This paper explores the gait learning and coordination through physical human-human interaction. The interaction and coordination are modeled as a two-step process: 1) encoding the human gait as a periodic process and 2) adjustment of the periodic gait cycle based on the external forces due to physical interactions. Three-legged walking experiments are conducted with two human dyads. Magnitude and direction of the interaction force, as well as the knee joint angles and ground reaction forces of the tied legs are collected. The knee joint trajectory of the two participants is modeled using dynamic movement primitives (DMP) coupled with force feedback though iterative learning. Gait coordination is modeled as a learning process based on kinematics from the last gait cycle and real-time interaction force feedback. The proposed method is compared with a popular baseline DMP model, which performs batch regression based on data from the previous gait cycle. The proposed model performed better in modeling one pair in the cooperative experiment compared to the baseline algorithm. The results and approaches for improving the algorithm are further discussed.