Convolutional Residual Network for Grasp Localization

Object grasping is an important ability for carrying out complex manipulation tasks with autonomous robotic systems. The grasp localization module plays an essential role in the success of the grasp maneuver. Generally viewed as a vision perception problem, its goal is determining regions of high graspability by interpreting light and depth information. Over the past few years, several works in Deep Learning (DL) have shown the high potential of Convolutional Neural Networks (CNNs) for solving vision-related problems. Advances in residual networks have further facilitated neural network training by improving convergence time and generalization performances with identity skip connections and residual mappings. In this paper, we investigate the use of residual networks for grasp localization. A standard residual CNN for object recognition uses a global average pooling layer prior to the fully-connected layers. Our experiments have shown that this pooling layer removes the spatial correlation in the back-propagated error signal, and this prevents the network from correctly localizing good grasp regions. We propose an architecture modification that removes this limitation. Our experiments on the Cornell task have shown that our network obtained state-of-the-art performances of 10.85% and 11.86% rectangle metric error on image-wise and object-wise splits respectively. We did not use pre-training but rather opted for on-line data augmentation for managing overfitting. In comparison to previous approach that employed off-line data augmentation, our network used 15x fewer observations, which significantly reduced training time.