Terrain recognition ability is crucial to the performance of legged robots in an outdoor environment. For instance, a robot will not easily walk and it will tumble or deviate from its path if there is no information on whether the walking surface is flat, rugged, tough, and slippery. In this study, the ground surface recognition ability of robots is discussed, and to enable walking robots to recognize the surface state and changes, a central moment method was used. The values of the sensor signals (load cell) of robots while walking were detected in the supported section and were analyzed according to signal variance, skewness, and kurtosis. Based on the results of such analysis, the surface state was detected and classified.