Generative neural data synthesis for autonomous systems
A significant number of Machine Learning methods for automation currently rely on data-hungry training techniques. The lack of accessible training data often represents an insurmountable obstacle, especially in the fields of robotics and automation, where acquiring new data can be far from trivial. Additional data acquisition is not only often expensive and time-consuming, but occasionally is not even an option. Furthermore, the real world applications sometimes have commercial sensitivity issues associated with the distribution of the raw data. This doctoral thesis explores bypassing the aforementioned difficulties by synthesising new realistic and diverse datasets using the Generative Adversarial Network (GAN). The success of this approach is demonstrated empirically through solving a variety of case-specific data-hungry problems, via application of novel GAN-based techniques and architectures. Specifically, it starts with exploring the use of GANs for the realistic simulation of the extremely high-dimensional underwater acoustic imagery for the purpose of training both teleoperators and autonomous target recognition systems. We have developed a method capable of generating realistic sonar data of any chosen dimension by image-translation GANs with Markov principle. Following this, we apply GAN-based models to robot behavioural repertoire generation, that enables a robot manipulator to successfully overcome unforeseen impedances, such as unknown sets of obstacles and random broken joints scenarios. Finally, we consider dynamical system identification for articulated robot arms. We show how using diversity-driven GAN models to generate exploratory trajectories can allow dynamic parameters to be identified more efficiently and accurately than with conventional optimisation approaches. Together, these results show that GANs have the potential to benefit a variety of robotics learning problems where training data is currently a bottleneck.