KESTREL: Knowledge-State-Driven Aerial Inspection of Electrical Energy Infrastructure

Abstract:

Aerial robots can already reach power lines and substations faster and more safely than ground crews, elevated platforms, or manned helicopters. In current inspection workflows, however, they mostly remain data collectors: they fly predefined or weakly adaptive routes, while the judgment of whether the collected evidence is sufficient for a reliable assessment is left to human operators after the flight. This project addresses that gap by formulating aerial inspection of electrical energy infrastructure as a closed-loop decision-and-control problem driven by an explicit, inspection-relevant knowledge state. The aerial robot will maintain a probabilistic representation of what is known and unknown about the inspected asset, including the defect likelihood and condition of individual components, observation quality, map completeness, and localization quality. It will use this representation to decide where to look next, when a component should be revisited, and when the available evidence is sufficient to stop. These decisions will be coupled with safety-critical motion generation based on model predictive control and control barrier functions, so that informative viewpoints remain dynamically feasible and respect stand-off constraints near live assets. Finally, the framework will be extended to persistent, multi-sortie operation with docking-based recharging and degraded communication. The methods will be validated progressively in simulation, hardware-in-the-loop setups, and experiments in representative power-line and substation scenarios.
 

Main goal

To establish the control-theoretic and computational foundations of self-directed aerial inspection, in which an aerial robot selects viewpoints, revisits, and stopping decisions according to their contribution to a reliable assessment of the asset, while operating safely and persistently near critical electrical infrastructure.

Research directions

• Inspection-relevant knowledge state: probabilistic models of component condition and defect likelihood, and observation models linking viewpoint, distance, and illumination to the reliability of what is observed.
• Information-driven inspection decisions: viewpoint selection, revisitation, and stopping criteria based on expected assessment value, with tractable approximations for multi-component assets.
• Safety-critical motion generation: coupling of inspection decisions with model predictive control and control barrier functions, so that only safely executable viewpoints are pursued.
• Persistent inspection: multi-sortie missions with docking-based recharging, mission resumption, and robustness to degraded communication.

In Media: