To Master Students: For interest in the proposed projects, contact us at: konstantinos.alexis@ntnu.no 

Projects for Master Thesis - Open for academic period 2023-2024. 

Multi-modal Learning-based Navigation
Supervisors: Kostas Alexis, Mihir Kulkarni
Available. 
Agile navigation has been demonstrated using only a lightweight sensor suite including RGB-D cameras and IMU sensor, harnessing the power of deep learning to directly map noisy sensor observations to low-level command output in a low-latency manner. Nevertheless, existing works often utilize either visual or depth data only. While visual data can offer rich information about the environments, navigation policies trained with visual image input only can suffer from generalizability issue when deployed on the real system, thus requiring extensive domain adaptation/randomization or appropriate intermediate abstract representation [1]. Depth images can be used as an abstract input representation for navigation task [2,3], allowing successful sim-to-real transfer in multiple scenarios. However, depth cameras can miss important details in the environment, such as thin obstacles or shiny surfaces (see above pictures), hence using depth camera alone is not sufficient in many real-world situations. This project and thesis aim to address this challenge by utilizing both visual and depth image inputs to train a collision-free navigation policy for aerial robots. ​
Environment Geometry-aware Probability Distribution Generation for Sampling-based Path Planning
Supervisors: Kostas Alexis, Mihir Dharmadhikari
Available. ​
Sampling-based path planning is commonly used in complex environments and high-dimensional path planning tasks due to their low computation time. These algorithms sample robot states from a fixed probability distribution to build a graph/tree data structure (sometimes a set of those), where the vertices and edges lie in regions of the state space where the robot does not collide with the environment, to search for a path. However, due to the fixed sampling distribution, in complex environments, they suffer from sampling inefficiency where majority of the sampled states do not lie in collision-free regions. This project and thesis aim to solve this problem by creating a model (may or may not involve learning-based techniques) that can output a sampling distribution based on the environment geometry. ​
 
Adaptive Underwater Camera Model for Integration in Visual SLAM
Supervisors: Kostas Alexis, Mohit Singh
Available. ​
Underwater Computer Vision has the potential to enhance the scalability and versatility of autonomous underwater robots due to its low sensor costs and rich perception. However, to deploy underwater robots while maintaining resilience, it is important to address challenges in using cameras for tasks of underwater localization, navigation, and spatial understanding. A major challenge is the varying camera parameters due to changes in physical properties of water. We aim to bridge the gap between camera calibration in air and the underwater counterpart using optic concepts to obtain a camera model that minimizes the need for re-calibration. The formulation will be then parameterized as a function of refractive index and depth to obtain a novel optimizable framework that will be integrated into a variable camera model Visual SLAM method. This will allow to reduce the preparation time for the deployment of underwater robots and enable long-term resilient operations.​
 
Developing Radar-Centric Methods for use in Estimation
Supervisors: Kostas Alexis, Morten Nissov
Available. ​
Vision and LiDAR perception sensors have been greatly explored for localization and estimation task and thus can provide very accurate estimation. However, the limitation of these modalities is clearly observed during degenerate conditions. FMCW (Frequency Modulated Continuous Wave) radar sensors offer promising perception possibilities in situations which would be detrimental for both the previously mentioned modalities. This potential, however, is limited by unintuitive, sparse data output. This project and thesis aim to improve existing radar-inertial methods by investigating and taking advantage of properties inherent to radar sensors (e.g. by statistical modeling of the false positive reflections as well as doppler and RCS-aware scan matching).  Ultimately building a full SLAM solution, thereby increasing long term estimation accuracy and mapping capabilities.​
 
