H∞ robust control is a powerful tool to design controllers robust against disturbances and model uncertainties. However, when a fault occurs in the system, H∞ controllers typically have reduced adaptability to the fault, possibly requiring costly system identification steps to adapt the controller to the unknown faulty model. In this research, integral reinforcement learning (IRL) is applied to the H∞ tracking control problem to adapt a controller online without any knowledge of system dynamics, where only system trajectory data is available to the learning algorithm. Both off-policy and on-policy IRL methods are applied to a pitch-rate tracker for the linearised F-16 short-period dynamics after a fault of 50% reduction in elevator actuator effectiveness has occurred. The adapted controller converges to the model-based controller for the faulty system within 15 s, and its tracking performance is compared to that of the controller for the nominal system.

The Generic Hypersonic Aerodynamics Model Example (GHAME) provides a practical benchmark for evaluating advanced control strategies for hypersonic vehicles. Its nonlinear dynamics and strong aero-propulsive coupling make it suitable for assessing nonlinear control methods. Although GHAME is the only publicly available hypersonic model derived from real flight data, Nonlinear Dynamic Inversion (NDI) has not previously been applied to it. This study develops a hierarchical control architecture based on time scale separation, combining NDI for attitude and position control with Incremental Nonlinear Dynamic Inversion (INDI) for angular-rate and velocity control. The implementation is carried out in MATLAB and Simulink, and the controller is tested under model uncertainty, atmospheric disturbances, and additionally tested under synchronized and desynchronized measurement time delays. The controller achieves accurate tracking when no delay is present. Under desynchronized delays, performance degrades rapidly and similarly in both axes. Under synchronized delays, degradation is more gradual: the lateral dynamics lose stability at roughly half of the allowable delay margin, whereas the longitudinal dynamics remain stable until the full margin is reached. This behavior reflects the very short natural time constants of the lateral subsystem in slender hypersonic configurations, which reduce the effectiveness of the assumed time scale separation. Overall, the study provides the first NDI application to the GHAME model and demonstrates that an NDI–INDI architecture can remain effective under uncertainty, disturbances, and realistic time delays when synchronization is maintained, although its application must be approached with caution due to the heightened sensitivity of the lateral dynamics. The findings naturally pave the way for future investigations, including flight envelope protection strategies, and integration with guidance and trajectory optimization methods for hypersonic missions.

As aircraft propulsion is transitioning to hydrogen fuel-cells, ensuring safe and reliable operation remains a key challenge. This thesis develops a control methodology to automate the startup and shutdown of a multi-stack hydrogen fuel-cell propulsion system for aircraft, whilst minimizing stack degradation. The work was conducted in collaboration with DLR, with contributions from Airbus, ZAL and HSU, focusing on air-cooled, open-cathode PEM fuel cells for aerospace applications.

Control-oriented models were developed using Multi-level Flow Modelling, Finite State Machine, and subsystem physical models, which were validated against experimental data in Simulink. Based on these models, subsystem controllers were designed to operate under the supervisory Finite State Machine automatic controller. The simulation results verify the defined control safety and reliability requirements, demonstrating that the multi-level control methodology is robust in automating the startup and shutdown operations, highlighting the use case in future aircraft fuel-cell propulsion systems.

Incremental Nonlinear Dynamic Inversion (INDI) is effective at controlling nonlinear aircraft dynamics by inversion of the control effectiveness matrix, but relies on accurate and time-synchronized measurements of the control input and the state derivative. Hybrid INDI address this limitation by blending sensor measurements with model-based estimates, most commonly through Complementary Filtering. In this work, an Extended Kalman Filter (EKF) is proposed as an alternative angular acceleration estimation method within a Hybrid INDI attitude control framework. A real-time EKF is developed and integrated into the control architecture of a nonlinear aircraft model, and its performance is compared against a baseline Complementary Filter approach.
The parameters of both estimators are tuned using a robust multi-model optimization procedure that accounts for sensor noise, delay, and model-plant uncertainties. Simulation results under nominal and degraded conditions demonstrate that the EKF-based Hybrid INDI approach provides estimation performance comparable to that of the Complementary Filter, with improvements in scenarios of combined sensor noise and delay degradation.

