Automatic 3D reconstruction of birds can aid researchers in studying their behavior. Recently there has been an attempt to reconstruct a variety of birds from single-view images. However, the common murre's appearance is different from the birds that have been studied. Moreover, recent studies have focused on side views. This thesis studies the 3D reconstruction of the common murre from single-view top-view images. A template mesh is first optimized to fit a 3D scan. Then the result is used to optimize a species-specific mean from side-view images annotated with keypoints and silhouettes. The resulting mean mesh is used to initialize the optimization for top-down images. Using a mask loss, a pose prior loss, and a bone length loss that uses a mean vector from the side-view images improves the 3D reconstruction as rated by humans. Furthermore, the intersection over union (IoU) and percentage of correct keypoint (PCK), although used by other authors, are insufficient in a single-view top-view setting.

This thesis attempts to solve the problem of estimating the extrinsic camera parameters (pitch and roll) from a wide-angle view image. The first contributionis a data generation pipeline capable of producing wide-angle distorted images with rotation and line segment annotations. This pipeline was used to produce four datasets with distortion and rotation in the range −5◦ to 5◦. The second contribution is two neural networks aiming to estimate the roll and pitch angles, one where line segments are used, and one where ResNet and DenseNet features are used. The roll and pitch angles are predicted both directly and with vanishing points as an intermediate representation in both networks. The line segment network managed to extract line segments from distorted images, and predict the roll and pitch angles with a mean error of 3.70◦ over all datasets. The network with features from ResNet and DenseNet performed the best with a mean angle error of 1.02◦ over all datasets.

This thesis shows the development of a quadruped platform inspired by existing quadrupled robot designs. A robot by the name of Mjukost was designed, built, and tested. Mjukost uses 12 Dynamixel AX-12a smart servos and can extend its legs up to 19 cm with an operating height of 16 cm. All the custom parts in Mjukost are ether 3d printable or easy to manufacture, and the total estimated cost of Mjukost is around 900$. Mjukost has a simple control system that can position its body freely in 6 DOF using an inverse kinematic model and walk on flat ground using an open-loop walking algorithm. The performance experiments show that its slow control loopcauses difficulties for the robot to follow precise trajectories, but its still consistent in its motions.

Robots are increasingly expected to go beyond controlled environments in laboratories and factories, to enter real-world public spaces and homes. However, robot behavior is still usually engineered for narrowly defined scenarios. To manually encode robot behavior that works within complex real world environments, such as busy work places or cluttered homes, can be a daunting task. In addition, such robots may require a high degree of autonomy to be practical, which imposes stringent requirements on safety and robustness. \setlength{\parindent}{2em}\setlength{\parskip}{0em}The aim of this thesis is to examine methods for automatically learning safe robot behavior, lowering the costs of synthesizing behavior for complex real-world situations. To avoid task-specific assumptions, we approach this from a data-driven machine learning perspective. The strength of machine learning is its generality, given sufficient data it can learn to approximate any task. However, being embodied agents in the real-world, robots pose a number of difficulties for machine learning. These include real-time requirements with limited computational resources, the cost and effort of operating and collecting data with real robots, as well as safety issues for both the robot and human bystanders.While machine learning is general by nature, overcoming the difficulties with real-world robots outlined above remains a challenge. In this thesis we look for a middle ground on robot learning, leveraging the strengths of both data-driven machine learning, as well as engineering techniques from robotics and control. This includes combing data-driven world models with fast techniques for planning motions under safety constraints, using machine learning to generalize such techniques to problems with high uncertainty, as well as using machine learning to find computationally efficient approximations for use on small embedded systems.We demonstrate such behavior synthesis techniques with real robots, solving a class of difficult dynamic collision avoidance problems under uncertainty, such as induced by the presence of humans without prior coordination. Initially using online planning offloaded to a desktop CPU, and ultimately as a deep neural network policy embedded on board a 7 quadcopter.

Image augmentation is a field that covers the subject area of altering existing data to create more for the use of model training processes. It may be seen as the practice of expanding upon existing data using a range of techniques that employ transformations to improve the diversity of training sets when applied to machine learning. In our case of image recognition, triplet loss is utilised to pair a reference image to a matching and non-matching input. However, since there are many single images, augmentation techniques are relied upon to expand upon our data set to improve the recognition of images and create true positives. True positives are created using standard augmentation techniques like perspective transformation, contrast, cropping, and more. Despite this, the same images may undergo other types of alterations, natural disturbances such as transformations and warping, that are not captured by standard augmentation techniques. Such instances constitute to the variance in identification. Therefore, the analysis of augmentations by artificial intelligence (AI) based recognition is proposed; AI is used in order to identify what contributes to realistic disturbances of single images that better imitate real-life transformations. Analysing existing standard image augmentation techniques should provide further insight within this scope, as to better determine ways of emulating natural disturbances, and the formulation of non-standard practices in tandem. How this is done is by the use of an image's identity, the pixels it's comprised of and their distributions. Through a methodology of inspecting image identities, the breaking down of augmentations, and the inquiry into practices of non-standard image augmentation techniques, we detect the variance in accuracy of generated models, analysing the comprised data sets. Our results show that augmentations improve accuracy on a basis of variance and divergence from the original image. Subsequent discussion expands upon the identities of images and how augmentations must still resemble true positives, with the potential of an augmentation gauged by its influence on the rate of growth and highest accuracy of a model. / <p></p><p></p><p></p>

