Robotics and Computer Engineering - Master's theses

Selle kollektsiooni püsiv URIhttps://hdl.handle.net/10062/42116

Sirvi

Nüüd näidatakse 1 - 20 114

INTUIT-VLNCE: Autonomous Navigation through Vision-and-Language
(Tartu Ülikool, 2024) Rodionov, Kirill; Roy, Kallol; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
The aim of this thesis was to developed an Embodied Agent with explicit intuition capable of navigating indoor Continuous Environments based on provided Natural Language instruction and Agent’s egocentric vision as part of a Vision-and-Language Navigation task. The thesis proposes creating explicit intuition by making an Agent predict not only an action to perform at a given time, but also predicting actions for the future. An Agent’s policy was trained mimicking the training procedure from LAW-VLNCE project [1]. Evaluations showed negative results after implementing proposed method.
VR-Enhanced Remote Inspection Framework for Semi-Autonomous Robot Fleet
(Tartu Ülikool, 2024) Reynes, Gautier; Valner, Robert; Norbisrath, Ulrich; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
This thesis presents the design and development of a Virtual Reality (VR)-enhanced user interface and communication infrastructure for remote inspection using a semiautonomous robot fleet. The core of this project is the creation of a VR interface that allows operators to immerse themselves in a digital twin of the remote environment, facilitating intuitive and efficient control over robot inspection. This interface supports both third-person and robot’s point-of-view perspectives, enhancing situational awareness and decision-making capabilities in hazardous environments. The software framework is built upon ROS 2 Foxy, and the VR application was designed with a new graphics engine called Wonderland Engine, particularly suited for lightweightWebXR experiences capable of running on a number of VR headsets, like the Oculus Quest 2. The communication between the interface and the robot fleet is tackled by a custom-built WebSocket server. The work is demonstrated using simulated robot scenarios in Gazebo. The demonstration serves as a proof of concept, showcasing the viability of the VR interface in a controlled environment and setting the stage for future real-world applications. This work contributes to the field of VR-enhanced remote inspection by providing an interface that bridges the gap between operators and remote environments. The integration of VR technology with robotic systems opens new possibilities for remote operation, offering a more immersive and intuitive control mechanism that can be adapted to various industrial and research applications.
Continuous Collaborative Mapping in Unknown Environments: A Multi-Robot System Approach
(Tartu Ülikool, 2024) Radigon, Malcom; Muhammad, Naveed; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Navigating and exploring unknown terrains remains a critical challenge within the field of mobile robotics. Achieving rapid and consistent exploration is crucial for the prompt generation of precise maps. This thesis introduces a cutting-edge distributed exploration system utilizing a multi-robot fleet. This innovative system is crafted to facilitate continuous exploration in unexplored areas by implementing a novel dronebased communication relay method. Additionally, it enables the synthesis of an integrated, comprehensive global map, providing crucial, rapid insights for human operators into the explored areas. The effectiveness of this system has been thoroughly evaluated through a series of simulated experiments encompassing various trials. These evaluations underscore the system’s capability in smoothly conducting exploration tasks, notably overcoming delays traditionally linked to communication challenges.
Small-scale cars for autonomous driving research
(Tartu Ülikool, 2024) Petrović, Uroš; Muhammad, Naveed; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Self-driving cars (autonomous cars, robo cars) are vehicles that can drive without human input. In the recent decades, this field of science has received a lot of attention. There is a growing need to improve the safety and effectiveness of traffic. Self-driving cars have the promise of being safe. The main reasoning is that by removing human error, and utilising algorithms, the number of mistakes and accidents that would arise in most situations could be drastically reduced. Autonomous driving is a heavily investigated area currently, but significant challenges remain open, thus requiring significant research during the decades to come. As the cost of life-size test platforms for autonomous-driving research is very high, the need to use smaller-scale vehicles to be employed as test platforms arises. This thesis investigates the viability of such a method, by evaluating the performance and suitability of a small-scale self-driving car platform, the Donkey Car, for different aspects of autonomous driving research. Perception, localization and mapping, planning and end-to-end capabilities of the platform were investigated and compared with publicly available real-life self-driving car experiments. The results achieved showed that small-scale cars are capable of performing autonomous driving research at an acceptable level in comparison with real-life cars.
PHOENIX: Revisiting Cloudlet Development with Recycled Phones.
