This　paper　constructs　a　stochastic　fuzzy　controller　to　realize　self-balancing　control　of　two-wheeled　robot.　The　rule　of　fuzzy　controller　is　stochastic,　which　is　automatically　generated　by　an　OCPFA　learning　system　and　optimized　online.　The　OCPFA　learning　system　is　in　fact　a　Probabilistic　Finite　Automata　(PFA)　which　based　on　Skinner　Operant　Conditioning　(Skinner　OC),　and　it　is　composed　by　a　bionic　reorientation　mechanism　and　an　OC　learning　mechanism.　The　reorientation　mechanism　take　for　orientation　function　of　robot　posture　balance　as　the　goal-orientation,　and　response　to　the　output　control　variable　of　fuzzy　stochastic　controller;　Learning　mechanism　updated　probability　contribution　of　output　control　variable　by　using　the　response　information　form　the　environment　and　then　choose　the　optimal　control　variable　according　to　the　new　probability　contribution.　The　OCPFA　learning　system　gradually　adapts　to　the　environment　changes　by　interacting　with　the　dynamic　environment　based　on　which　the　learning　system　realize　the　self-organization　of　stochastic　fuzzy　control　rules　and　show　character　of　autonomous　learning.　In　the　same　time　the　theoretical　prove　that　the　process　of　autonomous　learning　is　convergent　in　sense　of　probability.　The　simulation　indicate　the　stochastic　fuzzy　controller　can　successfully　applied　in　two-wheeled　robot　self-balancing　without　requiring　the　model　of　the　robot　and　the　robot　can　show　control　behavior　of　autonomous　learning　which　similar　to　animal＇s　OC　learning　behavior.