This paper presents a novel data-driven Reference Governor with Model Predictive Control integrating local motion replanning and path following for collision avoidance. Employing a model-free Reference Governor, the proposed solution utilises system knowledge through Bayesian Optimisation to augment predetermined evasive trajectories, minimising path-following errors and simultaneously ensuring obstacle safety margins. A single-track vehicle model in combination with non-linear tyre models is used to capture the vehicle’s dynamics. The optimised control action is the vehicle steering angle, whilst the Reference Governor optimises parameters of a sigmoid reference signal to minimise the tracking error and guarantee safety with respect to obstacles in emergency manoeuvres. The proposed approach is evaluated on a single lane change using a high-fidelity simulation environment and its performance is compared to a baseline controller integrating path following and obstacle avoidance. The results show a 14% reduction of safety critical overshoot, maximising obstacle safety distance and a four times lower controller cycle time compared to the baseline. Furthermore, through a robustness analysis, it is demonstrated that the proposed approach is more robust towards model mismatches and perception-based errors, as seen by average 30% and 40% reductions in near-miss and collision rates.

Autonomous driving technology aims to improve road safety, reduce energy consumption, and increase traffic efficiency. Developing a full-scale autonomous vehicle is costly and subject to strict regulations, making model-scale autonomous vehicle platforms more accessible alternatives for research. In recent years, an autonomous vehicle platform has been developed as a testbed for autonomous driving at the Delft Center for Systems and Control (DCSC).

However, several challenges are encountered during the development phase due to sensor malfunctions and inaccuracies in vehicle system modeling. The external vision-based localization system (Motion Capture System) generates inaccurate measurements, and the communication between the central controller and the sensors can be unstable, leading to packet loss. Additionally, there is a Radio Control (RC) input layer between the controller and the DCSC vehicles, and the dynamics of the RC inputs are unknown. These challenges have limited the implementation of general control architectures in the system.

To address these challenges, an Outlier-Robust Path-Tracking Framework is proposed in this thesis to perform path-tracking tasks under simulated sensor malfunctions and unknown vehicle dynamics. Within the framework, the motion of the vehicle is simulated using an identified DCSC vehicle model.
A first-order system is employed to model the relationship between the RC commands and the vehicle's rear wheel angular velocity and front wheel steering angle. Adaptive Outlier-Robust Kalman Filter is proposed to accommodate both the observational outliers and the model mismatch between the process model and the simulation model.

The performance of the framework is evaluated using synthetic datasets. Three types of observational disturbances—Gaussian noise, observational outliers, and packet loss—are introduced to the observations. The results show that, compared to the basic Kalman filter, the proposed filter provides more precise state estimation under observational disturbances. Additionally, compared to other non-adaptive outlier-robust Kalman filters, the proposed method can adjust the process noise covariance without requiring an accurate prior for the process noise covariance. The proposed framework demonstrated promising results for the path-tracking task under disturbances. Although the framework is tailored to our platform, the approach can be adapted to other systems with noisy measurements and model discrepancy.

As virtual technology rapidly advances globally, the integration of the sense of touch into virtual environments to enhance realism is becoming increasingly urgent, drawing significant attention from scientists and researchers. This has led to the development of innovative haptic devices designed to replicate tactile sensations, further immersing users in virtual experiences. However, despite significant research in this field, most of the current haptic devices appear to lack the ability to provide direct, flexible, and natural interactions. To address these challenges, this study introduces the FlexCube, a novel flexible multimodal haptic display capable of delivering three tactile sensations: softness, roughness, and temperature. The FlexCube is developed through the creation of two modules—a hydraulic-actuated module for temperature and contact area rendering and a roughness module using the electrovibration effect. Then, they are combined together and integrated with an available stiffness rendering device into one device - the complete FlexCube. Subsequent experiments were conducted to evaluate the performance and behavior of each module individually. The results demonstrate that the FlexCube is capable of rendering a step profile and real texture’s temperature profile, tracking a sinusoid contact area profile of 0.2 Hz with neglectable delay, and able to deliver salient roughness sensations with the maximum generated electrovibration force of 0.028N regardless of the possible surface deformation. Overall, the FlexCube can simultaneously deliver all the expected tactile sensations and appears to be a promising tool for applications in E-commerce, telepresence, and interactive simulations.

