Harpia: A Hybrid System for UAV Missions

• Ano de defesa: 2023
• Autor(a) principal: Vannini, Verônica
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Artifical intelligence; Autonomus systems; Inteligência artificial; Robótica; Robotics; Sistemas autônomos; UAV; VANt
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-08012024-111421/

This doctoral project presents Harpia, a hybrid artificial intelligence planning system for UAVs (Unmanned Aerial Vehicles) with a focus on autonomy. Harpia aims to perform tasks for general-purpose applications with minimal human intervention. To facilitate understanding, the problem addressed is set on a farm where the autonomous system must be capable of carrying out missions safely. The system architecture is implemented using the Robotic Operating System (ROS) and includes functionalities such as task re-planning and trajectory planning with obstacle avoidance. Re-planning can occur after real-time mission changes or due to unpredictable UAV behavior. Harpia combines the Planning Domain Definition Language (PDDL) for task planning, a Bayesian Network (BN) for evaluating mission execution, a K-Nearest Neighbors (KNN) algorithm for selecting a trajectory planner, Principal Component Analysis (PCA), and a Decision Tree (DT) to assess the health of the aircraft. Therefore, the novelty of Harpia focuses on robustness for autonomous planning and re-planning of the sequence of tasks and trajectories for regions of interest. The main contributions include an autonomous system architecture to plan missions with minimal human intervention, unconstrained by specific tasks, and computationally simple to operate in diverse scenarios. Computational tests report results for 220 simulated scenarios, in which Harpia adequately handled all situations, for example, making decisions about task re-planning with 97.57% accuracy based on battery health and choosing the best planning trajectory for each case with at least 95% accuracy.