REAVER (Resource Extraction Autonomous Vehicle for Environmental Recovery) is a reusable active debris removal mission for the geostationary Earth orbit (GEO) graveyard region, designed to capture five large non-cooperative GEO debris objects within one operational year and transport them to a recycling hub. ... Read more

The increasing traffic in Low Earth Orbit (LEO) necessitates effective and timely Collision Avoidance Manoeuvres (CAMs) planning. However, standard procedures neglect the epistemic uncertainty present in Conjunction Data Messages, thereby compromising the robustness of the resulting manoeuvres. While a method accounting for both epistemic and aleatory uncertainties exists, its high computational cost limits real-world scalability.

This thesis investigates the application of Deep Reinforcement Learning to address this limitation by developing a Deep Contextual Bandit agent for robust CAM planning. The proposed agent generates nearly optimal CAMs accounting for epistemic and aleatory uncertainty in under a second, significantly reducing the runtime compared to traditional numerical approaches. When evaluated on a synthetic database of nearly circular LEO orbits, the agent successfully minimised the worst-case probability of collision in over 90% of the test cases. ... Read more

Atmospheric drag remains one of the main sources of uncertainty in low Earth orbit propagation because the thermospheric density depends strongly on solar and geomagnetic activity and can vary substantially over time and space. This uncertainty directly affects orbit determination, orbit propagation and conjunction assessment and is therefore highly relevant for operational applications such as space traffic management. At the same time, the density model used in such applications must not only be sufficiently accurate, but also computationally efficient and compatible with near-real-time data availability constraints. This creates a practical trade-off between physical fidelity, operational robustness and computational cost.

This thesis investigates whether a neural network-based correction to the JB2008 empirical thermospheric density model can improve atmospheric density estimation and, more importantly, whether such improvements translate into better orbit propagation performance. Rather than replacing the empirical model entirely, a hybrid correction framework is developed in which JB2008 provides the baseline density estimate and a lightweight feed-forward neural network learns a correction using baseline density, solar, geomagnetic, spatial and temporally embedded input features. This design is motivated by
the desire to preserve the physical and operational utility of the empirical baseline while still allowing the model to capture systematic, time-varying deviations linked to changing space weather conditions.

To support operational relevance, the methodology is explicitly designed with near-real-time use in mind. The input feature set includes solar and geomagnetic drivers, physically motivated temporal embeddings that represent thermospheric memory effects and trust score features that quantify the reliability of outage-prone high-cadence space weather inputs. The model is trained on accelerometer-derived density data from the CHAMP, GOCE and GRACE missions. A key methodological choice is the use of a simple feed-forward architecture rather than a recurrent sequence model. By embedding
the relevant temporal history directly into the input features, the approach remains compatible with fragmented datasets and avoids dependence on continuous sequences, while still allowing delayed thermospheric responses to be represented.

Model performance is evaluated in two complementary stages. In stage A, the neural network is validated directly against accelerometer-derived density estimates. The results show that the correction model reduces the mean and spread of the log-residual error distribution relative to JB2008 and yields
reconstructed density time series that better match the reference densities in both phase and amplitude. These findings indicate that the model learns more than a static offset correction and captures part of the time-varying thermospheric variability that is not fully represented by the empirical baseline.
In stage B, the corrected density model is assessed in orbit propagation against Swarm A precise orbit determination data over seven validation arcs spanning quiet, moderate and severe geomagnetic conditions. The neural network correction model achieves the lowest overall 24-hour along-track RMS error among the tested density configurations and performs best during disturbed periods, when empirical density models tend to degrade most strongly. This suggests that the learned correction is not only beneficial when reconstructing density, but can also improve orbit propagation performance.

Overall, the results of this thesis indicate that a lightweight neural-network correction to JB2008 can provide a promising and operationally relevant improvement to empirical thermospheric density modelling. In particular, the findings suggest that a comparatively simple feed-forward architecture, when
combined with physically informed temporal feature design and explicit data-reliability indicators, is already capable of learning useful corrections to an empirical baseline. At the same time, the final orbit propagation validation is limited to a single satellite and a restricted set of validation arcs. The results
should therefore be interpreted as a promising proof of concept rather than as definitive evidence of broad operational superiority across all low Earth orbit regimes. Broader POD-based validation across additional satellites, altitude regimes and operational scenarios is required before more general conclusions can be drawn. ... Read more

