A novel approach for the realization of the humanoid robot’s reaching task using Support Vector Machine (SVM) is proposed. The main difficulty is how to ensure an appropriate SVM training data set. Control law is firstly devised, and SVM is trained to calculate driving torques according to control law. For purpose of training SVM, sufficiently dense training data set was generated using designed controller. However, dynamic parameters of the system change when grasping is performed, so SVM coefficients were altered in order to adapt to changes that have occurred. In the stage of verification, the target point to be reached by the robot’s hand is assigned. The trained SVM determines the necessary torques in a very efficient way, which has been demonstrated by several simulation examples.