We review recent research into trajectory planning for autonomous overtaking to understand existing challenges. Then, the recently developed framework Learning Model Predictive Control (LMPC) is presented as a suitable method to iteratively improve an overtaking manoeuvre each time it is performed. We present recent extensions to the LMPC framework to make it applicable to overtaking. Furthermore, we also present two alternative modelling approaches with the intention of reducing computational complexity of the optimization problems solved by the controller. All proposed frameworks are built from scratch in Python3 and simulated for evaluation purposes. Optimization problems are modelled and solved using the Gurobi 9.0 Python API gurobipy. The results show that LMPC can be successfully applied to the overtaking problem, with improved performance at each iteration. However, the first proposed alternative modelling approach does not improve computational times as was the intention. The second one does but fails in other areas. / Vi går igenom ny forskning inom trajectory planning för autonom omkörning för att förstå de utmaningar som finns. Därefter föreslås ramverket Learning Model Predictive Control (LMPC) som en lämplig metod för att iterativt förbättra en omkörning vid varje utförande. Vi tar upp utvidgningar av LMPC-ramverket för att göra det applicerbart på omkörningsproblem. Dessutom presenterar vi också två alternativa modelleringar i syfte att minska optimeringsproblemens komplexitet. Alla tre angreppssätt har byggts från grunden i Python3 och simulerats i utvärderingssyfte. Optimeringsproblem har modellerats och lösts med programvaran Gurobi 9.0s python-API gurobipy. Resultaten visar att LMPC kan tillämpas framgångsrikt på omkörningsproblem, med förbättrat utförande vid varje iteration. Den första alternativa modelleringen minskar inte beräkningstiden vilket var dess syfte. Det gör däremot den andra alternativa modelleringen som dock fungerar sämre i andra avseenden.​

The global transportation sector is currently shifting towards autonomous vehicles. This shift comes with challenges, such as; identifying obstacles, recognising its surroundings and acting safely based on these perceptions. To accomplish mentioned tasks, the vehicle is equipped with sensors, such as lidars and cameras. A lesser known, yet significant challenge lies in keeping these sensors clean from dirt and debris which tends to accumulate on the lens of the sensors when the vehicle is moving. This report investigates how lidar- and camera sensors can be cleaned more resource-efficient in comparison to the existing sensor cleaning systems on the market. The goal was to recommend a sensor cleaning system for the range of sensors of an autonomous heavy vehicle.The authors of the study developed and tested several cleaning methods which were evaluated among each other and existing systems, while considering a system perspective. The developed cleaning systems showed that enabling a low washer fluid consumption had a negative impact on the system’s scalability, durability, compactness and complexity, in comparison to the existing cleaning systems. When utilising a high-pressured fluid, the study found that a sweeping flat spray is more resource-efficient than a static cone spray, where the latter is being commonly used in conventional sensor cleaning systems. The concepts with a sweeping flat spray resulted in a fluid consumption 4-7 times lower than the best reference cleaning system. In the case of a lidar, when considering a system perspective, it is recommended to use two telescopic flat spray nozzles facing each other and placed in either corner of the lens. It is also recommended that the nozzles are activated one at a time and that fluid I sprayed immediately on activation and kept flowing during the entire stroke to achieve a shaving or ploughing effect on the dirt. This method of cleaning has been observed to be more resource efficient compared to the reference systems. The resource-efficiency of a sweeping flat spray exists for other lens sizes as well, such as cameras and headlamps, however the scaling effects need further investigating. Therefore, additional tests are suggested, such as stress tests to determine the long-term durability of the cleaning system. Additionally, more research is needed to understand the impact of dirt in different environments and how often the sensors need cleaning. This also includes investigating how dirty the sensors can become before losing functionality. / Den globala transportsektorn är på väg att skifta till autonoma fordon. Detta skifte medför flear utmaningar; som att göra fordonet medveten om dess omgivning, identifiera objekt och agera säkert baserat på dessa intryck. För att kunna utföra dessa uppgifter är fordonen utrustade med sensorer, såsom lidar och kameror. En mindre känd utmaning ligger i att hålla dessa sensorer rena från smuts som ansamlas på sensorernas lins när fordonet framförs. Denna rapport undersöker hur lidar- och kamerasensorer kan rengöras mer resurseffektivt i förhållande till befintliga sensorrengöringssystem på marknaden. Målet var att rekommendera ett rengöringssystem för sensorerna som krävs för autonom färd, nämligen lidar och kameror. Studien utvecklade och testade ett flertal rengöringsmetoder som utvärderades bland varandra och befintliga rengöringssystem, medan samtidigt ta hänsyn till ett systemperspektiv. De utvecklade rengöringssystemen visade att en låg vätskeförbrukning påverkade systemet negativt i aspekter som skalbarhet, hållbarhet, kompakthet och komplexitet, i jämförelse med the befintliga rengöringssystemen. Vid användning av högtrycksvatten fastställde studien att en rörlig platt stråle kan vara mer resurseffektiv än en statisk konisk stråle, där den senare är vanlig bland befintliga rengöringssystem. Koncepten med en rörlig platt stråle hade en vätskeförbrukning som var fyra till sju gånger lägre än närmaste referenssystem. Vid hänsyn till ett systemperspektiv resulterade det rekommendera rengöringssystemet i två teleskopiska munstycken placerade i motstående hörnor av linsen. En i taget utvidgar sig munstyckena samtidigt som de sprutar högtrycksvatten på linsen, därav möjliggörs en rörlig platt stråle och en resurseffektiv rengöringscykel. Att rengöra med en rörlig platt stråle anses även resurseffektiv när det gäller andra storlekar på linsen, såsom en kamera- eller strålkastarlins, däremot måste eventuella följder från skalningen undersökas i vidare arbete. Det föreslås även kompletterande tester, såsom stresstester för att kunna avgöra livslängden på systemet. Vidare, efterfrågas ytterligare undersökningar på inflytande av smuts i olika miljöer, samt hur ofta sensorerna behöver rengöras. Detta inkluderar även undersökningar kring hur smutsiga sensorerna kan bli innan de tappar funktionaliteten.

