Topics of student projects /  Témata studentských prací

[EN]: Below, you will find a list of all currently available topics for various university projects. For english speaking students: Some of the topics' details are written in Czech. However, do not be afraid to contact the supervisors for more information about the projects.

[CZ]: Některá témata jsou psána v angličtině a jiná v češtině. Nebojte, všechna témata je možné vypsat v obou jazycích. V případě zájmu o téma kontaktujte konkrétního vedoucího. 

Student must subscribe to a topic via the system: https://hub.fel.cvut.cz/topics/semestral_projects/all_topics_semestral

The topics managed by MRS group can be listed under department "13167". 

This project explores the potential of visual inertial navigation system called sqrtVINS [1], which is suitable for agile flight due to properties such as computational efficiency and numerical stability. Moreover, the method relies only on IMU and camera data, thus makes it suitable in environments, where GPS signal is not reliable. The project focuses on implementing the method and comparing it with current SOTA visual inertial algorithms on measurement data from agile flights.
Literatura
[1]Y. Peng, C. Chen, K. Wu and G. Huang, "sqrt(VINS): Robust and Ultrafast Square-Root Filter-Based 3D Motion Tracking," in IEEE Transactions on Robotics, vol. 41, pp. 6570-6589, 2025, , doi: 10.1109/TRO.2025.3626607.

The goal of this project is to develop adversarial reinforcement learning algorithms for optimizing drone interception strategies using game-theoretic methods. The interception problem is modeled as a two-player zero-sum game, where two competing agents, an intruder drone and an interceptor drone, iteratively refine their strategies through adversarial training. The focus is on handling realistic constraints such as drone dynamics, sensor limitations, and maneuverability, where finding Nash equilibria analytically is impractical.
[1] Li, S., et al. Robust Multi-Agent Reinforcement Learning via Minimax Deep Deterministic Policy Gradient. AAAI, 2019.
[2] Zhou, Z., & Xu, H. Decentralized Optimal Large-Scale Multi-Player Pursuit-Evasion Strategies. Neurocomputing, 2021.

The goal of this project is to develop reinforcement learning-based algorithms that enable drone swarms to cooperatively intercept high-speed intruder drones. Inspired by biological swarms such as bird flocks and insect colonies, the focus is on emergent collective behaviors that allow slower agents to outmaneuver a faster adversary through coordination. The developed strategies will be implemented and validated in simulation, explicitly accounting for realistic constraints such as limited sensing range, maneuverability, and inter-agent communication.
[1] Zhang R., et al. Game of Drones: Multi-UAV Pursuit-Evasion Game With Online Motion Planning by Deep RL. IEEE TNNLS, 2022.
[2] De Souza C., et al. Decentralized Multi-Agent Pursuit using Deep Reinforcement Learning. IEEE RA-L, 2021.

Cílem projektu je návrh metod posilovaného učení pro plánování pohybu bezpilotního vzdušného prostředku.
Klasické metody plánování a řízení dronu nedokáží využít plný potenciál dronů při letu vysokou rychlostí skrze prostředí s překážkami.
V rámci projektu bude úkolem využít metody strojového učení, jako například posilované učení, pro zlepšení plánování a řízení dronu v neznámém prostředí.
[1] Penicka, R., Song, Y., Kaufmann, E., & Scaramuzza, D. (2022). Learning minimum-time flight in cluttered environments. IEEE Robotics and Automation Letters, 7(3), 7209-7216.
[2] J. Eschmann, D. Albani and G. Loianno, "Learning to Fly in Seconds," in IEEE Robotics and Automation Letters, vol. 9, no. 7, pp. 6336-6343, July 2024.

The task In this project is to integrate the trajectory planning method [3] of agile quadrotor flight into Robot Operating System (ROS) so that this planner can be used during flight using the MRS system [2]. Furthermore, the planner will be extended by the possibility of using a grid map which can be created during the flight from on-board sensors. The planner will be tested on a realistic scenario of agile simulated quadrotor flight combined with appropriate predictive control method.
[1] Romero, Angel, Robert Penicka, and Davide Scaramuzza. "Time-optimal online replanning for agile quadrotor flight." IEEE Robotics and Automation Letters 7.3 (2022): 7730-7737.
[2] Baca, T., Petrlik, M., Vrba, M. et al. The MRS UAV System: Pushing the Frontiers of Reproducible Research, Real-world Deployment, and Education with Autonomous Unmanned Aerial Vehicles. J Intell Robot Syst 102, 26 (2021).
[3] Krystof Teissing, “Replanning of Collision-Free Trajectories for Unmanned Aerial Vehicle,” Master thesis, Czech Technical University in Prague, 2023.