Robust usage of Visual Information in Multi-modal SLAM
Supervisors: Kostas Alexis, Nikhil Khedekar
Available. ​
Robustness of state estimation to perceptual degradations can be tackled by utilization of data from multiple different sensors that have complimentary failure scenarios. Factor graphs allow for an intuitive representation for combining information from these sensors as factors and formulate the estimation as an optimization problem. Optimizing the constraints introduced by the sensors directly in a tightly-coupled manner instead of using the result of separate pipelines (loose-coupling) can allow for better detection and handling of degradations in the overall system. Taking LiDAR, Camera and IMU as the sensors available onboard a micro aerial vehicle, this project will investigate the robust usage of visual information for combination with LiDAR and IMU data to build a robust Multi-Modal SLAM method.​
Marsupial Walking and Flying Robotic Systems
Supervisors: Kostas Alexis
Available. ​
Ground and aerial systems, in particular walking and multirotor robots present unique advantages and disadvantages. Our experience in the DARPA Subterranean Challenge demonstrated that none of them was able to undertake the full exploration mission on their own. Simultaneously, beyond classical multi-robot teams where systems are deployed separately there is unique value in the ability to deploy the complementary navigation modality “on demand”. A solution is that of “marsupial” robotic system-of-systems, meaning in this case the ferrying of one or more small flying robots onboard the legged system. In this project you are invited to develop a comprehensive mechatronic-and-autonomy solution for the integration of an RMF-Owl flying robot onboard the ANYmal D legged robot. A first solution will be outlined by our team (depicted in the Figure above with an earlier version of RMF and the ANYmal C system). The goal is that the marsupial set of robots can share maps, plan in a combinatorial manner to inspect complex structures, while the flying robot can also safely land onboard ANYmal after the completion of its mission.     ​
 
Information-theoretic LiDAR Compression and LiDAR Saliency Maps
Supervisors: Kostas Alexis
Available
Information compression methods employ various approaches to achieve their goal and among others may allow the identification of the most salient features in the input data. In this project we want to use information compression over LiDAR data not only to reduce the input size but simultaneously achieve implicit identification and selection of the most (geometrically) salient features in the scene. Accordingly, this can support critical robotic tasks including that of Simultaneous Localization And Mapping (SLAM) especially in areas with weak geometries (e.g., road tunnels). The “palette” of methods considered includes both so-called “traditional techniques” (e.g., exploiting the wavelet transform on depth images, OpenCTM on 3D triangle meshes) either lossless or lossy, as well as neural networks-based techniques such as (variational) autoencoders. ​
Robotic Detection and Removal of Ghost Nets
Supervisors: Kostas Alexis
Available. ​
Ghost nets, essentially fishing nets that have been abandoned, lost or discareded, represent a major problem. Ghost nets and other fishing gear are responsible for trapping and killing significant amounts of marine life and can also cause further destruction including destroying shorelines, smothering coral reefs or even damaging boats. Unfortunately, the problem remains largely not addressed and gets greater by the day. Finding and collecting such ghost nets represents a complicated task that so far relies on sparse and small teams – often volunteers – that try to perform cleaning operations. As it is a time-consuming, challenging and often risky task robotics can play a significant role and undertake a major part of the activities. In this project, we seek to develop a robotic system-of-systems that involves a duo of units that can be tugged by a larger vessel and offer searching of the bottom of the sea in order to detect ghost nets, as well as cutting and lifting of such ghost nets. Within the project, you are tasked with the goal of designing such a system and demonstrating its potential use case in small-scale operations within a lab environment. ​
Learning and Evolution for Embodied Autonomy
Supervisors: Kostas Alexis
Available