This study explored how clients with distorted data affected the Federated Learning process using the FedAvg and FedProx algorithms. Different amounts of the three distortions, Translation, Rotation, and Blur, were tested using three different Machine Learning models. The models were a Dense network, the well-known convolutional network LeNet-5, and a smaller version of the ResNet architecture. The results of the study successfully showcases how different distortions affect the three models. Therefore, they also show that the risk of local data distortion is important to factor in when picking a Machine Learning model for Federated Learning.

Military developments suggest that autonomous weapons systems will be the future ofwarfare. Therefore, it is important to understand how to define the concept and how peopleexpress themselves around it. This paper will analyze how important actors in Sweden talkabout autonomous weapon systems. The concept of how autonomous weapons is constructedand what diversities and similarities there are in the expressions of autonomous weapons willbe examined in this paper. The question is if there is a hegemonic status in any discourse?Actors' expressions, such as political parties and researchers as FOI, SIPRI and The SwedishPeace and Arbitration Society, will be investigated. Social constructivism and discourseanalysis provide theoretical tools for the analysis. The discourses are organized byoperational and financial, human involvement, legal field and political discourses.The analysis shows how important characters try to give meaning to the phenomenon, forexample economic, human involvement and fear, which affect the different discourses. Theconclusions indicate that the juridical discourse has reached a hegemonic status, because allactors connect autonomous weapons to juridical frameworks.

Perceiving the environment is a crucial aspect of autonomous vehicles. To plan the route, the autonomous vehicle needs to be able to detect objects such as cars and pedestrians. This is possible through 3D object detection. However, labeling this type of data is time-consuming. By utilizing a virtual environment, there is an opportunity to generate data and label it in a quicker manner. This thesis aims to investigate how well three selected state-of-the-art models perform on a synthetic dataset of point cloud data. The results showed that the models attain a higher average precision compared to a dataset from the real world. This is mainly due to the virtual environment’s simplicity in relation to the real world’s detail. The results also suggest that models using different representations of point cloud data have different capabilities of transferring knowledge to the real world. / Att uppfatta miljön är en avgörande aspekt av autonoma fordon. Till planera rutten behöver det autonoma fordonet kunna upptäcka föremål som bilar och fotgängare. Detta är möjligt genom 3D-objektdetektering. Att märka denna typ av data är dock tidskrävande. Genom att använda en virtuell miljö, finns det en möjlighet att generera data och märka dem på ett snabbare sätt sätt. Denna avhandling syftar till att undersöka hur väl tre valda state-of-the-art modeller utför på en syntetiskt dataset av punktmolndata. Resultaten visade att modellerna uppnår en average precision jämfört med ett dataset från den riktiga världen. Detta beror främst på den virtuella miljöns enkelhet i förhållande till den verkliga världens detaljer. Resultaten tyder också på att modeller som använder olika representationer av punktmolnsdata har olika möjligheter att överföra kunskap till den verkliga världen.