The goalis project focuses on developing an advanced modeling approach to estimate the State of Charge (SOC) and predict the maximum permissible discharge current for drone batteries in real-time. By analyzing live data inputs—including battery temperature, voltage, and current load—the model will determine the current battery health and endurance capabilities. Furthermore, the system will utilize these calculations to predict necessary power inputs for the motors, ensuring optimal flight performance and preventing unexpected power failure due to thermal overloading or excessive voltage drop.
[1] Bauersfeld, Leonard, and Davide Scaramuzza. "Range, endurance, and optimal speed estimates for multicopters." IEEE Robotics and Automation Letters 7.2 (2022): 2953-2960.
[2] D. N. T. How, M. A. Hannan, M. S. Hossain Lipu and P. J. Ker, "State of Charge Estimation for Lithium-Ion Batteries Using Model-Based and Data-Driven Methods: A Review," in IEEE Access, vol. 7, pp. 136116-136136, 2019.

This project will aim to design a simulation pipeline for learning to fly multiple Unmanned Aerial Vehicles (UAVs) using imitation learning. The student will first study existing approaches to imitation learning and then select a suitable multi-UAV planning algorithm to imitate. Finally, the learned policy shall be shown to fly at least two UAVs inside a simulated environment.
Literatura
[1]R. Penicka, Y. Song, E. Kaufmann and D. Scaramuzza, "Learning Minimum-Time Flight in Cluttered Environments," in IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7209-7216, July 2022, doi: 10.1109/LRA.2022.3181755. [2]A. Gleave et al., “imitation: Clean Imitation Learning Implementations”, https://arxiv.org/abs/2211.11972 [2]Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT press. [3]Raffin, Antonin, et al. "Stable-baselines3: Reliable reinforcement learning implementations." Journal of Machine Learning Research (2021).

The project aims to design reinforcement learning methods for unmanned aerial vehicle trajectory planning. Classical drone planning and control methods fail to exploit the full potential of drones in high-speed flight through cluttered environments. The project will focus on the use of machine learning methods, such as reinforcement learning, to improve drone planning and control in unknown environments.
Literatura
[1]R. Penicka, Y. Song, E. Kaufmann and D. Scaramuzza, "Learning Minimum-Time Flight in Cluttered Environments," in IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7209-7216, July 2022, doi: 10.1109/LRA.2022.3181755. [2]Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT press. [3]Raffin, Antonin, et al. "Stable-baselines3: Reliable reinforcement learning implementations." Journal of Machine Learning Research (2021).

This project explores integrating constraints into reinforcement learning for quadrotor flight by encoding them as penalty terms directly in the reward function. The goal is to identify how different penalty formulations across flight phases like takeoff, navigation, and agile maneuvering affect the performance.

The aim of this project is to develop a physically accurate model of a UAV capable of high-speed flight, and to identify its dynamic parameters using real data. The focus is on understanding how high-speed flight affects UAV dynamics (e.g., increased aerodynamic drag, actuator saturation) and on using identification techniques to capture these effects in the model. The resulting model can be used for high-fidelity simulation or advanced control design.
[1] Bazzana, B., Brantjes, R., Gabellieri, C., & Franchi, A. (2024). An Experimentally Validated Model of the Propeller Force Accounting for Cross Influences on Multi-Rotor Aerial Systems. In International Conference on Unmanned Aircraft Systems, ICUAS 2024 https://doi.org/10.1109/ICUAS60882.2024.10556890

The goal of this project is to develop and test a control architecture that integrates obstacle avoidance directly into the feedback control loop of an autonomous UAV. Unlike traditional layered systems, where planning and control are decoupled, this approach aims to react to obstacles in real time based on current sensor data and system state. The focus is on achieving safe, smooth, and computationally efficient motion through cluttered environments.
[1] Krinner, Maria, et al. "Mpcc++: Model predictive contouring control for time-optimal flight with safety constraints." arXiv preprint arXiv:2403.17551 (2024).

The task of this project is to investigate the problem of trajectory re-planning during agile maneuvers in partially unknown cluttered environments. The algorithm should be able to quickly react to sudden changes in the environment (e.q., newly discovered obstacles, change in the task requirements) and re-plan the trajectory of the UAV while maintaining its agility and safety.

