Robotic Grasp Detection by Rotation Region CNN

Recently using deep learning methods for robotic grasping is a promising research. Many previous works used one- or two-stage deep learning methods to learn optimal grasping rectangles. However, these deep learning methods mainly detected vertical bounding boxes and performed post-processing for finding grasps. To avoid post-processing, we adopt the rotation region convolutional neural network (R2CNN) to detect oriented optimal grasps without post-preprocess. The modified R2CNN is divided into three stages: (1) feature extraction, (2) intermediate layer, and (3) gasp detection. In the second stage, we found that using a smaller set of anchor scale and a small IoU threshold were helpful to detect correct grasping rectangles. In our experiment, we used the Cornell grasping dataset as the benchmark and validated that using both axis-aligned and inclined bounding boxes in training. The results show that our modified R2CNN for image-wise detection reached up to 96% in accuracy.