Absolute Localization for Autonomous Robotic Inspection in Confined Spaces

The modern world is built upon a large industrial sector consisting of areas such as manufacturing plants, chemical processing, energy production, and a large marine-based global supply chain. Maintaining and ensuring optimal runtime of these large and complex assets requires frequent inspections and repairs. These inspection missions pose a series of potential health risks to human surveyors, and most critically confined space inspections result in several yearly deaths across the industry. It is clear that a robotic solution has the potential to mitigate some of these fatalities by removing humans from the danger zone. Due to the complex shape of the associated operational spaces, teleoperated robots do not perform well. Therefore, autonomous robotic solutions with the ability to self-localize and self-navigate are required. However, the localization of robots in confined industrial spaces poses a range of challenges. The environment is often feature-few, and poorly lit, and a high level of symmetry tends to be a common factor across most sectors. Due to these challenges, autonomous robot-based inspections with the ability to provide absolute positions of the
robot and defects are currently non-existent in the inspection of industrial confined spaces. To perform state estimation on autonomous robots, well-formed uncertainty estimates are required to be provided by all individual estimates. Therefore, accurate uncertainty estimation for 3D-3D registrations, in the context of localization within a given map of the environment, is also an area where further research is needed. Moreover, the ability to provide a good pose and uncertainty estimate of the defects within the critical areas of a vessel, is important to provide a proper evaluation of each vessel, whether evaluated by a human or by machine learning approaches. The first step in filling this gap in research and industry is the investigation of the optimal system to obtain absolute localization both for defect pose estimation and as an input for the navigation of the robots themselves. Therefore, the presented research aims to explore viable solutions for absolute localization in autonomous robotic inspection of confined spaces, using a mixture of compact lightweight visual and ToF cameras, with the goal of fitting the resulting system onto payload-limited
platforms in the shape of UAVs or small legged robots. The research presented in this dissertation is focused on the water ballast tanks of marine vessels and begins by tackling the absolute localization within these known environments using unmanned aerial vehicles. As part of the first presented method, a custom feature detector was developed to provide a robotic and defect pose estimate in a high-level map of the ballast tanks. The disadvantages of this approach were later addressed by employing more general deep learning-based feature descriptors in combination with a novel registration algorithm, Teaser++. To handle the computational load and ambiguity of the environment, specific attention was given to the design / execution of the developed algorithms onboard GPUs as well as to the inclusion of state-of-the-art uncertainty estimation. In addition, an exploration-based
inspection approach with a focus on defect detection was investigated and proved viable in a real-world simulated water ballast tank. With these findings, it was shown that it is possible to use small-scale aerial robotic systems to autonomously inspect confined spaces.