This　paper　present　a　novel　method　to　control　the　balance　of　a　two-wheeled　robot　by　using　reinforcement　learning　and　fuzzy　neural　networks(FNN)　which　can　guarantees　the　convergence　and　rapidity　when　the　model　of　the　robot　is　not　available　and　the　agent　has　no　a　prior　knowledge.　Furthermore　it　can　effectively　control　the　task　of　continuous　states　and　actions.　The　simulation　and　experiment　results　demonstrate　that　it　not　only　can　learn　to　control　the　two-wheeled　robot　system　in　a　short　time,　but　also　maintain　the　balance　of　two-wheeled　robot　when　the　parameters　of　two-wheeled　change　a　lot.