Automated driving is poised to transform the transportation landscape of the future, but several challenges remain before full automation is achieved. One of these challenges lies in managing perception uncertainties, such as those arising from radar and sensor measurements, while maintaining control in low tire-road friction conditions. These challenges often occur simultaneously in adverse weather conditions, but are typically researched separately. Their combined effect on safety and performance remains underexplored, even though addressing them together is critical for robust and reliable automated driving systems.

Lower tire-road friction limits the available tire force required for obstacle avoidance. Properly modeling these tire friction limits is particularly important in dynamic and uncertain environments to adequately account for the changing environment responsively. Moreover, addressing perception uncertainties and modeling low-friction conditions individually can significantly increase computational demands which poses challenges to achieve real-time performance required for real-world implementation. Therefore, this thesis jointly considers perception uncertainties and low tire-road friction conditions, accounting for their interacting effects. This is accomplished by addressing the following research question: "How can perception uncertainties be effectively integrated into motion-planning models for obstacle avoidance to improve performance and safety of automated vehicles in various road conditions?"

To address this question, a grid-based stochastic model predictive control framework is extended, implementing a non-linear bicycle model and a Fiala tire model (brush model) to consider realistic vehicle capabilities in low-friction conditions. Grid-based stochastic model predictive control reduces the uncertain obstacle environment into a set of linear constraints, utilizing an occupancy probability grid to effectively consider perception uncertainties. While the developed framework is presented as a proof-of-concept with a focus on safety and feasibility over real-time implementation, computational efficiency is not overlooked. By reformulating the constraints, the uncertainties are effectively accounted for while also reducing the optimization time by simplifying the probabilistic obstacle space to a deterministic convex region. If the nominal reformulation fails, a novel back-up method generates a conservative back-up set of constraints improving the safety and feasibility of the method. Two different back-up strategies are proposed, providing a trade-off between accuracy and computational effort.

To evaluate the contributions, three simulations were conducted. The first comparing the non-linear bicycle model and Fiala tire model to more simplistic models at various tire-road friction coefficients, highlighting the improvement in control of the proposed model in low-friction conditions. The second simulation evaluated the performance of the proposed back-up methods and environment representation by simulating tight scenarios that would fail using only the nominal approach. Feasibility rates increased compared to the baseline back-up method (43.8%) with feasibility rates of 62.5% for the precomputed back-up method and 75.0% for the current-state back-up method. These simulation results demonstrate that the capability of the proposed framework to compute feasible solutions and that the framework is able to compute valid hulls that are safe when the nominal approach fails to reformulate the constraints. The final simulation evaluated the performance of the complete proposed method by simulating both tight scenarios requiring a back-up method as well as various friction levels. These simulations demonstrated that the proposed framework is effective at lower friction levels, achieving high feasibility rates of 87.5% for the current-state back-up method, and 91.6% for the precomputed back-up method. This high accuracy is particularly promising for real-time applications using the precomputed back-up strategy, since this method leverages parallel computation of the back-up constraints.

This thesis demonstrated the effectiveness of the proposed extended grid-based stochastic model predictive control framework in considering perception uncertainties in low-friction road conditions. While presented as a proof-of-concept with an emphasis on feasibility and safety rather than real-time implementation, the proposed framework lays the groundwork for real-time applications, marking a step in solving all edge cases required to reach fully automated vehicles.

This thesis addresses the challenge of controlling automated vehicles performing evasive manoeuvres at the limit of handling. Special attention is paid to the development of nonlinear controllers, which can prioritise obstacle avoidance over path tracking objectives while considering vehicle stability constraints, to improve passenger safety. The thesis develops the entire pipeline for obstacle avoidance controllers, focusing on three aspects: vehicle state estimation, collision avoidance and control beyond the stable handling limits, e.g. drifting.

