In the automotive industry automation is popular and every year car OEMs advance their technology to be able to drive autonomously. Longitudinal control of the vehicles is an important part of the complete autonomous driving system. The difficulty of this control problem lies with changing longitudinal dynamics and the lack of full-state system information. This complicates controller design when using classic model-based approaches such as Optimal Control (OC). Currently the controllers are still manually tuned by control engineers in the vehicle. This is time consuming and expensive, therefore other methods for controller design such as learning are explored. Reinforcement Learning (RL) is one of those methods. To examine the potential benefits of learning a controller, this work will make a comparison between RL and OC. For RL, an actor-critic structure using deterministic policy gradient is applied. Due to partially observable system dynamics OC is used as an optimal output feedback controller. The comparison complies speed control of an autonomous vehicle. The RL agent will learn a controller by training on a nonlinear high fidelity vehicle model. In this work it was demonstrated that RL can reach the same performance as OC when all environmental settings are comparable. When environmental settings deviate, it was is found that RL outperforms OC. To verify the simulated results all controllers were confirmed in an experimental real-life setting.In conclusion, this proved a promising benefit of learning with respect to classical controller computation, when dealing with partially available system information.

Labour shortage in the logistic sector and the poor work conditions as a baggage handler are a major and highly relevant problem for airports around the world. A solution is provided by Vanderlande Industies: a baggage loading robot. Since the efficiency of the stacking process is sufficient, operators that overlook the robot must intervene often. To minimise these interventions, better packings have to be made. To make better packings, Vanderlande Industries can adapt the sequence in which the items are presented to the robot. However, literature and the company are not aware of the effects of sequencing. Thereby in this thesis, we conducted research into sequencing items in various ways for the Multiple Container Loading Problem (MLCP). We attempted to find out which methods enhance the fill rate of the packings. In order to understand the academic background of the MCLP, we first looked at the literature. Additionally, state-of-the-art techniques have been investigated. A trip to Schiphol has been undertaken to observe the real robot that is resolving the MCLP as a pilot project for Schiphol in order to put the issue in its proper respective. From this research, we concluded that there are various methods for solving the MCLP, however, sequencing methods applied to so far are mostly kept simple. Moreover, in practice, baggage items are placed in the order where size and weight are decreasing. The knowledge gap found was that sequencing for the MCLP is not researched. We presenttwo different types of algorithms to provide a well performing sequence. The first algorithm, the genetic algorithm, makes use of multiple generations to optimise the sequence. The other algorithm, the heuristic algorithm, makes use of a set of rules. The resulting packings formed with different sequencing methods show to be different. According to the simulation performed, the heuristic sequencing strategy produces the best results for the main scenario in terms of the layer fill rate. The performance of the packings can with the heuristic sequencer be increased by average with 11 % in comparison with packings generated with a random item sequence. Additionally, the heuristic sequencer’s packings are fairly consistent because it consistently manages to place 8 items in one layer of the container. Results in this study point out that sequencing the items with size decreasing performs worse than the packings formed from non-sequenced items, with a mean fill rate of 0.71 for the size decreasing items and 0.87 for the heuristically sequenced items. The alternate scenario produces fewer evident outcomes, but we can point out that simple strategies, such as sorting the items by short side, perform effectively. The results from this applied method: an increase of 7% in comparison with non-sequenced item packings. We assume the simple sorting approach worked well since it grouped all similar shaped objects, resulting in tight packings. The genetic algorithm produced slightly better results, with average outcomes being similar (0.87), but much more consistent packings showing a lower standard deviation of 0.00917 (GA) versus a standard deviation of (0.01747). The results from the heuristic algorithm were the least efficient. This demonstrates that sequencing using a heuristic is effective, but it is not a reliable strategy in all circumstances. The genetic algorithm, in comparison, generates generally worse results, but it can adjust to various situations and configurations. Although the results seem to be convincing, there rise discussion points. One of the major drawbacks is the fact that in this study, simulations are only done in two dimensions. This results in a few assumption regarding stacking behaviour. Thereby, a significant point of future work can be done by introducing the sequencing algorithm to three-dimensional simulations.