The task of this project is to investigate the problem of reliable state estimation for agile and high-speed UAV flights in environments where external localization systems (e.g., motion capture or GPS) are unavailable or unreliable. A key motivation for this work is drone racing, where UAVs must navigate complex tracks at high speed while relying only on onboard sensing - camera and IMU. The goal is to develop and evaluate algorithms capable of accurately estimating the UAV’s position, velocity, and orientation using onboard sensors during aggressive maneuvers. The work will build upon our state estimation framework developed for the A2RL Drone Racing Challenge 2025.
References
[1] F. Novak, M. Petrlik, M. Novosad, P. M. Gupta, R. Penicka, M. Saska, "Vision-only UAV State Estimation for Fast Flights Without External Localization Systems: A2RL Drone Racing Finalist Approach", arXiv preprint, 2026, https://doi.org/10.48550/arXiv.2602.01860
[2] S. A. Bahnam, R. Ferede, T. M. Blaha, A. E. Lang, E. Lucassen, Q. Missinne, A. E.C. Verraest, C. De Wagter, G. C.H.E. de Croon, "MonoRace: Winning Champion-Level Drone Racing with Robust Monocular AI", arXiv preprint, 2026, https://doi.org/10.48550/arXiv.2601.15222

The goal of this project is to advance the classical estimation methods for propulsion performance prediction. The project will focus on refining the work conducted in our upcoming paper to include first-order transient effects of batteries and motor-propeller systems to increase the accuracy of real-time predictions. Our current work ignores these transients for simplicity but the data required to employ better models is already available.

[1] D. Rakhmatov and S. Vrudhula, “An analytical high-level battery model for use in energy management of portable electronic systems.” 02 2001, pp. 488–493.
[2] J. F. Manwell and J. G. McGowan, “Lead acid battery storage model for hybrid energy systems,” Solar Energy, vol. 50, no. 5, pp.399–405, 1993. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0038092X93900602

The aim of this project is to use learning-based methods to accelerate and improve the identification of aerodynamics and propulsion parameters without utilising wind tunnels but instead with real-world open flight. The goal is to utilise several hours of flight data to achieve identification and make classical models more accessive.
[1] L. Bauersfeld*, E. Kaufmann*, P. Foehn, S. Sun, and D. Scaramuzza, “NeuroBEM: Hybrid Aerodynamic Quadrotor Model,” in Robotics:Science and Systems XVII. Robotics: Science and Systems Foundation, July 2021.

The aim of this project is to investigate the advantages of INDI control in real-world flight using only GNSS-based state estimation. This method will be compared against direct control of the UAV from a NMPC without a low-level controller.
[1] S. Sun, A. Romero, P. Foehn, E. Kaufmann, and D. Scaramuzza, “A Comparative Study of Nonlinear MPC and Differential-Flatness-Based Control for Quadrotor Agile Flight,” IEEE Transactions on Robotics, vol. 38, no. 6, pp. 3357–3373, Dec. 2022, doi: 10.1109/TRO.2022.3177279.

The aim of the project is to design and implement an interface for controlling an unmanned multirotor aerial vehicle using virtual reality and gestures to enable intuitive and fast real-time trajectory generation. The system should allow the user to define drone motion through natural spatial interaction and support immediate stopping as well as continuous trajectory modification during flight. The solution includes the design of a method for transforming VR gestures into a feasible trajectory and the implementation of a mechanism for its real-time updating. The project is primarily focused on the control of a single drone with the possibility of subsequent extension to multiple drones in a leader–follower architecture. The designed and implemented system will be verified in a robotic simulator and, if successfully realized, experimentally tested on real multirotor platforms of the Multi-Robot Systems group at the Department of Cybernetics.
Prerequisites: good knowledge of Unity C#/C++

The aim of the project is to design methods for mission state visualization of an unmanned system using virtual and augmented reality (VR/AR), with emphasis on clear representation of spatial data acquired from one or more drones. The solution focuses on constructing a mission overview display that enables the operator to efficiently monitor the environment, camera views, lidar data, and the spatial map of the mission. The system will focus on stable visualization of image and sensor data, design of a mission view interface, and integration of spatial environment representation, for example using a voxel map. The goal is to enable fast situation interpretation and support real-time decision making during robotic system control. The proposed solution should be able to process data from both single and multiple unmanned aerial vehicles and provide the operator with a comprehensive overview of the current mission state.
Prerequisites: good knowledge of Unity C#/C++