Model Predictive Control (MPC) is an effective reference tracking strategy for automated vehicle control, particularly useful during emergency evasive maneuvers such as double lane changes. This control method often requires a high-fidelity vehicle model to accurately capture nonlinearities and uncertainties, significantly increasing computational demand. Hybrid systems modeling frameworks have been developed to approximate these nonlinearities, thereby reducing computational complexities while maintaining satisfactory tracking performance. However, existing benchmarks that evaluate the impact of these hybrid approximations on tracking performance and computational demands are lacking. Establishing such comparative benchmarks is crucial for understanding how different levels of model complexity affect the overall efficiency and effectiveness of model predictive approaches in automated driving scenarios.

This research addresses this gap by presenting a comparative analysis of various levels of hybrid model complexity. It assesses their tracking performance and computational demand using both MPC formulation and Model Predictive Contouring Control (MPCC) formalism in different emergency maneuver scenarios.
Four hybrid approximations of the nonlinear single-track vehicle model with varying complexity levels are considered and employed as prediction models in both MPC and MPCC optimization problems. The closed-loop behavior of these control frameworks is simulated in double-lane change maneuver scenarios, evaluating tracking performance scenarios with varying levels of curvature. Additionally, variations in friction and velocity are evaluated in different scenarios to assess controller robustness to model uncertainty.

Results indicate that reducing the complexity of hybrid approximations can decrease computational demand, albeit at the expense of tracking performance. Moreover, MPC formalism offers a more robust approach to tracking performance and provides a feasible solution in a broader range of scenarios than the MPCC framework. By shedding light on the impact of different complexity levels for the hybrid approximation of the nonlinear model and the control optimization problem formulation, this work offers comprehensive guidelines for hybrid MPC applications for automated driving in emergency scenarios.

As part of the global trend towards carbon emission reduction, the road transport sector is motivated to produce solutions that reduce the use of fossil fuels. One of the posed solutions for this purpose is the so-called ‘e-axle’ for freight truck trailers. As a world-leading production company of air suspension systems for truck trailers, VDL Weweler B.V. wishes to provide suspension systems that can be used in combination with e-axles. An e-axle has as much as double the mass compared to a passive axle currently used in truck trailers. As the axle is part of the unsprung mass of the system, the increase will have a significant impact on the dynamics and forces within the air suspension system. Getting a better understanding of these impacts will help formulate the requirements that using an e-axle will pose to the suspension systems produced by VDL Weweler. Of particular interest is the impact on the durability of the trailing arms as a consequence of doubling the unsprung mass. The goal of this research project is therefore to quantify the stresses and the associated durability impacts as a first step towards finding the necessity of design adaptations.

During experiments conducted as part of this research project, an axle was fitted with a 500 kg ballast to reproduce the increased unsprung mass of an e-axle. From data analysis strain gauge data was concluded that the durability of the trailing arm could be reduced with at least 30.8\%. A significant impact to the dynamics of the system was observed in various dynamic indicators of the vehicle. A significant increase in the suspension deflections, which have a direct relation with the load on the trailing arm was observed. Furthermore, a significant reduction of the natural bounce frequency of the axle and higher transmission of axle vibrations to the chassis were found.

A novel kinematic vehicle model was derived to allow the creation of a digital twin model of the RRT durability test. A model validation was conducted to establish the ability of the vehicle and component models to reproduce the RRT. The numerical model proved to strongly reproduce the axle motions and suspension deflections in terms of Power Spectral Density of the dynamic response. The impact of unsprung mass was modelled with high accuracy in relative terms.

Based on the results of this research project, a significant impact to the durability of the trailing arm is expected. Furthermore, a solid start was made in the development of a model that could predict the durability impact of varying system parameters in an air suspension system of a truck trailer.

