Spatial object detection and environmental understanding are fundamental aspects of autonomous driving in mobile robots. In this work, a stochastic approach to 2D LiDAR and stereo camera sensor fusion for object detection is presented. By tracking LiDAR clusters 360° around the robot and continuously updating knowledge of their corresponding object types using Bayesian inference when they are within the camera’s view, uncertainties in the detection algorithm are addressed. Additionally, localisation and sensor uncertainties are accounted for by using Gaussian distributions to model the locations of entities. By estimating the state of these entities, dynamic objects can be tracked throughout the environment, while remembering their type from when they were last visible in the camera’s view. The results demonstrate correct linking of camera detections with LiDAR clusters, with an average positional inaccuracy of objects of 0.12m in simulation, and 0.25m across different experiments on the robot, as well as the ability to track humans even when they move out of the camera’s field of view with an average tracking error of 0.33m in simulation.

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.

Piping and Instrumentation Diagrams (P&IDs) are graphical representations utilized in chemical engineering plants. Due to confidentiality reasons and legacy drawings, these diagrams are sent in PDF format. Piping engineers need to make a material take-off (MTO), a document containing all the components of a P&ID from these drawings. Today, this is done manually, which proves to be time-consuming and laborious. A piping engineer spends approximately 36 hours per 10 P&IDs, with an average of 500-1000 P&IDs per project. Given the expertise and value of process engineers, this manual counting process incurs substantial costs and a repetitive workload. Consequently, there is a growing motivation to automate this process.

In response, this thesis introduces an innovative deep learning model, PandID-Net, designed specifically for P&IDs. PandID-Net uniquely integrates symbol detection, line detection, and text recognition into a single model, diverging from previous methods that relied on separate models and rule-based techniques. It is the first method that uses deep learning for the line detection task in P&IDs. This all in one approach not only simplifies the processing pipeline but also enhances computational efficiency in detecting and pinpointing symbols, lines, and text, as well as their interrelationships.

The optimal configuration of PandID-Net is found by an ablation study where the performance of individual components is tested in isolation. This optimized configuration is then evaluated and benchmarked against a prior study by Paliwal et al. on the same dataset. PandID-Net achieves a performance in F1 scores of 92.89 and 94.48 for line detection and keypoint detection respectively

Robots are increasingly deployed in various locations to automate tasks, including in barns. However, in barns cows can obstruct the sensors such as LiDAR or camera, leading to a lack of environmental information. As a result, the robot’s localization system only relies on odometry at those moments, introducing additional uncertainty to the robot’s pose. When the visibility of the environment is restored, the robot may mistakenly believe it is in a location that does not correspond to its actual position.
The first contribution of this master thesis is a novel method for improving the self-localization in barns by implementing a line detection algorithm which is called Line Adaptive Monte Carlo Localization (LAMCL). The novelty is that only line segments are used to detect a localization error instead of corners between different line segments. It also retains the robot’s current pose to filter out fault-detected localization errors. In addition, the new approach is applied in a dynamic environment. The proposed method combines the Split-and-Merge line detection algorithm with AMCL. The detected line segments are compared with the walls in the environment to identify localization errors. When an error is detected, the robot’s pose is adjusted by placing a section of the particles at the location of the error. In this way, the robot can find its true location again.
The second contribution is a new dataset. This new dataset, called DataCow, consists of four recorded routes in a barn with GT on a handful of spots to evaluate the self-localization. DataCow includes the pseudo-2D LiDAR scans and the odometry of a robot driving through a barn. This dataset is used to evaluate the new self-localization method LAMCL. Through the experiments, it has been discovered that this new method improves the system’s recovery ability, but the accuracy and precision are compromised.

