Committee:
     PRESIDENT: IVANOVA RADEVA, PETIA
     SECRETARI: BENITEZ IGLESIAS, RAUL
     VOCAL: ORTEGA RAMÍREZ, JUAN ANTONIO
Thesis abstract: Clinical pathologists focus on detailed morphological analysis of peripheral blood (PB) cells for the recognition of abnormal cells and contribute to the diagnosis of hematological diseases. While this traditional method is often precise, it requires significant expertise, is time-consuming, and is subject to variability, especially when discerning subtle cell characteristics. In light of these challenges, this dissertation proposes the innovative use of deep learning, specifically convolutional neural networks (CNNs) and generative adversarial networks (GANs), to add objectivity in morphological examination of PB cells. The research is developed through a three-fold strategy. First, a novel GAN-based model standardizes PB cell image visualization across varied staining techniques. Second, optimized CNN architectures are tailored for precise classification of the selected PB cell types, with emphasis on reactive and malignant blood cells. Results indicate that the introduced Stain Normalization Model effectively manages staining variations while preserving cell morphology. Furthermore, CNN models, based on these normalized data, contribute significantly to the automatic recognition of malignant cells. And third, a modular GAN system generates synthetic PB images to enrich model training, particularly of low-prevalence cell data with special diagnostic interest.This research emphasizes the transformative impact of combining the expertise of clinical pathologists with advanced computational strategies. While these systems serve as potent support tools, it is imperative to note that the clinical pathologist retains the ultimate authority in decision-making. The fusion of traditional expertise and modern technology is a promising trend in the present and near future, enhancing both objectivity and efficiency. This study signifies an advance towards the establishment of an integrated system for an automated screening tool for hematological diseases.

Committee:
     PRESIDENT: REY LÓPEZ, JUAN MANUEL
     SECRETARI: DORIA CEREZO, ARNAU
     VOCAL: ACOSTA RODRÍGUEZ, JOSÉ ÁNGEL
Thesis abstract: Nowadays, the environmental concerns for conventional fossil fuel-based power generation are transforming the energy landscape into a more sustainable system. To this end, distributed generation (DG) units have gradually become more important in modern energy systems with an increasing integration of renewable energy resources (RES). However, this transformation also implies new challenges for operating and controlling the network safely and efficiently, given the variability and intermittency inherent to these resources.Microgrids (MGs) are small-scale power systems that have emerged as promising solutions for integrating RES such as photovoltaic, wind turbines, fuel cells, biogas digestors, etc. However, given the heterogeneous nature of RES, power electronic devices are required to enable their integration into the AC systems. The incorporation of power converters increases the controllability of these resources but also brings new challenges regarding control and protection. In inverter-based AC MGs, the scope of this thesis, the DG units are interfaced with the MG by inverters, which require to be controlled to ensure the correct operation, stability, and voltage quality of the AC system. To solve these challenges, several inverter control approaches have been proposed in the state-of-the-art. Nevertheless, there remain significant challenges, limitations, and research gaps. In particular, the main shortcomings found in the state-of-the-art can be gathered into two main control areas: a) Power sharing and voltage regulation (Power management), and b) voltage unbalance compensation (Voltage quality).On one hand, power sharing and voltage regulation control approaches aim to ensure the voltage amplitude and frequency stability and its regulation within desired limits in islanded AC MGs. However, existing literature exhibits the following shortcomings: line impedance performance dependency, scalability issues, and control loop coupling. On the other hand, voltage unbalance compensation control approaches aim to increase the voltage quality of the MG. However, a lack of robustness and limitations on the place of compensation were found in the state-of-the-art strategies.This thesis aims to investigate novel control strategies to improve the power control performance and increase the voltage quality on inverter-based AC MGs, focusing on addressing the research gaps in the existing literature. The thesis is presented as a compendium of three peer-reviewed papers that collect the main contributions and results of the thesis.These papers propose three novel control approaches that outperform and overcome most of the identified state-of-the-art shortcomings and limitations. The main contributions of the thesis can be summarized as follows:On one hand, regarding power sharing and voltage regulation, the proposed approach exhibits the following key properties, a) the control performance is independent of the line impedance nature, b) scalability with respect to the number of inverters within the MG, and c) decoupling between the voltage regulation control loop and the power sharing control loop.On the other hand, regarding voltage unbalance compensation, the two proposed approaches exhibit the following outstanding properties, a) fair control effort distribution without control data exchange between inverters, b) robustness to connection and disconnection of inverters and loads, c) reduction of communication burden, and d) flexibility to compensate for multiple local/remote nodes.