The shift to sustainable energy sources has increased demand for Energy Transition Metals such as nickel, copper, cobalt, and manganese. To satisfy this need while reducing the negative social and environmental effects of conventional mining, Deep-sea Nodule Collection (DSNC) appears to be a feasible option. Despite its potential, DSNC encounters difficulties with path planning since it depends on imperfect global paths, which result in inefficiencies and safety issues. This thesis looks into the possibility of using a local path planning algorithm, specifically an Interpolating Curve Planner (ICP).

To assess the ICP's performance under various model hyperparameters and environmental scenarios, a simulation tool is created. The study examines how well the ICP performs in terms of area coverage, collection rate, and reliability along different hyperparameters and presurvey measurement resolutions. It integrates novel production rate objectives, previous track following, and real-time sensor measurements within an environment generated from bathymetric data.

The examined local path planner shows a step towards optimised DSNC. Further research is needed to refine objectives and the structure of the local path planning algorithm to maximize its potential in addressing DSNC challenges effectively.

Connected and Automated Vehicles (CAVs) have the poten- tial to revolutionize transportation systems, but their integration with human-driven vehicles (HDVs) in mixed traffic environments presents significant challenges, particularly in complex scenarios such as on-ramp merging. This paper addresses the challenge of on-ramp merging for CAVs in mixed traffic environments, proposing a novel approach called QMIX-QLambdaM. We formulate the problem as a Centralized Train- ing with Decentralized Execution (CTDE) Cooperative Multi-Agent Re- inforcement Learning (MARL) task, capable of handling dynamic sce- narios with both CAVs and HDVs. QMIX-QLambdaM enhances the QMIX algorithm by incorporating Q(λ) returns for improved value es- timation and an action masking mechanism for safer action selection. Our comprehensive experiments demonstrate that QMIX-QLambdaM consistently outperforms state-of-the-art algorithms, including QMIX, MAA2C, and COMA, across various performance metrics related to traf- fic efficiency and safety. The proposed method exhibits superior adapt- ability across different traffic densities, maintaining high performance in terms of safety, efficiency, and overall rewards. Furthermore, case stud- ies illustrate QMIX-QLambdaM’s ability to generate effective strate- gic control for both main-lane and merging-lane vehicles, showcasing smoother driving behavior and better collision avoidance compared to baseline methods. The learning curve comparison also reveals QMIX- QLambdaM’s advantage in credit assignment compared to other CTDE baselines for the formulated problem. The code are available at https: //github.com/ayton-zhang/MARL_qmix_merging.

This thesis presents a comprehensive approach to integrating a trajectory planner and follower for autonomous vehicles (AVs) using model predictive contouring control (MPCC). The planner generates collision-free trajectories with a kinematic bicycle model, while the follower tracks them using a dynamic bicycle model with a smaller integration step size and higher update frequency. The hierarchical architecture allows for a long planning horizon and a fast control loop. Mismatches between the planner and follower can result in tracking errors and conservative trajectories. To address this, a feedback-hierarchical interface is proposed, feeding back the mismatch error from follower to planner. Obstacles are then inflated with this error, minimizing harmful deviations and collision risks. The paper validates the Local Motion Planner using a simulator with a dynamic bicycle model, testing different follower-solver settings in urban scenarios. The results show a reduction in collision rate of 23\% compared to a single-layer MPCC planner, with similar levels of lateral error and task duration.

In the near future, travelling in vehicles will no longer be in regular vehicles, but in automated vehicles. The share of automated vehicles is predicted to increase significantly within 20 years. Passengers in automated vehicles will engage in non-driving tasks, such as sleeping, reading, working or just otherwise spending time on their phone. This will result in motion sickness becoming more prevalent, as passengers will no longer pay attention to the road. Therefore, there is a need for research in motion sickness. To further our understanding of motion sickness and possible mitigation strategies, mathematical models of motion sickness need to be developed. The temporal dynamics of motion sickness can be captured in the so called 'Oman model' \cite{Oman90}. However, most literature use group averaged parameters and motion sickness incidence to describe motion sickness. These methods do not capture well enough how individuals respond to sickening stimuli, as recent studies showed that individuals have strongly varying responses to various frequencies. Only, estimating individual motion sickness parameters is costly thus far, requiring multiple experiments to estimate the parameters. This study explains an optimal experiment design, where the input is varied in real-time closed loop manner such that the information content in the input is maximized for estimation of parameters, rusulting in the fact that individual motion sickness parameters could be estimated in a single experiment. Results show that on average, within the first 63 minutes, most parameter estimations have converged. The resulting RMSE is 1.06 on the MISC scale, comparing to other literature. This shows that the frequency and temporal dynamics of motion sickness and an individual level can be estimated at a drastically faster rate than previous methods. To our knowledge this is the first use of optimal experiment design techniques to asses the dynamics of human responses to stimuli in general, which is an important milestone for cybernetics research.

Bicycle safety is quickly becoming an increasingly important field as the number of electric bicycles on the streets grows faster each year. E-bikes are able to accelerate quicker and travel at faster speeds than conventional bicycles, increasing the severity of injuries in case of an accident. There are a number of ways to improve safety, such as building better infrastructure, implementing active safety systems, improving bicycling skills, or better protective gear. In this thesis, a controller based on Model Predictive Control has been designed to explore whether it could be used as a haptic guidance system that would improve motor learning of a cycling task, and whether it could assist the cyclist during cycling manoeuvres. The cycling task of lane change manoeuvres performed at a constant forward velocity was investigated. The controller was implemented on a desktop PC, which wirelessly controlled the TU Delft's Steer-by-Wire bicycle -- a bicycle where the steering is enabled by the use of electric motors instead of mechanical coupling between the front fork and the handlebars. To test the controller, ten participants took part in a pilot study. The study was designed following a counterbalanced measures design and the participants were split into two groups that experienced the controller's haptic guidance in different order. During the study, the participants were asked to ride the bicycle on a treadmill and hit virtual targets, shown on a display mounted in front of the treadmill, by carrying out lane change manoeuvres. The hypothesis stated that the controller improves performance while it is assisting the participants. It was found that the controller did not significantly improve immediate performance and this result is likely caused by too low of the task difficulty. However, it is likely that the controller was more effective at improving motor learning of lower skilled participants compared to higher skilled participants, but a small number of lower skilled participants limited the analysis. A short post-hoc no-hands riding test of the same cycling task was carried out to investigate whether the controller is able to carry out lane change manoeuvres with minimal rider input. A significant performance improvement was found during the no-hands test. In conclusion, due to limitations of the study, no performance or motor learning improvement caused by the controller was found. Yet the controller showed promising results in a no-hands riding test, which suggests that the controller could be used as a starting point for advanced safety systems for bicycles.

It is impossible to imagine modern day interaction with technology without the use of touchscreens. It is a go-to interface to use for many applications, because of the high stimuli-response compatibility and adaptability of the graphical user interface. But the haptic feedback one would have with physical buttons and dials, is lost with the use of touchscreens. High potential to improve the interaction with high resolution haptic feedback is often ignored. In this paper, the use of a haptic pseudo-potential field rendering method on a friction modulated touchscreen is proposed. With this method, the user is assisted in moving towards a target by lowering the friction and impeded in moving away by increasing the friction coefficient. In a human factors experiment, this rendering method is compared to a position-based friction modulation method. Subjects are instructed to find a target path, based on the haptic feedback. The results show that the position-based rendering method has a higher hit-rate and lower movement times. This demonstrates that the pseudo-potential field method is difficult to perceive, however it is expected that advancements in rendering larger friction coefficient ranges or even active lateral force feedback will improve this rendering method.