The detection of non-compliant low-thrust maneuvers in Geosynchronous Earth Orbit (GEO) is hindered by the similarity between thrust accelerations and uncertainties in solar radiation pressure (SRP) modelling. In this thesis, a hybrid architecture is proposed that combines a Long Short-Term Memory (LSTM) classifier with a Physics-Informed Neural Network (PINN) formulated as an inverse-thrust recovery solver. A synthetic dataset of 8400 GEO trajectories across three classes, nominal, low-thrust, and mismodeled SRP, is used for training and evaluation. The LSTM classifier achieved over 88% accuracy on a noisy test set, and thrust vectors were recovered by the PINN with a median magnitude error of 0.9% and directional error of 0.6◦. Superior performance over both a TLE sliding-window method and an Unscented Kalman Filter is demonstrated. ... Read more

Space debris is an ever-growing problem, posing increasing risks of collisions between space objects. To accurately predict a collision, the uncertainty of both objects must be propagated to the time of closest approach. This study compares 8 different uncertainty propagation methods for various challenging test cases, to determine the computational efficiency and accuracy of the uncertainty propagations. It is found that the Multi-Fidelity method (MF) is a promising method that scores high on both metrics. MF is then used to propagate the uncertainty of two space objects to the time of closest approach for realistic conjunction scenarios, and compared to a baseline using Monte Carlo samples. It is found that using MF with specific settings, the resulting collision probability remains within a 95% confidence interval, while the computation times are reduced by a factor of up to 10.000. ... Read more

Traditional space cataloguing approaches first combine consecutive measurements into "tracklets" and then associate these tracklets definitively to known objects, before updating the orbits accordingly. In contrast, multi-object tracking (MOT) methods consider multiple measurement association hypotheses simultaneously, but typically discard the pre-formed tracklets and avoid conclusive measurement-to-object assignments. This thesis presents a flexible MOT library based on finite set statistics (FISST) and proposes a modified multi-object filter that fully leverages the FISST framework while incorporating prior knowledge from existing tracklets. Additionally, a robust method is introduced to extract the most probable measurement assignments directly from the filtering recursion, enabling targeted single-object post-processing. The new tracklet filter is demonstrated to effectively discover and maintain state estimates for objects in low Earth orbit and geosynchronous orbit, using sparse optical observations from both ground-based and space-based telescopes with diverse pointing strategies. ... Read more

The Starfixers project addresses the urgent challenge of space debris mitigation in Low Earth Orbit (LEO), where inactive satellites from mega-constellations such as Starlink threaten sustainable space operations. Our mission is to design a cost-effective and environmentally responsible Active Debris Removal (ADR) spacecraft capable of de-orbiting at least ten failed satellites within one year.
The selected removal method employs plume impingement: a novel, contactless technique in which high-momentum gas jets are directed at target satellites to induce controlled trajectory changes. This avoids complex capture mechanisms and enables re-entry without requiring physical contact. The mission architecture ensures each target is individually approached and guided toward a controlled re-entry trajectory, before the "shepherd" spacecraft returns to the initial orbit for the next operation. The complete mission concept involves performing numerous controlled approach manoeuvres with each target debris gradually changing its trajectory using a custom-developed momentum transfer and transfer efficiency simulations until the debris reaches an elliptical orbit with a 381km perigee. At its final orbit the debris passively de-orbits within 7.5 months. The developed strategy is proven to be viable for debris ranging from 250- 500 kg and altitudes of 550-630 km. However, the detailed subsystem design is performed for 10 Starlink v1 satellites at 600 km circular orbit as this scenario was deemed most common for an ADR mission considering the abundance of Starlink. The mission is scheduled for deployment in January 2030 and is designed to stay within a 100 million total budget, covering all subsystem development, testing, and launch operations.
Subsystems have been developed in detail: the propulsion subsystem uses bi-propellant thrusters on a two-axis gimbal for precise plume control; the GNC and ADCS subsystems combine LIDAR, IR cameras, reaction wheels, and IMUs for autonomous attitude control, tracking, and detumbling. Power is provided by solar arrays and lithium-ion batteries to operate in the eclipse. Communication is handled via the ESA Estrack network, and all systems are designed for modularity and sustainability. During the final weeks of the DSE exercise, the team will finalise all subsystem designs, refine the plume-based and orbital targeting control algorithms, and integrate all components into a fully verified and validated spacecraft configuration.
... Read more

