Improved Deep Neural Networks for Generative Robotic Grasping

Abstract

This thesis provides a thorough evaluation of current state-of-the-art robotic grasping methods and contributes to a subset of data-driven grasp estimation approaches, termed generative models. These models aim to directly generate grasp region proposals from a given image without the need for a separate analysis and ranking step, which can be computationally expensive. This approach allows for fully end-to-end training of a model and quick closed-loop operation of a robot arm.

A number of limitations are identified within these generative models, which are identified and addressed. Contributions are proposed that directly target each stage of the training pipeline that help to form accurate grasp proposals and generalise better to unseen objects. Firstly, inspired by theories of object manipulation within the mammalian visual system, the use of multi-task learning in existing generative architectures is evaluated. This aims to improve the performance of grasping algorithms when presented with impoverished colour (RGB) data by training models to perform simultaneous tasks such as object categorisation, saliency detection, and depth reconstruction. Secondly, a novel loss function is introduced which improves overall performance by rewarding the network to focus only on learning grasps at suitable positions. This reduces overall training times and results in better performance on fewer training examples. The last contribution analyses the problems with the most common metric used for evaluating and comparing offline performance between different grasping models and algorithms. To this end, a Gaussian method of representing ground-truth labelled grasps is put forward, which optimal grasp
locations tested in a simulated grasping environment.

The combination of these novel additions to generative models results in improved grasp success, accuracy, and performance on common benchmark datasets compared to previous approaches. Furthermore, the efficacy of these contributions is also tested when transferred to a physical robotic arm, demonstrating the ability to effectively grasp previously unseen 3D printed objects of varying complexity and difficulty without the need for domain adaptation. Finally, the future directions are discussed for generative convolutional models within the overall field of robotic grasping.