Bimanual coordination is essential for the performance of daily activities, but the underlying motor control mechanisms are not yet fully understood. The goal of the present study is to identify the contribution of contralateral responses in the wrist joints to the performance of a bimanual task. Contralateral responses could possibly be used in rehabilitation therapy to activate hand functions that are affected by a neuromuscular medical condition. In our experiment, participants had to balance a tray in a virtual environment, while either the left or right hand was perturbed. Two identical robotic wrist manipulators intermittently applied force perturbations in flexion or extension direction. Following a perturbation, contralateral responses were present and operated towards stabilization of the tray, for example by allowing an overall faster correction of the perturbation. Notably, flexion perturbations resulted in much larger contralateral responses than extension perturbations. Contralateral responses occurred mainly in the time window for voluntary responses. Results were consistent with our hypothesis for discrete bimanual movements based on optimal feedback control theory: when both hands share one goal, functional coupling occurs in the wrist joints.

Socially-aware robot navigation strives to find efficient methods to autonomously navigate known environments while incorporating social metrics derived from human behavior. Several methods built for navigation on dynamic and uncertain environments have been adapted to resemble human navigation, but fail to appropriately reflect the social characteristics of human decision making. A reliable solution for this problem is found in model-based methods working as local or global planners. Model-based research on human aware navigation focuses on two specific alternatives. The first option corresponds to models that apply social-psychology and cognitive sciences to create human-like behavior, where the Social Force Model is the predominant approach. The second alternative is related to models that use machine learning to copy human-like characteristics into mathematical models. The former approach has proven to be efficient in human-aware navigation, but further studies are required to analyze the expansion of this method to more complex human-interactive navigation. In this thesis project, we look to test the Modified Extended Social Force Model (MESFM) to implement a guiding behavior on a humanoid robot. The MESFM incorporates a new force linked to the guided person to maintain a natural distance between both subjects, while generating smooth navigational maneuvers. In addition, whether caused by sensor failure or occlusion incidents, the event of losing track of the guided person has a good probability of occurring. This scenario is studied by extending the high level control of the robotic system with a searching mode. The architecture proposed to control the guiding behavior and the searching mode exploits the design of the humanoid robot Pepper, from SoftBank Robotics, to incorporate human-like gestures and increase the interaction between robot and human. We aim to apply this behavior as an Office-Guide Robot and test this solution to understand the reaction of people involved in the guiding task though subjective and objective metrics. Our system is developed on top of the open-software Robot Operative System (ROS) with features extracted from Pepper's Naoqi framework.