Committee:
     PRESIDENT: VALLÈS PERIS, NÚRIA
     SECRETARI: REPISO POLO, ELY
     VOCAL: TREJO RAMÍREZ, KARLA ANDREA
Thesis abstract: This research discusses elements to be considered for designing, developing, and validating a service robot that performs its task in different social environments. Due to the social focus of the provided services, technical considerations are demanded toaccomplish the task, and the acceptability of use for the people interacting with the robot.The first stage of the research considers previous cases on the implementation of service mobile robots, their analysis, and the motivation of how to solve their acceptability and use by people. The developmental part presents the technical and social considerations for implementing the CeCi (Computer Electronic Communication Interface) social robot. Two main problems of social robots and service robots in social environments currently on the market are addressed, which are the main focus of this research: First, their costs are not affordable for many companies, universities, or individuals in developing countries. The second is that their design is exclusively oriented to the functional part with a viewpoint inherent to the engineers who create them without considering the end users’ views, preferences, or requirements, especially for their social interaction. This last reason ends up causing a certain aversion to the use of this type of robot.In response to the issues raised, an affordable, low-cost prototype is proposed, starting from a commercial platform for research development and using open-source code. The robot design presented here is centered on the criteria and preferences of the end user, prioritizing acceptability for social interaction. This document details the selection process and hardware capabilities of the robot. Moreover, a programming section is provided to introduce the different software packages used and adapted for social interaction, the main functions implemented, as well as the new and original part of the proposal. A list of applications currently developed with the robot and possible applications for future research are discussed.As a final step, the complete implementation of the social robot is explained based on two design elements. The first element is the use of the design thinking methodology for the development and implementation of the robot. The second element, of a technical nature, is a previous taxonomy generated for defining socially-aware robot assistants. Therefore, in the construction process, special emphasis is placed on the realization of prototypes and their adjustment to the users’ preferences. Interviewswith users who were unaware of the robot’s capabilities were used to improve and validate the prototype. These prototypes and their evolution will be presented based on the adjustments performed. Beyond the users’ feedback, previous experiences exposed to the state of the art were considered for the evolution of this robot. The entire methodological process is validated with surveys, and results are presented as a SWOT (strengths, weaknesses, opportunities, and threats) analysis for future improvements.

Committee:
     PRESIDENT: MANSARD, NICOLAS
     SECRETARI: MORCEGO SEIX, BERNARDO
     VOCAL: LIPPIELLO, VINCENZO
Thesis abstract: Aerial manipulators, which commonly take the form of multirotors with attached robotic limbs, primarily employ their limbs for pure manipulation tasks and do not rely on them during aerial locomotion. Besides, their movement tends to be slow. This thesis aims to enhance an aerial manipulator’s agility by harnessing its limb’s capabilities to augment its overall motion. This objective involves investigating various modes of utilizing the limb: as a tail for aerial locomotion, as an arm for aerial manipulation, or as a leg for hybrid aerial-contact locomotion. The present thesis contributes to two specific domains: 1. Generation and control of agile motions for aerial manipulators, 2. Design and construction of a specialized aerial manipulator for executing agile motions.Generating agile motions requires predicting the movement of the robot considering its dynamics so that these dynamics can be used to favor the robot’s motion. Hence, we can achieve complex maneuvers with relative ease. Optimal control is a trajectory-generation technique that meets these requirements, and that is central to this thesis. We encode the robot’s tasks as cost functions of the optimal control problem (OCP) and use the whole-body dynamics as the constraints of the dynamic system. On the control side, to deploy such trajectories in a real robot, we use model predictive control (MPC) techniques, which is the closed-loop control extension of optimal control. To get the control command, an MPC controller solves the OCP in which we have encoded the agile trajectory, and then the controller applies the first command of the solution control trajectory. Thus, MPC requires solving an OCP at the control rate, i.e., within a few milliseconds. This forces us to use fast, specialized solvers based on