This thesis presents a novel navigation framework established to enable the exploration of planetary subterranean areas with Unmanned Aerial Vehicles (UAVs). The key contributions of this thesis work form a robot-safe rapid navigation framework that utilizes a novel bifurcating frontier-based exploration approach. UAVs (limited to quadrotors in this work) have superior navigation capabilities compared to ground robots in terms of 3D navigation as well as fast and versatile Traversability. Utilizing this advantage, this thesis investigates exploration and path-planning problems and presents novel mission behavior-oriented exploration strategies that are evaluated through either simulation with true physics and atmospheric models of planetary bodies or real-world deployment in subterranean areas.  The work included in this thesis is focused on two main research directions. The first direction establishes a novel coaxial quadrotor design that can operate in the thin atmosphere of Mars and utilize the Mars Coaxial Quadrotor (MCQ) to develop an energy-efficient exploration algorithm that leads to autonomously map Martian underground lava channel through true atmospheric model-based simulations. While the second direction establishes a Rapid Exploration Framework (REF) for the real-world deployment for the exploration of GPS-denied underground environments with UAVs. The contributions in the two directions are merged to develop a field-hardened autonomous exploration pipeline for UAVs that focuses on maintaining the heading vector of the UAV towards the most unknown area ahead of the UAV. While also bifurcating the exploration problem in local and global exploration for rapid navigation towards the unknown areas in the field of view and quickly globally re-positioning to a partially explored area. For navigating to the exploration goal of the UAV, it utilizes an expendable grid-based risk-aware path planning framework (D$^{*}_{+}$) that explicitly models unknown areas as risk and plans paths in safe space and for local obstacle avoidance and control the framework utilizes Artificial Potential Fields (APF) and a nonlinear Model Predictive Control based reference tracking scheme.Based on the learnings from field experiments and limitations of state-of-the-art grid-based planning methods on large-scale maps, the final contribution of the thesis establishes a Grid + Graph oriented Traversability-aware exploration and planning framework. The graph-based exploration method proposed in this thesis utilizes geometric shapes to define local traversable paths for the UAV to navigate to the local exploration goal. While utilizing a traversable graph that incrementally plans paths to the edge vertex of sub-maps in the direction of the global re-position goal. The strategy is evaluated extensively in simulations in subterranean urban, tunnel, and cave environments while it is also tested in real-world deployment at test mines of EPIROC and LKAB in Sweden.

Self-driving cars are steadily becoming a reality by a growing number of driver assistance functions enabled by smart perception sensors. The light detection and ranging (lidar) sensor show great potential for perception tasks due to its precise distance measurements. In order to take advantage of the high precision of a vehicle-mounted lidar, its position relative the vehicle needs to be calibrated. This is known as extrinsic calibration. The aim of this thesis is to investigate how to perform the extrinsic calibration of a static vehicle-mounted lidar in a static environment. In addition, the aim has been to develop a tool for running and customizing calibration simulations. The simulation tool CARLA, with its Python application programming interface (API), was chosen and developed to perform lidar simulations in a created virtual factory environment. The chosen calibration method uses a single vehicle-mounted lidar, targeting three spherical targets, whose known centre points act as reference points for the calibration. From the lidar point cloud a calibration algorithm is applied to find the position of the lidar. The algorithm estimates the centre points of the spherical targets and finds the lidar position by aligning the estimated centre points with the reference centre points. The algorithm includes functions that preprocess, cluster, fit spheres and perform point-to-point iterative closest point (ICP). The calibration method showed promising results in terms of point alignment and lidar position estimations. From 1000 simulations of random lidar positions, the average root mean square error (RMSE) of the point alignment was 0.33 mm with a standard deviation of 0.091 mm. The average absolute error of lidar position estimations was for translation [1.0 mm, 0.40 mm, 2.0 mm] with standard deviation [0.20 mm, 0.29 mm, 0.58 mm], and for rotation [0.11°, 0.11°, 0.10°] with standard deviation [0.098°, 0.098°, 0.094°]. Results also showed that uncertainties in the form of noise and point density have an impact on the accuracy of the calibration method. The developed simulation tool in CARLA can be used by other engineers at Veoneer to run customized calibrations or investigate other autonomous driving applications. Another conclusion is that the calibration method provides fast and accurate computations making it a potential candidate for extrinsic calibration.