Pile driving is a widely used technique for the construction of buildings and infrastructure. A popular technique is to vibrate the pile into the sediment. However, since building sites are increasingly being located in metropolitan areas, there is a growing concern about the environmental impact that vibrations may cause during driving. Therefore, pile driving assessment and prediction have become important for several reasons, including reducing the risk of damage to nearby structures and limiting disturbance to adjacent properties. In particular, pile driveability (i.e., the penetration rate) assessment is immensely useful prior to installation as it increases the construction performance and subsequently reduces costs and environmental impact. The important factors influencing the penetration rate include vibrator characteristics, pile properties, and soil conditions. However, due to assumptions and the lack of methods that accurately represent the complex phenomena at play during vibratory driving, a disparity is obtained between the predictions of modern pile behavior programs and the observed penetration rate. Recently, the registration of pile driving data has increased significantly. This extended amount of measurement data can potentially be leveraged for an improvement of the prediction of the penetration rate in future projects. Literature review on the application of machine learning (ML) within pile driving, geotechnical engineering and drilling revealed that the artificial neural network (ANN) is a promising alternative method for the prediction of the driveability of vibratory driven piles (i.e., vibro-driveability). In this work, machine learning methods and the traditional model were utilized to predict vibro-driveability. Promising ML techniques include the multilayer perceptron neural network (MLPNN) and radial basis function neural network (RBFNN). These neural networks were trained with the particle swarm optimization (PSO) algorithm. The backpropagation (BP) algorithm was also incorporated to train the MLPNN and RBFNN models as a conventional method. Based on results obtained with the aforementioned methods, we propose a new model, the Vibratory Driveability (VD) model, that combines the fruitful characteristics of the MLPNN and RBFNN. The performance of the five different models was compared with the performance of contemporary vibro-driveability prediction software for three test sets. This was done using different performance indices including the mean squared error (MSE), mean absolute error (MAE) and the weighted average percentage error (WAPE). Additionally, the desired characteristics of the predictions based on the geo-engineer's input were examined and compared with the obtained predictions. It was demonstrated that the ANN-based methods achieved drastic improvements in prediction performance and consequently outperformed the traditional model, making ANN-based methods the preferred alternative for the prediction of vibro-driveability. Among the ANN models, the VD model produced the highest performance, as it reflected the desired prediction behavior for all three test cases and showed competitive prediction performance in terms of the performance metrics. This work leads to the first-ever published research on the application of artificial neural networks for the prediction of vibro-driveability. As such, it could form the foundation for the development of new (vibratory) pile driving behavior assessment and prediction software. The development of these commercial applications could lead to a considerable reduction in costs and environmental impact.

This thesis tests the hypothesis that distributional deep reinforcement learning (RL) algorithms get an increased performance over expectation based deep RL because of the regularizing effect of fitting a more complex model. This hypothesis was tested by comparing two variations of the distributional QR-DQN algorithm combined with prioritized experience replay. The first variation, called QR-W, prioritizes learning the return distributions. The second one, QR-TD, prioritizes learning the Q-Values. These algorithms were be tested with a range of network architectures. From too large architectures which are prone to overfitting, to smaller ones prone to underfitting. To verify the findings the experiment was done in two environments. As hypothesised, QR-W performed better on the networks prone to overfitting, and QR-TD performed better on networks prone to underfitting. This suggests that fitting distributions has a regularizing effect, which at least partially explains the performance of distributional algorithms. To compare QR-TD and QR-W to conventional benchmarks from literature they were tested in the Enduro environment from the arcade learning environment proposed by Bellemare. QR-W outperformed the state-of-the-art algorithms IQN and Rainbow in a quarter of the training time.

Knowledge on adversaries during military missions at sea heavily influences decision making, making identification of unknown vessels an important task. Identification of surrounding vessels based on visual data offers an alternative to AIS information (Automatic Identification System), the current standard in vessel identification, which can be spoofed. One visual approach employs human expertise and manually identifies vessels guided by a ship catalog. In order to minimize or potentially eliminate human error and performance limitations, there is strong interest in developing an automated vessel classification pipeline. One such pipeline is currently being developed at TNO, capable of classifying over 500 separate classes. A crucial part of the classification pipeline is retrieving an accurate contour of a vessel from a digital image. To address this important challenge, this thesis proposes an advanced deep learning pipeline to automatically segment the vessel image into background (e.g. sky and sea) and the object of interest (a vessel). Deep learning models based on Fully Convolutional Neural Networks (FCNs) have achieved high performance on the task of semantic segmentation. Several networks such as CRF-RNN, PSPNet, DeepLab and Mask R-CNN are employed to determine a baseline performance. We will focus on identifying the cause of poor or failing segmentations and aim to construct a robust network capable of handling these challenges. By sampling disturbances, caused by ship distance and camera noise, augmented data sets are built to tune networks to input from on-site images. Additionally, experiments are done to evaluate the influence of different levels of disturbances. Previous approaches implementing the CRF-RNN network achieved top 1 and top 5 classification accuracies of 31.1% and 44.0% respectively. Employing the DeepLab network, trained to convergence on artificial noise augmented data, we report top 1 and top 5 accuracy of 68.9% and 88.8% respectively. Additionally, implementing an ensemble of classifiers, performance is increased to 73.0% and 91.7% for top 1 and top 5 accuracy respectively. This best result is comparable to the classification results with human annotated ship silhouettes. The human performance accuracy is 73.4% on top 1, and 91.3% on top 5 classification performance. Finally, we show that training on a collection of different levels of image disturbances results in a network that is robust against increasing disturbance in images, while retaining performance on clean images.