This research investigates the benefits of using a trajectory tracking controller based on a dynamic model for a four-Mecanum-wheeled vehicle (FMWV) over a kinematic-model-based controller. An FMWV was designed and built, incorporating both hardware and software components. Two Linear Quadratic Regulators (LQRs) based on kinematic and dynamic models were implemented. The dynamic model includes friction estimation, while the kinematic model assumes a no-slip condition. Simulation results indicate that the dynamic model reduces overshoot and improves trajectory tracking in low-friction scenarios compared to the kinematic model. High-friction scenarios show comparable performance for both controllers. Experimental results align with simulations, though some deviations highlight areas for further improvement. Overall, the dynamic-model-based controller demonstrates superior performance in low-friction conditions, reducing translational root mean square error (RMSE) and maximum path deviation (MaxE).

Abstract—This paper proposes a Nonlinear Model Predictive Control (NMPC) application for automated drifting, with exper- imental verification on a standard production vehicle, without hardware modifications. The controller stabilizes the vehicle on a high sideslip angle, which implies an equilibrium condition beyond tyre friction limits. The proposed control strategy shows feasible results that succesfully brings the vehicle towards a high sideslip state, for a variety of drifting scenarios. Simulations show that the control structure is able to sustain an automated drift along a desired path, with maximal lateral path deviation of 1 meter. An experimental implementation on a production vehicle testbench without hardware modifications has shown feasible control of bringing the vehicle into a high sideslip state, for both low- and high μ situations.

Commercially available Lane Keeping Assist systems fail to consider the driver's intentions since they mainly focus on minimising path tracking errors, resulting in conflicts between humans and automation. This often leads to users being unsatisfactory and turning off the assist, as a result diminishing the advantages such as reduced workload and increased road safety. Considering a driver model in the assist helps increase user acceptance. Therefore, we propose a torque-based hybrid controller for a human-centric haptic shared Lane Keeping Assist, pairing a data-driven driver model with a model-based controller to foster the collaboration between the driver and assist. First, the driver's non-linear steering wheel torque behaviour is modelled and predicted using a Bidirectional Long Short-Term Memory network with an accuracy ≥72.4% and a smoothness ≥0.85Nm/s over a 0.4s prediction horizon. Second, a Model Predictive Controller with a linear bicycle and steering model is developed, where it utilises the driver model's predictions as a time-varying reference. We developed three human-centric controllers for comparison and used a state-of-the-art commercial solution as the baseline controller. The experiments were performed in Toyota Motor Europe's fixed-base driving simulator, where 15 participants tested and evaluated the four controllers. The results show a 113.1% increase in collaborative ratio while maintaining a similar path tracking performance compared to the baseline.

In the coming decade, Advanced Driver Assistance Systems (ADAS) will play a crucial role in improving traffic safety within the European Union (EU). Notably, the European Commission mandates that from July 2024, all new vehicles must be equipped with Lane Keeping Assistance (LKA) systems. LKA systems provide force feedback to drivers when vehicles cross lane markings, employing cameras, radar, or LiDAR systems for lane detection.

To optimize LKA performance, this research focuses on understanding the influence of lane marking properties and adverse scenarios on detection ability.

Field Test and Setup:
A field test was conducted at a track in Lelystad, featuring various state-of-the-art lane markings: an old white paint lane marking as a reference, and three new lane markings (two tape types from 3M and one cold spray plast from Triflex). Lane markings were tested under dry and wet conditions.

Data Collection:
Measurements included luminance coefficient (Qd) and retroreflectivity values (Rl for dry and Rw for wet) of lane markings and asphalt, along with contrast ratios. Test runs were conducted during sunset, in complete darkness, and under different scenarios involving oncoming traffic and street lights.

Analysis and Findings:
After analyzing 414 valid runs, key findings were identified:

Lane markings are 3.3 times more likely to be detected in dry versus wet conditions.
Driving towards a light source reduces detection likelihood by 4.5 to 5 times compared to driving away from the light.
Higher wet retroreflectivity (Rw) improves LKA system performance.
Oncoming traffic with main beam headlights reduces detection likelihood significantly.
Bright light sources negatively affect contrast in camera-detected images.
New lane markings outperform old white paint markings by 2.1 to 4.3 times in detection likelihood.