(Tartu Ülikool, 2024) Ngoy, Perseverance Munga; Flores, Huber; Norbisrath, Ulrich; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
This work revisits the idea of re-purposing e-waste (aka old electronics) into cloudlets that can be used as general computing units for several applications, including edge computing, the Internet of Things (IoT), and pervasive computing applications. While the idea has been around for over a decade, the continuous evolution of personal electronics makes it difficult to keep up to date with general guidelines and principles for re-purposing. As a result, we investigate the latest developments in this topic and provide newer insights for recycling deprecated electronics using modern tools and frameworks. To do this, four Nexus 5 phones are adapted with PostmarketOS, a Linux-based mobile OS, and subjected to benchmark tests alongside a Raspberry Pi 4 and a laptop. Results indicate lower performance metrics in mobile phones, accompanied by substantial heat generation during intensive tasks. The study delves into the advantages, disadvantages, and techniques for optimization, including load balancing, task scheduling, parallel computing, cloud offloading, and energy-aware algorithms. Potential applications are explored, emphasizing community empowerment, low-cost microclouds, and disaster response. While mobile phone-based cloudlets show promise, the challenges and limitations must be addressed. The study underscores the environmental benefits of repurposing mobile phones and proposes sustainable practices, including the removal of batteries for a reduced carbon footprint. The thesis also advocates for the educational empowerment of underprivileged students through accessible technologies, highlighting the potential societal impact of this innovative approach to edge computing.
Using Machine Learning to Measure Digital Audiences
(Tartu Ülikool, 2024) Mison, Nathan; Badia, Josep; Sirts, Kairit; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
In recent years, the field of machine learning has witnessed an unprecedented surge in interest and application across diverse domains. Kantar Media, a leader in audience measurement, aims to apply machine learning to refine and revolutionize audience measurement techniques. This thesis demonstrates the application of recurrent neural networks and attention mechanisms. Despite existing solutions in audience measurement, they exhibit limitations in comprehensively addressing outlier data. The approach demonstrates that leveraging deep learning yields a remarkable test accuracy of 97 percent. However, it is also noted that certain outliers present persistent predictive challenges.
Development and Implementation of ESTCube-2 Star Tracker FPGA Design
(Tartu Ülikool, 2024) Komarovskis, Roberts Oskars; Allaje, Kristo; Eenmäe, Tõnis; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Attitude determination and control play a critical role in a spacecraft, where one solution for attitude determination can be star tracking. Star tracker systems can provide high-accuracy attitude information based on star identification, and such a system is also a part of the ESTCube-2 nanosatellite developed by the Estonian Student Satellite Foundation. One of the primary components in the ESTCube-2 star tracker is a field-programmable gate array (FPGA) responsible for processing the captured images. This thesis aims to develop and implement FPGA system design and necessary algorithms for star tracking and assess the performance of the obtained design. Different FPGA design components were made to interface with external devices and perform different functions related to star tracking.
Evaluation of image sequence (video) compression with an embedded processing system for future space applications
(Tartu Ülikool, 2024) Glorieux, Léon; Islam, Quazi Saimoon; Dengel, Ric; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
In recent years, the accessibility to Graphics Processing Unit (GPU)s and hardware accelerators for space missions has increased drastically with implementable Commercial Off-The-Shelf (COTS) solutions. As a result, the onboard computing power has increased for many space missions, and the performances of the different video compression methods have been improved. With such evolution in computing resources, it is crucial to analyze the different methods and record their performances, to find efficient compression methods in this new configuration. To find efficient compression methods, we developed a benchmark, to assess multiple methods on their compression ratios, power consumption, execution time and the difference of quality between the input and output images on a low resource configuration, adapted for space missions. From the analysis led in this thesis 4, on the standard codecs, the more lossless the compression, the higher the benefits from the acceleration. The machine learning approach shows promising results for the future, and the Consultative Committee for Space Data Systems (CCSDS) 122 despite the GPU acceleration was outperformed by Advanced Video Coding (H.264) and High-Efficiency Video Coding (H.265).
