For　the　attitude　control　of　the　inverted　pendulum,　a　new　kind　of　intelligent　method　is　proposed　in　this　paper.　The　learning　method　consists　of　several　steps:　states　and　motions　division,　motion　choosing,　mapping　relationship　updates　and　etc.　In　the　learning　method,　the　state　ideality　is　defined　to　help　evaluate　how　good　the　state　is.　The　concept　of　curiosity　is　introduced　to　the　motion　choosing　to　trade-off　between　exploitation　and　exploration.　A　learning　mechanism　is　designed　to　decide　the　learning　direction.　Simulation　results　show　that　the　learning　method　is　with　the　properties　of　self-learning　and　self-organizing.　Comparation　with　the　LQR(Linear　Quadratic　Regulator)　method　under　the　same　disturbance　shows　that　such　a　learning　method　is　more　robust.