An important durability question is whether military off-road vehicles can cope with the fatigue damage from extreme random, complex, and non-stationary loads for the design life of up to 30 years. An accurate life cycle fatigue damage estimation is essential for military off-road vehicles as the lives of military personnel can depend on a durable vehicle in future use. In addition, minimizing costly data acquisition time leads to a faster validation of a prototype vehicle, which in turn means the vehicle is taken into service faster. Hence, this study aimed to optimize existing time-domain load extrapolation methods and compare them in fatigue damage to frequency-domain models specifically for military off-road vehicles while minimizing data acquisition time and costs. An error below 5% between the fatigue damage estimations and the validation is deemed satisfactory.

Two time-domain and six frequency-domain fatigue damage models were applied on a varying fraction of the training dataset in a case study using off-road data from the Dutch Army. From 270 km of unique off-road strain measurements two subsets were created, i.e. a training subset of 90 km and a validation subset of 180 km. These measurements represent the actual use in all operating conditions, yet still lack the load cycles from extreme events which rarely occur. All eight models were optimized to fractions of the training set and compared to the fatigue damage of the full training set and validation data.

None of the models performed satisfactory( i.e.<5% error) if less than 36 km or 40% of the training dataset is used. Fatigue damage was underestimated up to 30% compared to the full training dataset. Above 36 km or 40% of the training set the error in fatigue damage estimations stabilizes indicating the training data contains sufficient information regarding rare and exteme loads. Yet, only one time-domain and one frequency-domain model achieved an error below 5% between the estimation on the training data and the validation data.

This shows that for a tail heavy off-road load distribution in the time-domain or wideband PSD in the frequency-domain only a selection of models provide accurate results when the measured signal contains sufficient extreme loads. The results give the Dutch Army a fatigue damage estimation method which provides good accuracy whilst minimizing the required data acquisition. However, it is recommended to expand the model to validate the results for other types of military off-road vehicles, a broad range of material parameters, and user profiles.

Full vehicle automation requires complete control over all driving scenarios that can be encountered on roads in order to ensure passenger safety at all times. This extends to dangerous situations such as losing control on slippery surfaces, commonly known as drifting. This work aims to make a step toward ensuring passenger safety in these situations by gaining control over vehicle behavior in high side slip regions. As these regions tend to be highly nonlinear and drifting dynamics happen at a millisecond scale, the time availability for generating proper commands is highly restricted. This work further proposes a method based on Nonlinear Model Predictive Controller (NMPC) that conveniently splits and reformulates the drifting problem to reduce the computational burden of the control structure and make it suitable for real-time implementations. The overall performance and robustness will be evaluated in obstacle-avoidance scenarios and in more general driving situations.

Autonomous driving is attracting growing attention because of the potential advantages it poses on safety, leisure, energy efficiency, reduced emissions, and traffic reduction. Current research is focusing on areas related to artificial intelligence to solve complex planning and decision-making tasks, object detection, or simultaneous localization and mapping.

However, training and testing these methods in the real world is rather unsafe and can be harmful to other road users. For that reason, the focus is set on developing such methods first in a simulated and controlled environment and then transferring the control strategy to the target domain. The issue with this approach is that the performance drops when the domain changes, a problem known as sim2real. In addition to this, certain methods cannot be deployed to real-time applications as the computational complexity of the techniques is too large to produce outcomes in a constrained time frame.

To date, research on reducing the sim2real for autonomous driving gap has focused on domain adaptation or domain randomization techniques, however, none has combined them with a high-fidelity vehicle dynamics model. This work aims to validate a safe transfer learning approach that reduces the reality gap in a zero-shot manner by combining the advantages of using simulated and real-world data, a digital twin of the vehicle, and a traffic scenario simulator.