The goal of this project is to design and implement an algorithm for solving the Kinematic Traveling Salesman Problem (KTSP) for multirotor UAVs similar to the Orienteering Problem in [1] . The solution will involve optimizing the sequence of waypoint fly-throughs to minimize flight energy while taking into account the UAV’s simplified point-mass model [2]. The combinatorial part of the TSP will be solved with a heuristic such as Variable Neighborhood Search (VNS) method [3] or similar to the one in [1]. The results will be validated on existing TSP datasets and with UAV flight in simulations.
Literatura
[1] F. Meyer and K. Glock, "Kinematic Orienteering Problem With Time-Optimal Trajectories for Multirotor UAVs," in IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11402-11409, 2022.
[2] Teissing, K., Novosad, M., Penicka, R., & Saska, M. (2024). Real-time Planning of Minimum-time Trajectories for Agile UAV Flight. IEEE Robotics and Automation Letters, 1–8. https://doi.org/10.1109/lra.2024.3471388
[3] Robert Pěnička, Jan Faigl, Martin Saska and Petr Váňa. Data collection planning with non-zero sensing distance for a budget and curvature constrained unmanned aerial vehicle. Autonomous Robots 43(8):1937–1956, December 2019.

The goal of this project is to design a learning-based method for planning paths over multiple targets solving, at least partially, the orienteering problem [1]. The project will focus on designing either an algorithm with or without a teacher, so that the learned neural network can create an effective plan for visiting specified goals [1], or is able to make the heuristic search for solutions more efficient [4]. As part of the project, the student will: 1) study existing methods for path planning over multiple targets that use machine learning, 2) propose a suitable learning-based method that will make the search for a solution to the kinematic variant of the orientation problem more effective [2], 3) compares the proposed method with existing approaches.
Literatura
[1] F. Meyer and K. Glock, "Kinematic Orienteering Problem With Time-Optimal Trajectories for Multirotor UAVs," in IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11402-11409, Oct. 2022.
[2] Xin L, Song W, Cao Z, Zhang J. NeuroLKH: Combining deep learning model with Lin-Kernighan-Helsgaun heuristic for solving the traveling salesman problem. Advances in Neural Information Processing Systems. 2021 Dec 6;34:7472-83.
[3] Kool, Wouter, Herke Van Hoof, and Max Welling. "Attention, learn to solve routing problems!." arXiv preprint arXiv:1803.08475 (2018).

The project aims to extend the classical Traveling Salesman Problem (TSP) with multiple robots to monitor large outdoor areas like fields, quarries and industrial sites. The student will extend the existing algorithm to solve the problem where drones can take-off and land at multiple depots during the monitoring mission. The majority of work will involve developing and testing heuristic methods like Genetic Algorithms and Local Search algorithms to find near-optimal solution.

The student will use multiple drones to collect data in a simulated environment. The drones should use local message passing to share information and reach a consensus on the collected data. This will be achieved using Gaussian Belief Propagation [1]. [1] Patwardhan, Aalok, Riku Murai, and Andrew J. Davison. "Distributing collaborative multi-robot planning with gaussian belief propagation." IEEE Robotics and Automation Letters 8.2 (2022): 552-559.

The aim of this project is to develop a system for automated coordination and task scheduling for multiple UAVs, where the role of human operator is the assignment of tasks, rather than commanding individual vehicles. Due to the nature of the target environments, emphasize is to be given to stochastic factors such as uncertain travel times, GPS-less vehicle localization accuracy and the presence of dynamical high-risk hotspots in the overall area of operation, which are to be planned for with some margin from worst-case scenario. Overall, the project is rather implementation oriented, with the end-goal of creating a meaningful contribution to an ongoing research effort with possibility of further cooperation after submission.

This project is about experimentation with with unconventional non-linear program formualtions for planning UAV trajectories and/or coordination of multiple vehicles in real-time. Unlike the topic above, the emphasis of the project is on "throwing mad science spaghetti and on the ceiling and seeing which strands stick". The single requirement is the willingness to perform individual diligent research work, e.g. seek out novelty in the literature, implementation, evaluation and acceptance that sometimes failure is also a reasonable outcome (and trying again). Previous experience with combinatorial optimization and/or off-the-shelf solvers such as Cplex, Gurobi, Ipopt etc. is a plus. Expected outcome of the work is a conference publication.

The task of disassembly planning is to disassemble an object consisting of several parts.
On the highest level, disassembly sequencing finds subsequent disassembly actions that can separate individual parts of an assembly.
This problem is usually formulated as a discrete search and optimization problem.
Disassembly path planning deals with computing collision-free paths (motions) for the individual parts.
The task of this project is to:
a) investigate state-of-the-art in the assembly or disassembly planning
b) implement (or use) suitable solver for simplified disassembly planning considering that only one piece can move at time
c) implemented suitable path planner to achive the movements of the objects.

Motion planning of robots/other objects leads to a search in high-dimensional configuration spaces.
This is typically solved by sampling-based planning, where random samples are drawn (usually uniformly) in the space, classified as free/non-free and the free ones are stored into a roadmap.
Path in the roadmap then describes motion of the robot/object.
The way of sampling influences the performance of the planning (speed, efficiency, quality..).
The goal of this project is to investigate usage of space-filling trees for sampling-based planning.

