Artificial Neural Networks (ANNs) have emerged as a powerful tool for classification tasks due to their ability to outperform traditional methods. Nevertheless, their effectiveness relies heavily on the availability of large, varied, and labeled datasets, which are often not available. To counter this constraint, data augmentation techniques have emerged, leveraging existing data to generate additional, variant data. Extending these techniques to multi-dimensional time series data, such as the transportation mode detection data considered in this thesis, however, introduces challenges. In response, generative models such as Variational Autoencoders (VAEs) have shown promising advancements. In this context, this thesis investigates the application of the Iterative Hierarchical Data Aug- mentation (IHDA) algorithm for ANNs, which represents a VAE-based data augmentation technique. The IHDA method utilizes VAEs not only to generate new data samples but also to map existing data to a lower-dimensional latent space, which is then utilized for identifying samples that might require additional training. The proponents of this method, Khan and Fraz, reported an accuracy elevation for the considered transportation mode detection classifier from 83% to 92%. However, due to the absence of publicly accessible code for this algorithm, the initial step of this thesis involved implementing the IHDA algorithm. Further, this research proposed and incorporated advancements like the σ-VAE, aimed to improve the generative capacity of the VAE and to refine its latent space mapping. Additionally, the Kullback-Leibler (KL) divergence was introduced as a similarity metric, aiming to optimize the identification process of samples that require retraining. Unfortunately, the results reported by Khan and Fraz could not be reproduced in this study. Furthermore, despite the potential shown by the σ-VAE to improve the generative capacity and refine the latent space mapping, along with the enhanced sample identification through the KL divergence, these enhancements did not lead to an overall improvement in the IHDA algorithm. This was primarily attributed to the low generative performance of the VAEs utilized, which also hindered a thorough evaluation of the effectiveness of the IHDA algorithm. Given these outcomes, it is suggested that future work should focus on employing more complex VAE models with the potential to enhance their generative performance, which, in turn, could improve the IHDA algorithm’s overall effectiveness.

With the rapidly growing world population and improving living standards of upcoming economies, the demand for fresh food is increasing vastly. Greenhouses have proven to be very effective in increasing crop yield since they offer a sheltered environment that can be controlled. Greenhouse growers need to take care of many processes, among which are maximising the crop yield while using as little resources as possible and controlling the indoor climate. An approach to alleviating demand for experienced growers is to automate processes within the greenhouse that aid in setpoint generation and setpoint tracking. Existing literature has presented multi-level architectures to cope with the two different timescales on which the two greenhouse subsystems act, i.e., the crop and greenhouse indoor climate subsystem. Current approaches in literature employ model-based techniques, which need tedious calibration per considered situation due to the influences of the local weather, environment and highly non-linear underlying physical processes. Therefore, these model-based approaches are scarcely used in the horticultural sector. Thus, this thesis proposes a data-driven control scheme that deals with long-term crop production control while taking into account resource usage. The proposed control scheme uses temporal and functional decomposition of the overall problem to accommodate the different timescales on which the greenhouse indoor climate subsystem and crop subsystem act. With the focus on the entire growing season, this thesis emphasises the importance of the climate strategy for crop production control. The aim of this thesis is threefold. Firstly, a non-linear simulator model for the crop dynamics has been selected, adapted, and calibrated, using the Autonomous greenhouse challenge (AGC) dataset. Secondly, the overall crop production control problem has been decomposed into a hierarchical structure, including two subproblems with different objectives, acting on different timescales and using different system representations. The communication protocol between the two layers has been determined, and the setpoint tracking controller is established for the lower layer and tested on a benchmark temperature reference trajectory. Thirdly, the Data-enabled predictive control (DeePC) algorithm is leveraged for the crop production control problem and compared in simulation against a conventional Economic model predictive control (EMPC) setpoint generating controller. Both use the aforementioned established setpoint tracking controller for controlling the greenhouse towards the generated climate strategies. In simulation, it is shown that the DeePC setpoint generating controller results in a climate strategy that results in a lower crop yield by 1.51% when compared to the EMPC climate strategy and 0.54% less when compared to the predefined benchmark reference trajectory. Furthermore, the resource usage needed for the climate strategy proposed by the DeePC setpoint generating controller uses 1.75% more heating resources when compared to the EMPC setpoint generating controller and 11.1% more when compared to a predefined temperature setpoint trajectory. The DeePC climate strategy results in 7.92% and 5.27% less electricity use when compared to the EMPC and predefined benchmark climate strategy, respectively. The DeePC climate strategy CO2 usage is 11.4% less when compared to the EMPC climate strategy CO2 usage and 20.8% less when compared to the predefined benchmark climate strategy. The decrease in resource usage compensates for the lower crop yield of the DeePC climate strategy. According to the metric of net economic profit, the DeePC generated climate strategy outperforms the benchmark climate strategy and EMPC climate strategy.