Sleep is an immensely important biological function and dysregulated sleep can have severe detrimental effects on health and well-being. Over 80 sleep disorders have been classified and of those sleep apnoea is one of the most common, having a prevalence of at least 20 percent in the general population. Sleep apnoea belongs to a family of disorders called Sleep-Disordered Breathing (SDB) which are characterized by breathing difficulties that occur during sleep. The underlying causes for SDB can be difficult to pinpoint making it important to develop effective and efficient detection methods. Sleep disorders are detected in sleep studies using one of two standard setups, Polysomnography (PSG) or Polygraphy (PG). A sleep study involves measuring physiological signals such as ventilation, heart rate and brain activity over a night of sleep. Recent studies have suggested that respiratory effort and the associated increase in negative intrathoracic pressure is a key variable for diagnosing and determining the optimal treatment for SDB. This variable is difficult to measure non-invasively and is not a part of a standard PG or PSG sleep study. The purpose of this MSc thesis was to explore and develop methods to infer respiratory effort from the non-invasive sensors used in a standard sleep study. A model-based approach was proposed to create a mapping between respiratory movements, measured by Respiratory Inductance Plethysmography (RIP) and the intrathoracic pressure. A grey-box model of the respiratory system was constructed based on the underlying physiology and the existing lung model literature. The model parameters were identified for each patient using a novel two step identification procedure, fitting the model to recorded sleep study data. The resulting model was used to simulate a system response where the desired intrathoracic pressure could be estimated. In addition to the model-based method two alternative methods were put forward, to explore the relationship between RIP and respiratory effort in the frequency and entropy domains. The modelling method and the two secondary methods were applied to recorded PSG data. Respiratory effort was observed via esophageal pressure, measured with esophageal manometry. The maximum performance of the modelling-, frequency domain- and entropy-method, evaluated as the Spearman correlation to the measured respiratory effort, were 0.53, 0.61 and 0.43 respectively. The performance of different methods varied greatly between patients in the dataset and at this point no method can conclusively be said to outperform another. This MSc thesis constitutes preliminary work in the development of non-invasive respiratory effort estimation methods. The results seem to validate the premise that there is a connection between respiratory movements and intrathoracic pressure. The modelling method is the most complicated of the three but has the most potential for improvements.

Learning capabilities are a key requisite for an autonomous agent operating in dynamically changing and complex environments, where pre-programming is not anymore possible. Furthermore, it is essential to guarantee that the learning agent will act safely by considering its stability properties. In this thesis, novel conditions are proposed, aiming to examine stability of the learned dynamics for two important model classes; namely Rectified Linear Unit (ReLU) Deep Neural Networks (DNNs) and Locally Weighted Learning (LWL). For the former method, a theoretical and computational framework is developed by establishing an equivalence between ReLU DNN models and Piecewise Affine (PWA) systems. This allows to leverage well-known tools of PWA system analysis, and consequently compute, characterize equilibria and determine their region of attraction for ReLU DNNs. Due to their increased complexity, a structured search for appropriate stability conditions was performed for LWL methods until the optimal trade-off between conservativeness and computational efficiency was obtained. These stability conditions are given as Linear Matrix Inequality (LMI) problems and they consist the first stability results in literature for these two model classes. Their efficacy is assessed in numerical and real-world dynamical systems and it is shown that the proposed LMIs are not unreasonably conservative, as they can evaluate accurately the stability properties of these two representations. Finally, this work demonstrates how to formulate appropriate stability conditions for learning methods in a principled manner.