Many recent robot learning problems, real and simulated, were addressed using deep reinforcement learning. The developed policies can deal with high-dimensional, continuous state and action spaces, and can also incorporate machine-generated or human demonstration data. A great number of them depend on state-action value estimates, especially the ones in the actor-critic framework. Deriving unbiased estimates for these values is still an open research question, mostly since the connection between accurate value estimates and system performance is not yet well-understood. This thesis work has three main research contributions. Firstly, it analyzes the connection between value estimates and performance for the TD3 algorithm. Secondly, it derives theoretical bounds for the true value function when dealing with environments where a reward is only given for successful completion of a task (sparse/binary reward). Lastly, a deliberate underestimation objective is added to the TD3 algorithm together with the theoretical bounds to improve system performance when using human demonstration data that only covers a specific part of the state and action space. All the algorithms are tested and evaluated using simulated robot manipulation tasks in the robosuite environment, where the robot is first trained on the demonstration data and then can gather more experiences in the simulation. Results show that the deliberate underestimation together with the value bounds enable the robot to learn from human demonstration, which was not possible for the standard TD3. Additionally, applying just the value bounds speeds up the learning process when using machine-generated datasets.

Deterministic predictions of wave induced ship motion are increasingly often used to enlarge the oper- ational envelope of ships that perform complex operations at open sea. The large amount of complex operations performed nowadays, motivates the relevance of motion predictions. The company Next Ocean provides these forecasts, based on linear seakeeping theory. While highly linear degrees of freedom are predicted accurately, roll motion predictions prove to be more demanding and are less accurate. This research aims to improve roll prediction accuracy by adding (non)linear damping to the equations of motion. The prediction optimization is performed on data of the platform supply vessel Acta Auriga.

To enable the use of nonlinear forces, the equations of motions are first implemented in the time domain using Cummins’ equations. The linear coefficients in this equation are determined from the frequency dependent coefficients in Acta Auriga’s hydrodynamic database. To ensure the correctness and investigate the limitations of the implemented Cummins equation, the results from the time and frequency domain are extensively compared. Both monochromatic excitations, as well as spectrum (JONSWAP) excitations are considered. In contrast to linear force coefficients, nonlinear force coefficients are generally not known for a given vessel. Implementing these forces into the equations of motion imposes the need for an approximation of these coefficients. Estimates for the linear, quadratic and cubic viscous damping forces are made, using both the measured motions of a vessel operating at sea and the predictions of the corresponding excitation force, made by Next Ocean. A multivariate regression algorithm is used for the identification.

The Cummins equation was successfully implemented. However, the translation from frequency to time domain was more error prone than anticipated; the frequency dependent coefficients need to meet requirements that are typically not met by databases meant for frequency domain calculations only. Furthermore, degrees of freedom without a restoring force showed running away behaviour that could not be negated without adding extra damping. The roll motion was susceptible to instabilities. The exact origin of this instability was looked for, but could not be found. Adding damping resolved the instability. Furthermore, the interpolation in the roll response amplitude operator introduced errors in the initialisation of time domain calculations, which led to inaccurate results when spectrum excitations corresponding to more severe sea states were used. Only under specific conditions, these high energy spectra led to accurate roll predictions. Adding damping made for a close match between frequency and time domain calculation for all degrees of freedom. The complexities mentioned made that an easily scalable algorithm could not be obtained using time domain calculations. Easy scalability be- ing important to Next Ocean means that running time domain simulations in Next Ocean’s product is deemed unrealistic. This also means that no nonlinear forces can be used in real time applications.

In an attempt to improve linear predictions, coefficients in the linear seakeeping model, including the linear damping coefficient, are identified and vessel motions are re-predicted with the added damping and updated linear force coefficients, using the Cummins equation. None of the identified parameters led to better predictions than were obtained by Next Ocean. Identifying and updating parameters was therefore concluded to be not beneficial to the quality of the motion predictions. Adding a fixed amount of linear roll damping, which was not identified from the field data, did lead to improved prediction quality; correlations between predictions and measurements increased by 9.6%.

While the motion prediction could not be improved using the Cummins equation, valuable information on the time domain simulations was obtained. The finding that better predictions were consistently obtained by adding a fixed amount of damping, sparks an opportunity for further research into the ideal amount of damping. This, and more elaborate identification schemes could provide meaningful insight to improve motion predictions.

