Modular Navigation and Autonomy for Aerial Robots

Abstract

A wealth of novel use cases and applications were made possible with the advent of Micro Aerial Vehicles (MAVs). Typical examples include aerial delivery, photogrammetry, inspection, and monitoring. The low hardware cost and general availability sparked an explosion in adoption not only in industry but also in the interest of the scientific community. A large body of research on flight control, state estimation, perception, navigation, and planning for these agile and cost-effective vehicles has been published. However, the standard quadrotor is limited in its physical interaction capabilities, as the under-actuated nature constrains the force and torque envelope. To overcome these limitations, a novel type of aerial robot has been developed - the Omnidirectional MAV (OMAV). By adding thrust-vectoring capabilities, the OMAV is able to counter and exert forces and torques in all directions. These aerial robots enable novel use cases that need aerial-physical interaction, such as contact-based inspection, painting, drilling, milling, or general manipulation. Physical interaction entails a distinctively different set of requirements concerning control, state estimation, and navigation, compared to the free-flight nature of traditional MAVs. One major difference is the discrete nature of physical contact - ensuring safety by applying conservative simplifications in navigation and mapping to stay away from obstacles is not effective. Similarly, control algorithms need to take into account different sources of disturbances. An aerial robot is subject to desired disturbances from physical interaction as well as undesired and unknown disturbances from the airflow regime close to structure. The combination of two essential flight modes, the safe free-space flight in potentially challenging environments, and the precise and accurate physical interaction, is especially difficult to accommodate in classical, monolithic, and fixed-resolution mapping and navigation systems. In this doctoral dissertation, we address many of the frontiers preventing the real-world use of aerial robots. In the first work, we extend the classical control methodology with a perception-in-the-loop system that uses geometrical information of the robot's surroundings to disentangle interaction forces from disturbances. This lays the foundation for stable and safe physical interaction, by allowing the robot to react fiercely against disturbances while staying compliant to the object it interacts with. Drawing from the experiences of a wealth of real-world tests in large, difficult, and diverse indoor and outdoor environments such as glaciers, deserts, and construction sites, we motivate and contribute a novel modular navigation architecture tailored to aerial robots. We explore the idea of a polylithic navigation system, that uses many individually and parallely executable and testable behaviors. Each of these behaviors may rely on a different sensing or estimation process. By combining many such behavior policies, the system is able to actively choose, blend, and employ different strategies for different flight modes or environments. The set of available strategies can then be adopted based on the available sensing modalities and estimation quality, drastically increasing the system's robustness and adaptability. As part of this thesis, we contribute many policy building blocks for collision avoidance, surface following, servoing, and interaction. In different publications, we show how a massively parallel set of purely reactive policies compares favorably to state-of-the-art navigation algorithms in terms of navigation capability, but at vastly lower computational cost and without the need for a complete map. We demonstrate this on a variety of representations, such as volumetric maps on GPUs, hierarchical multi-resolution representations on CPU, and in a data-driven, NeRF-based approach. Additionally, we show how the manifold-like structure of surface meshes embedded in 3D space enables high-accuracy surface traversal at minuscule computation cost. In an extensive study, we apply a polylithic navigation system in a construction environment. In this study, we show how an aerial robot can navigate and interact robustly and accurately without a globally consistent map in a purely reactive fashion. Finally, we contribute a large number of improvements to the hardware and low-level architecture of aerial robots. These improvements are specifically tailored to interface with a polylithic navigation system and future data-driven architectures. Additionally, we present a novel concept that enables OMAVs to fly as a fixed-wing vehicle - without needing any additional actuators in comparison to the pre-existing aerial robot. This allows the robot to fly large distances and retain its manipulation capabilities - a combination that enables many practical applications in remote locations. Overall, this thesis advances the state-of-the-art in navigation, hardware, control, and conceptualization of aerial robots. Albeit demonstrated and designed for aerial robots, the scientific work on navigation presented in this thesis also applies to walking, driving, and swimming robots. Our contributions are another important step towards robots that perform work and difficult tasks in locations inaccessible to humans.