Confidence Estimation for Robot Skills
In a semi-autonomous setting where humans and robots work together and humans provide support to robots for difficult aspects of task execution (e.g. help with identification of relevant objects) transparency in the communication is crucial for trust of humans. Thus, a robot should be able to explain why it needs support by the human or it is currently not able to complete a given task. In order to allow the robot to ask for human support before it fails in executing a task, the robot needs an estimator about its confidence in completing a task. This is similar to human behavior. Starting from a moderately complex setting in navigation, methods for estimating the skill confidence will be investigated.
contact: steinbauernoSpam@ist.tugraz.at
Automated Reconfiguration of Production Lines
Production lines are usually configured manually and are quite rigid. In this thesis we would like to investigate how production lines can be automatically reconfigured for new products or a set of new products. The basis for this flexibility are smaller production islands that provide a particular production skill and can be relocated within a production line automatically. This thesis will be conducted in cooperation with the Smart Factory of Graz University of Technology, where prototypes of such flexible production islands and robots for automatically moving them around already exist. The challenge is to describe the product, the provided production skill, and their arrangement as optimization problem and solve it efficiently. Moreover, the reconfiguration should be executed using a mobile robot system.
contact: steinbauernoSpam@ist.tugraz.at
Scene Graphs for Offroad Navigation
Scene graphs are a rich way to describe a scenery and the relations of objects and regions in an image. Data sets like Visual Genome are the basis to learn models for the generation of scene graphs. Such datasets and models are usually tailored towards urban environments. In this thesis we like to investigate how scene graphs can be generated for complex natural scenes and if they can be used in context-based navigation in complex off-road environments.
contact: steinbauernoSpam@ist.tugraz.at
Software Engineering Workflow for Autonomous Robots
Autonomous robots are complex systems comprising a number of hardware and software components and showing quite complex behaviors. While in other domains like aerospace or automotive elaborated software frameworks and processes are well established this is hardly realized for the autonomous robot domain. In this thesis we like to investigate what good software development practices exists in other engineering domains and develop a concept for such a software development process to produce better and more reliable software. In particular we like to look into requirement engineering, automate testing of robots and its software, verification and software product lines. This thesis will be conducted in cooperation with Prof. Bernhard Aichernig who is an international expert for formal methods in software engineering.
contact: steinbauernoSpam@ist.tugraz.at
Intelligence Test for Autonomous Robots
In classical test theory (aka intelligent test) one aims to estimate the capabilities of individuals on a relative (or absolute) scale using a set if items (tasks) and the rate of solving items by the individuals. The test theory is a well established methodology and uses probabilistic approaches in order to consider the nondeterministic nature of task solving (lucky guesses and careless mistakes). The construction of good and discriminative items with different defined difficulty is a main challenge in this context. While this is well researched in the context of humans there is little work in the context of autonomous robots. We are interested in such items designed for robots because we also want to rate the capabilities of autonomous systems (e.g. robot A is more capable than robot B) but also for predicting how well a robot will perform in a new environment and task (new item). In this thesis in cooperation with researchers from psychology we like to develop an approach and a set of test items for tasks of autonomous robots such as navigation or manipulation in order to be able to apply test theory from psychology to robots. In a simulated environment proper items (environments and tasks) will be generated (automatically) and rated according to their difficulty for different robots. This set of validated items can later be used to rate and compare the capabilities of the different robots. This aspect is of increasing interest in the context if the certification pf autonomous systems.
contact: steinbauernoSpam@ist.tugraz.at
Weather Dependent Cost Representation for Routing of a Mobile Robot
Cost-based representations of the environment are frequently used in the path planning domain to obtain an optimized path based on various objectives, such as traversal time or energy consumption. However, obtaining such cost representations still heavily depends on the environment the robot navigates is, particularly in outdoor settings with diverse terrain types and slope angles. Furthermore, depending on the weather condictions, the conditions of the environment might change, and therefore also the capabilities of the robot on certain terrain. For example, while the robot might be able to traverse a grassy area on a sunny day, it might have problems after a heavy rainfall or when there is snow. This topic aims to address this problem by using weather dependent collected data of an outdoor environment to generate a cost-based representation fo the environment which a robot can use to navigate in. As basis, a data-driven approach (Deep Neural Networks) which generates a cost representation for various outdoor terrains on a dry day is provided. Using recorded data on a new weather condition (snow), the cost representation is to be adapted and the knowledge of the traversability of the terrain is to be transferred to the new weather condition.
