In the last few years, a branch of machine learning (ML), known as deep learning (DL), has greatly revived the entire field of research, achieving major achievements in areas as diverse as Computer Vision (CV), Natural Language Processing (NLP), speech recognition and automated reasoning. This is made possible by its intrinsic capability to learn not only the mapping function from pre- processed input to output, but also the data representations itself. Indeed, distributed representations learned by deep learning architectures are much more disentangled and representative to effectively solve tasks that require high levels of abstraction.

However, up until now, deep learning has always been associated with large quantities of data and especially high computational capacity, drastically reducing its use and fields of application. For example, precision agriculture is one of the several fields that could largely benefit from the application of ML and DL algorithms. Indeed, diverse sources of data coming from remote sensing instruments such as satellites, aerial vehicles (UAV) and terrestrial robots (UGV), if efficiently exploited, they could extremely reduce costs and increase production yield. Moreover, as is common in a considerable amount of domains, getting data from agricultural field is extremely difficult since there is often no power and Internet connection.

In this context, the presented research primarily focuses on ML algorithms for Service Robotics and on Edge AI, finding new effective training objectives and strategies in order to create lightweight and more resilient DL algorithms able to be processed locally on embedded low power hardware devices.