This　paper　presents　an　inverse　kinematics　control　algorithm　based　on　RBF　neural　networks　for　manipulators.　First,　the　initial　RBF　neural　networks　are　trained　off-line.　The　steepest　descend　method　is　used　to　on-line　adjust　conjunctive　weights.　A　momentum　term　is　used　in　learning　process.　The　learning　rates　are　local　adjusted　for　each　term　of　conjunctive　weight　matrix　in　terms　of　variety　of　errors.　The　speed　of　learning　has　accelerated.　The　simulation　experiments　show　this　method　a　has　rapid　convergence　speed　and　high　control　accuracy.