Space debris is becoming an ever increasing problem in space operations. Mitigation techniques exist, but require knowledge about debris objects in order to track them and predict where they will be in the future. One technique that is commonly used to estimate this information is called light curve inversion. The apparent brightness of an object passing overhead is measured from the ground, and information like orbit and attitude states can be estimated and object shapes can be characterized. This thesis focused on improving existing techniques for state estimation and shape characterisation for LEO objects. The state-of-the-art method, Multiple-Model Adaptive Estimation (MMAE), was implemented, tested and improved upon. Additionally, Variable-Structure Multiple-Model methods were implemented and tested. Testing involved simulating measurements for LEO satellites and running estimations. The improved MMAE algorithm was able to correctly identify the shapes of test objects for almost all test cases, with low attitude estimation errors. ... Read more

Nowadays, space-based systems for navigation, communication, Earth observation, meteorology and many other applications are indispensable for services critical for society. In order to preserve these vital applications among the increased risks in space, the Space Surveillance and Tracking segment of the Space Situational Awareness program has been set up in the last decade by the European Space Agency. This segment is occupied with one core activity: maintaining and building up a space object catalog. The object states within it can be used for conjunction assessment, collision avoidance and fragmentation event detection, allowing active satellites to be protected.

For objects in Low Earth Orbit, the Space Surveillance and Tracking segment primarily employs radar, which uses active signals, whose signal strength degrades with the fourth power of the distance. Beyond this orbit regime, passive optical sensors are primarily used, because they do not require an active signal and are therefore able to observe beyond Low Earth Orbit without large power requirements. As a consequence of the sparse optical sensors which have to track the large population of space objects, the optical observing arcs are typically very short: minutes at most. These short optical arcs are called tracklets.

Determining an orbit from a short arc is a challenging problem and is commonly referred to as the too short arc problem. Specifically, this problem entails that a single short arc tracklet is insufficient to construct an orbit with realistic state estimates. In order to obtain accurate state estimates, the object corresponding to the tracklet must be re-observed during a later pass to obtain six independent parameters of high quality. The initial problem to solve, however, is to identify which tracklets are correlated and thus associated with the same object. Thus, tracklet correlation is a critical step in the cataloging of objects. Evaluating and improving the performance of tracklet correlation methods is the key subject of this thesis. The performance is assessed on three key metrics: the association accuracy, the computational efficiency and the robustness.

In the first part of this thesis, tracklet correlation methods are evaluated. The most promising ones have been identified to be the Adapted Gooding method, the BVP optimization and IVP optimization. The performance of these methods is evaluated for different data distribution characteristics and orbit regimes, first without incorporating perturbations in the tracklet correlation methods.

In the second part, the tracklet correlation methods are modified to improve the performance. A major gap in the state of the art has been successfully addressed, namely to incorporate perturbations in the tracklet correlation methods. Results have shown a significant increase in association accuracy. Specifically, tracklet pairs with arbitrary time gap can be associated with the same true positive rate. Moreover, tracklet pairs from different orbit regimes can be associated with a similar true positive rate if the most important perturbations are incorporated.

Finally, the BVP and IVP optimization are found to be robust to noise, while the Adapted Gooding method severely suffers from the perfect measurement assumption. The performance of the latter has been improved by improving the accuracy of the angles through a Least Squares fit. Although the BVP and IVP optimization are robust to noise, the performance is limited by the accuracy of the angle rates, according to the Chi-squared distribution. A novel tracklet correlation method called the BVP-DC2-DC3 is developed to overcome this limitation, achieving an excellent TPR of 100% and a TNR of 99% for the analyzed objects in GEO for tracklets with a measurement noise of one arcsecond. This accuracy is superior to literature. Moreover, the computational efficiency is superior or comparable to literature. Additionally, the BVP-DC2-DC3 has a high robustness, because for the analyzed GEO tracklets with a measurement noise of five arcseconds, a TPR of 100% and a TNR of 95% can still be achieved. ... Read more

The growing space debris environment poses a significant threat to operational satellites and the future of spaceflight. To assess the collision risk between space objects, often probabilistic methods are used, which suffer from the fact that very uncertain states will lead to a low probability of collision, described as the dilution effect. Apart from this drawback, satellite operators are required to decide whether to mitigate a risk well before the time of closest approach. However, the time horizon available for this is often too short. This research aimed at studying whether novel risk assessment metrics could increase the reliability of risk quantification and extend the time horizon available for decision making. Outer probability measures have proven to be a reliable method for mitigating the dilution effect, whereas a geometry-based metric has proven more challenging to use. Together with the theoretical analysis of the metrics, their operational application has been studied. ... Read more

