Learning deep representations for robotics applications

[摘要] In this thesis, two hierarchical learning representations are explored in computer vision tasks. First, a novel graph theoretic method for statistical shape analysis, called Compositional Hierarchy of Parts (CHOP), was proposed. The method utilises line-based features as its building blocks for the representation of shapes. A deep, multi-layer vocabulary is learned by recursively compressing this initial representation. The key contribution of this work is to formulate layerwise learning as a frequent sub-graph discovery problem, solved using the Minimum Description Length (MDL) principle. The experiments show that CHOP employs part shareability and data compression features, and yields state-of- the-art shape retrieval performance on 3 benchmark datasets. In the second part of the thesis, a hybrid generative-evaluative method was used to solve the dexterous grasping problem. This approach combines a learned dexterous grasp generation model with two novel evaluative models based on Convolutional Neural Networks (CNNs). The data- efficient generative method learns from a human demonstrator. The evaluative models are trained in simulation, using the grasps proposed by the generative approach and the depth images of the objects from a single view. On a real grasp dataset of 49 scenes with previously unseen objects, the proposed hybrid architecture outperforms the purely generative method, with a grasp success rate of 77.7% to 57.1%. The thesis concludes by comparing the two families of deep architectures, compositional hierarchies and DNNs, providing insights on their strengths and weaknesses.

[发布机构] University:University of Birmingham;Department:School of Computer Science

[关键词] Q Science;QA Mathematics;QA76 Computer software