To Master Students: For interest in the proposed projects, contact us at: [email protected] 

Projects for Master Thesis - Open for academic period 2024-2025 

Rapid Wildfire Detection with a Team of Aerial Robots
Supervisors: Kostas Alexis, Tor Arne Johansen | Keywords: XPRIZE Wildfire Challenge
Available. 
This project aims to develop a sensing payload that enables the detection and localization of a wildfire virtually as soon as it is ignited, meaning at a very early stage of its evolution. This in turn shall allow rapid response to such potentially catastrophic events. The focus is both on the hardware development of the sensing load - which should be integrated on a small fixed-wing unmanned aerial vehicle - as well as on the robot perception and fusion algorithms for detection and localization from heights.  ​
Extinguishing Wildfires with Guided Airdrops
Supervisors: Kostas Alexis, Tor Arne Johansen | Keywords: XPRIZE Wildfire Challenge
Available. ​
This project aims to develop a specialized payload to be ferried by a fixed-wing unmanned aerial vehicle (UAV) that can be used to suppress a wildfire detected at its very early stages after ignition. The envisioned payload shall be released by the ferrying UAV at high speeds and using control surfaces (w/o propulsion) shall guide itself accurately to drop on the desired point of the fire area. Upon impact, the payload shall release a specialized fire retardant or other technology to eliminate the early-detected small wildfire area using a limited number of such airdrops. ​​
 
Toward Intelligent Robot Navigation Using Scene Graphs and Semantic Scene Understanding
Supervisors: Kostas Alexis
Available. ​
 Semantic scene understanding is crucial in robotic perception, aiming to perceive both the structure of the surrounding environment and the objects within it, along with their relationships. 3D scene graphs are structures that organize this knowledge in a graph structure depicting objects as nodes and their relationships as edges. To assist in real-world robotic tasks, Hydra has introduced layered versions of 3D scene graphs where the multiple layers of abstraction are aligned to assist simultaneously with different robotic needs. These layers are 3D metric-semantic mesh, objects and agents, places, rooms, and buildings, allowing for a simultaneous structural and semantic representation of the environment. Currently, our path planning algorithm, GBPlanner, guides the exploration of both aerial and ground robots in unknown environments. As illustrated in Figure 1b, GBPlanner generates collision-free paths by evaluating exploration gain based solely on the environment's geometry. However, information about features such as doors, windows, and other passages—typically missed by semantic-agnostic methods—could be integrated into the planning process to enable more efficient, task-driven exploration. Furthermore, mapping these objects along with their states (e.g., open or closed) can support semantically-aware robot interactions with the environment.
Photo from source. 
 
Multiobjective Optimization of Reward Functions for Smooth and Precise Control
Supervisors: Kostas Alexis, Mihir Kulkarni | Keywords: Aerial Robotics
Available. ​
 Learning: Multiobjective Optimization of Reward Functions for Smooth and Precise Control
Abstract: A key challenge in reinforcement learning is designing an effective reward function, as it significantly impacts learning efficiency and performance. Traditionally, reward function tuning is a manual and time-consuming process, relying on human intuition and trial-and-error adjustments. However, it is also possible to tune the reward function automatically. This thesis proposes the use of Bayesian Optimization (BO) to automatically tune the reward function for RL, reducing the need for manual intervention. Specifically, we aim to optimize the reward function to balance precision (how well the drone can track a set of waypoints) and smoothness (minimal abrupt changes in motor commands). By formulating the tuning process as a multi-objective optimization problem, we seek to improve the overall performance and stability of drone control, and understand the tradeoffs present in the parameters of the reward function.
Photo from source. 
 
Embedded Radar-Inertial Navigation for Resilient Odometry
Supervisors: Kostas Alexis, Morten Nissov | Keywords: Perception, SLAM
Available. ​
Recent literature has become interested in applying small, mm-wave FMCW radars for inertial navigation purposes. These sensors offer the possibility for incredibly robust inertial navigation solutions which do not succumb to the association problems which can plague vision-based methods. Furthermore, these sensors enable operation in a wider array of environmental conditions due to their longer wavelength, making ambient lighting and the presence of obscurants a non-issue. Project will involve experience with inertial navigation systems, the inertial navigation equations as they pertain to EKF-estimation, error-state formulations, compute-efficient derivations, embedded programming, and ROS; all to enable high-rate, reliable estimation. Sensors of interest include (but are not limited to) mm-wave FMCW radars, inertial measurement units, and the potential for additional sensor input (e.g., barometer and magnetometer).​
Enhancing Point Cloud Alignment with Hierarchical and Open-Vocabulary 3D Scene Graphs
Supervisors: Kostas Alexis | Keywords: Learning for Robotics
Available. ​
Scene alignment is a fundamental challenge in robotics, enabling key tasks such as localization and navigation. Traditional methods approach this problem in a hierarchical way through visual cues, point cloud registration, and feature matching at the point level. More recently, 3D scene graphs have been used to align partially overlapping scenes. A 3D scene graph represents a scene as a directed graph, where nodes correspond to semantic entities (e.g., objects and rooms) and edges capture semantic and geometric relationships. Recent advancements incorporate vision-language models (VLMs) to enrich object nodes with open-vocabulary features, and similarly, open-vocabulary relationships can be inferred using VLMs. Prior work has explored two main approaches to point cloud alignment using scene graphs. A decoupled method [5] first matches graph nodes before performing feature-based point cloud registration on the found matches. In contrast, a coupled method refines feature matching by incorporating closed-vocabulary semantics and geometry information embedded in the scene graphs. However, these methods do not take full advantage of open-vocabulary features, inferred relationships, or the hierarchical structure of scene graphs. This project aims to advance point cloud alignment by integrating hierarchical scene graph structures with open-vocabulary features, enhancing both robustness and accuracy in 3D scene graph-based registration.
 