In September 2021, the finals of the DARPA Subterranean Challenge took place and Team CERBERUS – led by the PI – won the competition. Both for CERBERUS and for the other top-scoring teams, flying robots presented scouting capabilities but otherwise the bulk of the mission was carried by ground robots owing to their extended endurance and effective range. This contradicts our intuition regarding the benefits of flying robots in demanding subterranean environments with first being the fact that they do not depend on the challenges of the otherwise perilous terrain. A host of species, such as bats or insects, demonstrate the analogy of benefits of flying systems in natural underground environments but flying robots still need fundamental improvements to reach similar performance. This master thesis is motivated by the observation that despite the vastly different missions that flying robots undertake, and the different environments in which they operate, certain types of vehicles – primarily multirotors and fixed-wing systems – dominate across use cases. We claim that this is not only counterintuitive, as a higher variety of embodiments and autonomy solutions would follow the natural paradigm, but also we claim that this likely represents the key reason for the limitations observed on autonomous flying robots. Specifically, we claim that an environment-specific co-synthesis and co-optimization of the robot’s “body” (airframe, sensing, propulsion) and “brain” (autonomy functionality especially relying on data-driven learning methods for navigation) is necessary. Reflecting upon this analysis, in this master project you are tasked with employing the combined benefits of evolutionary algorithms and learning to facilitate the automated environment-specific adaptation of flying robot embodiments. 
TrogloBots: Designing Deep Subterranean Robots
Supervisors: Kostas Alexis (NTNU)
Available. ​
Subterranean robotics and associated autonomy have been investigated with the recent DARPA Subterranean Challenge representing the most notable relevant effort. In that, Team CERBERUS won the challenge and throughout its three years employed primarily legged and flying robots all equipped with multi-modal localization and mapping fusing LiDAR, visible-light cameras, thermal vision, and inertial measurements. However, it is possible that such designs are not optimally tailored to “deep subterranean operations and environments”. By Deep Subterranean Robots (DSR) we refer to systems operating autonomously – without the support of infrastructure – in underground environments that are tens and hundreds of kilometers away from the closest access point from the surface. In such environments, access to power resources are extremely limited (if any) and definitely unconventional. Thus, it is essential that such systems are extremely power efficient. Accordingly, this likely necessitates different locomotion and perception systems. Pertinent observations can be acquired by the study of “troglobites (troglobionts)”, animal species strictly bound to underground habitats, such as caves. Unlike species that spend time underground only for a subset of their activities (“eutroglophiles”, “subtroglophiles”, “trogloxenes”) troglobites have employed unique and extreme paths in evolution including the complete, or almost complete lack of eyes. Albeit the fact that in robotics we can have a set of extremely accurate active sensors (e.g., LiDAR) their continuous use is very power demanding. Likewise, their size and energetic characteristics are tailored to the sparsity of underground resources. Without biomimetic approaches not being the only possible path, in this master project the challenge is to design a true “TrogloBot”, a DSR that could operate for long operations (e.g., several hours) underground with as minimalistic power consumption as possible. Example applications may include deployments in the Earth’s deep subterranean voids (e.g., the 680km+ passages of the Mammoth Cave) or for exploratory scientific studies in the underground voids of other planets such as the lava tubes of Mars. 
RL-based Safe Navigation using Depth Images
Supervisors: Kostas Alexis, Martin Jacquet
Available. ​
Exploiting aerial robots in unknown complex environment requires using onboard sensors, such as lidars or depth cameras, to generate safe motions which account for observed obstacles. However, the analytical evaluation of safety of all possible trajectories cannot be performed. Pre-sampled motion primitive libraries can be used to reduce the number of trajectories to evaluate, but the problem is still too complex to solve analytically, since it relies on building an accurate volumetric map of the environment from sensor data. This project and thesis aim to investigate the use of reinforcement learning (RL) to tackle the problem of collision-free navigation to predict control-action sequences or trajectories. While supervised learning has been leveraged to evaluate a safety criterion on pre-sampled trajectories directly exploiting the sensor data, it requires a labeling process which can be cumbersome and limits the generalization of the learned policy. On the other hand, RL has been widely employed toward safe navigation over the past few years and seems to provide better generalization capabilities and a smaller sim-to-real gap. ​
 
