Shallow neural networks for autonomous robots

Abstract

The use of Neural Networks (NNs) in modern applications is already well established thanks to the technological advancements in processing units and Deep Learning (DL), as well as the availability of deployment frameworks and services. However, the embedding of these methods in robotic systems is problematic when it comes to field operations. The use of Graphics Processing Units (GPUs) for such networks requires high amounts of power which would lead to shortened operational times. This is not desired since autonomous robots already need to manage their power supply to accommodate the lengths of their missions which can extend from hours to days. While external processing is possible, real-time monitoring can become unfeasible where delays are present. This also applies to autonomous robots that are deployed for underwater or space missions. For these reasons, there is a requirement for shallow but robust NN-based solutions that enhance the autonomy of a robot. This dissertation focuses on the design and meticulous parametrization complemented by methods that explain hyper-parameter importance. This is performed in the context of different settings and problems for autonomous robots in field operations. The contribution of this thesis comes in the form of autonomy augmentation for robots through shallow NNs that can potentially be embedded in future systems carrying NN processing units. This is done by implementing neural architectures that use sensor data to extract representations for event identification and learn patterns for event anticipation. This work harnesses Long Short-Term Memory networks (LSTMs) as the underpinning framework for time series representation and interpretation. This has been tested in three significant problems found in field operations: hardware malfunction classification, survey trajectory classification and hazardous event forecast and detection.