Designing controllers for complex robots is not an easy task. Often, researchers hand-tune controllers for humanoid robots, but this is a time-consuming approach that yields a single controller which cannot generalize well to varied tasks. This thesis presents a method which uses the NSGA-II multi-objective optimization (MOO) algorithm with various training trajectories to output a diverse Pareto set of well-functioning controller weights and gains. The best of these are shown to also work well on the real Talos robot. The learned Pareto front is then used in a Bayesian optimization (BO) algorithm both as a search space and as a source of prior information in the initial mean estimate. This learning approach, which combines the two optimization methods, is capable of finding a suitable parameter set for a new trajectory within 20 trials and outperforms both BO in the continuous parameter search space and random search along the Pareto front. The few trials required for this formulation of BO suggest that it could feasibly be applied on the physical robot using a Pareto front generated in simulation.

The use of machine learning (ML), especially neural networks, in modeling control systems has shown promise, particularly for systems with complex physics. However, applying these models in safety-critical areas requires reliable verification and control synthesis methods due to their inherent complexity. Formal methods, using stochastic finite state models like interval Markov decision processes (IMDPs), provide a way to analyze and verify these systems against detailed safety and performance specifications defined using linear temporal logic over finite traces (LTLf). Abstraction of ML models into such IMDPs, allows the deriving of formal guarantees on the IMDP that carryover to the underlying ML model. This thesis focuses on designing a switched controller for a cart-pendulum system using neural network dynamic models (NNDM) by formal control synthesis, validating it through formal verification methods. The methodology includes modeling the system behavior under different controllers, abstracting these models into IMDPs, applying the respective formal methods, and validating the approach through experiments. The aim is to demonstrate the framework's utility in a practical context, comparing different neural network architectures and researching the applicability of formal guarantees to both the models and the actual system. The main contributions are a practical application of the framework to a specific system, a comparison of neural network architectures for dynamic modeling, and an experiment-based validation of the framework's effectiveness. It confirms that the formal guarantees for abstracted models are relevant to the actual system, providing insights into the framework's potential for real-world applications. The findings suggest areas for further research, particularly in making such frameworks more accessible for practical deployment in safety-critical systems.

Many autonomous navigation tasks require mobile robots to operate in dynamic environments involving interactions between agents. Developing interaction-aware motion planning algorithms that enable safe and intelligent interactions remains challenging. Dynamic game theory renders a powerful mathematical framework to model these interactions rigorously as coupled optimization problems. By solving the resultant coupled optimization problems to equilibrium solutions, the game-theoretic models explicitly account for the interdependence of agents’ decisions and achieve simultaneous prediction and planning. Coupled constraints between players, such as collision avoidance, can also be handled explicitly. However, most existing game-theoretic motion planning approaches rely on known objective models of all agents. This assumption presents a key obstacle to real-world ego-centric planning applications of these methods, where only local information is available. This thesis investigates solution approaches to relax this assumption and explicitly account for the ego agent’s uncertainty about other agents’ objectives while adaptively conducting game-theoretic motion planning. The main contribution of this work is an online adaptive model-predictive game-play (MPGP) framework that jointly infers other players’ objectives and computes corresponding generalized Nash equilibrium (GNE) strategies. These strategies are then used as predictions for other players and control strategies for the ego agent. The adaptivity of the proposed approach is enabled by differentiating through a trajectory game solver whose gradient signal is used for maximum likelihood estimation (MLE) of opponents’ objectives. Compared with existing objective inference solutions in dynamic games, the proposed approach handles general inequality constraints in games and further supports direct integration with other differentiable modules, such as neural networks (NNs). Two simulation experiments indicate that the proposed approach performs closely to solving games with known objectives and outperforms the game-theoretic and model-predictive control (MPC) baselines. Two hardware experiments further demonstrate the real-time planning capability of the planner and its real-world applicability. In addition to this main contribution, the second contribution of this work is a variational autoencoder (VAE) pipeline built upon the proposed differentiable game solver. This contribution aims at going beyond the point estimation in the first contribution and inferring potentially multi-modal beliefs about players’ objectives based on observations. The main idea is to employ variational inference (VI) to approximate Bayesian inference of players’ objectives. The variational autoencoder (VAE) framework is utilized for amortization to avoid per-sample optimization. Initial results on a single-player example show that after training, the proposed pipeline can: (i) generate a game objective distribution that resembles the underlying training data distribution and (ii) accurately predict a narrow, uni-modal posterior objective distribution when the observation is unambiguous based on seen data in the past and (iii) generate a multi-modal belief distribution of player’s objective to capture mostly likely modes in case of high uncertainty.