the dynamic programming principle, such as differential dynamic programming (DDP). In their original form, these solvers cannot consider the control bounds. These bounds are important to create trajectories compatible with the real robot. To tackle this problem, in this thesis, we propose two DDP-based methods to consider the control bounds: one is based on a squashing function, and the other is based on a projection method. Even with these solvers, we face challenges in meeting the solving rate and are forced to reduce the MPC horizon. Reducing the MPC horizon implies that the MPC can only see a portion of the original OCP, possibly leaving out some of the tasks. This affects the predictive capability of the controller and compromises the accomplishment of the tasks, especially those that require an elaborate and dynamic maneuver. To overcome this difficulty, in this thesis, we propose to update, at each MPC iteration, the terminal cost function in the MPC with a function that encodes the part of the trajectory that remains unseen by the controller.Regarding robot design, deploying agile motions becomes difficult with existing aerial manipulators, which are generally big-size multirotor platforms with non-compliant, high-gear ratio limbs. In this thesis, we present Borinot, an open-source aerial robotic platform designed to research hybrid agile locomotion and manipulation using flight and contacts. This platform features an agile and powerful hexarotor that can be outfitted with torque-actuated limbs of diverse architecture, allowing for whole-body dynamic control. We present experiments with this robot showcasing different agile motions.In addition to the stated contributions, this thesis contributes in other areas required to operate the robot, such as a procedure for identifying the dynamical parameters based on factor-graph estimation or a hardware enhancement that allows for direct thrust control of Borinot’s rotors.

Committee:
     PRESIDENT: DE SCHUTTER, BART HENRI
     SECRETARI: GRIÑO CUBERO, ROBERTO
     VOCAL: OLARU, SORIN
Thesis abstract: This doctoral thesis explores the utilization of generalized Nash equilibrium (GNE) seeking methods for evolutionary games over networks in designing optimization-based controllers for large-scale complex systems (LSCSs). The focus lies on population games, modeling strategic interactions among large populations of agents with bounded rationality. These games are deemed suitable for modeling large-scale optimization-based control problems either directly or by employing evolutionary dynamics models (EDMs) to update optimization variables. The concept of GNE emerges as a fitting solution for non-cooperative optimization-based control scenarios, where multiple subsystems maximize their utilities subject to coupling constraints. Consequently, this thesis frames optimization-based control as a GNE seeking problem and introduces novel game-theoretical methods for achieving GNEs in population games over networks. Applications in congestion games, charging coordination of plug-in electric vehicles (PEVs), dynamic resource allocation in solar collector fields, and residential demand response are discussed.The first part focuses on GNE seeking methods for population games under full information schemes, where non-local information estimation is not explicitly modeled. Two approaches are proposed: one dealing with capacity-constrained population games, introducing novel strategy revision protocols to ensure capacity constraint satisfaction, and another addressing population games with affine equality and convex inequality constraints, devising a payoff dynamics model (PDM) steering agents towards GNEs. Formal asymptotic stability guarantees are provided, and these methods are applied to design controllers for dynamic resource allocation in LSCSs.In contrast, the second part formulates GNE seeking methods for population games under partial information schemes, explicitly modeling non-local decision information estimation. It tackles reaching Nash equilibria without coupling constraints and achieving GNEs with affine equality constraints coupling multiple agents' decisions. Novel PDMs are formulated to estimate non-local information and compute dual Lagrange multipliers associated with coupling constraints. Asymptotic stability guarantees are provided, and the methods are applied to large-scale congestion games.The third part introduces hybrid PDMs combining continuous and discrete-time dynamics. Two event-triggered PDMs are devised to accommodate practical control features like logic implications and occasional communication. One mechanism focuses on PEV charging coordination, considering charging completion and random arrivals/departures of PEVs, while another aims to reduce communication costs in residential demand response. Formal convergence guarantees towards GNEs are deduced for both cases.In summary, this thesis contributes theoretically and practically to GNE seeking methods for population games over networks and to the design of optimization-based control strategies for LSCSs.