To test this framework reinforcement learning agents are trained to track different paths. Further evaluation includes the influence of using a high-fidelity model and real-world data. Furthermore, the effect of the reward function on the control strategy of the learning agents is also examined. The training process of this transfer learning framework starts with virtually generated scenarios, which are simpler than their real-life counterparts and noise-free. The motion of the learning agent is simulated with the digital twin and in every episode, its parameters are randomized. Moreover, noise is added to the control action. These randomizations are useful for the controller to operate under uncertainty and avoid overfitting to model inaccuracies. Once the performance saturates, real-world logged data is included so that the learning agent adapts to the target distribution, i.e. on noise levels and driving style. After the performance stops improving, the results are tested in Model-in-the-Loop and Vehicle-in-the-Loop with the SimRod, an all-electric drive-by-wire vehicle.

The results of this study indicate that using this zero-shot transfer learning framework yields much better tracking accuracy (about 30% on average) in the physical world with respect to controllers trained only with lower-fidelity models or synthetically generated data. This study stresses that a combination of the three sim2real methods and synthetically generated and real-world data are necessary to reduce the reality gap. Additional results include the influence of the reward function design and using different reinforcement learning algorithms (proximal policy optimization and soft actor-critic).

The pursuit of faster and more precise mechatronic systems necessitates the use of inventive mechanical designs and advanced control schemes. In recent years, flexible elements have seen increased use as they enable the development of high-precision motion systems in a wide variety of applications ranging from semiconductor fabrication to precision surgery. Although these flexures provide ample benefits for systems that require lightweight elements and predictable behavior, structural vibrations can restrict their effectiveness in achieving precise positioning. This motivates the need for vibration suppression in motion systems, particularly those with flexible components. To that end, smart structures with embedded transducers and a suitable control algorithm can be used to actively damp the underlying structure’s resonance modes, although time-varying parameters and uncertainties from various sources can degrade the performance of the vibration controller. In this thesis, an adaptive control scheme is developed to maintain the desired damping performance regardless of system variations. The method employs adaptive notch filters in a cascade arrangement to track multiple modal frequencies of a smart flexible structure, then tunes a positive position feedback controller for multimodal damping with a straightforward adjustment rule. The adaptation is shown to be fast, accurate, and efficient, with clear advantages in suppressing the vibrations of time-varying and uncertain systems. The method also provides key features absent from other adaptive damping implementations, including the ability to effectively estimate modal frequencies using brief transient signals typical of damped structures, as well as signals buried in noise. Finally, the adaptive scheme is validated experimentally with a flexible beam, showcasing its strong potential for practical applications.

2getthere specialises in autonomous people transport through their GRT vehicle, used for transporting people at the airport from the parking to the terminal to ensure the GRT can operate comfortably and safely in a mixed traffic environment. The vehicle needs to plan a smooth and collision-free path. An essential aspect of a safe path is to predict the motion of the surrounding vehicles of the GRT to guarantee a smooth adaption to the constantly changing traffic scenario. The smooth adaption to the changing traffic scenario is essential for the GRT. The GRT cannot brake with the same magnitude as surrounding traffic due to standing people inside the vehicle. The importance of the smooth adaption to the traffic scenario resulted in the following problem statement; is it possible to create a motion prediction model for the surrounding vehicles of the ego vehicle to allow for pro-active velocity planning of the ego vehicle itself. From the literature, motion prediction models can be divided into physics-based motion models, manoeuvre-based motion models, and interaction-aware motion models. Each level has a certain amount of building blocks, with each level being an extra addition of situation awareness to the previous level.
An interaction-aware motion model is the most advanced prediction model taking the infrastructure and interaction between the vehicles in the traffic scene into account. Therefore this principle of motion prediction is chosen for the motion prediction of the surrounding vehicles of the GRT. First, all possible state and route information of the surrounding vehicles are gathered. Secondly, a tree is created based on the vehicles' routes in the traffic situation, with the unique combination of manoeuvres known as a single branch of the scenario tree. Hereafter, the possible conflict areas between the individual manoeuvres are determined based on the desired trajectories for a specific route. The first vehicle passing the conflict area will influence the second vehicle's trajectory for passing the same conflict area, which will influence the third vehicle's trajectory and so on. Each vehicle can avoid a collision by passing in front or behind the previous vehicle(s) at the conflict area. Evaluating all passing possibilities will result in the velocity trajectory with the lowest cost, constrained by comfort. The sum of all costs for each branch of manoeuvres is used as a metric to determine the possible velocity profiles for each vehicle in the current traffic situation, resulting in a better understanding of the traffic scenario and ensuring a comfortable adaption to a changing traffic scenario. This thesis will evaluate several methods found in the literature to built a motion prediction model suitable for the application at 2getthere. The chosen motion model will be evaluated by two scenarios in a T-junction intersection.