One of the most promising ideas in autonomous vehicle control systems is letting the vehicle drive autonomously outside the normal, linear, operating region and letting it "drift". By doing so, the maneuverability of the vehicle could be enhanced. To enable systems that can control this behaviour, estimation of certain vehicle states is needed with high accuracy and high frequency.In this project, a new solution to this problem is proposed by combining a mixed dynamic-kinematic observer with a single camera that produces velocity measurements based on tracking the ground plane. To improve filtering of the camera velocity measurements, the measurement error covariance matrix is updated online based on a model of the camera measurement error. Evaluation of the new methodology was done on data recorded from a 1:10 scale test vehicle and performance was assessed based on ground truth data obtained using a Motion Capture System.In normal driving conditions with correctly identified vehicle parameters, an observer without camera performs better by 25% in terms of RMSE on lateral velocity and sideslip angle estimation. However, the online adaptation of the covariance matrix results in an estimate that is at least 45% more accurate in terms of RMSE than the same observer without online covariance adaptation. Next to that, experiments show that the proposed observer with camera has better robustness to uncertainty in model parameters by almost a factor five in terms of RMSE than the observer without camera.When the grip of the tires is physically lowered and the vehicle is drifting, the proposed observer can track large sideslip angles (>30°), where the state-of-the-art observer without camera is not able. The state-of-the-art observer has an increase in RMSE of 75% on all estimated quantities in comparison to the proposed methodology. These results show that adding a camera to an existing sideslip angle observer greatly enhances robustness of the observer to uncertainty in model parameters and violation of model assumptions. This comes dat the cost of losing some accuracy in normal driving conditions.

This MSc. thesis explores the design and implementation of a motion planner for non-holonomic autonomous vehicles in parking spaces. The planner must avoid collisions with static obstacles, satisfy performance, comfort and safety constraints for the motion, satisfy the non-holonomic constraints of the vehicle model, consider the dynamics of the vehicle actuators and be fast enough for real-time implementation. The motion is planned by solving an Optimal Control Problem (OCP) that is discretized in time in order to obtain a non-linear, non-convex, multi-variable optimization problem. The thesis addresses how to solve this optimization problem so that it results in motions that satisfy the requirements. Specifically, the motion is split into a number of waypoint tracking sections, planned by correspondingly simplified optimization problems. The thesis also addresses methods to further simplify the optimization process in order to reduce the implementation time. The results show the planner satisfies the constraints but does not quite work in real time. Recommendations for improving the planner in the future are given in the report.

The highly competitive nature of Formula one makes every team eager to find areas of improvement on their cars. In the case of Scuderia Toro Rosso, it was found that there is a gap to improve performance while driving on kerb stones. Analysis of competitors showed that they can take wider driving lines with more speed. In order to improve this performance, the need for a high fidelity dynamic car model arose. This model can be used to understand the car dynamics, which can be used to choose a better suspension setup. The baseline model used for this study was developed in MSC.Adams. After a discussion with field experts, two areas of model improvement were identified. In the baseline model, the tire is modeled using a MF-Tire model, where the vertical dynamics was modeled as a single spring and damper. In order to simulate the tire dynamics on kerb stones, a more advanced model is necessary. For this research, FTire was chosen as the best option. The other area of improvement is modeling the compliance of the suspension system. In the baseline model, all suspension members are modeled as rigid links, without any compliance. To model this compliance, an ’equivalent stiffness’ model was proposed, where all stiffnesses are combined into flexible joints. To use the FTire model, a set of parameters needs to be found for the Formula one tires. To find these parameters, measurements were done to capture the dynamics of the tire. A method is developed to filter out any rig disturbance, and make the data useful for parameter estimation. The estimation was done using Matlab, with MSC.Adams and FTire in-the-loop, and made use of a Genetic Algorithm (GA). Comparing the simulated behavior with the measurement showed good correlation for various conditions. Since every part in the car contains some compliance, modeling and measuring every single part would be an extensive task. In order to reduce the complexity, all compliances are combined into a simplified model. In this simplified model, all compliances are modeled in the joints. A Pattern-Search algorithm in Matlab is used in combination with MSC.Adams to find the joint stiffnesses in such way the model matches measurements of the overall car compliance. To measure the overall improvement, both models are combined into a full car model. The baseline and proposed model are validated against measured track data. From this compare, it can be concluded that the proposed model shows a significant better correlation on kerb simulations. This is mainly due to the fact that the baseline tire model is not capable of simulating the harsh road input. The proposed model provides a more powerful tool for engineers to better understand the car dynamics, and to find the best suspension setup for kerb riding.

