Pixel-Level Grasp Detection for Unknown Objects with Encoder-Decoder-Inception Deep Network

Abstract

In the past, robot grasping mostly used the extraction of candidate grasping rectangles. This method is discrete sampling, time-consuming, and may ignore potential best grasping pose, which has limitations. To solve the above problems, a pixel-level grasping detection method is proposed in this paper, which uses deep neural network to directly generate grasping pose on pixels. Firstly, the Encoder-Decoder-Inception Convolution Neural Network (EDINet) is proposed to predict the grasping pose. In our finding, EDINet uses encoder, decoder and inception modules to improve the speed and robustness of grasp detection in unstructured scenes. Then, the proposed deep neural network structure was evaluated on the published Cornell grasp dataset, our structure achieves 98.9% test accuracy. Finally, we carry out the grasping experiment on the actual objects. Experiments show that the average success rate of our network model is 97.2% in a single object scene and 92.0% in a cluttered scene. In addition, EDINet only completes a grasp detection pipeline within 25ms.