Implications and Recommendations:
The study confirms the importance of lane marking detectability for LKA systems reliant on cameras. Enhancing retroreflection and contrast improves detection ratios, thereby enhancing traffic safety and potentially preventing accidents. Further research is recommended on the impact of oncoming traffic and street lights, including light type and brightness.

For future lane marking developments, testing under adverse conditions with varying light sources and wet surfaces is advised to optimize LKA system performance and ensure safety on the roads.

The speed of the unsprung mass has a direct effect on the comfort of the ride. Too much unsprung mass movement and high speeds can make the ride rough and uncomfortable by sending vibrations and other disturbances to the people inside the car. By getting an accurate estimate of the speed of the unsprung mass, the suspension system can adjust the damping forces to better absorb uneven road surfaces, reducing vibrations and making the ride more comfortable. In active suspension, controllable actuators are installed between the unsprung and sprung mass. However, incorporating sensors into suspension introduces complexity to the system and increases the vehicle’s mass, production, and maintenance costs. Further, there are unmeasurable or complex quantities which make sensing more difficult. The use of virtual sensing offers a useful opportunity for estimating suspension parameters that are not directly measurable or complex to measure. This research aims to reduce the complexity of a system through the development of virtual sensor using parameters that can be easily measured. This developed virtual sensor is then incorporated in a suspension controller to estimate the effect on the ride comfort of a vehicle.

This thesis introduces a novel model predictive controller (MPC) that integrates both torque vectoring and path following into one controller. Due to a need to improve vehicle safety, systems are being developed in order to improve vehicle handling. One system that is able to improve the vehicle handling is torque vectoring (TV). With torque vectoring, the magnitude and the direction of torque can be controlled by either applying the motor or brake torques. Additionally, in order to eliminate human error as a cause of accident, automated drive (AD) vehicles are being developed. A key task for AD vehicles is to perform path following (PF), where the vehicle follows a predetermined reference path generated by path planning.

Beforehand, these tasks were performed by separate controllers, where one controller performed path following and the other controller focused on torque vectoring. The disadvantage of this method is that it leads to sub-optimal results as both controllers have opposing objectives. The TV controller is able to decrease the steering angle in order to improve the vehicle handling, whereas the PF controller could require a higher steering angle in order to follow the path. By integrating both tasks, the novel
controller is able to optimise the control output such that both objectives are achieved.

The use of model predictive control strategies with TV have been studied and its ability to deal with hard constraints, while decreasing the state errors and control input, makes it a suitable choice to use it for TV. When the MPC strategy is compared to more common control strategies it is clear that the MPC TV algorithm provides better results in terms of responsiveness, lateral acceleration and vehicle handling. Furthermore, due to its ability to integrate multiple applications and its ability to handle a greater level of complexity, a nonlinear model predictive control (NMPC) formulation will be used to perform both torque vectoring and path following.

In order to test the new MPC controller, benchmark controllers have been created for comparison. The benchmark controllers are two controllers that are able to perform torque vectoring by generating a corrective yaw moment to follow the yaw rate reference. The torques are then allocated based on the size and direction of the corrective yaw moment. Then, the NMPC controller that performs path following by using both the steering angle and torques as an input will be compared to a controller that is able to perform path following by using the steering angle as an input.

The controllers are compared by using the sine with dwell test and the double lane change manoeuvre. These manoeuvres are used to test the lateral performance, vehicle handling, responsiveness and tracking performance of the vehicles. Key performance indicators (KPI) are used in order to evaluate the results regarding tracking performance and the vehicle handling.

The results show that the NMPC controller has an increase in performance regarding both path following and vehicle handling. When compared to the benchmark torque vectoring controller, the vehicle handling is increased by 5% and the lateral performance is increased by 6 %. Additionally, compared to the path following controller, by adding torque vectoring, the NMPC controller has improved the path
following by 4 %, the vehicle handling has been improved by 5 % and the responsiveness has been improved by 11 %.