Launch vehicles traditionally rely on gain-scheduled PID controllers, which are time-consuming to develop, and require extensive modelling. Incremental Nonlinear Dynamic Inversion (INDI) may serve as an alternative to reduce model dependency, and streamline launch vehicle control system design and tuning. This work applies a cascaded INDI controller, with an inner attitude loop and an outer position loop, to the model of a small-scale launch vehicle testbed. Such testbeds are used in the context of Guidance, Navigation & Control (GNC) to develop and test algorithms on a physical system in a cost-effective and low-risk fashion. Wind disturbance, propellant sloshing, and tail-wags-dog, i.e. the effect of the engine’s motion on the rest of the vehicle, are considered in the model. The results show that INDI provides a simple way to develop a nonlinear control law that maintains robust position reference tracking in the presence of disturbance and model uncertainty. Finally, it is determined that a small-scale testbed is not suitable for studying the impact of sloshing on the closed loop control systems of launch vehicles, since at this scale the sloshing mass is too small and the natural frequency of the sloshing dynamics is too high.

Fault-tolerant flight control remains a major challenge as aircraft systems become increasingly autonomous and must operate under uncertain conditions and potential actuator failures. Reinforcement learning has shown strong potential for learning control policies directly from interaction with the environment, but purely offline-trained agents often lack the ability to adapt once deployed.

This thesis proposes a hybrid reinforcement learning framework that combines offline deep reinforcement learning with online adaptive control. A novel RUN-DSAC-IDHP controller is developed, integrating an uncertainty-aware offline policy learned with RUN-DSAC with an online Incremental Dual Heuristic Programming (IDHP) adaptation layer. The offline component provides a high-performance baseline policy, while the IDHP actor continuously adapts the control policy online when the aircraft dynamics change.

The proposed approach demonstrates how combining offline deep reinforcement learning with online adaptive control enables controllers to maintain strong baseline performance while adapting in real time to faults and changing flight conditions.

Sustainability is a key commitment for future innovation and improvement of the aerospace industry with active research invested towards advanced and new aircraft designs such as the FLYING V, which is a flying wing design for commercial aviation, promising higher efficiency against conventional tube-and-wing aircraft. The Flight Control System (FCS) must be designed to prove the airworthiness of the aircraft. In this work, a fault-tolerant FCS is designed with adaptive Incremental Dynamic Inversion (INDI) rate control law and outer loop with longitudinal, lateral and directional control laws. The study evaluates the fault tolerance of the adaptive INDI with a structural fault case that results in a loss of effectiveness, and it is shown that the adaptation allows the aircraft to cope with the faults and maintain satisfactory tracking performance at different flight conditions.

Interest in oceanic climate change and to better understand the oceanic dynamics, a key interest in oceanic topography pushes for cheaper and smaller EO satellites which can achieve similar resolution and measurement accuracy. Alticube+ lies at the forefront of a new frontier, aggregated system of CubeSats connected by booms to fulfil the scientific objective of providing accurate ocean height measurements on par with large monolithic satellites. This configuration introduces strong coupling between attitude dynamics, flexible structural modes, and reaction wheel jitter, posing a challenge for attitude controller design.

This research aims to design a centralised LQR attitude controller that enables effective utilisation of reactions wheel to satisfy scientific motivated pointing control and knowledge requirements. The second objective is to research how reaction wheel jitter interacts with the large aggregated structure to degrade the pointing control, affecting the measurement accuracy. To achieve this, a comprehensive simulation framework was developed in MATLAB/Simulink, capable of modelling both rigid-body and flexible spacecraft dynamics within a unified environment. A centralised Linear Quadratic Regulator (LQR) combined with a Kalman filter was designed to address multi-axis attitude regulation and pointing knowledge requirements under realistic actuator and sensor constraints.