This report presents various methods for solving the inverse kinematic problem for a non-conventional robotic manipulator with 6 degrees of freedom and discusses their respective advantages and disadvantages. Numerical methods, such as gradient descent, Gauss-Newton and Levenberg-Marquardt as well as heuristic methods such as Cyclic Coordinate Descent and Forward and Backward Reaching Inverse Kinematics are discussed and presented, while the numerical methods are implemented and tested in simulation. An analytical solution is derived for the Saab Seaeye eM1-7 and implemented and tested in simulation. The numerical methods are concluded to be easy to implement and derive, however, lack computational speed and robustness. At the same time, the analytical solution overcomes the same issues, but will have difficulties in singularities. A simple path planning algorithm is presented which plans around singular intervals, making it viable to use the analytical solution without encountering problems with singularities.

As the power distribution system grows and more sensors are added, more data is created every day. This data can be crucial for finding faults, but there is now so much data that it ends up being unused. This presents a valuable opportunity to gain crucial insights into the continuously expanding and increasingly complex power distribution system.  This thesis aims to utilize this valuable resource by finding a feature extraction method that can find valuable features in real-world data, use these features to cluster the data, separate different faults into different clusters, and develop a method for how these clusters can be classified, making it possible for an expert to classify large amounts of data quickly.  In the end, an autoencoder was used for the feature extraction. The features could be used to cluster both labeled and unlabeled real-world data. The clustering also made it possible to find errors in the labeled data, as the data from one class were clustered into two clusters. A method was developed that allowed the clusters of 32454 unlabeled datapoints to be accurately classified in approximately 30 minutes.  This thesis has successfully developed a method that can be used to get insights from large amounts of data, helping experts within the field of power engineering build the power distribution system of the future.

Accurate and robust perception of surrounding objects of interest, such as onroad obstacles, ground surface, curb and ditch, is an essential capability for path planning and localization in autonomous driving. Stereo cameras are often used for this purpose. Comparably, 3D LiDARs directly provide accurate depth measurements of the environment without the need for association of pixels in image pairs. In this project, disparity is used to bridge the gap between LiDAR and stereo cameras, therefore efficiently extracting the ground surface and obstacles from 3D point cloud in the way of 2D image processing. Given the extracted ground points, three kinds of features are designed to detect road structures with large geometrical variation, such as curbs, ditches and grasses. Based on the feature result, a robust regression method named least trimmed squares is used to fit the final road boundary. The proposed approach is verified with the real dataset from a 64-channel LiDAR mounted on Scania bus Klara, as well as the KITTI road benchmark, both achieving satisfying performances in some particular situations. / Exakt och robust perception av omgivande föremål av intresse, såsom hinder på vägar, markytor, trottoarkanter och diken, är en väsentlig förmåga för vägplanering och lokalisering vid autonom körning. Stereokameror används ofta för detta ändamål. I jämförelse, 3D LiDAR ger exakta djupmätningar direkt av miljön utan att behöva matcha pixlar i bildpar. I detta projekt används skillnaden för att överbrygga klyftan mellan LiDAR och stereokameror, och därmed effektivt hitta markytan och hinder från ett 3D-punktmoln genom 2Dbildbehandling. Givet att markytan har hittats, tre typer av funktioner undersöks för att upptäcka vägkonstruktioner med stor geometrisk variation, som trottoarkanter, dike och gräs. Baserat på funktionsresultatet används en robust regressionsmetod, least trimmed squares, för att passa den slutliga väggränsen. Det föreslagna tillvägagångssättet verifieras med två dataset med data från 64-kanalig LiDAR, en från Scania-bussen Klara och KITTI, och uppnår tillfredsställande prestanda i vissa givna situationer.

When pharmaceutical companies develop new drugs or vaccines there are large amounts of data in the form of images that need to be analysed, and any automation of that process is helpful to reduce time. These analyses could be for example concentration, decomposition, or classification, and essential to all these is high-quality particle localisation and segmentation. Therefore this thesis will focus on semantic segmentation of images of particles and viruses from a TEM microscope.  Various kinds of CNNs have been shown to give promising results in this area, however, there is still a need for improvement of these methods. Therefore it is interesting to see if combining a modern CNN with a graph-based model would improve its performance.  This thesis proposes a combination of a CNN, a modified version of the U-net, and a graph-based method, the CRF-RNN. The CRF-RNN is appended to the modified U-net and they are merged to create an end-to-end trainable segmentation network. This is then compared to using the standalone modified U-net to see if the CRF-RNN improves the accuracy.  Various loss functions, activation functions, and dropout rates are then tested to see which gives the best results under the given conditions. For both models, the best out of the tested hyperparameters were the dice loss and no dropout layers at all. For the standalone modified U-net the optimal activation function was the sigmoid function. However, for the network with the CRF-RNN addition, both softmax and sigmoid were good in different aspects.  Experiments done show that the model with the CRF-RNN addition performs slightly better than the model without, based on both measured metrics and visual inspection of the predicted outputs. Therefore it can be concluded that the CRF-RNN does improve the network it is attached to in this case. There is still much that could improve the networks though, for example parameter tuning of more hyperparameters using a GA.