This dissertation is dedicated to understanding the potential of improving the motion comfort of automated vehicles and explores multiple options that serve this purpose. Comfort is usually prioritized behind factors such as safety and efficiency but is nevertheless influential to the acceptance of automated vehicles. The goal of enhancing motion comfort overlaps with the need to overcome challenges brought by the motion sickness phenomenon. Motion sickness is found to impact a significant portion of travelers in all types of transport. It tends to develop faster among occupants who are not engaged in the driving task. Its symptoms can cause difficulties for non-driving-related tasks (NDRTs) to be performed effectively by the passengers. Therefore, a part of the research in this dissertation is directed specifically toward mitigating motion sickness in automated vehicles...

This thesis describes the concept of a reconfigurable wheelbase for a car as a way to improve agility at higher velocities. Altering the longitudinal position of each wheel with respect to the centre of gravity, leads to a change in lateral wheel forces. This affects the car's yaw rate, therefore making it possible to influence its agility. Available literature has shown that this subject is still quite weakly investigated, which makes it an interesting topic of research. To investigate this influence of actively reconfiguring the wheelbase, simulations were done using IPG Carmaker together with MATLAB/Simulink, during which a simplified vehicle model of a Toyota Camry was used. A proportional derivative controller combined with control allocation was designed to regulate the reconfigurable wheelbase system. It features a reference generator that returns a desired yaw rate and activation logic which ensures that the controller is only activated in a velocity range of 60 km/h to 110 km/h and when a steering input is given. Furthermore, a benchmark study is designed, using a fuzzy logic controller that is based on a simplified version of one presented by Soltani et al. (2017). This controller uses the same reference generator and activation logic for consistency. The rules and membership functions of the fuzzy logic block are adopted from the mentioned paper. The performance of both these controllers and a baseline vehicle without any control, are evaluated for a step steer manoeuvre, an increasing circle manoeuvre and a double lane change, at different velocities and/or steering amplitudes. Based on the results of these tests, it is shown that both controllers are able to influence the lateral dynamics positively during most of these manoeuvres and are therefore able to improve the vehicle's agility in the velocity range in which the controller is active.

Motion sickness is a common phenomenon, with close to two-thirds of the population experiencing it in their lifetime. With the advent of automated vehicles in the market, it is anticipated to become an even greater problem as the passengers face a lack of predictability of motion and loss of control over the vehicle. This could nullify the host of possible benefits that automated vehicles propose to offer, and therefore affect their acceptance among potential users.

It is well known that the nauseogenicity of imposed motion is dependent on the frequency content of endured accelerations, with low-frequency accelerations being the primary contributor. This thesis presents a motion planning algorithm targeted towards minimization of motion sickness among passengers of automated vehicles, through targeted reduction of low-frequency accelerations. A Deep Reinforcement Learning (DRL) framework was utilised along with the design of a custom environment and a reward function which incorporates a measure of nauseogenicity of the planned trajectories. The frequency shaping effect of the reward function was evaluated by comparing against a DRL agent trained to optimize general motion comfort described by total acceleration energy. It was found that the nauseogenicity was reduced by 9.6% with the proposed DRL agent.

Further, on-road trials were performed with human drivers to establish a benchmark of driving comfort. The performance of the DRL agent was compared to human drivers as well as against an optimization-based motion planner that computationally maximizes the reward function. The DRL planner displayed comparable performance to the human drivers, and was within 10 to 15% of the discomfort levels of the optimization-based planner for a range of travel times. Meanwhile, the DRL planner offered notable improvements in computational efficiency, taking 1-2 ms to generate a sub-optimal trajectory, as opposed to approximately 5 s as taken by the optimization-based planner.