The modern day growth in satellites and debris in orbit around Earth is resulting in ever increasing requirements for state estimation systems used to avoid collisions and ensure safe operations as part of space situational awareness initiatives. While there are many methods to perform state estimation, the optimal control based estimator (OCBE) provides unique benefits, such as the ability to perform in the presence of mis-modelled accelerations while allowing for reconstruction of unknown maneuvers from the target spacecraft.
However, the method suffers from high computational costs. This thesis covers the implementation of a new variant of the OCBE, focused on propagation and estimation of state in terms of regularized "EDromo" elements.
The method showcases benefits in terms of estimated state error, and reduced sensitivity to the time gaps between measurements, at the cost of further computational complexity. For future work, suggestions to improve efficiency of the method are also presented. ... Read more

Cislunar space is emerging as a critical regime for future space missions. However, dedicated Space Situational Awareness (SSA) capabilities beyond Earth orbit remain limited. This thesis develops a simulation-based framework to design and assess space-based cislunar observers together with their sensor tasking strategies. Using the Earth-Moon Circular Restricted Three-Body Problem, representative catalogs of target and observer orbits are modeled, and an angles-only optical sensor with realistic exclusion constraints is simulated. Target states are estimated with an Extended Kalman Filter, while greedy schedulers based on information gain (IG), age-of-information, and finite-time Lyapunov exponent rewards are compared. The analysis shows that cislunar observers, particularly L2 halo orbits with IG tasking, dramatically outperform Earth-based sensors and reveal strong couplings between observational geometry and estimation accuracy. Age-of-information yields a simple, robust baseline, whereas the FTLE-based reward performs poorly in this formulation. ... Read more

The growing population of objects in Low Earth Orbit (LEO) presents challenges for space situational awareness and catalogue maintenance. Ground-based optical surveillance strategies for LEO catalogue generation were evaluated, comparing fixed-pattern scanning methods against a two-phase Probabilistic Admissible Region (PAR) approach using SPOOK simulations and Airbus Robotic Telescope validation.

For LEO targets, baseline scanning achieved 0.0116% detection rates and zero redetections, while PAR achieved 23.71% redetection rates across 194 attempts, enabling initial orbit determination. Medium Earth Orbit (MEO) validation revealed regime-dependence. Baseline methods achieved 49-62% MEO redetection versus 0% LEO, while PAR achieved 91.94% MEO versus 23.71% LEO.

Results establish that redetection effectiveness depends on orbital regime. For LEO, where rapid motion prevents fixed-pattern redetections, PAR-based approaches provide redetections for catalogue generation. For MEO, both strategies succeed though PAR maintains superior performance. ... Read more

This research aims to advance the field of Space Situational Awareness (SSA) by optimizing multi-sensor multi-target tracking (MSMTT) algorithms within the Random Finite Set (RFS) framework. The project focuses on addressing the computational challenges posed by the increasing number of Resident Space Objects (RSOs) through the development of an efficient, scalable estimator capable of handling dense object environments and ambiguous data. Key components of this research include a detailed evaluation of existing RFS methods showing promise in the field of SSA, such as Labelled Multi-Bernoulli methods (LMB). By integrating and enhancing these techniques, the project aims to improve tracking accuracy and computational efficiency by implementing the filters using C++ source code, thereby supporting both current space operations and future mission planning. ... Read more

The rapid growth in the population of objects orbiting the Earth has led to increased congestion and collision risk. Solar-sail missions have been proposed as a means of debris removal by harnessing the perpetual force from sunlight to perform maneuvers; however, their capability to avoid collisions under the combined effects of solar radiation pressure and atmospheric drag remains to be investigated.

To address this gap, a framework was developed to simulate conjunctions between a sail and debris using representative uncertainties to compute collision risk. Analytical and numerical locally-optimal control laws were applied to steer the sail away from conjunctions and minimize maneuver durations while safely reducing the collision risk. The results revealed patterns in the applicability of specific control laws, with maneuver durations ranging from minutes to hours and showing strong dependence on orbital, physical, and conjunction parameters. ... Read more

This thesis investigates methods to enhance the robustness of space object cataloguing pipelines, focusing on tracklet correlation and orbit estimation using angular measurements from short observation arcs. The cataloguing robustness is defined as achieving high true positive and negative rates for tracklet correlation to allow for the build-up of an accurate object catalogue. The study addresses the main research question: How can the robustness of the cataloguing pipeline be improved when applying orbit estimation methods to the full angle set of short observation arcs?