Accurate models of driver steering behavior are essential with ever-growing automation in road vehicles. In this project, the effects of driving speed on steering behavior on winding roads are investigated and modeled. Data were collected in a human-in-the-loop curve driving experiment with fifteen participants, who were asked to drive at five different constant speeds between 30 and 70 km/h and at variable speed, using gas and brake pedals. This experiment was performed in the fixed-base simulator of the Human-Machine Laboratory at the Faculty of Aerospace Engineering. The collected data show that drivers do not use a constant portion in time or distance of the road ahead at different speeds. Instead, they put their visual aim point at an approximately constant time away, while they filter a constant distance after this. Moreover, drivers’ response time delays, lead time-constants, and position-to-heading feedback gains decrease with increasing driving speed. In a next step, a speed-dependent linear-parameter-varying (LPV) model was constructed to describe variable speed steering behavior. For some participants and parts of the trajectory, measured steering wheel deflections are modeled more accurately by this model than linear-time-invariant (LTI) estimates. However, simpler LTI models on average explain the measured variable speed steering better. Overall, this research clearly shows that drivers adapt their steering behavior to driving speed and how key steering behavior parameters are adjusted with speed. Modeling this adaptation for realistic driving, however, is challenging due to the still unknown dynamics between speed and parameter adaptation.

Fully automated vehicles have the potential to increase road safety and improve traffic flow by taking the human element out of the driving loop. They can also provide mobility to people who are unable to operate a conventional vehicle. Safe automated vehicles must be able to respond in emergency situations or drive on slippery roads in bad weather conditions. Therefore it is crucial to have a safe and robust control strategy that can use the full handling capabilities of the vehicle.

This thesis presents how safe reinforcement learning can be used to design a steering policy that can drive an automated vehicle at the limit of friction.

The steering policies are trained using the Lyapunov Safe Actor-Critic (LSAC) algorithm. LSAC is a combination of the Soft Actor-Critic (SAC) algorithm and a Lyapunov stability analysis to solve constrained control problems.

The performance of LSAC is tested in a vehicle simulator against SAC and Model Predictive Control (MPC) in a series of tests that include changing lanes at different speeds, recovering from a destabilizing collision, and driving on a race track at the limit of friction.

The experiments show that LSAC outperforms MPC and SAC control strategies in terms of safety and vehicle stability. LSAC can recover from larger disturbances than MPC and SAC. A control strategy is presented that will keep the vehicle stable when driving at the limit of friction but can use the maneuverability of an unstable vehicle when is it necessary to avoid dangerous situations. Additionally, a policy is presented that can find the fastest way around a race track while staying within the track limits.