A vehicle that is travelling at high sideslip angles can still be controlled by drifting. Implemented into a vehicle, this phenomenon could lead to increased vehicle safety and performance. Additionally, it could lead to higher acceptance rates of autonomous driving.
In this work a three state vehicle model is used. Using this model, simulations are performed with varying vehicle parameters and tyre models. First, the mathematical descriptions of all used models are stated, after which phase plane representations for the identification of drift equilibria are elaborated on. Next, different steering characteristics and their influence on the drift equilibrium points are considered. These characteristics are achieved by varying the lateral rear tyre stiffness. The effect of
three different types of tyre models on the drift equilibria is discussed. The models that are considered are the linear tyre model, Dugoff tyre model and Magic Formula model.
In addition to this, the effects of the location of the centre of gravity, the vehicle mass and the cornering stiffness are investigated. Finally the outcome of the Dugoff tyre model and the Magic Formula model for construction of drift equilibria is shown and discussed.
Contrary to the Dugoff and Magic Formula model, no drift equilibria were found using the linear tyre model. Using the Dugoff model, drift equilibrium points were found for the understeering and neutral steering vehicle, whereas the oversteering vehicle showed drift equilibrium ranges. Finally, drift equilibria depend on vehicle parameters, showing change in behaviour when the latter are varied.

Driving simulators have been used in the automotive industry for many years now. They have been vastly employed for conducting tests in a safe, reproducible and controlled immersive virtual environment. The ability of the simulator to recreate the in-vehicle experience for the occupant is established through motion cueing algorithms. Such algorithms have consistently been developed with model predictive control (MPC) acting as the main control technique. Currently available MPC based methods either compute the optimal controller online or derive an explicit control law in an offline setting. These approaches limit the applicability of MPC for real-time applications due to online computational expense and offline memory storage issues.

This thesis report presents a solution to deal with issues of offline and online solving through a combined/hybrid approach. For this, explicit MPC is used to provide an initial guess as warm start for the implicit MPC based motion cueing algorithm. From the simulations, it was observed that the presented hybrid approach was able to reduce online computational load by shifting it offline using the explicit MPC. Further, braking constraints and adaptive washout weights were implemented in the hybrid motion cueing algorithm, to improve the specific force tracking performance and reduce any false cue occurrences. Finally, emulator studies were performed to realize driving simulator performance with the hybrid MPC approach. The thesis concludes by showing the improvements made with the developed algorithm and proposes recommendations for future work.

Electric vehicles are a cleaner and more efficient means of transport. However, the sub-energy-optimal acceleration and deceleration inputs of drivers result in speed trajectories that cause superfluous expenditure of the stored electrical energy in battery. Optimising the speed trajectories to minimise the consumption of stored energy is a potential strategy for the efficient operation of electric vehicles. In this thesis, we propose a numerical solution to the eco-driving problem by optimising the speed trajectories via Pontryagin’s Minimum Principle. The solution is robust to the vehicle parameters and the driving conditions, and is used to generate energy-aware driving advice for near-straight line driving manoeuvres. In the end, we test the global optimality of the speed trajectories by convexification of the problem.

Cooperative driving controllers are becoming interesting subjects for future research in automated driving with the increase in connectivity. Using true-scale
autonomous vehicles to properly test new control algorithms can be a challenge
due to three main factors: costs, safety, and space requirements. To overcome
these problems, scaled-down vehicle platforms or so-called Remote Control (RC)
car platforms can be used to test algorithms in a safe environment without the
large costs involved with true-scale platforms. At present, existing RC platforms
are unnecessarily expensive or not able to drive autonomously.

This thesis presents the design of a low-cost 1/12 scale RC platform, consisting
of modified JetRacer Pro AI’s called R2C Car’s. The sensor suite consists of
a 160FoV 8MP camera and an added 2D LiDAR. These are used in a sensor
fusion and filtering framework to detect and locate other R2C Cars. Furthermore, wheel encoders are added to retrieve a velocity estimate of the R2C Car.
To showcase the possibilities of the R2C platform, a longitudinal cooperative
controller is implemented. This controller is based upon a Distributed Model
Predictive Control algorithm, whereby 5Ghz WiFi is used for the communication between R2C Cars. Evaluation of the R2C platform together with the
cooperative controller demonstrates that the platform is suitable for cooperative
driving experiments. Furthermore, information is presented on the performance
of the distributed algorithm.