Physical A*: Graph-Based Search Algorithm for Robot Navigation On-the-Go
(Tartu Ülikool, 2024) Shrestha, Anish; Matiisen, Tambet; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Robot navigation is commonly viewed as a trajectory planning problem, relying on a preexisting map. However, the availability of a prior map can be problematic, especially in military or rescue scenarios. This thesis elaborates on the concept of a two-level planning and navigation algorithm called physical A* to address this problem, focusing typically in use cases where a prior map is not known. Physical A* is an A* graph traversal where the robot physically drives along the nodes of the graph. The graph is constructed on-the-go. A lower level planning component proposes multiple waypoints stored as graph nodes. A higher level planner, with a broader understanding of the geographical or spatial context computes the goal heuristic for these nodes. Based on the goal heuristic, the waypoint with least cost is selected to explore towards the goal. Physical A* mainly concentrates on exploring the waypoints that would lead the robot towards the goal in the most optimal form.
Optimal agent positioning for dynamic event monitoring and analysis.
(Tartu Ülikool, 2024) Naghma Afreen; Paul, Aditya Savio; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Events like natural hazards and celestial occurrences are dynamic in nature, they are unpredictable and change instantaneously. In order to study events better, it is essential to develop methods for better understanding. However, observing such events is challenging due to their changing nature compared to observation of events that do not change with time. Moreover, traditional camera systems developed for observing static events tend to generate results with less accuracy and information loss. Dynamic event observation requires dynamic setup of camera systems to improve the study of such phenomena. We propose a method for optimizing camera configurations for observing dynamic events. The observation space is sampled using a volumetric sampler. The cameras are iterated over these samples and the event is captured; an optimal solution then evaluated using a scoring criteria. The scoring function assists in analyzing the optimal configuration of camera setup for observing the desired event. The research aims to develop and improve observational techniques for dynamic events.
Asymmetric Deep Multi-Task Learning
(Tartu Ülikool, 2024) Maharramov, Ali; Matiisen, Tambet; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Recent developments make deep neural networks a valuable asset for autonomous driving. They can be deployed as an end-to-end system or part of more complex systems for specific tasks. If a system needs several tasks by neural networks, using multi-task learning (MTL) introduces few benefits compared to deploying several single-task learning (STL) models, such as better time and space complexity on deployment and potentially increased generalization on the backbone network. However, MTL often faces unique challenges. Many existing MTL datasets have limited labels or lack the required labels for specific tasks, and generating labels for these tasks leads to resource and time consumption for researchers. Training the model on an asymmetric labeled dataset, a dataset where labels for specific tasks are unavailable for a subset, can cause a biased gradient, reflecting an unbalance in the accuracy of tasks. In this thesis, asymmetric MTL were investigated and compared to symmetric MTL and STL methods.
Multi-Vehicle Path Planning using Shared Data
(Tartu Ülikool, 2024) Kozjutinskis, Aleksandrs; Roy, Kallol; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Vehicle navigation is a problem without a real solution. Every action taken can be disturbed by an accident, making the suggested path not optimal. As a result, the only moment it is possible to state that the selected road was optimal is after the vehicle has completed it and all possible disturbing events have occurred. Due to the high impact of the navigation system on traffic flow, this research suggests sharing information between vehicles to make decisions as a team, giving way for those who benefit most. It also allows them to renavigate vehicles if they face traffic jams or congestion on the road. The newly developed algorithm proved to be useful, overperforming other common path-finding strategies by 10-15% and providing results 5% close to the optimal path ll possible disturbing events have occurred. Traffic flows after the developed algorithm also provided a lower traffic jam rate compared to other algorithms. Algorithms were tested using the simulation developed for this research.
Exploring Smartphone-Based Reinforcement Learning Control for Educational Robotics: Implementation on OpenBot
(Tartu Ülikool, 2024) Gras, Lilou; Muhammad, Naveed; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
This research explores the feasibility of implementing Reinforcement Learning (RL) algorithms entirely on a smartphone to control an educational robotic platform, OpenBot. This study aims to determine if RL can be executed on Android smartphones without simulated environments and whether it would be accessible for students and enthusiasts as a practical RL project. Initially, Deep Q-Learning (DQL) and Policy Gradient (PG) algorithms were tested on standard RL scenarios, Cartpole and Pong. This allowed to gain insights on both algorithms and what to expect in a successful RL training. The policy gradient algorithm was then implemented entirely on the smartphone controlling OpenBot to drive across a track for 15 seconds. In general, after approximately 400 episodes of training using policy gradient, the agent was able to successfully navigate the track for the aimed 15 seconds in half of its attempts. Despite the encouraging results of the study, some technical challenges remain open, such as, exploding gradients, the randomness of weight initialization, and engineering challenges such as high battery consumption.