A tiny house is a building for permanent human habitation that is specifically designed to have a limited ground surface. The tiny house design discussed in this report has a strong focus on circularity, sustainable material usage, smart systems, and affordability. To achieve an overall self-regulating and ecological concept, the aim is to combine and optimise the different flows that go through the tiny house - i.e. electricity, waste, and water. These flows are also smartly integrated and made more efficient on a network scale. Several tunus tiny houses are combined in a village because sustainable living environments can be created more effectively when collaborating in communities. Eventually, the goal is to obtain a network with such flexibility that its principles can be implemented on any collection of tiny houses or even terraced houses and flats.

Large-scale baggage handling systems, or large-scale logistic networks, for that matter, pose interesting challenges to model-based control design. These challenges concern computational complexity, scalability, and robustness of the proposed solutions. This thesis tackles these issues in a collection of papers organized in two overlapping parts. The first part concerns modeling and Model Predictive Control (MPC) design of large-scale baggage handling systems (BHSs), where a modeling framework for BHSs is proposed that is subsequently used to develop an MPC scheme for control of large-scale BHSs. The MPC controller optimizes for the timely arrival of pieces of baggage at their destination within the BHS network under capacity constraints while minimizing the overall cost of transporting pieces of baggage. Several formulations for the resulting constrained optimization problem are proposed, and they are compared with each other in terms of closed-loop performance and computational complexity. It is shown, via simulation studies, that the proposed solutions can outperform a heuristics-based approach commonly used for control of BHSs while scaling well to larger BHS network instances. In its second part, the thesis focuses on robustness of control design in the face of a partially known disturbance input (i.e., input baggage demand), and especially on developing a scalable tube-based MPC scheme. For this purpose, considering the BHS model essentially as a linear positive system, a linear-programming-based approach is proposed for the joint calculation of a robustly positively invariant subset and a constrained state feedback controller that minimizes the disturbance-driven L∞ norm of the output over this set. A tube-based MPC control scheme is finally developed by coupling the state feedback controller with a nominal MPC controller, guaranteeing recursive feasibility and asymptotic stability. It is shown via simulation studies that the proposed tube-based approach is effective against unpredictable disturbances. In addition, since the design of both the nominal MPC controller and the state feedback controller involves only linear programs, the proposed tube-based approach scales well to BHS networks of larger size. Linear positive systems are of interest in several branches of engineering, logistics, biochemistry, and economics. As a spin-off topic and inspired by the applications of the theory of linear positive systems to modeling and control design of systems in the mentioned domains, the third part of the thesis focuses on the reachability analysis of discrete-time linear positive systems. More specifically, we revisit the problem of characterizing the subset of the state space that is reachable from the origin for discrete-time linear positive systems. This problem is of interest in topics such as optimal control of linear positive systems and realization theory of linear positive systems. It is established in this thesis that the reachable subset can be either a polyhedral or a nonpolyhedral cone. For the single-input case, a characterization is provided of when the infinite-time and the finite-time reachable subsets are polyhedral. Finally, for the case of polyhedral reachable subsets, a method, based on solving a set of linear equations, is provided to verify whether a target set can be reached from the origin using positive inputs. @en

