Learning data-efficient hierarchical features for robotic graspable object recognition

Robotic graspable object recognition is a crucial ingredient in many exciting autonomous manipulation applications. However, identifying complex image features from limited data remains largely unsolved. In this paper, we leverage the advantages of two kinds of feature representation approaches, kernel descriptors and deep neural networks, to present a novel hierarchical feature learning framework for robotic graspable object recognition. This work enables the recovery of sparse and compressible features from limited data examples. Firstly, we design multiple kernel descriptors from the raw RGB-D images to adequately capture the discriminative structure of the object. Then, the extracted abstract representations are transferred to a four-layer deep neural network to generate more representative features for final graspable discrimination. Our network obtains impressive generalization capability with limited training data. Extensive experiments are carried out to validate the proposed method and the results show the state-of-the-art performance in discriminating graspable object task under limited-data.