The flexible spacecraft model was formulated using Kane’s equations and a lumped-parameter representation of the dominant structural modes. Reaction wheel jitter was modelled via static imbalance effects, enabling realistic excitation of flexible modes. The simulation framework was verified through analytical comparisons for the rigid-body case and validated for the flexible model by comparison with a finite element model, including modal frequency alignment and structural response under controlled torque excitation.

Simulation results demonstrate that the centralised LQR controller achieves stable attitude convergence within the allocated \(450\,\mathrm{s}\) operation window for the majority of initial conditions. Monte Carlo analysis shows that approximately \(80\%\) of the simulated cases satisfy the absolute pointing error requirement of \(0.2\,\mathrm{deg}\), with performance primarily limited by reaction wheel saturation. Flexible dynamics introduce oscillations, particularly in the pitch axis, where reaction wheel jitter and inertia uncertainty result in pointing errors up to \(0.3\,\mathrm{deg}\) under worst-case conditions. Despite this, internal antenna misalignment and pointing knowledge errors remain within mission requirements for nominal operating conditions.

The results indicate that centralised LQR-based control remains a viable and effective solution for flexible assembled CubeSat systems such as Alticube+, provided that actuator saturation, structural flexibility, and uncertainty effects are explicitly accounted for during design. This work provides insight into the achievable pointing performance envelope of Alticube+'s architecture.

Driven by the need to enhance safety, improve efficiency and address labour shortages, autonomous vessel operations are increasingly moving beyond open-water navigation toward more complex missions demanding integration of multiple control strategies. Dock-to-dock operations represent a critical mission encompassing the full spectrum of motion control challenges: long-distance transit, switching between control phases, and precise low-speed maneuvering in confined waters under environmental disturbances.

This thesis evaluates control strategies for autonomous dock-to-dock sailing on inland waterways. A benchmark system using industry standard PID controllers is compared to an all-in-one Reinforcement Learning (RL) controller and a third, hybrid system is proposed trading off the improved performance of the RL controller with the inherent stability guarantees of the benchmark system. Simulation results show all three controllers can successfully perform the mission. The RL controller docks significantly faster while rejecting higher lateral wind forces but struggles to generalise to unseen docking scenarios, while the hybrid system improves interpretability at the cost of performance. Furthermore, initial real-life testing of the benchmark system validates the simulation results.

This thesis presents a simulation of a Martian lander using Reinforcement Learning. The objective is to train an agent to land on Mars using the Proximal Policy Optimization (PPO) method. The spacecraft is controlled by control allocation, where the translations and rotations are controlled independently. Also a PD controller is made to control the same lander.

The PD controller is found to be more accurate in comparison with the Reinforcement learning controller. A reinforcement learning thruster model is also made. This spacecraft is controlled by five thrusters and the motions coupled. This lander needs more control effort, and is less accurate in tracking a reference velocity. It is also less accurate with landing on the designated landing spot.

Advancements in deep reinforcement learning (RL) open the door to the development of robust flight control systems (FCS) that have the potential to improve both safety and performance during off-nominal flight conditions. Simulation-based work on offline-RL FCS has already demonstrated robustness to adverse weather conditions, mechanical failures, and a wide range of operational conditions. However, it has neglected important dynamical phenomena that limit its applicability to reality. In anticipation of a future flight testing campaign of similar RL-based FCS, this research emulates the transition from simulation to reality by modelling prevalent sensor and actuator dynamics, and introduces a method to incorporate a long short-term memory (LSTM) artificial neural network (ANN) into the policy of a Soft Actor-Critic (SAC) agent. The approach is found to largely diminish the sensitivity of the controller to sensor noise and actuator dynamics, while increasing its robustness to delays in comparison with the ubiquitous feed forward deep neural network (DNN) and a traditional linear controller.