Integration of a High Speed Communications System into ESTCube-2
(Tartu Ülikool, 2024) Doğan, Melis; Allaje, Kristo; Allik, Viljo; Eenmäe, Tõnis; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
ESTCube-2 was a low Earth orbit CubeSat demonstrator that was launched on 9th of October, 2023. One of the main payloads of the satellite was the Earth observation payload, which was capable of producing images of tens of megabytes. Thus the spacecraft required a dedicated high speed communications system to downlink the images. At the beginning of the thesis, an S-band transmitter had been bought by the ESTCube team for this purpose. This thesis presents the integration of the HISPICO transmitter from IQ Spacecom into the ESTCube-2 platform. This integration involved writing low-level embedded device drivers to control the transmitter. The device drivers were written in C, using FreeRTOS on an STM32L4 microcontroller. Furthermore, the application level logic was created for the reception of images from the imaging payload, dividing the received images into RF frames, and adding forward error correction to the frames - a functionality that exists on HISPICO, but the ESTCube team were not given access to. For the ground station, a reception pipeline using a LimeSDR Mini and GNURadio Companion was created. To test the chain of communication, an image of 76.8 kB was transmitted over 318 HISPICO frames and was successfully received.
Validation of NoMaD as a Global Planner for Mobile Robots
(Tartu Ülikool, 2024) Allik, Robert; Singh, Arun Kumar; Kruusamäe, Karl; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
As autonomous mobile robots are becoming increasingly common in real-world applications, like warehouses and self-driving cars, so is the need for robust navigation methods growing. NoMaD, a vision-based navigation architecture, was recently presented and showed very good metrics in obstacle avoidance tested in ”challenging environments” but the exact level of challenge was unclear. This thesis aimed to measure that performance in a standardized way and implement an improvement by augmenting it with a LiDAR sensor. This new solution is based on ROS Navigation, where NoMaD acts as the global planner guiding the robot, leaving the obstacle avoidance to the local planner. Both NoMaD and the new solution were tested in environments inspired by the standardized navigation environment dataset BARN. While NoMaD reached the proclaimed success rate (90%) in simple baseline tests, it failed to do so in more complex environments, even when the hardware limitations of the setup were compensated for, with success rates ranging from 3.3% to 53.3%. The new solution, however, achieved all-around good results (83%) with no collisions. While it has its own drawbacks the new approach shows some merit.
ESTCube-2 Mission Control System: Preparation for In-orbit Operation
(Tartu Ülikool, 2023) Toomast, Cathy; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Space missions rely on the mission control system (MCS) for spacecraft health monitoring, commissioning, routine operations, and emergency procedures. However, often the MCS stays in the shadow of the satellite itself, creating a situation where the satellite is launch-ready, but the system necessary to operate the satellite from the ground is not. This is also the case with ESTCube-2, a satellite that started development in 2016 and is soon ready for launch. ESTCube- 2 is a spacecraft mainly developed by student volunteers, and its main mission is to demonstrate deorbiting with plasma brake technology [1]. To operate the spacecraft, there is a need for a functional mission control system. The MCS has been developed through the years, but when it was initially built, the satellite was not ready to be tested with the system. For this reason, the system is not yet widely used and has unresolved issues. The author of the thesis investigates and lists the actions that need to be taken to have an operational MCS by the start of the mission. Furthermore, to understand the needs of mission operating systems, the author used a qualitative research method, interviewing four people with experience with operating satellites. For the final system to be useful for the mission, the author made changes in the mission control system during the thesis and implemented suggestions from interviews with spacecraft operators. Initial tests on the ground were performed with the ESTCube-2 engineering model. Additionally, the author will list ideas and notes regarding what could be done better in the future at the end of the thesis.
XR Teleoperation Demo Development
(Tartu Ülikool, 2023) Zorec, Matevž Borjan; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
This thesis designs an educational real-time visual feedback teleoperation demonstration. The importance of a good user experience is highlighted while showcasing the feasibility of using open-source solutions such as Godot Engine version 4 for teleoperation setups. Reviewed literature narrowed design requirements, outlining that a representative teleoperation demonstration could provide a positive experience, intuitive movement control, direct real-time visual feedback for teleoperation and be open-sourced, with user and video stream evaluations as research objectives. Employing design thinking, 'RoverXR' is iteratively developed with M5 RoverC-Pro for movement and serving WebSocket protocol real-time Motion JPEG high-definition video from Raspberry Pi v2.1 Camera Module via a Raspberry Pi Zero. Custom MPV player and Godot scenes were prepared, featuring video stream playback and providing a virtual user interface on the Meta Quest 2 headset. User evaluation participants report a positive, engaging experience and provide helpful feedback, showcasing the potential of low-latency, high-quality video streaming, and virtual scene representation in teleoperation demonstrations for educational purposes.