Robotic manipulators are used for industrial automation and play an important role in manufacturing industry. Increasing performance requirements such as high operating speed and motion accuracy conflict with demands on heavy pay-loads and light-weight design with reduced structural stiffness. The motion control system is a key factor for dealing with these requirements, particularly for increasing the robot performance, improving safety and reducing power consumption. Most industrial robot control systems rely on current and angular position measurements from the motors, meaning that the actual controlled variable, that is the position of the robot’s end-effector, needs to be calculated using a model. Therefore, the mathematical model used for motion control must accurately describe the system’s dynamic behavior. Based on physics equations, the model contains unknown parameters that are usually identified from experimental data. This identification is a challenging problem, since the equations are nonlinear in the parameters, the system is highly resonant and experiments can only be done in closed loop with a controller.  Assuming a real robot is available for experiments, data-driven identification is common in order to obtain the most accurate description of the real system’s behavior. The method applied in this thesis estimates the dynamic stiffness parameters by matching the model’s frequency response function to the system’s frequency response, which is obtained from measurements done with the closed-loop robot system. The main focus of this thesis are strategies for increasing the process efficiency such that the time it takes to do the experiments is reduced, while the quality of the model is maintained or improved. Two strategies related to experiment design are presented: First, the number of quasi-static robot configurations for data collection is decreased by choosing the most informative configurations from a set of candidates. Second, less data-demanding methods for estimating the system’s frequency response are considered. The effectiveness of the presented approaches is demonstrated both in simulation and with real data.  If no robot is available for experiments, e.g. in the development phase, a model must be built based on specification data of components and other information available to the designer, such as CAD data. This thesis contains a modeling approach that derives a high-fidelity robot model of low order (lumped parameter model with few degrees of freedom) by combining results from test-rig measurements of isolated components with carefully reduced finite element models of the robot’s structural parts. / Robotmanipulatorer används för industriell automation och de spelar en viktig roll inom tillverkningsindustrin. Ökande prestandakrav som hög hastighet och noggrannhet hos robotens rörelse står i konflikt med trenden att bygga lättviktsrobotar som kan hantera tunga laster och som samtidigt är säkra för att jobba nära människor. Robotens styrsystem är en nyckelfaktor för att hantera dessa krav, särskilt för att öka robotens prestanda, förbättra säkerheten och minska strömförbrukningen. I de flesta tillämpningar är styrsystemets uppgift att säkerställa att robotens hand gör den önskade rörelsen, d.v.s. att handens position och hastighet motsvarar användarprogrammet. Positionen och hastigheten hos robotens hand är inte mätbara med sensorerna som är inbyggda i vanliga industriella robotar, vilket gör att de måste beräknas med hjälp av en matematisk modell. Denna modell måste beskriva det komplicerade sambandet mellan robotarmens rörelser och de motorer som orsakar rörelsen. Modellen är baserad på fysikaliska samband och innehåller okända parametrar som vanligtvis tas fram med hjälp av mätdata. Det som mäts är position och moment hos robotens alla motorer och det som är eftersökt är parametrarna relaterat till robotens styvhet. Metoden som används i denna avhandling tar fram styvhetsparametrarna genom att matcha modellens frekvenssvarsfunktion med det uppmätta frekvenssvaret för den verkliga roboten. Huvudfokus är strategier för att öka processeektiviteten så att tiden det tar att utföra mätningarna minskar, samtidigt som modellens kvalitet bibehålls eller förbättras. Två strategier presenteras: Den första minskar antalet robotkonfigurationer för mätdatainsamling genom att välja de mest informativa konfigurationerna från ett antal kandidater. Den andra strategin bygger på mindre datakrävande metoder för att skatta robotens frekvenssvar. Effektiviteten av de presenterade strategierna visas både i simulering och med verklig mätdata. Att få fram en bra matematisk modell är svårt om ingen robot är tillgänglig för mätningar, t.ex. i utvecklingsfasen av en ny robot. I så fall måste en modell byggas baserat på specifikationsdata för komponenter, t.ex. leverantörens information om växellådans styvhet, eller materialegenskaper för robotens struktur-delar. Styvheten av robotens strukturdelar kan beskrivas mycket noggrant med den så kallade finita element-metoden som delar strukturen i små delar och kombinerar ekvationerna för varje del till ett stort ekvationssystem. Detta ekvationssystem måste reduceras för att vara användbart i styrsystemets robotmodell. Denna avhandling innehåller ett modelleringssätt där man får fram en noggrann robotmodell genom att kombinera en reducerad styvhetsbeskrivning av robotens strukturdelar med specifikationsdata för komponenter. / <p><strong>Funding:</strong> Vinnova competence center LINK-SIC.</p><p></p><p>2023-05-04: ISBN (PDF) has been added in the E-version.</p>