Sample efficiency is a critical metric in intelligent control systems as it directly influences the feasibility and effectiveness of learning-based approaches. This paper presents the study of how Randomized Ensemble Double Q-Learning (REDQ), a sample-efficient model free algorithm, can be used in flight control applications. Three controllers were developed for: pitch, roll and combined biaxial attitude tracking tasks and tested on a high fidelity Cessna Citation 550 model. For each control task, three agents were trained offline: two using REDQ enhanced Soft Actor Critic (SAC) architectures and one using a standard SAC architecture for comparison. REDQ agents showed statistically significant improvements in sample efficiency during initial learning. Average accurate tracking convergence (error < 1◦) occurred within 5,500 training steps for pitch, 6,400 for roll and 11,500 for biaxial control. The gains in sample efficiency were shown to have drawbacks in learning stability and robustness when deviated too far from nominal conditions.

This study explored the application of particle subswarm optimisation - a non-gradient-based swarming intelligence method - in neuroevolution for solving the powered descent landing problem. It was compared to a gradient-based Soft Actor-Critic (SAC) reinforcement learning algorithm under sparse and non-sparse reward settings. A simulation environment was developed incorporating aerodynamic modelling, grid fins, thrust vector control, and variable inertia. Results show that a Soft Actor Critic struggled with non-sparse rewards due to its inability to generalise across the long-horizon powered descent task due to non-informative gradients from high-variance state-action values. In contrast, learning with the sparse rewards became trapped in a local optimum due to a weak reward signal. Particle subswarm optimisation achieved rapid convergence to feasible solutions and reduced propellant consumption, outperforming SAC due to an episodic gated reward structure, local minima avoidance, and gradient independence. This study establishes particle subswarm optimisation neuroevolution as a viable alternative to reinforcement learning for the assignment of throttle during powered descent of a launch vehicle's first stage.

Demand for fully autonomous VTOL aircraft with high agility has exposed the shortcomings of Euler‐angle models: limiting factors such as the presence of singularities and the need for expensive trigonometric operations impact the reliability and safety of autonomous systems and impair the execution of highly-aggressive manoeuvres. Quaternions offer a singularity‐free alternative, yet their use on full‐scale helicopters remains underexplored. Addressing this gap in research, this paper introduces a methodology for converting existing Euler‐based linear models of a Bo105 helicopter to quaternion form, preserving model fidelity. A hierarchical flight control system is then designed, controlling attitude with a novel implementation of Linear Quadratic Integral (LQI) control with quaternions, and commanding velocity and position using PI and P controllers. The controller is tuned using a structured approach implementing the Particle Swarm Optimization (PSO) algorithm. The system is then augmented with a Finite State Machine (FSM) autopilot to follow offline-generated trajectories. The implemented system was evaluated by simulating a slalom and a pop-up manoeuvre: the simulations demonstrated maximum tracking errors of 5 m for slalom and 7 m for pop‐up, primarily due to velocity‐loop delays, while achieving zero steady‐state error in the position. These results confirm the effectiveness of quaternion‐based modelling and control for agile, autonomous VTOL operations.

Attitude Control and Momentum Management are inherently coupled in the context of a space station that relies on Momentum Exchange Devices (MEDs) for attitude control. By exploiting environmental torques, attitude deviation can be minimized while limiting the saturation of the MEDs. Existing studies often simplify the aerodynamic torque contribution or omit it from controller design. In this work, a linear mean aerodynamic torque model was identified for a Right Ascension of the Ascending Node (RAAN) corresponding to the maximum aerodynamic torque magnitude, using a least-squares approach. An H∞ controller was designed for disturbance rejection for different space station configurations, where flexible-body effects were considered by reducing controller authority at high frequencies. The main findings of this work indicate that the nominal configuration of the space station lacks gravity gradient control authority for aerodynamic torque compensation, resulting in an increased attitude/momentum trade-off. The control approach was validated for a more asymmetric configuration through time-domain simulations using the derived linear model – perturbed with a representative disturbance profile constructed from data reflecting the station geometry and orbital environment – and a high-fidelity simulator. These simulations showed that the controller achieves stable performance, small attitude deviations and bounded MED momentum. Additionally, simulations were performed by varying the RAANs around the design point to identify the controller's validity region.