Sampling-based Bi-level Optimization aided by Behaviour Cloning for Autonomous Driving
(Tartu Ülikool, 2023) Shrestha, Jatan; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Autonomous driving has a natural bi-level structure. The upper behavioural layer aims to provide appropriate lane change, speeding up, and braking decisions to optimize a given driving task. The upper layer can only indirectly influence the driving efficiency through the lower-level trajectory planner, which takes in the behavioural inputs to produce motion commands for the controller. Existing sampling-based approaches do not fully exploit the strong coupling between the behavioural and planning layer. On the other hand, Reinforcement Learning (RL) can learn a behavioural layer while incorporating feedback from the lower-level planner. However, purely data-driven approaches often fail regarding safety metrics in dense and rash traffic environments. This thesis presents a novel alternative; a parameterized bi-level optimization that jointly computes the optimal behavioural decisions and the resulting downstream trajectory. The proposed approach runs in real-time using a custom Graphics Processing Unit (GPU)-accelerated batch optimizer and a Conditional Variational Autoencoder (CVAE) learnt warm-start strategy and extensive experiments on challenging traffic scenarios show that it outperforms state-of-the-art Model Predictive Control (MPC) and RL approaches regarding collision rate while being competitive in driving efficiency.
Comparison of Water Detection Models for an Off-road Unmanned Ground Vehicle
(Tartu Ülikool, 2023) Rustambayli, Fidan; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Water hazards can cause unmanned ground vehicles (UGVs) to become stuck or break down during an autonomous mission, damage electronic components and sensors, and require costly repairs or replacements, making it crucial for UGVs to identify water hazards in real-time, determine secure path around them, or reduce their speed when appropriate to cross them safely. This thesis proposes a water detection system for UGVs in off-road environment. The proposed approach combines convolutional neural networks (CNNs) with transfer learning, leveraging their capabilities for effective water detection. The thesis includes a comprehensive review of traditional sensor-based methods and recent deep learning-based techniques. Real-world data collected in off-road environments are utilized to evaluate the proposed approach, and the method achieves a 0.50 Mean-IoU score and 92.74% accuracy on the test dataset. We also include a comparative analysis of the method with a previous deep learning-based semantic segmentation method for water detection. The comparison provides insights into the relative strengths and weaknesses of these approaches for water detection in off-road environments. Overall, this thesis provides valuable insights into the use of deep learning for semantic segmentation in challenging environments.
LEAN metoodika rakendamine aktsiaselts Chemi-Pharm tootmises
(Tartu Ülikool, 2023) Nigul, Mihkel Erich; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
LEAN metoodika baseerub Toyota tootmissüsteemil, mis loodi 20. sajandil Jaapani töötusinsener ja ettevõtja Taiichi Ohno poolt. Taiichi Ohno poolt loodud süsteem oli revolutsioon masstootmises, sest see võimaldas tõsta märkimisväärselt tootmise efektiivsust ja vähendada praaktoodete arvu. Toyota tootmissüsteem on tänaseni ajakohane, mistõttu on ka käesoleva lõputöö eesmärk Taiichi Ohno meetodite rakendamine kaasaegses tootmisettevõttes. Käesolevas lõputöös uuriti ja rakendati LEAN metoodika tööpõhimõtteid ja automatiseeriti ebaefektiivsed tootmisprotsessid. Lõputöö praktilises osas teostati villimisliinile parendustöid, kasutades PDCA probleemi lahendamise tsüklit. Kõigepealt fikseeriti algseis, uuriti villimisliini tööd ja kaardistati kitsaskohad, leiti probleemide juurpõhjused ja pakuti välja võimalikud lahendused. Seejärel viidi ettepanekud ellu, analüüsiti parendustööde kasu ettevõttele ja standardiseeriti lõputöö käigus lõputöö praktilises osas arendatud parendused.

Sirvi

Viimati lisatud