World Models for Semantic Inspection
Supervisors: Kostas Alexis | Keywords: Embodied Autonomy
Available
Abstract: World models have emerged as a powerful paradigm for learning compact representations of an environment and predicting future states, enabling efficient decision-making in autonomous robotics. This project explores the application of world models to semantic inspection tasks, where a robot equipped with visual sensors must construct a semantic understanding of its environment while maximizing inspection efficiency. By leveraging learned latent space representations, predictive modeling, and reinforcement learning, the system aims to generate optimal inspection paths that balance exploration, occlusion handling, and semantic information gain. ​
One Control to Rule them All
Supervisors: Kostas Alexis (NTNU), Mihir Kulkarni, Welf Rehberg Martin Jacquet | Keywords: Learning for Robotics
Available. ​
Reinforcement learning is increasingly being utilized in robot control tasks. Conventionally, an RL policy is attuned to a particular vehicle by exploiting its particular dynamics in simulation-based learning. In this work we seek a fundamental shift in this paradigm: we aim to create a "single policy for the flight control of multirotors" such that one unified policy can control any arbitrary multirotor aerial robot. We are looking to investigate the correct definition of the observation space of the policy, an expressive neural architecture and an appropriate strategy for learning (including reward shaping) such that this unified control policy can be derived and reach at least on-par performance with the current state-of-the-art (but without vehicle-specific tuning requirement).​
Jumping and in-flight-stabilization of a Jumping Quadruped in Mars Gravity Using Reinforcement Learning
Supervisors: Kostas Alexis, Jørgen Anker Olsen | Keywords: Legged Robots, Learning for Robotics​
Over the last decade's satellites, telescopes, landers, and wheeled rovers have been the main form of space exploration. As the field of legged robotics has developed and matured significantly in recent years, we now see the opportunity to explore more diverse and interesting terrain in space using specialized quadruped robots optimized for challenging off-world planetary environments, such as craters, caves, and lava tubes. Legged robots, such as the Boston Dynamics Spot and the ANYbotics ANYmal, present a set of advantages in mobility and versatility in complex environments over traditional wheeled robots and rovers. Jumping legged robots may be able to traverse the geometrically complex subterranean voids of lava tubes on planets such as Mars. A jumping legged robot for Martian surface and lava tube exploration will retain the key advantages of quadruped systems in overcoming rough terrain, while also being able to coordinate its actuators and exploit the low gravity environment of Mars and compliant leg designs to jump for significant height and thus overcome large obstacles. This project thesis aims to contribute to the modeling, control, and simulation of jumping legged robot that is currently being built and tested by our team at NTNU.   The main goal of the project thesis is to study and develop reinforcement learning techniques to enable the robot to perform mid-air stabilization using its legs. The second objective is to incorporate jumping and landing safely into this control policy. If progress is sufficient, deployment of the developed control policy to an actual legged robot may be possible. This continues work of previous master students with similar topic.​
Aerial Gym 3.0 - Blazing fast high-fidelity simulation for aerial robots
Supervisors: Kostas Alexis, Jørgen Anker Olsen | Keywords: Legged Robots, Learning for Robotics
Available. ​
At the Autonomous Robots Lab we have built and maintain "Aerial Gym", which is currently the most widely used NVIDIA Isaac-based aerial robot simulator. Currently in its second version, Aerial Gym is limited to multirotor aerial robots and requires a transition to the "latest-and-greatest" tools of NVIDIA. In this thesis we seek to a) upgrade Aerial Gym to exploit the latest developments from NVIDIA, alongside b) augment its functionality to fixed-wing aerial robots. This would render Aerial Gym the tool-of-choice to the deliver the benefits of reinforcement learning in the fixed-wing community as well as the researchers investigating convertible platforms (e.g., VTOL tailisitters). ​
MPC-based Occlusion Avoidance for Object Monitoring
Supervisors: Kostas Alexis, Martin Jacquet | Keywords: Model Predictive Control
Available. ​
