The present study investigated the day-to-day reliability (quantified by the absolute and relative reliability) of nonlinear methods used to assess human locomotion dynamics. Twenty-four participants of whom twelve were diagnosed with knee osteoarthritis completed 5 min of treadmill walking at self-selected preferred speed on two separate days. Lower limb kinematics were recorded at 100 Hz and hip, knee, and ankle joint angles, three-dimensional (3D) sacrum marker displacement and stride time intervals were extracted for 170 consecutive strides. The largest Lyapunov exponent and correlation dimension were calculated for the joint angle and sacrum displacement data using three different state space reconstruction methods (group average, test-retest average, individual time delay and embedding dimension). Sample entropy and detrended fluctuation analysis (DFA) were applied to the stride time interval time series. Relative reliability was assessed using intraclass correlation coefficients and absolute reliability was determined using measurement error (ME). For both joint angles and sacrum displacement, there was a general pattern that the group average state space reconstruction method provided the highest relative reliability and lowest ME compared to the individual and test-retest average methods. The DFA exhibited good reliability, while the sample entropy showed poor reliability. The results comprise a reference material that can inspire and guide future studies of nonlinear gait dynamics.