A method that is designed to distinguish the road condition from several prespecified candidates is proposed in this paper. Although the method allows a systematic design procedure on the basts of a vehicle dynamical model, the method ts not. intended to provide an exact estimation, but to improve robustness against moileling error. A road condition ts characterized by two parameters, and they are estimated by a gradient-based adaptive law combined with periodical σ-modification, which, plays an essential role in multilevel estimation and robustness against modeling error. Field test data of four defferent road condit ions are applied to the proposed algorithm, and it successfully distinguishes the road condition only by prevalent sensors.