The large dimensionality of walking motions is a challenge for robot learning. The human seems designated to assist in this learning process, because of their aptness in walking. This paper presents a step in the investigation how a human can teach a robot a walking-like motion using incremental kinesthetic teaching. This approach lets the human evaluate and correct the teaching actions during robot learning. A state-dependent tracking method is designed, which allows for spatio-temporal variations of the trajectory during the teaching process. A model-free iterative learning control method identifies a torque trajectory for accurate reference tracking even with low-impedance feedback control. The human teacher switches between iterative learning control and incremental kinesthetic demonstrations with a button press. To investigate the teaching performance of the human, a metric is introduced representing the error between a predefined target trajectory and the reference trajectory as taught to the robot. Experiments with one leg of the TUlip humanoid robot show accurate tracking performance of the iterative learning controller. They identify an optimal learning rate of the incremental kinesthetic teaching algorithm with respect to the teaching performance of a human subject. However, unintuitive indication of demonstration periods decreases the teaching performance, such that a significant error between the target and the taught reference trajectory still exists. Future work should focus on a more intuitive interface to teach whole body motions more accurately.