| dc.description.abstract | This dissertation is structured around three core research problems, each addressing a fundamental challenge in enabling safer operation of aerospace vehicles. While the primary motivation stems from the growing use of multirotor Unmanned Aerial Vehicles (UAVs) in civilian and industrial applications, the methodologies developed apply to a broader class of autonomous systems. The three focus areas include: 1) obstacle avoidance (collision safety) -- learning-based obstacle avoidance for real-time motion planning in cluttered environments, 2) noise control (acoustic safety) -- phase and crossover angle control for propellers to reduce aerodynamic interference and noise, and 3) robust guidance (safety from sensitive parameters) -- desensitization of optimal control problems to parametric uncertainties using desensitization techniques. Each of these research topics is explored independently with the goal of advancing autonomy, guidance, and control strategies for reliable and safe operations of aerospace vehicles.
The first part of this dissertation addresses the challenge of collision safety by developing a learning-based framework for real-time UAV motion planning and guidance in cluttered environments. In this section, Neural Networks (NNs) are configured and trained on a dataset of optimal trajectories generated using a convex optimization-based planner. The convex optimizer generates dynamically-compliant, energy-optimal trajectories. To ensure effective learning across various obstacle configurations, hyperparameter tuning is performed for the NN models using Optuna. The proposed motion planning framework employs a cascaded planning architecture in which a high-level path planning NN model first predicts a set of waypoints around detected obstacles, followed by a second NN to generate a smooth, dynamically feasible trajectory using B-spline representations. This two-stage design improves interpretability and post-processing of the waypoints generated. A safety margin correction step is also introduced, which adjusts the predicted trajectories to enforce a buffer zone around obstacles and guarantee collision-free paths. Additionally, an object detection and localization (ODL) software stack is also developed, which leverages stereo vision and a tiny YOLOv3 network to estimate the 3D positions of obstacles. The complete perception-guidance pipeline, including object detection, depth-based obstacle localization, and trajectory generation, is deployed onboard a quadcopter equipped with a stereo camera and a Jetson Nano as a companion computer. Flight experiments validate the system's ability to navigate safely and autonomously in an indoor arena with up to two obstacles.
The second part of this dissertation focuses on controlling the phase angles of rotating propellers, which improves the acoustic signatures of multi-rotor UAVs. Initially, a phase control algorithm is developed that enables precise adjustment of the phase angles. This controller is experimentally tested on various propeller configurations, including single, side-by-side, and coaxial, to validate its effectiveness in different aerodynamic coupling conditions. The phase controller is extended to specifically address counter-rotating configurations, where the two propellers spin in opposite directions and periodically overlap in angular space. This region of overlap, referred to as the crossover, influences the noise propagation characteristics of the system. A crossover position controller is developed to enable the shifting of the angular alignment of the propellers. To assess the controller’s accuracy and performance, a non-invasive image-based validation method is introduced using a high-speed video system and centroid tracking of reflective markers affixed to the propellers. This technique allows for precise measurement of phase and crossover angles without the need for onboard sensors or encoders. Experimental results demonstrate that the controller can achieve desired crossover alignments with an average error of less than 4 degrees across a wide range of propeller speeds and configurations.
The third part of this dissertation addresses the challenges associated with uncertainties in dynamical systems and focuses on the development of a novel trajectory optimization framework that minimizes sensitivity to parametric uncertainties. The investigation begins with a standard low-thrust spacecraft trajectory optimization problem, where an initial attempt is made to improve robustness of the cost (i.e., the final mass of the spacecraft) to the uncertainties in the thrust magnitude of the propulsion system using a costate-based (a.k.a. Lagrange multipliers) desensitized optimal control approach. However, this method does not yield any improvements in the robustness to uncertainties. Further analysis revealed that the costate profiles arising from the hybrid nature of the direct-indirect formulation are non-unique. Using this property of the costates, a new approach, called the Reduced Desensitization Formulation (RDF), is developed for desensitizing optimal control problems. To investigate the effect of desensitization at specific intervals during the flight, a time-triggered variant of RDF is also introduced. The base and time-triggered RDF is applied to a broad range of optimal control problems. Simulation results demonstrate that RDF reduces the dispersion of the trajectory cost functional under parametric uncertainties. | en_US |