The volatility of renewable energy sources pose a significant challenge for Distribution Network Operators (DNOs) as it makes planning and maintaining a reliable electricity grid more complex. An essential tool in dealing with the uncertain behavior of renewable energy resources is the load flow simulation, i.e., the standard electricity network simulation in network design and operation. There is, however, still much untapped potential of applying these kind of simulations. The thesis presents improvements to the theory on linear load flow approximations. The resulting algorithms are then applied to various real world problems: control of a community battery, handling very large simulations, coping with low sensor coverage and evaluating strategic scenario's with high uncertainty. Firstly, theory is presented for the control of a community battery. It is shown how such a battery can be used for grid congestion reduction, backed up by a live experiment. A charge path optimization problem is posed as a linear problem and subsequently solved by an Linear Programming (LP) algorithm. It was found that the voltages and currents can be controlled to a great degree, increasing the grid capacity significantly. Network design formulas are described with which a DNO can quickly estimate the potential (de)stabilizing effect caused by a community battery on the steady-state voltages and currents in the grid. Next, load flow simulations are improved by applying numerical analysis techniques and the accuracy and efficiency of a linear load flow approach is investigated. The resulting fast load flow algorithm is then applied to a very large problem: integrally simulating the low and medium voltage network of Alliander DNO, a grid with over 22 million cable segments with a total combined length of over 88,000 km, built according to international standards. It is shown that this integral simulation can identify voltage problems much more accurately. Next, Bayesian state estimation is considered. A mathematical model is proposed to complement a limited set of real-time measurements with voltage predictions from forecast models. This method relies on Bayesian estimation formulated as a linear least squares estimation problem. The model is then applied to an IEEE benchmark and on a real network test bed. An observability analysis suggests strategies for optimal sensor placement. Next, theory is presented on coping with uncertain long-term scenarios for strategic simulations. A stochastic profile model is proposed based on copulas which can be calibrated by technology adoption data. Using a Monte Carlo approach, the stochastic profiles of all DNO assets are then simulated, identifying parts of the network with heavy loads. Finally, the thesis concludes by demonstrating additional applications of the presented methods, such as fast network capacity checks and reducing losses via network reconfiguration. It concludes by giving suggestions for future research. @en

ZF Friedrichshafen AG is developing its own active suspension system designed as a single unit per corner to optimize the dynamic behavior of a vehicle. The innovative active damper design integrates all the essential components into one damper unit, with the aim to facilitate implementation. This master thesis in cooperation with ZF Friedrichshafen AG presents the development of a disturbance compensation controller for the active suspension unit. The goal is to suppress forces induced by road irregularities on the passenger cabin. The emphasis throughout this work lies on the implementation and applicability of the control algorithm.