contact: matthias.edernoSpam@ist.tugraz.at
Designing and Implementing a Road Follower for Driving on Forest Roads
This thesis explores the development of an autonomous driving system capable of navigating unpaved, unstructured forest roads. The project focuses on designing and implementing a "road follower" system that utilizes various sensor data and computer vision techniques to detect the roadway and steer the vehicle safely. This thesis tackles the challenges of safely detecting forest roads, maintaining awareness of the robot’s relative position on the road, and planning a safe and efficient path along the road in the absence of road/lane markings. The successful implementation of this system would contribute to advancements in autonomous driving technology, particularly in applications like forestry, search and rescue, and agricultural operations.
contact: matthias.edernoSpam@ist.tugraz.at
Global Robot Localization using Aerial Photogrammetry and On-Board Sensing
The increasing demand for autonomous robots in various sectors, like agriculture, search and rescue, and delivery services, necessitates robust and reliable localization techniques. This project investigates the fusion of aerial photogrammetry and on-board sensing for global robot localization, overcoming the limitations of traditional methods like GPS in challenging environments. In this thesis, your mission is to craft a framework that leverages aerial photogrammetry to create a detailed map of the robot's operating environment. This map will serve as a foundation for real-time localization, achieved by fusing data from on-board sensors like LiDAR or cameras. You will develop and test robust algorithms that account for sensor noise and environmental changes, ensuring accurate global localization for your robot.
contact: matthias.edernoSpam@ist.tugraz.at
Development of a Simulation Pipeline for Multi Robot Systems
Multi Robot Systems (MRS) can be abstracted at different levels, from the physical interaction with the world to abstract agent models. The development of algorithms to control such systems can be greatly aided by simulation, but there is a wide variety of existing simulation software focusing on different aspects and offering different modelling power and computational requirements. This project focuses on reviewing existing tools in order to design and develop a pipeline for simulating MRS at different levels (from multi-agent task management and coordination to individual robot control), while allowing the integration of arbitrary robot systems and behaviors.
contact: laurent.freringnoSpam@ist.tugraz.at
Verification for Intelligent Agents
Bringing autonomous systems to the field is a challenging task, especially in the context of disaster response where requirements in terms of reliability, usability, and performance are very high. Agent-oriented programming and more specifically Belief-Desire-Intention (BDI) software agents is a proven solution to develop reactive and adaptive systems in such dynamic environments. However, the complexity of such software makes it difficult to provide guarantees that it would meet user and safety requirements even though this would be very beneficial in reducing testing time and increasing end-user acceptance. This master thesis focuses on exploring existing methods for verification in BDI agents including Model Checking and Runtime Verification, developing and integrating them into an existing software stack and evaluating them through simulation and field experiments.
contact: laurent.freringnoSpam@ist.tugraz.at
Sensor Calibration for Autonomous Mobile Robots
This master's thesis aims to develop a sensor calibration package, to accurately compute the transformation matrices among various sensors used in autonomous mobile robots, including LiDAR, IMUs, and cameras. These sensors are critical for localization and perception tasks, enabling the robot to effectively navigate and understand its environment. Efficient fusion of data from these diverse sensors requires a precise understanding of their spatial relationships. The proposed calibration package addresses this need by providing algorithms to determine the transformation matrices, which describe the positional and rotational relationships between sensors.
contact: hamid.didarinoSpam@ist.tugraz.at
Terrain classification and obstacle detection using LiDAR and camera data
This master's thesis aims to explore the integration of LiDAR and camera data for terrain classification and obstacle detection in outdoor environments, drawing inspiration from already existing research. The research will concentrate on the development and evaluation of algorithms and techniques for utilizing LiDAR and camera data to accurately classify terrain types and detect obstacles in real time. The results of this research are expected to enhance autonomous navigation capabilities for mobile robots operating in complex and dynamic environments.
contact: hamid.didarinoSpam@ist.tugraz.at
Fusing Multisensor Data for Enhanced Pose Estimation in Autonomous Driving Systems
One of the key components of an autonomous driving system is pose estimation. Various sensors, including GPS, cameras, and LiDAR, can be employed for this purpose. Each sensor has its unique strengths and weaknesses. There are instances where each may fail; for example, GPS signals can weaken in forested areas or near tall buildings, and camera and LiDAR odometry can experience drift over time. This master's thesis aims to develop a data fusion package inspired state-of-the-art systems to integrate data from these diverse sensors and accurately estimate the robot's position.
contact: hamid.didarinoSpam@ist.tugraz.at