Autonomous inspection plays a vital role in modern robotics, supporting routine site monitoring to mapping unstructured environments. To achieve this, field robots leverage onboard perception, planning and control systems to interpret their environment and execute complex tasks autonomously and in real time. As robots become increasingly adopted for long-term and large-scale inspection missions, the need for a resourceful agent becomes evident. From the integration of autonomous robots with mining infrastructure to rapid deployment of the systems for enhancing situational awareness to extended interaction between humans and robot on inspection scenes, these applications underscore the importance of a scalable, interpretable and adaptable inspection autonomy. To this end, this thesis investigates the key enablers for autonomous inspection: view planning and scene representation. Specifically, it focuses on: (a) developing a reactive view-planning strategy, (b) constructing interpretable inspection scene representation.

The first contribution introduces the method of reactive view planning for mobile robots. Primarily, the First-Look inspection framework establishes the foundation for visual inspection. It utilizes instantaneous 3D point-cloud data alongside desired photogrammetric constraints to generate an inspection view pose. The planner recursively evaluates this process to provide visual coverage of the inspection target. Building on this, the First-Look planner is extended with an exploration behaviour to form the First-Look Inspect-Explore (FLIE) framework. Through this integration, the FLIE framework addresses visual inspection of distributed and unknown inspection targets in unknown environments. Finally, the thesis investigates a hierarchical inspection framework, combining global view planning over historical maps with adaptive local replanning capability for inspection under environment uncertainty. The second contribution explores the method of constructing an actionable, hierarchical scene graph representation of inspected targets, namely 3D Layered Semantic Graph (3DLSG). Specifically, the inspection scene is composed as an ordered set of nodes and edges, organized into abstract layers and structured in a hierarchical manner. The method encodes sensor, semantic and planning information from the scene as attributes of the nodes and edge set. Through this actionable and interpretable hierarchy, the method addresses semantic-geometric path planning for navigation and informed target-selection during inspection planning.  

Overall, this thesis demonstrates deployment of autonomous robots: Unmanned Aerial Vehicles (UAVs) and Quadrupedal platforms, in subterranean and open-pit mines and outdoor urban environments. In all scenarios, field robots operate autonomously, relying solely on onboard perception, planning and control to complete their inspection tasks.