The startup company Fleet Cleaner has developed a mobile robot, specialized in the hull cleaning of large cargo vessels. Navigation and localization of this robot is currently performed manually. This is a difficult process that is greatly complicated during operation. This is mainly due to the availability of relative positioning sensors only, which are prone to error build-up and noise, and to the difficulty of interpreting optical underwater images in turbid water conditions. Instead, operators must rely on acoustic images from a forward-looking sonar. In the field of mobile robotics, Simultaneous Localization and Mapping (SLAM) is an often used technique to improve navigation and localization by utilizing visual information. The objective of this thesis is to develop a sonar-based SLAM framework, tailored to working environment of the Fleet Cleaner robot. The thesis scope has been restricted to the conceptual design of such a framework and the implementation of one of the subsystems, visual odometry.  A conceptual design of a SLAM system is proposed using a systematic approach. Different working principles are evaluated according to operating conditions and requirements that specify desired behavior. Analysis of operating conditions reveal the limitations of sonar imagery, such as a high signal-to-noise ratio and inhomogeneous intensity patterns. In addition, the environment is sparse, with few distinct recognizable landmarks, limiting feature-based approaches. Because of these limitations, visual odometry is essential to reduce error build-up between loop closure corrections. A Fourier-based approach to visual odometry is implemented, taking the whole image view into account instead of extracted features. By analyzing the dominant peak in the phase correlation matrix, the in-plane sonar motion between consecutive image frames can be estimated. Several image processing steps are necessary to improve peak sharpness, increasing the quality of registration. To validate the proposed method, an experiment was conducted during cleaning of the Pioneering Spirit, the world’s largest construction vessel. Under normal circumstances, visual odometry showed less error build-up in the position estimate than wheel odometry. However, outliers appear when driving near the waterline, caused by reflections and wave reverberations. Ultimately, the proposed visual odometry method improves the current positioning system and serves as a basis for an integral SLAM implementation.

Swarm robotics (SR) is an emerging field of research that utilizes a swarm of robots that work together as a team to solve complex problems. One such challenge in SR is the exploration of unknown environments to find multiple targets. This problem finds applications in various domains, such as archaeology, underground exploration, signal source localization, and more. In line with this, the objective of this thesis is to develop a scalable SR algorithm for efficiently finding clustered targets in large unknown environments. Similar to how artefacts in archaeology are often found clustered around old settlements in large areas. Inspired by Bee Swarm Optimization (BSO), the proposed algorithm leverages the strengths of BSO in balancing exploration and exploitation. However, modifications are made to adapt the proposed algorithm to incorporate limited communication range scenarios, common in large-scale environments. Also, the exploration-exploitation properties of BSO are redesigned. The proposed algorithm is named Modified Bee Swarm Optimization (MBSO). Here, robots assume different roles (scout, onlooker, experienced forager), similar to BSO, to optimize search and exploitation tasks. To address the issue of limited communication range, the robots establish an ad hoc network, truncating target information throughout the swarm. Additionally, an Artificial Potential Field (APF) is introduced to guide robots towards targets, and away from readily travelled clusters. To further aid the balancing of exploration and exploitation, a swarming architecture is introduced. This architecture is called the Architecture Multi-robot systems heterogeneous robots with Emergent Behaviour (AMEB) and aids with the decision-making of individual robots. The AMEB architecture is used to determine whether scouts should become onlookers, and the speed at which experienced foragers change back to scouts while considering real-robot physical limitations and individual performance levels. This architecture facilitates the continued advancement of the algorithm by allowing for the integration of additional sensory inputs, which in turn influences individual decision-making and, consequently, the emergence of the swarm. Lastly, cluster recognition is added to the algorithm, resulting in robots not transferring to readily travelled areas. The characteristics of MBSO are evaluated. The target finding performance is benchmarked against a generic random walk method that stems from Lévy walking. Furthermore, this study investigates the effects of scaling the algorithm with an increasing number of robots and varying specific control parameters of MBSO on various aspects such as performance, redundancy, scalability, stability, and robustness. The results of this research have implications beyond archaeology, as the algorithm can be applied to various multi-target search problems in large unknown environments, such as minefield detection. By adjusting the proposed input variables, the algorithm can be optimized for these different scenarios. The developed SR algorithm shows promise in efficiently finding and truncating target information, leveraging the short communication range of the individual robots. Overall, this thesis presents a novel approach to autonomous multi-target searching, in cases where targets are spread out in clusters, using scalable robotic swarming algorithms. In a field of 0.36 ha with 34 targets spread out over 5 clusters, robots that transfer with a speed of 6.4 km h−1 and optimal parameters for MBSO, scaling from 10 to 15 robots leads to 7.74 % more targets being found. Scaling from 15 to 20 robots resulted in 13.22 % more found targets...