Committee:
     PRESIDENT: VALIÑO GARCÍA, LUIS
     SECRETARI: COSTA CASTELLO, RAMON
     VOCAL: MORÉ, JERÓNIMO JOSÉ
Thesis abstract: Fossil fuels have played a pivotal role in driving global economic development. However, their rampant use has brought about severe environmental consequences, particularly air pollution and greenhouse gases (GHG) emmissions, which poses substantial threats to human health and the planet's well-being. To address these pressing issues, the development of efficient and sustainable technologies like Proton Exchange Membrane Fuel Cells (PEMFCs), powered by hydrogen from renewable sources, has become paramount. PEMFCs hold immense promise as a viable solution for decarbonization and mitigating air pollution and GHG emissions associated with transportation and vehicles. Nvertheless, overcoming challenges related to their efficiency and longevity remains a crucial step forward.Specifically, this thesis explores the application of PEMFCs in an automotive fuel cell electric vehicle. The longevity and durability of a fuel cell system, together with its efficiency, are pivotal factors in the advancement of fuel cell vehicles, as they significantly impact the cost and consumer acceptance of these vehicles. In the domain of automotive applications, the primary contributors to fuel cell degradation include challenging operating conditions (involving temperature, humidity, pressure, mass flow rate, high and low power densities), dynamic load profiles, improper handling during start-stop and idle modes, and the presence of contaminants in the reactant gases. A distributed-parameter model of the stack is presented in this thesis, accurately depicting the internal dynamic behavior of the PEMFC. Additionally, the models of balance-of-plant subsystems, including anode, cathode, and thermal subsystems are introduced and implemented in Simulink. All these models are experimentally validated. Based on these models, a model predictive control strategy is designed to obtain the optimal operating conditions for the stack and generate from these the balance-of-plant subsystem setpoints.A supervisory controller for a fuel cell system is designed. This supervisory controller comprises three key components: a setpoint generator, a state machine, and a power limit calculator. The supervisory controller is subjected to rigorous simulation testing using a standard automotive driving cycle.These sophisticated control methodologies and algorithms are then applied to the automotive system developed in the European INN-BALANCE project, which served as the foundation for this PhD Thesis. Therefore, this thesis presents experimental results obtained from the implementation of the proposed controller in a prototype vehicle at PCS and CEVT facilities, encompassing startup, shutdown, dynamic load tracking, and subsystem local controller performance.Finally, this thesis provide a comprehensive overview of the contributions made by the research and its outcomes. The conclusion also proposes potential avenues for future research, marking the culmination of the investigative journey undertaken in this thesis.

Committee:
     PRESIDENT: SERRATOSA CASANELLES, FRANCESC ASSIS
     SECRETARI: GRAU SALDES, ANTONI
     VOCAL: VENTURA ROYO, CARLES
Thesis abstract: Surface inspection of coated surfaces in the automotive industry, traditionally has been a manual process in charge of keen eye operators in charge of inspecting the whole car body. However, as might be deducted, manual inspections often lack repetitiveness and reliability, very much desired in a such a strict sector. Computer vision tackled this problem with the first automated defect detection systems, pinpointing the defects in the car body and their size. Nevertheless, as these systems are constrained to just detection, the operator is still in charge of properly labelling the defects to rework them correctly. Additionally, there is a lack of traceability between the process and the defects themselves, taking longer to identify the root causes of faulty paint shop facilities.In this thesis, we address the multidisciplinary problem of defect identification in specular surfaces, with two main research lines. In the first one, we developed a novel illumination approach based on indirect diffuse lighting, in contrast with the conventional specular reflection. Together with a high resolution camera, we demonstrated an important improvement in terms of defect recognition with respect to the existing defect detection systems. These results are assessed with specialized auditors from the SEAT Martorell factory.The second research line, oriented to computer vision, explore the possibilities of implementing a deep learning solution for industrial defect identification. We developed a fast and reliable context aggregation model, featuring dilated convolutions and residual connections between opposite layers. This model is then trained following a loss leverage between classification and segmentation, for a smoother training procedure. Additionally, in order to cope the frequent class imbalances in the industrial datasets, we developed a guided-crop image augmentation strategy, based on cropping real defect randomly into non-defective images to generate synthetic new samples. The results state that the combination of this model with this augmentation strategy is able to outperform well-known segmentation models.Eventually, for data scarcity situations, we resorted to image synthesis methods to generate new fake samples. Models like Pix2pix have proven to be able to generate close to real im- ages, helping the segmentation model to converge faster than with the previous guided-crop image augmentation technique. Later, this generative method will be surpassed by a more sophisticated one, which features spatially-adaptive normalization layers that help to synthe- size images even without an encoder. Overall, it demonstrated good capabilities in multiple industrial datasets.