Single Molecule Localization Microscopy (SMLM) has enabled researchers to breakthrough the diffraction limit and obtain nanometer resolution images of macromolecular structures. But due to the time involved in obtaining ample data for proper image, the technique is venerable to many problem including fluctuations due to thermal gradients from surrounding which cause the frames to drift. SMLM relies on the stochastic blinking of fluorophore probes. Thus drift in SMLM could be explicitly modelled as a stochastic state space process. These models could be used to perform drift correction. Two state space models are proposed relying on different properties of SMLM.

The first model utilizes shifting of underlying image structure. The state space model for this property is constructed using shift matrices. A system identification method along with image reconstruction method is also derived to form the drift compensation algorithm for this model. This algorithm is further developed to provide adequate performance within low computational time. The second model relies on the position of emitter molecules and utilizes linking or pairing of fluorophore probes in succeeding frame to obtain the output data. Drift compensation algorithm for this model is constructed using Prediction Error Methods (PEM) and Kalman (RTS) smoother. The drift correction algorithm for these two models are also bench-marked with existing algorithms to obtain insight into performance. Furthermore, other properties of these algorithms are explored using simulation dataset and recommendation are provided for improvement and further research.

In order to improve the autonomy of construction robots, a Simultaneous Localization and Mapping (SLAM) solution is needed that localizes the robot using solely on-board sensing with sub-5 mm accuracy. As pointed out by the results of the Hilti SLAM Challenge, the state-of-the-art in SLAM currently does not offer a solution that comes close to this requirement. The winning algorithm FAST-LIO2 achieved an accuracy typically around 4-10 cm. When considering the literature, it can be noted that the state-of-the-art SLAM algorithms are designed with generality in mind, and limited attention has been paid to SLAM that focuses on increasing accuracy by targeting a specific scene structure.

In order to improve SLAM accuracy, this thesis proposes a novel 2D Lidar SLAM algorithm focused on indoor environments, hereafter called Ray-SLAM. The algorithm is centered around the assumption that walls are available, which is valid in many cases on the building construction site. Ray-SLAM introduces several novelties which are (1) a sparse map representation to facilitate a joint pose-map optimization scheme, (2) observation-to-map alignment using a non-iterative procedure (contrary to the Iterative Closest Point algorithm) and (3) a stop-and-go strategy to prevent Lidar motion distortion to corrupt the map.

The algorithm was extensively tested using six real-world indoor datasets recorded with the Ouster OS0 Lidar in rooms with various shapes and sizes and with a total trajectory length of 307 m. The motion was constrained to the horizontal plane and a stop-and-go motion pattern was applied. Ground-truth was recorded at static points to evaluate the accuracy. In the majority of the test cases, Ray-SLAM was able to estimate the trajectory successfully. Failure cases led back to either a lack of unoccluded walls in the scene or violated model assumptions about these walls (e.g. incorrect referencing to clutter close to the wall or to a door being opened).

Ray-SLAM's accuracy was compared with two 3D Lidar SLAM algorithms, LOAM and FAST-LIO2 respectively. The compared algorithms solved a harder 6DOF estimation problem, but had access to 64 Lidar scan rings instead of one and (optionally) the IMU. An overall Mean Absolute Error of 5.7 mm is reported by Ray-SLAM in the successful cases, which is 5.0x more accurate than LOAM and 2.4x more accurate than FAST-LIO2. Further research is suggested to improve the robustness of Ray-SLAM and extend it to a full 3D SLAM algorithm.

This thesis proposes a novel algorithm, Wi-Closure, to improve computational efficiency and robustness of map matching in multi-robot SLAM. Current state-of-the-art techniques connect maps with inter-robot loop closures, that are usually found through place recognition. Wi-Closure decreases the computational overhead of these approaches by pruning the search space of potential loop closures, prior to evaluation by a typical place recognition algorithm. Wi-Closure achieves this by identifying where trajectories are close to each other through sensing spatial information directly from the wireless communication signal. Then, place recognition is only performed on scans taken at locations close to each other. Wireless sensing provides information even when operating in non-line-of-sight or without existing communication infrastructure. The validity of Wi-Closure is demonstrated in simulation and hardware experiments. Results show that using Wi-closure greatly reduces computation time, by 54% in simulation and by 77% in hardware, compared with a multi-robot SLAM baseline. Importantly, this is achieved without sacrificing accuracy. Using Wi-Closure reduces absolute trajectory estimation error by 99% in simulation and 89% in hardware experiments. This improvement is due in part to Wi-Closure’s ability to avoid catastrophic optimization failure that typically occurs with classical approaches in challenging repetitive environments.