Sorting systems form an example of event driven systems. These types of systems are referred to as discrete event systems (DES), and they consist of jobs that need to be performed at available resources. In an autonomous sorting system, jobs consist of robots receiving and delivering parcels at the correct locations. With scheduling, optimal allocation of the jobs to those resources over time is computed, where the decisions that need to be made are routing, ordering and synchronization. The behaviour of DES is often described by non-linear models, but max-plus linear (MPL) systems are a class of DES that can be described by a model that is linear in the max-plus algebra. This algebra uses two operators maximization and addition. Allowing different routes and switching between orders of jobs extends an MPL system to a switching max-plus linear (SMPL) system. Robots in a sorting system often have many routes to choose from, and need to make order choices with respect to other robots in the system. In this thesis, a general SMPL model is made for the autonomous sorting system at software company Prime Vision, which can be applied to any sorting area design. The solution to the scheduling problem for the model results in a time schedule for the active robots at the correct locations in the sorting area, as well as the optimal decisions on routing, ordering and synchronization. The optimization problem is solved with a model predictive scheduling (MPS) approach and recast as a mixed integer linear programming (MILP) problem. The model is created in Python and the optimization problem is solved with Gurobi. The resulting schedule is visualized with a simulation, in which the decisions of the robots are clearly shown. An idea for implementation of the optimization into the sorting system is given as well.

Researchers have been interested in studying the connection between emotion and memory for decades but much remains unknown due to the elusive nature of the human brain. Furthering our understanding of the phenomenon is crucial for improving the treatment of neurological disorders associated with emotion dysregulation, as well as for enhancing autonomous systems that interact with humans, such as robotic prostheses and artificial intelligence. A promising approach to studying cognitive processes is developing computational models of brain activity, which have the potential to uncover the underlying neural mechanisms. This master's thesis presents a novel dynamical model of the neural activity underpinning the emotional associative memory encoding process. This type of memory is especially complex and poorly understood as it requires memorizing the relationships between various environmental stimuli that may elicit distinct emotions. To model this phenomenon, the thesis proposes a network model using the Dynamic Causal Modeling (DCM) framework, focusing on the Amygdala (Amy), the Hippocampus (Hip), and the Orbitofrontal Cortex (OFC) due to their central roles in emotional associative memory. Furthermore, the thesis provides an analysis of network state coordination and state error stability, deriving analytical bounds for state error dynamics that illustrate how the model's properties are linked to improved memory encoding. The dataset used to estimate the DCM model comes from a functional Magnetic Resonance Imaging (fMRI) study conducted by Zhu et al. at Donders Institute for Brain, Cognition and Behaviour, where participants were asked to memorize pairs of two emotionally potent images (group EE), two neutral images (group NN), or one neutral and one emotional image (group NE). The DCM model developed in this thesis reflects how the neural dynamics differ among the groups and the modeled brain regions, corroborating many of the behavioral results of the experiment and bringing novel insights. Based on stability and error dynamics analysis, the model of group NE is found to exhibit the best overall information flow, which is consistent with the experiment, where this group achieved the best memorization results. Furthermore, the coupling between Amy-Hip is shown to play a key role in improving memory encoding. Another important contribution of this thesis is the extensive overview of DCM in Chapter 2, which discusses the mathematical foundations, the system identification algorithm, criticism raised against the method, studies of its statistical validity, and advice on implementation. Additionally, Appendix B presents the first guide for DCM script development in the SPM12 toolbox in MATLAB. Furthermore, this thesis compares the performance of four different variations of DCM, proving that Stochastic DCM achieves superior results in experimental paradigms with brief emotional stimuli. The success of this approach is believed to stem from its capacity to handle model misspecification and neuronal noise. Overall, the Stochastic DCM developed in this thesis is considered to provide the most comprehensive representation of neural dynamics governing the encoding of emotional associative memory compared to other existing dynamical models.

