Robot Grasping using Dilated Residual Convolutional Neural Network

Grasping detection for multiple objects is an important task in the application of robots. Despite the achievement obtained by several encoding-decoding deep learning models, grasping detection for objects with much different sizes is still not well solved. In this paper, a Dilated Residual Connection Neural Network (DR-ConvNet) is proposed to achieve the multi-object grasping detection in RGB-D images. On one hand, the detection and location for small objects can addressed well due to the encoding-decoding structure. On the other hand, the proposed dilated residual module can capture more complete features for large objects by its larger receptive field than ordinary convolutions. To verify and highlight the advantages of the proposed DR-ConvNet, an expanded new dataset called S-Cornell is built based on the Cornell grasping dataset. On this dataset, the proposed method can achieve 9S.7% and 95.4% accuracy in single-object and multi-object grasping detection respectively. Several experiments prove that the proposed model can outperform some current popular grasping detection methods in RGB-D images.