To enable communication for patients who have lost the ability to speak due to severe neuromuscular diseases, covert speech based brain-computer interfaces (BCIs) might be used. These system use neural signals arising from covert speech and translate them into text or synthesised speech. Covert speech is imagining to speak without moving any of the articulators and therefore does not rely on actual motor activity. As recognizing covert speech from neural signals is extremely challenging, machine learning algorithms are deployed. To make use of the full potential of machine learning approaches in the field of decoding covert speech and to accommodate real-world deployment of a BCI, a large number of training samples is required to train the networks.
In this study, a novel database is presented containing EEG and audio data from 20 subjects recorded during the covert and overt pronunciation of 15 Dutch prompts. To validate the recorded data, two speaker-independent classification tasks were performed using a ResNet-50 algorithm as classifier with spatial-spectral-temporal features extracted from the EEG signals. The speaker-independent three-class classification of pre-stimulus (rest) trials versus covert speech trials versus overt speech trials obtained an average accuracy of 70.6% and the speaker-independent five-class classification of five covert vowels (“aa”, “ee”, “oo”, “ie”, “oe”) obtained an average accuracy of 19.6%. Even though the five-class classification task did not reach an above chance level accuracy, the high performance reached by the three-class classification task provides support of the existence of discriminative information in the covert speech segments to decode covert speech in the future.
Future research should focus on EMG artifact detection and on determining the performance per subject to improve the dataset. Furthermore, subject normalisation strategies should be investigated to address the challenges of subject-independent covert speech decoding.

To improve Active Vibration control methods like Independent Modal Space Control (IMSC), an estimate of the modal contribution of vibrational eigenmodes of compliant structures is required. In this work, three Recursive Bayesian input and state estimation algorithms previously introduced in civil engineering are evaluated for use on on high-tech compliant mechanisms to estimate modal contributions for use in Active Vibration Control. The difference in environment when applying these algorithms from civil engineering into high-tech compliant motion stages allows for a significantly different sensor configuration possibly improving estimation performance.

The three algorithms, namely, the Augmented Kalman Filter (AKF), Dual Kalman Filter (DKF) and Gilijns de Moor Filter (GDF) are implemented in simulation and on an experimental compliant motion stage setup. The modal contributions of the guidance flexures are estimated through a noisy acceleration measurement at the stage and strain measurement at the flexure base.

For validation, the filters are evaluated on overall fit quality and system acceleration dependency through the quality of estimation of the flexure tip location. The DKF is shown worst performance with lowest fit scores overall and high acceleration dependency while the AKF and GDF perform comparably well. The GDF is shown to transmit less noise into the estimate however, and thus performed overall best.

Accurate vehicle localization is considered to be a key element in future automated driving systems. A network of multiple sensors is employed to deliver information for this localization process. Loosely coupled integration of global navigation satellite systems (GNSS) and inertial navigation systems (INS) data is a common sensor-fusion method for such positioning. One of the problems of this approach, is that exact knowledge of the process- and measurement noise covariance matrices is often not available. The GNSS measurement noise uncertainties, in particular, are highly dynamic and, depending on the specific environment, might follow a non-Gaussian distribution. Since particle filter are known to be superior in non-Gaussian environments, a hybrid filtering variant is proposed: adaptive particle-aided cubature Kalman filtering. This algorithm compromises between a particle filter with kernel density estimation algorithm in periods of non-Gaussian GNSS noise, and a standard cubature Kalman filter in case of Gaussian GNSS noise. The results of GNSS/INS-based localization simulations indicate that the proposed adaptive particle-aided cubature Kalman filter outperforms traditional filtering methods in terms of minimal localization errors.