Autonomous robotic exploration must balance fast mapgrowth with robustness under uncertainty. Among frontier-based methods, information-theoretic variants (e.g., Shannon/Rényi) are fast and reliable. This thesis integrates Behavioral Entropy (BE)—which models human-like risk perception—into a planner-in-theloop deep reinforcement learning(DQN) framework that learns the BE risk parameter 𝛼 online.
The policy selects 𝛼 from a discrete action set, uses occlusion-aware visibility to compute BE-based information gain, and delegates motion to a classical A* planner; reward shaping couples coverage/entropy reduction with motion cost. Action-usage analysis reveals an adaptive risk schedule—aggressive early, conservative mid-episode, selectively aggressive late—enabling faster cleanup of residual area. Overall, a reinforcement-learning method with risk-attitude tuning yields a robust, planner-compatible explorer.

Autolanding remains a challenging task due to changing environmental conditions and the sensitivity of conventional control methods to model uncertainties and delays. This work investigates the application of Incremental Nonlinear Dynamic Inversion (INDI) to an autoland scenario using the PH-LAB Cessna Citation aircraft model. INDI reduces the model dependency by utilising sensor-based feedback, thereby enhancing robustness to modelling inaccuracies and external disturbances. Emphasis is placed on addressing variable sensor and actuator delays, which can degrade control accuracy and stability. The proposed INDI controller integrates multiple feedback loops and Pseudo-Control Hedging (PCH), complemented by a
hybrid filter for angular acceleration estimation and a novel altitude estimator. The latter utilises linear accelerations measured by the IMU, which are subject to smaller delays than altitude measurements from the Digital Air Data Computer (DADC). This estimate is fused with the DADC altitude through a Kalman filter for improved accuracy and mitigation of cumulative errors. Simulation results demonstrate that the controller achieves safe landings under external disturbances and variable delays. Sensitivity analyses further show consistent performance across multiple scenarios, while identifying the controller’s operational
limits.

Incremental Dual Heuristic Programming (IDHP) is a successor to the Dual Heuristic Programming (DHP) algorithm that uses an online identified incremental system model, this algorithm showed promising flight control performance and tolerance of faults in simulation experiments. This paper studies the potential for extending IDHP through augmenting the computation of agent updates and returns, more specifically, by using eligibility trace updates and multi-step temporal difference error. This results in the IDHP(𝜆), MIDHP, and MIDHP(𝜆) algorithms, which are compared against IDHP in several simulated flight control scenarios with faults introduced mid-flight. The results demonstrate that the proposed algorithms have improved flight control performance and fault tolerance in terms of tracking errors when controlling a nominal aircraft and an aircraft with faults introduced, with the most improvement observed in MIDHP(𝜆)

Exploring planetary bodies using robot swarms can potentially increase the value of the exploration missions; enabling the execution of novel measurements and explorations previously deemed impractical or unattainable. Despite its potential, the technology readiness level of planetary swarms is not very mature. This work uses multi-agent reinforcement learning to find control policies that allow swarms to autonomously explore unknown areas in a decentralized manner, contributing towards the technology readiness of the field. A multi-agent proximal policy optimization (MAPPO) algorithm is proposed for this end, where the policy uses LIDAR perception information, and the input of the value function contains local and global environment information. The algorithm finds control policies that achieve cooperation behaviors and generalize to unseen swarm sizes and environments learning with simple, sparse reward functions. Moreover, different types of reward functions, value inputs, and environment configurations are investigated. Compared with the state-of-the-art in the field, MAPPO can learn with a larger number of agents, more complicated environments, and using sparse rewards instead of dense ones.