Towards Fully Autonomous Inspection with Quadruped Robots in Unstructured and Safety-Critical Environments

Inspection operations in industrial and search-and-rescue (SAR) environments often expose workers to hazardous conditions, including unstable terrain, confined spaces, and partially inaccessible areas. In this context, autonomous robotic systems represent a promising solution for reducing human exposure to risk while increasing operational efficiency. Among mobile robotic platforms, quadruped robots offer significant advantages due to their mobility and ability to operate in complex and unstructured environments. This thesis investigates the problem of autonomous inspection using quadruped robots in hazardous industrial and emergency scenarios. The work begins with a stakeholder-driven analysis involving first responders and industrial operators, with the objective of identifying operational requirements, expectations, and potential robotic applications for inspection tasks. Based on the collected requirements, the research focuses on the development of autonomous exploration and feature detection strategies for inspection-oriented missions. To support autonomous inspection in multi-floor environments, the thesis proposes an exploration strategy aimed at maximizing environmental coverage under time constraints while accounting for the costs and risks associated with traversing different floors. In parallel, a probabilistic feature detection and localization framework is introduced to support the identification of relevant environmental features in inspection and SAR scenarios. The proposed method prioritizes robust and operationally meaningful detections over highly precise localization, reflecting the practical needs of first responders operating in time-critical conditions. The thesis further investigates the integration of autonomous exploration and semantic feature detection, evaluating the influence of different exploration strategies on both environmental coverage and detection performance. The results demonstrate that exploration policies can significantly affect explored area coverage while maintaining comparable detection effectiveness. To address more realistic operational conditions, the work proposes scenarios involving uneven terrain, dynamic obstacles, and task-level planning constraints. In particular, the thesis introduces a navigation framework that accounts for terrain slope and robot safety during path planning, as well as an emotionally modulated local planner for navigation in dynamic environments. Finally, a PDDL-based planning framework is proposed to support inspection tasks in multi-floor emergency scenarios involving blocking obstacles and task coordination requirements. The proposed approaches are validated through extensive simulation experiments and real-world testing with the Spot robot. Overall, the results demonstrate the feasibility of autonomous inspection using quadruped robots under realistic operational constraints, including limited time, partial observability, uneven terrain, dynamic obstacles, and high-level planning requirements. Collectively, this thesis contributes toward the development of autonomous quadruped robotic systems capable of supporting inspection operations in safety-critical real-world environments.