Committee:
     PRESIDENT: NGUYEN, ANH-TU
     SECRETARI: BLESA IZQUIERDO, JOAQUIN
     VOCAL: IFQIR, SARA
Thesis abstract: Autonomous driving is a promising solution for current transportation systems that would provide sustainable traffic flow, efficient mobility, and pollution reduction. However, the interaction of multiple vehicles presents several engineering challenges, such as vehicle-to-vehicle communication and critical decision-making. Robustness is another challenge that is crucial to ensure the proper design of autonomous driving technology. Vehicles are subject to significant uncertainty due to modeling errors, discretization of sensor signals, or external signals that adversely affect their nominal behaviour. This thesis proposes robust and optimal model-based approaches to address motion planning and control problems. Robustness is addressed from a set theory perspective. Specifically, the advantages of zonotopic sets are used to exploit the design of tube constraints in planning and control, as well as to quantify and propagate uncertainty in the modeling and estimation phases. In the field of modeling, this thesis utilizes control/planning-oriented models using the Takagi-Sugeno (TS) and linear parameter-varying (LPV) paradigms to represent the non-linear model of the vehicle. These representations are obtained from the physical equations of motion or directly from data. The data-driven dynamic model is trained using an adaptive neuro-fuzzy inference system (ANFIS). Meanwhile, the uncertainty associated with the model parameters is estimated through a zonotopic recursive least-square (ZRLS) approach.In the field of motion control, this thesis proposes a two-layer trajectory tracking robust strategy based on model predictive control (MPC) in the outer layer. In the inner layer, a combination of a linear quadratic zonotopic controller (LQZ) and a zonotopic Kalman filter (ZKF) is used. The vehicle's energy is managed by the MPC through the state of charge of its battery while ensuring performance indexes in terms of longitudinal, lateral, and angular velocities. To achieve this, the strategy combines a data-driven dynamic model with kinematic and state-of-charge models.Subsequently, the two-layer robust configuration is utilized for planning purposes. In the first step, the model predictive planner (MPP) generates feasible references online while maximizing both the longitudinal velocity and the saving energy through the battery state of charge. This is different to the control stage, in which the controller follows the reference previously computed offline. Additionally, the planner integrates a data-driven prediction model and a computationally efficient driving-corridor-like method based on track width constraints to handle collision avoidance for both static and moving obstacles. In the second step, an MPP is proposed to address coordination between connected autonomous vehicles. This planner utilizes the LPV representation of the non-linear vehicle model, which is derived from the physical equations of motion using the non-linear embedding approach. Furthermore, it incorporates mixed-integer linear inequalities to ensure safe overtaking maneuvers and avoid non-linear constraints imposed by obstacle boundaries. To assess the effectiveness of the proposed control and planning approaches, a small 1/10 scale electric car is utilized. Furthermore, simulations are performed in aggressive autonomous driving regimes, which involve steep curves and driving at maximum velocity.