A baseline tracklet correlation approach, based on the Boundary Value Problem (BVP) within the Admissible Region framework, is implemented. This method uses angular observations and hypothesized ranges to estimate an object's state, with correlations evaluated via a cost function based on the Mahalanobis distance. Classical IOD methods are employed to investigate their application toward validation of tracklet correlation when reconsidering the full angle set. The considered methods include the angles-only Gauss method, a multiple angles least-squares Gauss approach, Gooding’s method, as well as a Batch Least Squares (BLS) orbit determination (OD) method. The BVP and IOD methods consider two-body dynamics, and the BLS Earth's zonal harmonics and third body effects from the Sun and Moon. Simulated measurements are derived from Two Line Element sets (TLE) for initial reference states for LEO, MEO, and GEO objects, propagated with the SGP4 model accounting for Earth's atmospheric drag, zonal harmonics and third body Sun and Moon effects, providing the test data.

Results show that the BVP method performs best for GEO, achieving ~90% true positive rates with reasonable uncertainty gating. For LEO and MEO, higher thresholds and cost-function minima are required due to greater observation complexity and force-model discrepancy. Gooding’s method, making use of a Lambert solver, demonstrated robust performance across multiple orbital revolutions, while Gauss’ methods were less effective for large time gaps. Additionally, BLS struggled with sparse data and large time steps, offering limited state refinement despite higher computational expense.

The findings suggest gating based on chi-squared distribution thresholds for GEO and higher magnitudes for LEO and MEO to optimize true negative rates. While the BVP method provides sufficient accuracy for re-observation scenarios, classical IOD methods and BLS exhibit limitations under sparse tracklet conditions. This work highlights challenges in cataloguing lower-altitude objects, for ground-based optical observations, and suggests the application of the BVP method on lower altitudes requires inclusion of force models for the primary perturbations.
... Read more

This thesis addresses the issue of space debris tracking through the development and evaluation of advanced radar-based Initial Orbit Determination (IOD) algorithms and optimised measurement strategies. Space debris presents a challenge to space missions, necessitating tracking methods that do not require cooperation from the objects being tracked. The batch least squares estimator (BLSE), extended Kalman filter (EKF), unscented Kalman filter (UKF), and unscented batch estimator (UBE) were evaluated using both simulated and real radar data from the Sentinel-1A satellite. The UBE algorithm showed the lowest root-mean-square error (RMSE) and fastest convergence, effectively managing the non-linear complexities of orbit determination.
Four radar measurement strategies were examined to determine the minimum number of observations required for accurate orbit prediction and reacquisition. Strategies that provided full coverage and focused on the closest point of approach (CPA) demonstrated the most reliable results such as the inbound-CPA-outbound method and the changing radar update time. The study also investigated the impact of excluding specific radar measurements, such as range, radial velocity, azimuth, or elevation, on the accuracy of orbit estimation. It was found that a combination of range and angle measurements provided the most accurate results. ... Read more

The congestion of orbits around the Earth, particularly in Low Earth Orbit (LEO), necessitates the implementation of space-based Space Situational Awareness (SSA) missions. This increasing orbital congestion, driven by the proliferation of mega constellations, poses significant challenges to satellite operations and collision avoidance. In response, commercial entities are beginning to plan and launch space-based SSA missions. Despite this emerging interest, there is a notable gap in the literature regarding the scheduling of operations for these missions, which are characterised by the inclusion of multiple operational modes.

Additionally, existing literature on operational scheduling often lacks high-fidelity modelling of onboard resource dynamics. This research addresses this gap by focusing on energy-constrained scenarios, developing novel operational scheduling frameworks suitable for assigning data collection and data downlink tasks for optical space-based SSA missions under such constraints.

To enhance the fidelity of the scheduling process, this research employs both low-fidelity and high-fidelity simulators to model the dynamic behaviour of onboard resources. The developed frameworks are tested in various scenarios to validate their practicality and effectiveness. A significant contribution of this research is the open-source availability of part of the developed scheduling frameworks, utilising software accessible under the TU Delft license. This open-source approach ensures that the research can be extended and built upon by future researchers, fostering ongoing advancements in the field of space-based SSA operations. ... Read more