Fjärrstyrning av industrirobotar är en efterfrågan som ökar inom den svenska industrin. Detta är en efterfrågan dels för att följa mål 9 i Agenda 2030 att hålla personal borta från farlig arbetsmiljö. Studien kommer att genomföras på företaget IndustriAutomation (IA) i Sandviken och framöver vara ett underlag för arbetsuppdrag.Lösningar som togs upp i studien gäller både mjuk- och hårdvara med fokus på lösningar som fungerar med KUKA som robotleverantör. KUKA UK har tidigare tagit fram en joystick-lösning som har ett högt pris, den ligger som grund till tanken bakom studien där det bland annat undersöks om den ekonomiska kostnaden är hållbar kontra andra alternativ.Intervjuer med företag, en undersökning av marknaden och samtal med kunder genomförs för att ta fram de lösningar som används i studien. Fokus läggs på lösningar som innebär att roboten kan fjärrstyras i drift.Lösningarna som studien tar upp håller sig mycket lägre i pris än KUKAs joystick och dessutom erbjuds det flera olika lösningar som kan komma att passa olika kunder bättre. Styrdon med haptisk feedback till användare för ett billigare inköpspris återfinns bland alternativen. Kostnadsberäkningen är genomförd på inköpspris och tar inte hänsyn till de arbetstimmar som kan tillkomma för att leverera lösningen till kund, vilket innebär att det i slutändan kan vara mer lönsamt med KUKAs joystick. Den lösningen innefattar allting, installation och setup. I stort sett en plug-in lösning för en robotcell att komma åt fjärrstyrning.

What if stents could be easier removed when complications arise or removed once theirjob has been done? No matter how perfect a stent design becomes, the fundamental issueis that a foreign object in the body under extended periods eventually will cause devicedegradation and complications. So instead of making a long-term solution of a perfectstent, the focus in this paper is redirected from expensive materials to a simple design,with simple materials. The goal of this master’s thesis was, therefore, to design a stentthat incorporates bistability with the use of origami so it has the possibility of beingremoved after actuation. The result is a family of designs where the bistable behavior,cylinder roughness, and cylinder length are some of the parameters that can be adjusted.The design family is made out of a cylinder body based on Waterbomb origami and twobistable star elements at the cylinder’s endpoints. This design’s bistability is observedmathematically, with a projection approach, and then in simulation by doing an FEAAbaqus. Prototypes are manufactured with laser-cut designs and an assembly consistingof laying two stif l ayers w ith a n a dhesive l ayer i n b etween. T hese p rototypes a re thenused in an experimental setup that results in a clear display of the bistability, scalability,and adaptability of the design. The paper also gives a step-by-step guide in reproducingthe designed stent and suggestions for further research where applying the design toa medical setting is the most prominent one. With the design’s very fexible nature,further study also includes other actuation uses that could beneft from this radially andlongitudinal expanding bistable actuator.

This report describes the progress made in designing and constructing an unmanned surface vehicle. The focus is on the development of the vessel's perception system, which includes three LiDAR sensors and a camera. To process and integrate the data from these sensors, a sensor fusion architecture was implemented using a method called Point Painting. This approach involves labelling the points of a point cloud with class labels by projecting the pixels from a semantically segmented image onto the cloud. Since the initial publication of the algorithm, more modern segmentation networks have been invented. To improve the model's accuracy, the segmentation network is replaced with a more modern architecture. Using the CityScapes dataset, the implementation achieved accuracy comparable to the state-of-the-art, which is competitive with the state-of-the-art models for semantic segmentation.