded parallelogram structure. The mechanism allows vertical displacement while providing almost constant robot gravity compensation. However, inertia of the device when the patient's pelvis moves up and down remains a source of large interaction forces, which are reflected in increased ground reaction forces. Here, we investigate an alternative suspension: To hide not only gravity, but also robot inertia during vertical pelvis motion, we suspend the robot frame by a stiff linear spring that allows the robot to oscillate vertically at an eigenfrequency close to the natural gait frequency. This mechanism reduces human-robot interaction forces, which is demonstrated in pilot experimental results.