The goal of motion planning is to find a collision-free path/trajectory for an object in an environment with obstacles. Motion planning algorithms are most commonly used in robotics (planning the motion of manipulators, mobile robots, helicopters, etc.), in analysis within CAD systems, and also in computational biochemistry. In the latter case, planning algorithms are used, for example, in protein docking and protein folding tasks. Part of the protein docking problem may also involve analyzing the passability of small molecules through so-called tunnels in proteins.
The aim of this thesis is to accelerate motion planning algorithms for 3D objects (robots, general 3D objects) using machine learning algorithms. Classical motion planning methods are based on sampling the configuration space. This can be problematic in environments with many obstacles, where the planner repeatedly causes the object to collide with obstacles.
In this semester project, the task is to design a method to prevent these frequent collisions by leveraging previously available data, to which selected machine learning algorithms will be applied.
Prerequisites: good knowledge of C/C++/Python and experience working in a Linux/Unix environment; knowledge of Blender is an advantage. Robot and molecular models will be provided by the thesis supervisor.

In the motion planning problem, the task is to find a path (trajectory) to move an object between two locations; for complex objects/environments, this can be solved by sampling the configuration space. An optimal path can be found using methods based on the RRT* algorithm.
In this project, we will focus on using machine learning to improve (accelerate) the search for optimal paths. Student will implement a RRT*-based planner and design its extension using machine learning to improve its convergence.

Robotic simulators are widely used for test and developing various robotic systems. Simulations allow the user to run multiple experiments under different conditions without physical interaction between the robot and the environment. For example, simulations can be used to learn behavioral policies using AI methods such as reinforcement learning. Learning behaviors using camera data requires a good - fast and realistic - simulation of camera sensors, which is still tricky in many simulators. This semestral project aims to create a system to simulate the behavior of unmanned aerial vehicles (drones) and simulate basic sensors such as cameras or lidars using Unreal Engine 5.
The student will extend our UE5-based simulator (FlightForge) by new functions like automatic generation of environments, integration of path planners etc. The actual task will be adjusted based on actual needs of the simulator users.

In motion planning, the goal is to find a collision-free path for an object moving in a workspace with obstacles. Classical sampling-based methods such as PRM and RRT can generate feasible solutions, but they often struggle in cluttered or complex environments, especially when different homotopy classes of paths exist. Understanding and classifying these fundamentally distinct paths is important for tasks where multiple alternative solutions are required. Recent advances in machine learning, particularly variational autoencoders (VAEs), offer a way to capture compact latent representations of workspaces. Such representations can simplify reasoning about environments with variable and nonconvex obstacles, and can support efficient classification of paths into homotopy classes. Assignment Goals The student will: a) Develop a method to obtain a latent space representation of 2D workspaces using a variational autoencoder (VAE). b) Consider scenarios with a variable number of obstacles, including nonconvex obstacle shapes. c) Design and implement a system to cluster paths according to their homotopy classes. Output of the work: a) implemented code in Python using Torch b) created dataset for learning c) code will be on a school gitlab repository d) technical report (aprox. 5-10 pages in english)

Motion planning aims to compute collision-free motions for robots. A widely used class of methods are sampling-based motion planners such as RRT and PRM, which explore the robot’s configuration space by generating samples and building a graph of collision-free configurations. A feasible path in this graph corresponds to a valid motion of the robot. The objective of this semester project is to investigate how variational autoencoders (VAEs) can be leveraged to improve sampling in configuration space. Specific goals: Study relevant literature on the use of VAEs in robot motion planning, e.g., [1], [2] Create datasets suitable for training VAEs Apply VAEs as sampling modules within planners such as RRT or PRM Experiment with different VAE architectures and training strategies Analyze the impact of VAE-based sampling on planning performance References [1] S. Park, S. Jeon, and J. Park, A Constrained Motion Planning Method Exploiting Learned Latent Space for High-Dimensional State and Constraint Spaces, IEEE/ASME Transactions on Mechatronics, vol. 29, no. 4, pp. 3001–3009, Aug. 2024. doi: 10.1109/TMECH.2024.3399594 [2] B. Ichter and M. Pavone, Robot Motion Planning in Learned Latent Spaces, IEEE Robotics and Automation Letters, vol. 4, no. 3, pp. 2407–2414, July 2019. doi: 10.1109/LRA.2019.2901898