Committee:
     PRESIDENT: HARO ORTEGA, GLORIA
     SECRETARI: SÁNCHEZ RIERA, JORDI
     VOCAL: PORZI, LORENZO
Thesis abstract: 3D scene modeling refers to the process of creating a digital representation of real-world environments for better understandingand further manipulation. Early works involved the utilization of pure structured representations such as meshes or voxel grids.The recent success of deep learning has enabled implicit representations of the scenes using deep neural networks, whichtypically are of high efficiency in automatically learning from data and handling high-resolution complex scenes. However, theseimplicit representations are often not explainable with an automatic feature learning process, hindering their further applicationsin practical scenarios. For example, to reduce potentially catastrophic failures in high-risky fields such as healthcare orautonomous driving, the uncertainty associated with the modeling results must be included into the decision-making process.Aiming at more interpretable 3D scene modeling, this thesis explores various methods to quantify the uncertainty in the processof 3D scene modeling that utilizes implicit neural representation. Firstly, we propose stochastic neural radiance fields to model aprobabilistic framework for capturing the uncertainty on rendered RGB images and estimated depth from the learned scenemodel, with a drastically reduction of model complexity compared to previous Bayesianbased probabilistic methods. To enhanceits capability of handling more complex scenes with varying geometry and appearance, we next extend the modelexpressiveness by incorporating a flow-based generative models to automatically learn arbitrarily complicated densitydistributions in a flexible manner. While these methods are able to achieve accurate 3D modeling with reliable uncertaintyestimation, they still suffer from the time-consuming optimization process, impeding their further applications in practicalscenarios. Therefore, we then explore a hybrid representation with incorporated voxel grids to explicitly learn the volumetricdensity of each 3D position, dramatically improving the model efficiency in both optimization and inference. So far, the currentframework is inherently designed for estimating predictive uncertainty for the areas of the scene that can be observed in thetraining images, and fail to output reliable uncertainty involving unobserved scene contents such as occlusion for roboticexploration in unknown experiments. For this purpose, we finally propose to introduce a 3D uncertainty field based on the trainedmodel, exhibiting consistently high uncertainty for predictions from those unobserved scene regions.In summary, we design an efficient and reliable probabilistic framework for 3D scene modeling to quantify the comprehensiveuncertainty associated with the predictions both from observed and unobserved scene regions. This framework provides a solidtool for uncertainty-aware decision-making and analysis in a variety of applications that rely on 3D scene modeling.

Committee:
     PRESIDENT: ANGULO BAHON, CECILIO
     SECRETARI: MORENO RIBAS, ANTONIO
     VOCAL: UGUR, EMRE
Thesis abstract: The physical world exhibits a wide range of obstacles to the application of robotics in a large number of tasks. Scripted behavior and/or teleoperated programs are still used in many industries (e.g. car assembly) because they are robust and reliable. However, their scope is limited and falls short in less controlled environments. We explore the possibilities of task planning, or Artificial Intelligence (AI) planning, for solving tasks in a more flexible, non-scripted way. In contrast to motion planning, AI planning focuses on high-level decision-making, rather than on concerns such as computation of trajectories and dynamics. Task planning is a very powerful tool for virtual situated agents (e.g. videogames or web services). One of its main advantages is that it allows deliberative, explainable, and adaptive behavior as long as a reliable model of the environment is available. However, the physical world presents some challenges that make the application of AI planning more difficult. This thesis has the following objectives, each one tied to a different challenge: (O1) integration of AI and motion planning; (O2) handling the unintended effects of actions taken by the robot; (O3) performing tasks even when the robot is not aware of all the relevant information; and (O4) automatic learning of action models to avoid the need for handcrafted ones. Our first contributions revolve mainly around objectives O1 to O3, which involve planning and acting. We propose hierarchical paradigms of planning, exploitation of topological properties of a problem for simplifying Markov Decision Processes, and planning alongside physical simulators to minimize the risk of unintended effects. The second part of our contributions focuses on O4, and consists of different algorithms for learning and recognizing STRIPS action schemata. Published results and findings are provided to support each contribution, alongside examples from manipulation scenarios, such as automatic disassembly of electromechanical devices, and socially assistive interactions.

Committee:
     PRESIDENT: PALACÍN ROCA, JORGE
     SECRETARI: SANFELIU CORTES, ALBERTO
     VOCAL: RIBAS XIRGO, LLUÍS
Thesis abstract: This dissertation aims to investigate two crucial tasks in the field of signal and image processing: signal denoising and image enhancement. Firstly, for signal processing, we propose three innovative denoising algorithms tailored specifically for one-dimensional EEG signals. These algorithms combine the strengths of deep learning and traditional signal processing techniques to effectively adapt to various noise types associated with different cognitive tasks, thereby enhancing the quality and accuracy of the signals. Secondly, in the domain of image processing, we introduce three novel image enhancement algorithms designed to tackle multiple noise types in natural scenes. By integrating deep learning methodologies with prior knowledge, these algorithms enhance image sharpness, contrast, and detail reproduction, demonstrating adaptability and reliability across different lighting, weather conditions, and photographic equipment. Lastly, we conduct a comprehensive analysis of the similarities and differences between image enhancement and signal denoising tasks. Through comparing the methodologies employed by each in handling diverse noise types, we derive meaningful conclusions to guide future research. These contributions are poised to significantly advance technological capabilities and theoretical understanding in both domains.