Aircraft ground operations, i.e., taxiing, takeoff roll, rejected takeoff and landing rollout, are among the most routine yet demanding phases of flight: a recognized source of incidents and a major contributor to pilot workload, they have nonetheless received far less attention in the industry than the flight phases. Two complementary needs follow: assisting the human pilot during high-speed landings, where controllability of the aircraft is safety critical and a stability system is useful only if accepted in the cockpit; and automating taxiing: where autonomy on the ground rests heavily on perception, yet the airport scene differs enough from the road that off-the-shelf solutions in the automotive industry do not transfer directly, and public datasets of airport operations are few and unevenly labeled. This thesis addresses both needs, bringing machine learning-based control and perception to the ground operations for crewed and uncrewed aircraft. Firstly, the proposed data-driven method tunes the Aircraft Electronic Stability Control (A-ESC), a system that assists the pilot through the high-speed rollout, injecting a corrective yaw moment whenever the aircraft leaves a safe yaw-rate envelope. Secondly, a family of machine learning vision-based systems is developed, which lets the aircraft detect obstacles, read the markings that govern where it must stop, follow the taxiway centerline, and localize itself when satellite navigation is denied. All the systems are validated in closed-loop on a common, pre-existing high-fidelity testbed. A single methodological thread runs through the four contributions. On the crewed side, the pilot’s preference is learned from pairwise judgments through preference-based optimization. On the uncrewed side, every perception network is domain-specialized rather than borrowed, and trained leak-free either on real airport imagery re-labeled with foundation models or, for the markings, the centerline and the intersections, on synthetic imagery whose ground truth is computed analytically from the airport’s surveyed geometry.
Le operazioni di terra degli aeromobili, rullaggio, decollo, decollo interrotto e atterraggio, sono tra le fasi più comuni e insieme più impegnative: pur essendo una fonte riconosciuta di incidenti e un fattore primario nel carico di lavoro del pilota, hanno ricevuto meno attenzione delle fasi di volo. Ne derivano due esigenze complementari: assistere il pilota durante atterraggi ad alta velocità, in cui la controllabilità è critica per la sicurezza e un sistema di stabilità è utile solo se accettato in cabina; e automatizzare le operazioni di terra, dove la percezione visiva è necessaria; tuttavia, lo scenario aeroportuale rende le soluzioni proposte dal settore automobilistico non direttamente utilizzabili, e i set di dati pubblici delle operazioni aeroportuali sono pochi e annotati in modo disomogeneo. Questa tesi affronta entrambe le esigenze portando controllo e percezione basati sull'apprendimento automatico alle operazioni di terra per aeromobili con e senza pilota. In primo luogo, viene proposto un metodo di messa a punto del sistema Aircraft Electronic Stability Control (A-ESC) basato sui dati, che garantisce assistenza durante le operazioni ad alta velocità applicando un momento di imbardata correttivo quando l'aeromobile esce dall'inviluppo sicuro di velocità di imbardata. In secondo luogo, si sviluppa una famiglia di sistemi di visione basati sull'apprendimento automatico che consente all'aeromobile di rilevare gli ostacoli, leggere la segnaletica che indica dove fermarsi, seguire la mezzeria della via di rullaggio e localizzarsi quando la navigazione satellitare è negata. I sistemi sono validati ad anello chiuso su un banco di prova già esistente. Un filo conduttore metodologico attraversa i contributi. Sul versante con pilota, la preferenza è appresa da giudizi a coppie tramite ottimizzazione basata sulle preferenze. Sul versante senza pilota, ogni rete di percezione è specializzata sul dominio anziché mutuata, e addestrata senza fughe di dati sia su immagini aeroportuali reali rietichettate, sia per segnaletica, mezzeria e intersezioni, su immagini sintetiche il cui riferimento di verità è calcolato analiticamente.
Machine learning-based control and perception for crewed and uncrewed aircraft ground operations
D'ADDAZIO, LUCA;URSI, LORENZO
2025/2026
Abstract
Aircraft ground operations, i.e., taxiing, takeoff roll, rejected takeoff and landing rollout, are among the most routine yet demanding phases of flight: a recognized source of incidents and a major contributor to pilot workload, they have nonetheless received far less attention in the industry than the flight phases. Two complementary needs follow: assisting the human pilot during high-speed landings, where controllability of the aircraft is safety critical and a stability system is useful only if accepted in the cockpit; and automating taxiing: where autonomy on the ground rests heavily on perception, yet the airport scene differs enough from the road that off-the-shelf solutions in the automotive industry do not transfer directly, and public datasets of airport operations are few and unevenly labeled. This thesis addresses both needs, bringing machine learning-based control and perception to the ground operations for crewed and uncrewed aircraft. Firstly, the proposed data-driven method tunes the Aircraft Electronic Stability Control (A-ESC), a system that assists the pilot through the high-speed rollout, injecting a corrective yaw moment whenever the aircraft leaves a safe yaw-rate envelope. Secondly, a family of machine learning vision-based systems is developed, which lets the aircraft detect obstacles, read the markings that govern where it must stop, follow the taxiway centerline, and localize itself when satellite navigation is denied. All the systems are validated in closed-loop on a common, pre-existing high-fidelity testbed. A single methodological thread runs through the four contributions. On the crewed side, the pilot’s preference is learned from pairwise judgments through preference-based optimization. On the uncrewed side, every perception network is domain-specialized rather than borrowed, and trained leak-free either on real airport imagery re-labeled with foundation models or, for the markings, the centerline and the intersections, on synthetic imagery whose ground truth is computed analytically from the airport’s surveyed geometry.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_D_Addazio_Ursi_Tesi.pdf
non accessibile
Descrizione: Tesi
Dimensione
95.15 MB
Formato
Adobe PDF
|
95.15 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_D_Addazio_Ursi_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
1.06 MB
Formato
Adobe PDF
|
1.06 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260627