Autonomous Exploration of Large-Scale Natural Environments

Abstract

This thesis addresses issues which arise when using robotic platforms to explore large-scale, natural environments. Two main problems are identified: the volume of data collected by autonomous platforms and the complexity of planning surveys in large environments.
\n
\nAutonomous platforms are able to rapidly accumulate large data sets. The volume of data that must be processed is often too large for human experts to analyse exhaustively in a practical amount of time or in a cost-effective manner. This burden can create a bottleneck in the process of converting observations into scientifically relevant data.
\n
\nAlthough autonomous platforms can collect precisely navigated, high-resolution data, they are typically limited by finite battery capacities, data storage and computational resources. Deployments are also limited by project budgets and time frames. These constraints make it impractical to sample large environments exhaustively. To use the limited resources effectively, trajectories which maximise the amount of information gathered from the environment must be designed.
\n
\nThis thesis addresses these problems. Three primary contributions are presented: a new classifier designed to accept probabilistic training targets rather than discrete training targets; a semi-autonomous pipeline for creating models of the environment; and an offline method for autonomously planning surveys. These contributions allow large data sets to be processed with minimal human intervention and promote efficient allocation of resources.
\n
\nIn this thesis environmental models are established by learning the correlation between data extracted from a digital elevation model (DEM) of the seafloor and habitat categories derived from in-situ images. The DEM of the seafloor is collected using ship-borne multibeam sonar and the in-situ images are collected using an autonomous underwater vehicle (AUV).
\n
\nWhile the thesis specifically focuses on mapping and exploring marine habitats with an AUV, the research applies equally to other applications such as aerial and terrestrial environmental monitoring and planetary exploration.