In motion planning, the goal is to find a collision-free path for an object moving in a workspace with obstacles. Classical sampling-based methods such as PRM and RRT can generate feasible solutions, but they often struggle in cluttered or complex environments, especially when different homotopy classes of paths exist. Understanding and classifying these fundamentally distinct paths is important for tasks where multiple alternative solutions are required. Recent advances in machine learning, particularly variational autoencoders (VAEs), offer a way to capture compact latent representations of workspaces. Such representations can simplify reasoning about environments with variable and nonconvex obstacles, and can support efficient classification of paths into homotopy classes. Assignment Goals The student will: a) Implement basic sampling-bsased motion planner (e.g., RRT) b) Use existing tools (will be provided by the supervisor) to obtain latent-space of 2D workspace or 2D paths and a method to detect homotopy classes in 2D workspace. c) Extended the method from b) to 3D workspace Codes will be implemented in Python. Student has to write a short (max 10 pages) report in the end in english.

The task of this project is to investigate the problem of data collection planning with agile UAV in complex environments. The goal is to plan a path for a UAV that collects data from multiple locations while avoiding obstacles and minimizing the total flight time. The project will involve developing algorithms for collision-free path planning and data collection optimization (Traveling Salesman Problem, Orienteering Problem etc).

The task of this project is to develop a mission plan for a team of cooperatiing UAVs to collect data from multiple locations in a cluttered environment. The goal is to design an algorithm that assigns tasks to each UAV and plans their trajectories which avoid collisions with each other and the environment and minimizing the total flight time. The project will involve developing algorithms for task allocation, collision-free path planning and multi-robot coordination.

The goal of this project is to investigate the problem of path planning with distinct topology classes in environments with obstacles. The task is to design and implement an algorithm which in a short time frame identifies as many paths as possible between two (or more) points, traversing different directions around obstacles.

The deployment of autonomous aerial robots in large-scale outdoor missions requires robust environmental understanding. In scenarios where Global Navigation Satellite Systems (GNSS) are inaccurate or intentionally jammed, robotic systems must rely on onboard sensors and local odometry, which inevitably accumulate drift over the travelled distance. To correct this drift, global localization against a known reference is required. The goal of this project is to generate a robust topological and semantic representation (specifically, a scene graph) from onboard sensor data and available prior maps (e.g., satellite or aerial imagery). This prior scene graph will serve as a global map to support drift-free localization and autonomous navigation in complex, large-scale outdoor environments.

Requirements:
Solid foundation in programming (Python, C++ is plus plus), basics of linear algebra, computer vision and machine learning. The ability to learn new things and solve problems.

This project is inspired by the problem of aerial safety and defense from malicious UAVs. The goal of the project is to improve an existing LiDAR-based algorithm for onboard detection and tracking of potentially non-cooperating UAVs or to develop a completely new approach that can rely on inputs from an RGB-D camera, thermocamera, or other sensory modalities. The existing detection algorithm is currently deployed on the real autonomous aerial interception system called Eagle.One and the results of this project will be evaluated for usage within that system in real-world experiments.

Requirements: programming in Python and/or C++, basics of computer vision, machine perception and deep learning

Literature:
[1] Matouš Vrba, Viktor Walter, Václav Pritzl, Michal Pliska, Tomáš Báča, Vojtěch Spurný, Daniel Heřt and Martin Saska. "On Onboard LiDAR-Based Flying Object Detection." Transactions on Robotics 41:593–611, 2025.
[2] Michal Pliska, Matouš Vrba, Tomáš Báča and Martin Saska. "Towards Safe Mid-Air Drone Interception: Strategies for Tracking & Capture." Robotics and Automation Letters 9(10):8810-8817, 2024.
[3] Matouš Vrba and Martin Saska. "Marker-Less Micro Aerial Vehicle Detection and Localization Using Convolutional Neural Networks." IEEE Robotics and Automation Letters 5(2):2459-2466, April 2020.

Although the agility of autonomous multirotor UAVs has made giant leaps recently, this progress has not yet been reflected in multi-UAV tasks such as swarming or formation flight. This project focuses on addressing the missing pieces necessary to realize a truly agile cooperative multi-UAV flight using only onboard equipment for mutual sensing between the team-members, and without relying on communication. This involves implementing new methods to detect not only the relative position of the other team-members, but also their full 3D pose using an RGB camera or even the current frequency of their propellers using a specialized event camera.

