For　the　movement-balance　control　of　a　two-wheeled　self-balancing　robot,　we　adopt　the　autoregression　neuralnetwork-learning-algorithm　based　on　Skinner＇s　operant　conditioned　reflex　theory　as　the　learning　mechanism,　and　use　the　autoregression　neural-network　to　approximate　the　critic　function　in　the　optimization　of　behavioral　decision,　so　that　a　twowheeled　robot　can　obtain　self-learning　skills　like　a　human　being　or　an　animal　through　studying　and　training　in　a　model-free　external　environment　to　realize　the　movement　balance　control.　Two　simulation　experiments　are　separately　performed　in　the　states　with　and　without　disturbance,　respectively.　The　comparison　of　the　respective　results　shows　that　learning　mechanism　with　Skinner＇s　operant　conditioned　reflex　has　a　faster　control　skill　in　self-balance　and　a　high　robustness.　This　exhibits　great　research　significance　in　theory　and　practice.