There is growing interest to control cyber-physical systems under complex specifications while retaining formal performance guarantees. In this thesis we present a framework for formal control of uncertain systems under complex specifications. We consider dynamical systems with random disturbances, whose probability distribution is unknown. When it comes to the specifications, we focus on those given as syntactically co-safe linear temporal logic (scLTL) formulas. Such formulas resemble natural language, and allow us to reason over complex behaviours of the system. We follow an abstraction-based approach: we abstract the original system to a finite-state Markov model, in which the state discretization error as well as the distributional ambiguity, are embedded as uncertainties. To do so we make use of tools from optimal transport and ambiguity sets of probability distributions. After that, we synthesize a strategy for the abstraction and obtain probabilistic guarantees that the abstraction satisfies the specifications. Finally, we correctly refine the strategy to one that the original system can follow, and prove that the guarantees we obtained for the abstraction also hold for the original system. We propose two approaches to obtain the abstraction. First, we propose a data-driven approach to abstract the system into an interval Markov decision process (IMDP) when samples from the unknown distribution is available. Then we use already existing algorithms to synthesize a strategy for the IMDP. Secondly, we propose an approach to abstract the original system into a robust Markov decision process (robust MDP). This second approach is applicable to more general uncertainty models besides the data-driven one, and reduces conservatism of the abstraction. Furthermore, we propose an algorithm to synthesize robust strategies for robust MDPs, which also renders the guarantees that the abstraction satisfies the specifications. Finally, we demonstrate the usefulness of our proposed approaches through several case studies that involve both linear and nonlinear systems.

Stability, safety and optimality are often sought-after properties in the field of controller synthesis. In the last century, linear control theory has matured to a level where scalable algorithms are widely available that are able to synthesize controllers with stability and optimality guarantee. However, the synthesis of safe controllers largely remains an open question. Furthermore, when nonlinear systems are considered, these methods often result in sub-optimal performance, or fail to provide a stabilizing solution at all. Current nonlinear controller synthesis methods typically lack stability, safety and/or optimality guarantee, or require computationally expensive online optimization. This thesis presents a novel method that is able to overcome these limitations. The presented controller synthesis method combines techniques from Approximate Dynamic Programming, Lyapunov theory and barrier theory, to synthesize an optimal controller in the form of a Neural Network. Alongside the controller a second Neural Network is deployed, which is trained to serve as a certificate function to provide stability and/or safety guarantee. To assure the correctness of the procedure, Satisfiability Modulo Theory and Counterexample-Guided Inductive Synthesis are utilized. We subject the method to a number of case studies, through which the versatility of the method is demonstrated. We show the method is able to work with linear and nonlinear systems, as well as systems with input and state constraints. Furthermore, we show the method yields controllers that are able to outperform linear controllers in terms of cost minimization.

