This thesis investigates a hybrid framework for vision-based relative navigation of spacecraft operating in close proximity with an uncooperative target, combining classical Kalman filtering techniques with deep learning models. The motivation arises from the growing need for accurate and autonomous guidance during critical phases of on-orbit servicing (OOS), rendezvous, and active debris removal (ADR) missions, where the target spacecraft is non-responsive and its dynamics are only partially known. In such scenarios, traditional model-based filters often face limitations due to model mismatch, uncertain noise statistics, and strong nonlinearities. The proposed approach integrates recurrent neural networks (RNNs) into Kalman-based architectures to enhance adaptability, robustness, and estimation accuracy without requiring explicit knowledge of the system dynamics or noise statistics. Two complementary hybrid pipelines were developed: a KalmanNet-based translational filter, in which a neural network learns to infer the Kalman gain directly from data, and a GRU-based rotational model, designed to infer the target’s angular velocity directly from sequences of quaternion measurements by learning the underlying rotational dynamics in a fully data-driven manner. This model is then embedded within a Multiplicative Extended Kalman Filter (MEKF) to provide more accurate information on the rotational motion and improve attitude estimation accuracy. Both systems were validated through high-fidelity simulations based on the PRISMA mission, reproducing realistic orbital conditions and vision-based measurements of an uncooperative target. The hybrid filters significantly outperformed classical counterparts (EKF, UKF, H∞, and MEKF + QuateRA) in estimation accuracy, noise resilience, and generalization across multiple scenarios, confirming the potential of deep learning-enhanced Kalman filtering as a promising direction for next-generation autonomous spacecraft navigation around uncooperative targets, where physical interpretability and adaptive learning coexist effectively.
La presente tesi esplora un framework ibrido per la navigazione relativa di veicoli spaziali basata sulla visione, operanti in prossimità di un target non cooperativo, combinando le tecniche classiche di filtraggio di Kalman con modelli di deep learning. La motivazione nasce dalla crescente necessità di sistemi di guida accurati e autonomi durante le fasi critiche di missioni di on-orbit servicing (OOS), rendezvous e active debris removal (ADR), in cui il satellite target è non responsivo e la sua dinamica è solo parzialmente nota. In tali scenari, i filtri basati su modelli tradizionali mostrano spesso limiti dovuti a discrepanze dinamiche, incertezze nelle statistiche del rumore e forti non linearità. L’approccio proposto integra reti neurali ricorrenti (RNN) all’interno di architetture basate su Kalman, migliorando adattabilità, robustezza e accuratezza della stima senza richiedere una conoscenza esplicita della dinamica del sistema o del rumore. Sono state sviluppate due pipeline ibride complementari: un filtro traslazionale basato su KalmanNet, in cui una rete neurale apprende a ricavare direttamente dai dati il guadagno di Kalman, e un modello rotazionale basato su GRU, progettato per stimare la velocità angolare del target a partire da sequenze di quaternioni, apprendendo in modo completamente data-driven la dinamica rotazionale sottostante. Quest’ultimo modello è poi integrato all’interno di un filtro esteso moltiplicativo (MEKF), fornendo informazioni più accurate sul moto rotazionale e migliorando la stima complessiva dell’assetto. Entrambi i sistemi sono stati validati mediante simulazioni ad alta fedeltà basate sulla missione PRISMA, riproducendo condizioni orbitali realistiche e misure visive di un target non cooperativo. I filtri ibridi hanno mostrato prestazioni significativamente superiori rispetto ai corrispettivi classici (EKF, UKF, H∞ e MEKF + QuateRA) in termini di accuratezza di stima, resistenza al rumore e capacità di generalizzazione in diversi scenari, confermando il potenziale del filtraggio di Kalman potenziato dal deep learning come approccio promettente per la navigazione autonoma di veicoli spaziali intorno a target non cooperativi, in cui l’interpretabilità fisica e la capacità di apprendimento adattivo coesistono efficacemente.
Deep learning-enhanced Kalman filtering for vision-based relative navigation
Carone, Silvia
2024/2025
Abstract
This thesis investigates a hybrid framework for vision-based relative navigation of spacecraft operating in close proximity with an uncooperative target, combining classical Kalman filtering techniques with deep learning models. The motivation arises from the growing need for accurate and autonomous guidance during critical phases of on-orbit servicing (OOS), rendezvous, and active debris removal (ADR) missions, where the target spacecraft is non-responsive and its dynamics are only partially known. In such scenarios, traditional model-based filters often face limitations due to model mismatch, uncertain noise statistics, and strong nonlinearities. The proposed approach integrates recurrent neural networks (RNNs) into Kalman-based architectures to enhance adaptability, robustness, and estimation accuracy without requiring explicit knowledge of the system dynamics or noise statistics. Two complementary hybrid pipelines were developed: a KalmanNet-based translational filter, in which a neural network learns to infer the Kalman gain directly from data, and a GRU-based rotational model, designed to infer the target’s angular velocity directly from sequences of quaternion measurements by learning the underlying rotational dynamics in a fully data-driven manner. This model is then embedded within a Multiplicative Extended Kalman Filter (MEKF) to provide more accurate information on the rotational motion and improve attitude estimation accuracy. Both systems were validated through high-fidelity simulations based on the PRISMA mission, reproducing realistic orbital conditions and vision-based measurements of an uncooperative target. The hybrid filters significantly outperformed classical counterparts (EKF, UKF, H∞, and MEKF + QuateRA) in estimation accuracy, noise resilience, and generalization across multiple scenarios, confirming the potential of deep learning-enhanced Kalman filtering as a promising direction for next-generation autonomous spacecraft navigation around uncooperative targets, where physical interpretability and adaptive learning coexist effectively.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Carone_Thesis.pdf
accessibile in internet per tutti
Dimensione
19.06 MB
Formato
Adobe PDF
|
19.06 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Carone_Executive_Summary.pdf
accessibile in internet per tutti
Dimensione
2.24 MB
Formato
Adobe PDF
|
2.24 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/10589/246908