Over the past few decades vehicles have become more safe than ever. Not only has the chassis design improved substantially, the addition of things such as airbags and seat belts has significantly reduced the fatality rate of passengers involved in traffic accidents. A more recent development are driver assistance systems such as the anti-lock braking system (ABS) and the electronic stability program (ESP). ABS prevents wheel lock during emergency braking and ESP preserves lateral stability. These are active systems designed to help the driver maintain control of the vehicle and prevent dangerous situations from occurring.

One type of accident that has not been addressed enough by these driver assistance systems is the rollover accident. Though a relatively infrequent occurrence, rollover accidents are extremely dangerous. According to statistics from the NHTSA the passenger fatality rate for rollover accidents is almost fourteen times as high compared to the overall accident fatality rate. In fact, despite its infrequent occurrence, rollover accidents account for 33\% of all passenger vehicle fatalities. This highlights the need for a rollover prevention system in modern vehicles.

The aim of this thesis is to design an accurate yet computationally light rollover prevention controller. Previous research shows that measuring or estimating rollover state is incredibly difficult. Several methods exist, each with their own advantages and limitations. In this work, the novel Predictive Load Transfer Ratio (PLTR) is used to estimate rollover state. This rollover detection method has good accuracy and does not require special measuring equipment. Moreover, the PLTR has slight predictive properties as well. This is then coupled with a model predictive controller (MPC) in order to actively prevent rollover, as opposed to a more reactive feedback controller. The controller is designed using differential equations derived from a two-track vehicle model in order to keep it as computationally light as possible. A benchmark controller is also designed using the same MPC framework with another rollover detection method.

The controllers are then tested in a virtual environment, with the test vehicle performing the fishhook maneuver which induces rollover in the vehicle. Analysis from the results of the simulation show that the proposed rollover prevention controller has better rollover prevention properties compared to the benchmark controller while being more efficient as well. However, this comes at the cost of reference yaw rate tracking. Based on these results, future work can be done to merge this controller with an ESP system where it only activates in critical rollover scenarios.

Accurate and robust vehicle state estimation is important for proper operation of vehicle control systems. For lateral stability control accurate estimation of the Vehicle Sideslip Angle (VSA) is of utmost importance. This thesis aims to develop an accurate and highly robust model to estimate the VSA. Common vehicle sensors such as the Inertial Measurement Unit (IMU), Global Positioning System (GPS) and steering angle sensor are used to provide measurements of vehicle states that are relatively easy to directly measure. Additionally, wheel bearing strains measurements are collected by the Load-­Sensing Bearing (LSB), which can be mapped to the tire forces that are measured by wheel force transducers. Insight regarding the tire forces could be beneficial to VSA estimation. The focus lies upon neural network estimators and hybrid structures, which combine the neural network estimator with an observer model.
A large-­scale experimental dataset composed of standardised vehicle manoeuvres is used to develop and evaluate different estimation architectures. The dataset consists of 216 manoeuvres, which corresponds to approximately two hours of driving time.

Neural network estimators are data-­driven and therefore rely on high quality data. Various neural network architectures exist for the purpose of time series estimation. In this thesis the Feedforward Neural Network (FFNN), Recurrent Neural Network (RNN) and Transformer are examined in detail. The FFNN
and RNN yield similar performance in terms of Root Mean Squared Error (RMSE) and Maximum Error (ME) on the test set. The Transformer is unable to match this performance level due to the absence of a proper positional encoding of the measurements. Due to the simpler structure and lower data requirement, the FFNN is selected for application in the hybrid structures.
To create a hybrid estimator, the uncertainty level of the VSA estimate from the neural network is required. To obtain the uncertainty using a neural network, various methods exist such as Monte Carlo Dropout (MCDO), Monte Carlo Batch Normalisation (MCBN) and the Uncertainty Deep Ensemble (UDE). These methods are compared using three metrics, the RMSE, Predictive Loglikelihood (PLL) and Continuous Ranked Probability Score (CRPS). A combination of these metrics does not only evaluate the estimation accuracy, but also the quality of the corresponding estimated uncertainty level. The UDE consisting of
FFNNs provides the best performance and even outperforms the single FFNN in terms of RMSE by a decrease of 8.6%. Therefore, the UDE is used to develop the different hybrid structures.

The Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) are built on a nonlinear bicycle model. Different sets of inputs for the observers, as well as constant or adaptive covariance matrices create different variations. These observer variations are combined with the UDE to form hybrid estimation
models. The different variations show equal performance due to the high accuracy of the UDE on the test set. The high performance of the UDE causes the measurement noise to be tuned to a very low level, which results in the observer ’trusting’ the UDE estimations. Therefore, the performance of the different
hybrid structures is almost equal to that of the original UDE.

To investigate if the observer is actually able to correct neural network estimations, a suboptimal neural network is created. This network is trained only on measurements with an absolute lateral acceleration lower than 5 m/s2. This mimics sparse data of high sideslip angles, a common problem for data­-driven
approaches in this setting. This causes the uncertainty of the neural network estimations to be higher in these operating regions. A linear scaling of the uncertainty level to match the variance of the noise on the other measurements does not yield satisfactory results. However, an exponential scaling of the
uncertainty level causes a higher differentiation between low and high confidence levels. This exponential scaling decreases the RMSE of the hybrid structure with 11.4 %. Furthermore, a method for adapting the process noise based on the quality of the observer model estimation is implemented. This equates to
similar performance in terms of RMSE, but introduces a number of additional parameters to tune.

This thesis presents a new MPC controller which integrates path tracking and stability control into one controller. Previously these tasks were done by separate controllers, where one controller handled the path tracking while another controller ensured the vehicle was kept in the stable operating region. A drawback of this method is that the controllers have opposing objectives. The path tracker could require a higher steering wheel angle to follow the path, while the Vehicle Stability Controller (VSC) might require a lower angle to keep the vehicle stable. By integrating these two controllers into one controller, the new controller is able to take both tasks into account and optimise the control output such that both objectives are satisfied. This is achieved by implementing two extra yaw rates into the MPC model. These are the expected yaw rates based on the steering wheel angle and lateral acceleration of the vehicle. By comparing these two yaw rates to the actual yaw rate, the stability of the vehicle can be determined. The MPC controller is then able to prioritise path tracking or vehicle stability. This is achieved by actively varying the weights in the cost function depending on the vehicle state. To compare the new MPC controller, 8 benchmark controllers have been created. These controllers can be divided into two groups of four controllers. The first group is able to use differential braking in the control output, while the second group can only output an equal brake torque for all wheels. The benchmark controllers use different methods for path tracking and stability control, to get an understanding of the performance benefits of each method.
These different methods include: adding an extra target yaw rate based on path curvature and speed for tracking, adding constraints to ensure vehicle stability and using a separate stability controller to stabilise the vehicle. All controllers are evaluated using the industry standard Moose test as well as a double lane change in simulations. These manoeuvres are used in industry to evaluate stability and can also be used to evaluate path tracking. Furthermore the robustness of the controllers was evaluated by changing various parameters. These variations include: changing vehicle speed, adding extra weight to the vehicle, lowering the road 𝜇 level and performing a lane change where each lane has a different 𝜇 level. The results were evaluated using objective Key Performance Indicators regarding tracking performance and vehicle stability. The results show that the new MPC controller with the combined path tracking and stability control improves performance in both objectives. The new controller improves path tracking by 8% compared to the pure path tracking controller. While the stability is improved by 11% compared to the controller with a separate VSC. Furthermore the new controller was able to keep the vehicle stable at higher speeds and was more robust to varying conditions.