In the field of Systems and Control, optimal control problem-solving for complex systems is a core task. The development of accurate mathematical models to represent these systems’ dynamics is often difficult. This complexity comes from potential uncertainties, complex non-linearities, or unknown factors that might affect the system. Because of these challenges, there is a need for methods that can understand the dynamics using available data and control strategies that can work with such models without relying too much on expert knowledge or task-specific insights. These methods are essential for creating efficient and reliable solutions in a wide variety of applications within the discipline. The need for models that do not require expert knowledge has spurred the interest in applying machine learning methods to control problems. Probabilistic Inference for Learning COntrol (PILCO) is a model-based Reinforcement Learning (RL) algorithm known for its probabilistic approach to model-based RL. By employing Gaussian Process (GP) dynamics models, PILCO integrates uncertainties into its learning process, allowing it to derive control policies from limited data. PILCO’s use of the Squared Exponential (SE) kernel in its GP can restrict the learning capacity. Especially in higher-dimensional spaces, due to the SE kernel’s inherent smoothness assumption that might not capture complex or non-smooth dynamics effectively. The algorithm’s reliance on moment matching for approximating posterior distributions introduces another weakness, which can lead to inaccuracies in non-Gaussian or multi-modal contexts. These shortcomings may limit PILCO’s efficiency and scalability in more complex, higher-dimensional tasks or in situations where the underlying dynamics are not well-captured by the chosen kernel and approximation methods. This thesis introduces Deep Kernel PILCO (DKL PILCO), a novel framework that uses Deep Kernel Learning (DKL) for learning the dynamics, and the Unscented Transform (UT) to propagate the uncertainty. The effectiveness of this approach is demonstrated across various tasks, highlighting the potential of DKL and UT to enhance the scalability and efficiency of model-based RL methods such as PILCO, making it a promising candidate for real-world control applications.

The rapid advancement in autonomous driving technology underscores the importance of studying the fragility of perception systems in autonomous vehicles, particularly due to their profound impact on public transportation safety. These systems are of paramount importance due to their direct impact on the lives of passengers and pedestrians. Additionally, their reliability can be easily compromised given the complexity and unpredictability of driving environments. However, current research and existing regulations often fail to adequately address the adversarial robustness of autonomous vehicle perception systems. This thesis delves into the adversarial robustness of camera-based perception systems of autonomous vehicles. Our research concentrates on developing and implementing evasion attacks that use black-box gradient estimation, as well as physical attacks in traffic sign detection and classification systems. Our findings indicate that even minor perturbations can impact the accuracy of these systems, leading to detection and classification errors. This finding highlights a critical vulnerability in the perception system's robustness against adversarial attacks. Moreover, the study extends to assess the transferability of adversarial examples across diverse perception models. Our results also expose significant gaps in the current regulatory frameworks of autonomous vehicles, necessitating the establishment of more rigorous and comprehensive safety standards.

Over nine billion people will have to be fed fresh and healthy food in 2050 according to United Nations. This puts pressure on the horticulture sector, responsible for a large portion of the world’s food production. The literature has shown that automatic optimal control algorithms are able to make better use of resources and can even outperform growers in terms of resource efficiency. While applications of automatic optimal control methods show promising results, few are made adaptive to the greenhouse-crop system, assuming the system to be time-invariant. That is, the prediction models or techniques used in the controllers are not updated to adapt better to the evolving greenhouse-crop system. This thesis sets a first step towards this on-line system adaptation: it investigates the possibility to use a linear state-space model to estimate the non-linear greenhouse-crop systems dynamics around an operating point, for the purpose of automatic optimal climate control. In this work, a linear state-space model is identified and is implemented in an automatic optimal controller based on Model Predictive Control (MPC). Moreover, a greenhouse-crop simulation is built from state-of-the-art models based on first-principles. This ground-truth simulation is used to generate data and test the proposed control method.

Noise's impact on biochemical systems has long been a focal point of investigation, given its potential to compromise signal accuracy and disrupt system functionality. This paper conducts a comprehensive exploration into the noise characteristics within a set of signal differentiators recognized for their high precision. Noteworthy for their modularity, swift computation, and ease of implementation, these differentiators play a pivotal role in computing concentration changes and bear the potential to regulate the dynamics of biological systems. This study establishes a comprehensive simulation framework to examine the noise characteristics of these differentiators across diverse input signal scenarios. Furthermore, we also apply noise suppression techniques such as noise filters to mitigate excessive noise and enhance noise performance. Our findings reveal that these differentiators significantly amplify the system noise level, surpassing both the Poisson level and the original system noise level. Moreover, while noise filters demonstrate notable success in noise reduction, achieving Poisson-level noise without compromising signal integrity remains a challenge. This investigation yields invaluable insights into the noise properties of biochemical differentiators, shedding light on their inherent limitations. Additionally, it presents a viable pathway to enhance noise behaviour, thereby extending the scope of applications for these differentiators.