Pose estimation provides accurate position and orientation information of the intelligent agents in real time. The accuracy of the estimation directly affects the performance of sequential tasks such as mapping, motion planning, and control. EKF (Extended Kalman Filter) is a standard theory for nonlinear pose estimation by modeling state uncertainty to Gaussian distribution. However, EKF has requirements for proper initial estimate and system noise to obtain bounded optimal estimate. Meanwhile, model nonlinearity and non-gaussian noise modeling affect the performance of EKF significantly in practical applications. In this thesis, we focus on improving the performance of nonlinear pose estimation by reinforcement learning. By formulating an EKF measurement update as a Markov Decision Process (MDP), reinforcement learning agents can be trained to learn the estimator gain through data samples and executed as the online estimator for pose estimation tasks.

Based on the above idea, we propose a novel reinforcement learning-compensated EKF estimator (RLC-EKF), where the RL agent serves as a second-time measurement update that subsides the residual error from the standard EKF estimate. The estimator is developed and testified on two specific pose estimation scenarios. Firstly, as a continuous work from the previous study, a framework for 3 DOF orientation estimation using inertial sensor and magnetometer is replicated. Then, the framework is extended by different RL algorithms training and multi-scale robustness validation. Besides, we implement the estimator on a feature-based 2D plane localization framework. The proposed framework shows the feasibility of the underlying algorithm on a localization task with a known map. As a result, the RLC-EKF estimator gives superior performance and convincible robustness compared to classical methods in severe conditions such as varying initial states, degree of noise intensities, and model covariance.

A common problem in robotics is the simultaneous localization and mapping (SLAM) problem. Here, a robot needs to create a map of its surroundings while simultaneously localizing itself in this map. An unknown environment is assumed. Traditionally, it has been approached through filtering solutions. This paradigm has shifted to pose graph optimization (PGO). This method scales well with large maps and is fast and accurate. Furthermore, it is especially suited to the distributed SLAM problem as existing distributed optimization methods can be leveraged. One such method is the alternating direction method of multipliers (ADMM), which has been used in distributed PGO. ADMM has a simple implementation and can achieve high accuracies in distributed PGO. A solution of good quality can be acquired in a few iterations with ADMM. However, ADMM is slow to converge to high accuracies. This thesis introduces an algorithm which implements Nesterov acceleration in the ADMM algorithm for distributed PGO in SLAM problems. Such an implementation will be novel. To create the proposed Nesterov accelerated ADMM (N-ADMM) algorithm, the current literature is adapted and extended based on the choices made in this thesis. The main research question is how to make these choices. The proposed N-ADMM algorithm is implemented in C++ and compared with unaccelerated ADMM and the state of the art. N-ADMM has shown better performance in some scenarios. To further research what these scenarios are characterized by, two models are introduced to create new datasets of which the parameters can be controlled. The effects of graph size and bad initial guesses are investigated.

The free energy principle is a recent theory that originates from the neuroscience. It provides a unified framework that combines action perception and learning in the human brain. This research aims to implement the perception aspect of the free energy principle into robotics. This is achieved via the dynamic expectation maximisation (DEM) algorithm. DEM is derived from the free energy principle and provides a novel solution for state and input observation for LTI systems under the influence of coloured noise. This thesis provides an overview of how the DEM observer is derived from the free energy principle. Thereafter, an experimental design is presented in which data is collected to validate the performance of DEM on experimental data. The data is collected from a quadcopter drone flying in wind conditions. A detailed overview of the DEM observer settings is presented and motivated. \\ This research shows that because of its use of generalized coordinates DEM is able to leverage the coloured nature of the noise for better state estimation. This is demonstrated by the fact that DEM obtains a higher state estimation accuracy than other coloured noise state observers, such as state augmentation and SMIKF. Moreover, in input estimation, DEM is able to obtain similar results as an unknown input observer. Finally, the accuracy vs complexity trade off of DEM is highlighted.