Tiny Visual-Inertial Odometry System
Supervisors: Kostas Alexis | Keywords: Visual-Inertial Odometry
Given - Not Available. ​
Visual-inertial odometry systems have long been researched with significant success. Exploiting relevant advances in vision systems and processors, as well as breakthroughs in the knowledge domains of state estimation and computer vision, the community has developed robust and high-performance visual-inertial estimation systems presenting drift that grows as slowly as 1% over the robot path. However, such systems are often computationally demanding and they rely on high-quality camera sensors. This prohibits the potential for their ubiquitous utilization in particularly miniaturized systems. In this project you are asked to develop a miniaturized visual-inertial system that exploits an integrated system that offers a monocular camera, an Inertial Measurement Unit, a 1D distance sensor (time-of-flight), alongside a processing solution integrating 1x  Arm® Cortex® M7 @ 480 MHz and 1x Arm® Cortex® M4 core @ 240 MHz. Your solution should fit well in the M7 core of this system and perform visual-inertial odometry at 20FPS.
In-air Stabilization of a Jumping Quadruped in Mars Gravity using Reinforcement Learning
Supervisors: Kostas Alexis, Jørgen Anker Olsen
Available. ​
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 Marian 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 such as jumping legged robot that is currently being designed 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 must 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.​
Image from relevant project: https://www.spacehopper.ethz.ch/ ​
MPC-based Occlusion Avoidance for Object Monitoring
Supervisors: Kostas Alexis, Martin Jacquet
Available. ​
MPC has 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.​
Design, modeling and realization of a compliant enclosure for swashplateless coaxial aerial vehicles colliding at high speeds
Supervisors: Kostas Alexis, Paolo De Petris
Available. ​
Coaxial aerial vehicles relying on swashplateless mechanism have proven to be able to emulate full actuation over forces and torques (six degrees of freedom) without a complex actuation design. Exploiting their smaller size, lightweight and controllability peculiarities, compared to standard quadcopters, this work focuses on the collision-tolerant properties of such design. The goal is to realize a compliant embodiment that permits the system to collide at very high speed without any repercussions.
Radiation Mapping for the Autonomous Characterization of Nuclear Sites
​Supervisors: Kostas Alexis, Paolo De Petris
Available. ​
Robotic systems can be utilized for the purposes of radiation mapping and spectroscopic characterization of nuclear sites. Such deployments can prove critical during emergency response (i.e., when a nuclear accident takes place) or for a host of other activities including nuclear decommissioning. Motivated by the above, in this project you are tasked to develop a miniaturized sensing unit involving a scintillator system such that an autonomous flying robot can enter nuclearized environments and return with a dense map estimate of distributed radiation fields further characterized for the spectroscopic characteristics (i.e., what material) within each region.
Real-time Underwater Motion Planning with Safety Guarantees in the Presence of Dynamic Obstacles
​Supervisors: Marios Xanthidis, Eleni Kelasidi, Kostas Alexis
Available. ​
Autonomous underwater operations is a domain of increasing interest in many sectors (such as aquaculture), but also many important challenges. Autonomous Underwater Vehicles (AUVs) need to be able to compute where they are (Localization), how the environment looks like (Mapping), and decide quickly on how to act in order to reach a goal safely (Motion Planning). This thesis will focus on Motion Planning, which becomes very challenging, when robots are operating in cluttered environments and in proximity to other moving objects. The goal of this thesis is to produce novel motion planning methods, which will enable robots to navigate such environments in real-time and with hard safety guarantees for the entire duration of the operation. Ideally, the produced planners would be able to accommodate geometric descriptions of the robot and environment with reasonable accuracy, instead of oversimplified approximations (i.e. spheres, boxes, etc.). ​
Assessment of DVL-accuracy with biomass interference in aquaculture settings
​Supervisors: Marios Xanthidis, Eleni Kelasidi, Kostas Alexis
Available. ​
Autonomous underwater operations is a domain of increasing interest in many sectors (such as aquaculture), but also many important challenges. Autonomous Underwater Vehicles (AUVs) need to be able to compute where they are (Localization), how the environment looks like (Mapping), and decide how to act in order to reach a goal safely (Motion Planning). This thesis will focus on the interjection of localization and mapping, by studying the accuracy of DVL sensors in the presence of biomass. The DVL is a very popular acoustic sensor used in underwater operations, that can measure the relative linear speed of the robot and its distance with respect to the surroundings. DVLs have been utilized for operations in aquaculture settings, and researchers have studied its utility and challenges, with interference from biomass being a major one. The goal of this thesis is to extend the understanding of the robustness of DVLs in fish farms by collecting data, developing models that correlate the presence of biomass with observation noise, and producing uncertainty prediction capabilities that could inform future localization and mapping techniques. ​
Real-time 3D Reconstruction by Fusing Multi-beam Sonars
​Supervisors: Marios Xanthidis, Eleni Kelasidi, Kostas Alexis
Available. ​
Autonomous underwater operations is a domain of increasing interest in many sectors (such as aquaculture), but also many important challenges. Autonomous Underwater Vehicles (AUVs) need to be able to compute where they are (Localization), how the environment looks like (Mapping), and decide on how to act in order to reach a goal safely (Motion Planning). This thesis will focus on Mapping, which becomes very challenging in the underwater domain due to electromagnetic degradation invalidating the utility of popular sensors in robotics (cameras, lidars, etc). On the other hand, acoustic sensors (such as multibeam sonars) could provide range information on the surroundings, but are often bounded to 2D observations, accompanied with noise.  However, recent studies have introduced the use of two different perpendicular multibeam sonars that provide 3D observations with a limited field of view. The goal of this thesis is to produce a flexible framework that could be used to test different arbitrary mounting configurations of multibeam sonars, provide the methods for real-time 3D reconstruction of the observed scene by fusing them, and study challenges and limitations that could be risen in different configurations. 

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!