One Image, Many Insights: A Synergistic Approach Towards Enabling Autonomous Visual Inspection

Abstract: Visual inspection in autonomous robotics is a task in which autonomous agents are required to gather visual information of objects of interest, in a manner that ensures safety, efficiency and comprehensive coverage. It is, therefore, crucial for identifying key landmarks, detecting cracks or defects, or reconstructing the observed object for detailed analysis. This thesis delves into the  challenges encountered by autonomous agents in executing such tasks and presents frameworks for scenarios ranging from operations by multiple spacecrafts in close proximity to celestial bodies in Deep Space to terrestrial deployments of Unmanned Aerial Vehicles (UAVs) for inspection of large-scale infrastructures. The research thus pursues two main directions: Firstly, a novel formation control strategy is developed to enable autonomous agents to perform proximity operations safely, efficiently, and accurately in order to map the surface of Small Celestial Bodies (SCBs). This investigation encompasses control and coordination strategies, leveraging a realistic astrodynamic model of the orbital environment to navigate safely around SCBs. Along this direction, the contributions focus on enabling a distributed autonomy framework in the form of a cooperative stereo configuration between two spacecraft, allowing acquisition of 3D topological information of the candidate SCB. The framework employs a Leader-Follower approach, treating the maintenance of the desired stereo-formation as a 6 Degree-of-Freedom (DoF) nonlinear model predictive control (NMPC) problem.The second research direction focuses on addressing the problem of enabling robotic inspection for terrestrial applications. With the growing demand for efficient and reliable inspection techniques to improve in-situ situational awareness, the research concentrates on addressing the problem of obtaining detailed visual scan of available structures without any a priori knowledge of either the environment nor the structures. Thus, the key contributions of the presented work reside in the implementation of a unified autonomy, with the unification drawing it's root from the merging of two distinct research perspectives: Inspection and Exploration planning. The contribution establishes a novel solution by introducing a map-independent approach with a synergistic formulation of a reactive profile-adaptive view-planner coupled with a hierarchical exploration strategy and an environment-invariant scene recognition module. By integrating exploration and inspection methodologies, this research seeks to enhance the capabilities of UAVs in navigating and inspecting unknown structures in unfamiliar environments. Through theoretical developments, extensive simulations and experimental validations, this thesis contributes to the advancement of the state-of-the-art in visual inspection with autonomous robots. Moreover, the findings extend current capabilities of autonomous agents in the field of space exploration as well as in disaster response and complex infrastructure inspection.