The ever-increasing industry desire for improved performance makes linear controller design run into its fundamental limitations. A nonlinear controller, such as Reset Control (RC), is needed to overcome these. RC is promising since, unlike other nonlinear methods, it easily integrates into the PID design framework preferred by industry. Thus far, closed-loop behaviour of RC has been analysed in the frequency domain either through Describing Function analysis or by direct closed-loop numerical computation. The former method computes a simplified closed-loop RC response by ignoring all harmonics, an approach which literature has found to inflict significant modelling errors. The latter method gives an accurate solution but does not provide understanding of how open-loop RC design affects closed-loop performance. No methods link these aspects, which impairs RC design and tuning. The main contribution of this work is aimed at providing this link, while achieving an accurate closed-loop RC model. A novel approach for modelling RC is considered, which uses state-dependent impulse inputs. This approach is shown to permit an accurate computation of closed-loop RC behaviour starting from an open-loop model, thus linking both aspects, enhancing system understanding. A frequency-domain description for closed-loop RC is obtained, as needed for the PID design framework, which is solved for analytically by inserting several well-defined assumptions. This solution is verified using a simulated high-precision stage, critically examining sources of modelling errors. The accuracy of the proposed method is further substantiated using controllers designed for various specifications.

Attitude estimation is one of the basic processes of the attitude and orbit control system hosted by every satellite. The spacecraft should be orientated properly during the orbit, according to the mission objectives. The estimation is based on a plethora of sensors, like rate gyroscopes, star trackers, sun and earth sensors, magnetometers (for low earth orbits). However, during the orbit the spacecraft operates according to some protocol, which is characterized from its different modes. During safe mode conditions, gyroscopes and/or star trackers might be turned off for energy saving reasons. Then the satellite senses only the geomagnetic field's direction and directions of other celestial bodies. In addition, the motion of a rigid body is described from well-known dynamic equations up to model uncertainties. The goal is to estimate how the satellite faces the Earth at each instant of time, by utilizing the two aforementioned sources of information. Limited research has been undertaken for this problem; on the contrary, the vast majority of published works provide solutions to the problem at the presence of rate sensors. Within a general non-linear framework, the problem of orientation and rate estimation can be set under the Bayesian formulation of estimation. The available tools provided by this approach, are the widely-known Kalman-Based filters. An alternative compelling approach considers the dual optimal control problem thus being more compatible with the geometric nature of the problem. This approach is also more promising, since it overcomes some critical difficulties of the former one and it is applicable in a coordinate free fashion. In this thesis, the aforementioned observation is emphasized and looked into. Therefore, two specific filters are studied: the second-order-optimal minimum energy filter on the special orthogonal group; and the predictive filter on, which is derived on this work. The two filters perform very well with the predictive filter to be less tolerant in measurement noise. Both filters are analyzed and simulated in various scenarios and a comparison is demonstrated, given particular emphasis to their structure. Finally, the predictive filter on is extended and used in an attempt to provide an alternative approach for the problem of the orbital position estimation by utilizing a kinematic model for the Earth's geomagnetic field. Towards this direction it is proved that under some realistic assumptions this later problem is equivalent to two separate problems and therefore the resulting filter can be decomposed in two distinct filters interconnected in series. The optimal correction term which is determined through the estimation process provides information regarding the orbital rotation rate.

The visualization of objects within or beyond a turbulent medium is hampered by the aberrations the medium induces in the wavefront. The sharpness of an image is maximal when the incoming wavefront is flat, with aberrated wavefronts yielding distorted images of limited utility. In the particular case of astronomy, the flat wavefront of the light from a distant target is aberrated as it moves through the atmosphere, before reaching our telescopes. The field of Adaptive Optics (AO) is dedicated to the correction of these aberrations, with the goal of enabling the imaging of objects through turbulent media with as much detail as possible. Due to a delay inherent to the AO control loop, each control input is used to correct an aberration slightly ahead in time, which means that, for the correction to be effective, the controller must be equipped with predictive capabilities. Prediction demands a model, and fortunately, the dynamics of atmospheric turbulence can be effectively modelled using a linear system. This makes the Kalman filter, the statistically optimal state estimator for linear systems, a natural choice for prediction. However, the steady-state Kalman filter requires a solution to the computationally intensive Discrete-time Algebraic Riccati Equation (DARE), which precludes its application to large-scale systems. Furthermore, the Kalman filter assumes a precise model with particular properties is available, which often is not the case in practice. The advent of extremely large telescopes demands prediction algorithms that can handle likewise extremely large system dimensions, which means that application of the steady-state Kalman filter, in its conventional form, is unfeasible. This thesis proposes a data-driven approach to Kalman filtering that exploits the intuitive sparsity pattern of the matrices involved in the modelling of an AO system. The developed algorithm replaces the DARE and, with knowledge of the system matrices, estimates the Kalman gain from measurement data, with the exploitation of sparsity providing a substantial drop in complexity when compared to the DARE, and with its data-driven nature inherently compensating, to a certain extent, for modelling errors. This thesis further proposes a two-stage approach to reduce the burden of the online prediction operation. While the measurement and state-transition matrices of the AO system are sparse, the Kalman gain is generally dense, consequentially slowing down prediction, which should accompany the sampling period of the loop. Our proposal splits prediction into a sparse stage for prediction of local structures in the wavefront phase, and a low-dimensional dense stage for prediction of the remaining low-frequency aberrations. The two-stage predictor is inherently suboptimal, but allows for quicker prediction and identification for extremely large systems. Via tuning, the user strikes a balance between performance and execution time.