Requirements: programming in Python and/or C++, good knowledge of computer vision methods

Literature:
[1] Spetlik, Radim, Tereza Uhrová, and Jiří Matas. "Efficient Real-Time Quadcopter Propeller Detection and Attribute Estimation with High-Resolution Event Camera." Scandinavian Conference on Image Analysis. Cham: Springer Nature Switzerland, 2025.
[2] Yin Zhang, Zian Ning, Shiyu Zhao. "Observability-Enhanced Target Motion Estimation via Bearing-Box: Theory and MAV Applications." arXiv:2601.06887, 2026.
[3] Zhang, Yin, et al. "EvDetMAV: Generalized MAV Detection From Moving Event Cameras." IEEE Robotics and Automation Letters, 2025.

In biological systems, such as bird flocks or fish schools, fast coordinated maneuvers in cluttered environment with obstacles are a common phenomenon. In robotic systems however, this is still an open challenge. The goal of this project is to gain new insights about how these biological systems achieve such impressive behavior and transfer these findings to robotics. The project will primarily utilize our data of bird flocks captured using a novel method utilizing stereo event-based cameras.

Requirements: programming in Python and/or C++, good knowledge of computer vision methods, basics of multi-robot systems

Literature:
[1] Y. Ai, H. Zhai, Z. Sun, W. Yan, and T. Hu. "FlockSeer: A portable stereo vision observer for bird flocking." IET Cyber-Systems and Robotics, vol. 6, no. 3, p. e12118, 2024, doi: 10.1049/csy2.12118.
[2] M. Ballerini et al. "Interaction ruling animal collective behavior depends on topological rather than metric distance: Evidence from a field study." Proceedings of the National Academy of Sciences, vol. 105, no. 4, pp. 1232–1237, Jan. 2008, doi: 10.1073/pnas.0711437105.
[3] Vásárhelyi, Gábor, et al. "Optimized flocking of autonomous drones in confined environments." Science Robotics 3.20 (2018): eaat3536.

The goal of this project is to develop a new UAV platform fully compatible with the MRS UAV System. Design, build and experimentally validate both the mechanical and electrical parts of the UAV, with emphasis on a compact, durable frame, high flight efficiency, reliable power distribution, and integration of a companion computer for advanced autonomy. The design will be driven by three primary use-cases:
a) high payload capacity for carrying various sensors and equipment
b) agile, fast flight
c) long-duration / efficient cruise (optionally via VTOL concept)

Field experiments with UAVs often struggle due to missing infrastructure: slow GNSS/RTK setup, improvised power. This project creates a portable “base station in a suitcase” that can be deployed anywhere (e.g., a field), providing networking, RTK, telemetry, power, and experiment recording in one integrated system.
Design and build a portable base station (large suitcase) that contains everything needed to run UAV experiments in the field
- Long-range Wi‑Fi (AP + directional antenna support)
- Configurable on-site network (router/switch + managed configuration)
- RTK GNSS base station integrated and easy to operate
- Integrated battery/power distribution for hours of operation
- Camera system for filming experiments
- User interface: settings and status monitoring (RTK, Wi‑Fi AP setup)
The result should be a robust “unpack, power on, fly” station that reduces setup time and increases reliability and reproducibility of experiments.

Code-based computer-aided design allows a consistent, open, modular, and extensible approach to designing mechanical structures such as UAVs. Students working on this topic may choose from multiple aspects to work on: the implementation of a bidirectional code-graphical user interface is a software project that includes human-machine interaction. Extending the existing `build123things` Code-CAD with advanced design features, possibly implementing modern Rust-based geometry kernels, is more of a manufacturing-oriented topic. Besides, with the advent of large language models, finding a way to join the rigorous CAD with fuzzy natural language is an obvious topic to explore. Students interested in these topics are expected to be exceptionally motivated and proactive.
Reference: https://zenodo.org/records/17495691

This project focuses on developing a wireless charging landing pad for UAVs used in MRS robotic experiments. The key idea is to reuse the magnetic coils that are already being developed for relative localization between drones. These coils are mounted on the UAV and will be provided to the student. They are responsible for magnetic-based ranging, and in this project they will also be exploited for wireless energy transfer. The student will therefore concentrate on designing and developing the charging pad placed on the ground. The goal is to create a system that can efficiently transfer power to the existing onboard coils. A major part of the work will be the mechanical and electromagnetic design of the landing pad so that it naturally facilitates alignment between the drone-mounted coil and the pad coil. This includes studying coil geometry, resonant tuning, efficiency optimization, and tolerance to misalignment during autonomous landing. The robotic objective of this project is to enable “infinite-flight” missions, where drones can autonomously land, recharge, and take off again without human intervention. The system is also relevant for heterogeneous robotics scenarios, where aerial and ground robots cooperate.