MPC have been used in the context of so-called perception aware applications, where vision-based constraints and objectives are integrated into the underlying optimization problem and solved in a receding horizon fashion. These approaches rely on the existence of a line of sight. However, this assumption is not easily verified in realistic scenarios, in uncontrolled and often cluttered environments. Occlusion avoidance is often tackled with geometric ray-casting to simulate the line of sight. But its evaluation is often computationally heavy, in particular when several occlusion obstacles are considered, and rely on a good knowledge of the environment. Moreover, these approaches rarely consider uncertainties associated with sensor measurements.  On the other hand, chance-constrained MPC are stochastic optimal control technics which have been successfully employed to tackle collision avoidance in receding horizon schemes. This thesis aims to address the problem of real-time occlusion avoidance in object monitoring in cluttered environments, exploiting stochastic constraints. 
Transforming Aerial Robots
Supervisors: Kostas Alexis, Paolo De Petris | Keywords: Aerial Robotics
Available. ​
This project aims to develop a novel reconfigurable multirotor for extreme maneuverability. In particular, we seek to design and built a multirotor with each motor being on the tip of a 5-bar manipulator mechanism thus allowing seamless reconfiguration of its shape and control allocation. Alongside the design, we seek to derive an adaptive control policy to optimally maneuver such a robot along agile trajectories demonstrating superior maneuverability compared to conventional fixed-configuration aerial robots. ​
  • Details - this project is open ended to be co-designed with the student. Contact us. ​​
  • Project Relevance: SPEAR (European Commission Horizon Europe)
  • Contacts: Kostas Alexis ([email protected])
Soft Aerial Robots for Hard Tasks
​Supervisors: Kostas Alexis, Etor Arza | Keywords: Embodied Autonomy
Available. ​
 This project aims to develop soft aerial robots in order to explore the advanced potential such systems can enable when it comes to navigating through highly-cluttered environments (e.g., dense forests). ​
  • Details - this project is open ended to be co-designed with the student. Contact us. 
  • Project Relevance: SPEAR (European Commission Horizon Europe)
  • Contacts: Kostas Alexis ([email protected])
Control Barrier Functions for Safe Underwater Navigation​
​Supervisors: Kostas Alexis, Mohit Singh | Keywords: Underwater Robotics
Available. ​
Control barrier functions (CBFs) provide an avenue to ensure safety for critical systems and operations. Previous work of our lab has related to the problem deriving CBFs to ensure the collision-free flight of aerial robots given the immediate observation of a depth sensor (e.g., from stereo vision, RGB-D or LiDAR data). In this thesis we seek to augment the design and utility of CBFs in the underwater domain. Specifically, we seek to design novel CBF-based safety filters ensuring collision-free navigation of underwater robots exploiting onboard visual sensing and odometry estimates. ​
  • Details - this project is open ended to be co-designed with the student. Contact us. ​
  • Project Relevance: AUTOASSESS (European Commission Horizon Europe)
  • Contacts: Kostas Alexis ([email protected])
Everything is a Robot: Making an autonomous robot in the form of everyday house items
​Supervisors: Kostas Alexis | Keywords: perception, Semantics
Available. ​
 This project aims to investigate the potential of turning every day house items, such as a coffee table or a desk lamp to robots and identify the possible use space and potential of such novel products. By the end of the thesis the goal is to have designed such a prototype and investigate its applicability and relevance in the real-world. This is a particularly open-ended project calling for the student creativity.
  • Details - this project is open ended to be co-designed with the student. Contact us. ​
  • Project Relevance: Open-ended
  • Contacts: Kostas Alexis ([email protected])
Design, Development and Control of an Underwater Swimming Quadruped
​Supervisors: Kostas Alexis, Mohit Singh | Keywords: Underwater Robotics, Legged Robots
Available. ​
Legged robots go swimming! This project aims to develop a quadruped legged robot for underwater operations. The design shall present a host of gaits that allow the system to swim and also walk on the sea floor. Your task is to design the mechatronic solution, realize it in practice and develop control for its baseline leg gaits. 
Modeling the Mechanisms of Human Visual Attention towards Objects of Interest
​Supervisors: Kostas Alexis| Keywords: Perception, Saliency
Available. ​
This project aims to model the mechanisms of human visual attention (bottom-up and top-down) towards objects of interest within the framework of inspection operations. To that end, this project shall exploit specialized glasses that capture the human gaze and localize where it points on a camera frame. By building a dataset appropriate for supervised learning a computational approach to model how inspectors attend and assess objects and structures of interest shall be developed.​

Own ideas - high risk projects!

Do you have your own idea about a robotics project? Are you willing to discuss a high-risk project with the understanding that things might not always work? Contact me and schedule a meeting!