The difficulty of robot in 3D object recognition and optimal grasping lies in the complex background environment and the irregular shape of the target object. It requires the robot to determine the position and pose of the optimal grasping part of the target while recognizing different 3D targets like human. One kind of deep learning method based on the cascaded faster regionbased convolutional neural networks (RCNN) model is proposed to identify the target object and its optimal grasping pose. The improved Faster RCNN model is proposed at the first level, which can recognize small target objects and accurately locate them in the image. Then, a faster RCNN model at the second level is designed to find the optimal grasping pose of the target object recognized by the previous level to realize the optimal grasping of the robot. Experimental results show that the method proposed in this paper can find the object accurately and determine its optimal grasping pose.