This work addresses the problem of exploration and coverage using visual inputs. Exploration and coverage is a fundamental problem in mobile robotics, the goal of which is to explore an unknown environment in order to gain vital information. Some of the diverse scenarios and applications in which exploratory robots can have a significant impact include search and rescue missions, environmental monitoring, and space exploration. Specifically, we focus on the aspect of finding areas of interest, also referred to as targets, in the environment. In this thesis, we propose a deep reinforcement learning based approach relying solely on visual observations. In particular, our method builds upon the off-policy, actor-critic, Importance Weighted Actor-Learner Architectures (IMPALA) framework by including a set of novel auxiliary tasks, i.e. Pose Estimation and Local Map Prediction. These auxiliary tasks are inspired by Simultaneous Localization and Mapping (SLAM) approaches to exploratory robotics problems. The intuition is to assist internal representation learning and build locale specific knowledge by teaching the agent to predict its position and orientation, as well as transfer the visual information to information about the its local proximity. Experiments conducted in the DeepMind Lab simulation environment show improved performance over the base IMPALA agent and demonstrate the effectiveness of these auxiliary tasks. Furthermore, we investigate the performance of the agent, trained through various stages of curriculum, compared to a human controlled agent. The trained agent is shown to outperform the human in the majority of tested scenarios.

Misalignment in prosthetic legs can lead to bad posture, back pain and stump problems. A big influence therein is the alignment in the frontal and transverse plane of a prosthetic knee. This work aims to create an auxiliary tool for the prosthetist to align the prosthesis in an optimal way for the patient, with the use of inertial sensors. The goal is to estimate the current alignment in the frontal and transverse plane of the individual’s knee and therewith identify the changes that should be made to achieve an optimal swing of the shank.
A forward kinematic model of the swing phase of a prosthetic knee is combined with an inverse kinematic model to estimate the adjustment setting of the individual’s knee. This is done with the data from two inertial sensors on thigh and shank of the prosthetic leg of the patient. Furthermore, the desired alignment that creates an optimal swing phase is estimated. With the comparison of the estimated and the desired alignment, an adjustment proposition is calculated. In addition, a sensitivity analysis on the sensor-body orientation is conducted. The results for the current alignment setting show a rough accumulation around the expected linear trend. Deviations and outliers are explained with mistakes during the measurement and errors in the data processing. Also, the calculation of the optimal alignment angles and proposed changes show promising results. The results for the sensitivity analysis on the sensor- body orientation show a linear trend. However, the slope is much smaller than the expected 1. This means that disturbances in the sensor-body orientation have a smaller influence on the results of the estimated alignment angles than assumed. The influence on the adjustment angle in the transverse plane is even smaller than on the one in the frontal plane. These results lead to the conclusion that there are additional factors with an impact on the calculations.The basis towards a working algorithm is laid out. Future work on eliminating sources of error in the data processing is suggested. Among other things, a robust approach to define the walking direction has to be established. Further, an additional measurement with a motion capture system is recommended to create a better foundation for further analysis.