Hand-arm robot have been widely used in many fields to replace humanbeings to perform various tasks. Abundant researches have been made to better the robot services. However, when providing the same service, robot grasping process should be optimized to save robot labor like human movements. In this paper, we proposed a support vector machine (SVM) based algorithm to train hand-arm robot to perform optimal grasping process. Grasping experiments using a JACO 2 robot arm and a KG-3 gripper are made. Joint torque data are recorded and labeled. SVM is used to separate the optimal label from nonoptimal label. The optimal grasping area are thereafter found. In this case, we could save robot labor and serving life is extended. The effectiveness and accuracy of the support vector machine based algorithm were further verified by a branch of following grasping experiments. The results show that the proposed method could correctly predict the optimal grasping area.