To promote automation in vehicles, autonomous driving should feel comfortable. To achieve low discomfort, a comfort oriented nonlinear model predictive controller is created. We know humans are sensitive for discomfort in certain frequencies in acceleration. By penalizing the frequencies for discomfort a higher comfort performance can be achieved. Two band pass filters are created to penalise the frequencies. A band pass filter (0.03-0.2 Hz) for the frequency of motion sickness and a band pass filter (1-2 Hz) for general discomfort. Due to the MPC framework the filters can be implemented on the predicted accelerations. The filtered accelerations are penalised within the MPC. The MPC is made for path following control. To test the MPC a reference generator is built. The reference generator creates reference signals about the path ahead for the controller. To test the performance of the filters, tests are done with different controllers. In the different controllers the filters are penalised individually and together and compared against other controllers. The controllers are tested on multiple scenarios (e.g. double lane change and a sinusoidal trajectory). On the scenarios multiple disturbances are tested (e.g. wind disturbance and sensor noise). We conclude that the MPC design that relies on the motion sickness filter has a significant decrease of motion sickness in scenarios where a lot of motion sickness is present, with improvements up to 30.7% compared to the basic controller. The MPC design that relies on the general discomfort filter helps bring the general discomfort down. On the double lane change maneuver with a changing velocity the general discomfort filter has improvements up to 10.7% compared to the basic controller.

An issue in driving simulation is that behaviour displayed in simulation does not exactly replicate behaviour in real-life. For example, roadside vegetation density impacts a driver’s speed and lateral position in on-road studies, but not in driving simulator studies. In this study it is investigated if the increase in fidelity and presence that head mounted display based virtual reality bring forth, yields behavioural adaptation as shown in real-life by evaluating the effect of three different roadside vegetation density conditions in a novel head mounted display based virtual reality driving simulator. Twenty-nine participants drove a 2m wide car over a 11km long 3.5m wide winding road. Participants completed three trials driving through three different roadside vegetation density conditions per trial, the density conditions being light (one tree per 40m), medium (one tree per 20m) and dense (one tree per 5m). The effect of the density conditions was evaluated with respect to speed, lateral position, steering reversals, subjective workload and self-reported risk. Increased vegetation density increased self-reported risk, but did not affect speed or lateral positioning. This finding is congruent with findings in con- ventional simulator studies, which could indicate that despite the advantages of head mounted display based virtual reality regarding fidelity and presence the resulting driving behaviour still has a closer connection to conventional simulated driving than real-life driving. In future work the built and implemented simulator can be improved and optimized, furthermore the relative validity of the simulator should be investigated.

Commercial heavy vehicles are especially prone to rolling over due to their inherent properties, such as the high centre of gravity - track width ratio and compliant chassis frame. Autonomous trucks cannot become widespread without eliminating this danger by guaranteeing rollover-free vehicle motion. Currently existing Roll Stability Control implementations rely on the assumption of having a responsible driver behind the steering wheel -- therefore, unsupervised driving will need a higher level of roll safety than what the currently used methods can provide. Fortunately, there are two main attributes of self-driving vehicles that can be utilized to achieve this goal: Information about the reference path ahead of the truck will be available and the used algorithms may have a full control authority over the available actuators.

This thesis project developed two, redundant rollover mitigation techniques to be run in parallel, for a tractor-trailer combination: The proactive and reactive Roll Stability Control methods. These vehicle motion controllers are separate, independent functional entities.

While the proactive approach attempts to prevent upcoming events, the reactive Roll Stability Control is designed to mitigate imminent rollovers that could not be anticipated based on motion reference information. This controller is placed within the motion control paradigm of Control Allocation. The objective of Control Allocation is to coordinate different actuators to achieve both longitudinal and yaw accelerations as desired by the higher-level tracking controller. Roll stability is achieved by extending the set of functionalities of this framework, using both brakes and steering to realize the needed interventions. Emphasis is put on accurate wheel lift-off detection, using lateral acceleration, steering angle and accurately estimated roll angle signals.

Design choices during syntheses of both controllers were made based on the conclusions of a thorough analysis of roll dynamics, carried out using a high-fidelity vehicle model, provided and validated by Volvo Group Truck Technologies. Subsequently, both controllers were implemented within Volvo's real-time framework. While the proactive method's performance was only assessed using simulations, the reactive controller was tested on Volvo's proving grounds.

The thesis concludes that achieving a higher level of roll stability of (autonomous) heavy vehicles is possible, whilst having a less conservative overall behaviour compared to traditional approaches. This research contributes to making a step towards the next generation of rollover prevention for automated trucks.