For a manipulator to grasp the novel objects with variable sizes, different shapes, and arbitrary poses, a single-stage grasp pose detection algorithm based on multi-level features is designed by taking the grasp position detection problem of objects as the grasp angle classification and grasp position regression processing, and performing a single prediction for grasp angle and grasp position. The RGD image is generated by replacing the blue channel of RGB image with depth data, and the lightweight feature extractor VGG16 is used as the backbone network. For the problem that the feature extraction ability of VGG16 is weak, the Inception module is used to design a network model with stronger feature extraction capability. Then, grasp position is sampled using the method of priori box on the feature map of different levels, and the adaptability of the model to the objects with variable sizes is improved through the combination of shallow features and deep features. Finally, the detection result with the highest confidence is output as the optimal grasp pose. The evaluation results of the proposed algorithm on the image-wise dataset and the object-wise dataset are respectively 95.71% and 94.01%, and the detection speed is 58.8FPS, and the accuracy and speed are improved compared with the current methods.