Socially compliant robot navigation in pedestrian environments remains challenging owing to uncertainty in human behavior and varying pedestrian preferences in different social contexts. Local optimization planners like Model Predictive Control can incorporate collision avoidance constraints, but they can only lead to socially compliant trajectories if the cost function embeds information about the desired social behavior. The same holds for Reinforcement Learning, where a sophisticated reward function needs to be defined. However, formalizing social behavior through a reward or cost function is difficult due to the complex nature of pedestrian behavior. Imitation learning allows for inferring the desired behavior by learning from human demonstrations, making them suitable for learning socially compliant navigation policies but without any safety considerations. In this work, we propose to learn a socially compliant navigation policy directly by observing surrounding pedestrians’ trajectories from a commonly available detection and tracking pipeline and combine it with a local optimization planner to enhance safety. A Subgoal Recommender policy is developed to guide the local optimization planner to generate socially compliant trajectories by providing intermediate subgoals. To adapt the policy to changing social contexts without forgetting previously learned information, we train the Subgoal Recommender in a Continual Learning setup exploiting new pedestrian data. We demonstrate in simulation that our method can learn a policy that has similar performance metrics as that of the observed trajectories with 95% confidence estimated from a t-test, resulting in a lesser number of collisions. Further, the policy can adapt to different social preferences exhibited by pedestrians, while being able to remember the learned behavior in a previously encountered social context. Furthermore, we show that our proposed method can learn navigation policies from actual pedestrian data recorded using the onboard perception pipeline of a Clearpath Jackal robot.

The platooning of trucks provides significant benefits in the existing transportation systems regarding social, economic and environmental aspects. Due to the platoon's high dependence on sensors, it is critically important to have fault-tolerant countermeasures that deal efficiently with failures by achieving minimal risk condition. This thesis tackles designing an emergency escape manoeuvre for a faulty truck that has lost its ability to monitor the driving environment while moving in a platoon formation on a highway. A fallback strategy is proposed that leads the faulty truck to park on the shoulder of the road and allows the rest of the platoon participants to continue their journey. To solve this problem, first, a functional platooning system should be designed. Therefore, a nonlinear bicycle dynamic model is employed, while a new Bidirectional (BD) Cooperative Adaptive Cruise Control (CACC) system is utilized for control purposes. The emergency escape manoeuvre has to be pre-computed and ready to be triggered when the failure occurs, as concerning defined specifications, limitations and constraints. It is generated using a quintic splines methodology, while a new optimization approach regarding the minimization of jerk peaks is adopted, enhancing the efficiency of the design procedure. Furthermore, a gap-closing controller is employed and tuned accordingly to reform the platoon after the faulty truck abandons the formation. Moreover, a suitable decision logic is defined to achieve smooth transition among regular performance and fallback state at both platoon and inter-vehicle level. The functionality of each component is firstly evaluated individually, based on specific performance metrics. Finally, all the mentioned components are put together and the fallback strategy that executes emergency escape manoeuvre in platoon context is derived. Several simulation tests illustrate proof of principle for the proposed framework. It is shown that the proposed strategy indeed achieves minimal risk condition for the faulty truck achieving minimal tracking errors. A global applied strategy is adopted for the gap-closing controller by making healthy follower(s) travel with the maximum possible acceleration. At the same time, the leading truck of the formation maintains a constant speed until the time gap between them becomes $0.8s$, before the switch back to the CACC, to complete this action as fast as possible and safety issues in the driving performance. Additionally, the CACC system is evaluated to ensure that the string stability criterion is satisfied in the mode of operation. Moreover, a virtual leader mechanism is introduced to fill in faulty truck's blind spot in the road that arise from the sensor failure. It is shown that it can provide a sufficient window of movement for the platoon before switching to the fallback strategy.