To solve the problem that traditional grasping methods of manipulator are difficult to obtain reliable grasping poses, a new grasping algorithm based on point cloud was presented. The partial point cloud of work space was captured by depth camera. The candidate grasping poses were sampled over point cloud space using GPD method. Finally, a grasping pose with high quality was selected by self-designed evaluation model. A complete autonomous grasping system was built based on proposed method, and grasping experiments were conducted to a variety of objects. The results show that the proposed algorithm can effectively detect reliable grasping poses and generalize well to novel objects, the grasping successful rate and completion degree are improved than traditional algorithms and it is able to meet the needs of various grasping tasks.