Prototypes and one-off integrations often end up with an Arduino “in heatshrink” to provide extra I/O, level shifting, power rails, and bus bridging. This is fragile, hard to reproduce, and not well integrated into the MRS UAV System. This project aims to replace these ad‑hoc solutions with a compact, robust, reusable I/O module that plugs into the UAV X500 module (or USB) and provides standardized power and interfaces with proper software support.
Board features:
- Regulated power rails: 3.3V, 5V, and 12V outputs with current monitoring and protection
- Multiple I/O interfaces: GPIO, I2C, SPI, UART, and ADC channels
- Servo Outputs (PWM)
The result should be a functional, documented hardware module plus firmware and system integration, eliminating the need for temporary microcontrollers in field deployments.

The goal of this project is to design and validate a universal RPM meter module for UAV platforms capable of measuring RPM of 1–8 motors. The device should be easy to install on a wide range of multirotor and VTOL UAVs, provide reliable and synchronized RPM measurements for all channels, and stream the data to the onboard system for monitoring, logging, and diagnostics.
Choose and justify an RPM sensing approach (e.g., ESC telemetry, back‑EMF/phase sensing, optical or magnetic sensing). Implement a robust interface to the UAV system (e.g., UART/CAN/I2C/USB), including a simple protocol/message format for publishing per‑motor RPM data. Test accuracy, latency, and robustness under realistic conditions (different motor/prop combinations, varying throttle, vibration, EMI), and document limitations and recommended installation practices.

In this project, the student will design and build a custom thrust stand to test the performance of drone motors and propellers. The student will use a compact force-torque sensor (normally used in robotic manipulators) to precisely measure how much thrust and torque a motor produces during operation. The system must be able to handle high power levels, measuring voltages from 0 to approx. 80V and currents up to 250A, while also tracking motor RPM using optical sensors, magnets, or emf signals. The student will be responsible for choosing all electronic parts, designing the physical device using 3D printing or other tools, and creating an easy-to-use software interface for data collection. The student will ensure the stand supports both a manual mode for direct control and an automatic mode that runs pre-set test patterns like power ramps or step sequences. The system needs to be fast enough to take samples at 1000 Hz. For an extra challenge, the student can add a video recording feature to film the experiments or process the collected data to find specific motor parameters for the MRS UAV system. The student is encouraged to improve the project with their own ideas, such as adding better safety or completely new features.

This project focuses on the design and implementation of methods that enable a team of heterogeneous robots — an Unmanned Aerial Vehicle (UAV) and an Unmanned Surface Vehicle (USV) — to collaboratively manipulate an object on the water surface using tethers. The goal is to control the system such that the manipulated object follows a predefined trajectory on the water surface while both robots coordinate their motion despite their different dynamics and sensing capabilities.
References
[1] F. Novák, T. Báča and M. Saska. Collaborative Object Manipulation on the Water Surface by a UAV-USV Team Using Tethers. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). 2024, 3425-3432.

This project focuses on the design and implementation of a Model Predictive Control (MPC) framework for an Unmanned Surface Vehicle (USV), i.e., an autonomous boat. The goal is to develop a controller that enables the USV to accurately follow a desired trajectory while respecting the system dynamics and operational constraints.

Obstacles and the relative pose of both UAVs determine the availability of communication in the field. Whereas static modeling of the communication channel is a well-researched discipline, dynamic robotic scenarios still require further understanding. Students will explore lightweight data-driven communication quality models in practical aerial robotic scenarios. This hands-on topic entails theoretical analysis, novel implementation and real-world validation.

Dense swarms of UAVs easily overload the radio channel. However, with a proper model of communication availability, one could optimize the constellation to mitigate the cross-talks by exploiting radiation patterns or constellation spacing. Besides, behavior in the event of communication loss, a strategy for reconnection is of research interest as well. This hands-on topic entails theoretical analysis, novel implementation and real-world validation.

This project is part of the UVDAR project, which uses UV-LEDs mounted on the legs of UAVs to detect and track their 3D pose. The camera on the observing UAV uses special UV filters to detect the LEDs. Currently, UVDAR relies on image preprocessing and Particle Filter-based 3D pose estimation. The project would include the replacement of this method with a deep learning model that can handle occlusions and complex environments more robustly.

[1] V. Walter, N. Staub, A. Franchi and M. Saska, "UVDAR System for Visual Relative Localization With Application to Leader–Follower Formations of Multirotor UAVs," in IEEE Robotics and Automation Letters, vol. 4, no. 3, pp. 2637-2644, July 2019 [Online]. Available: https://ieeexplore.ieee.org/abstract/document/8651535

This project investigates how Optical Camera Communication (OCC) can be leveraged on UAVs by combining two complementary signaling mechanisms. Rolling-shutter cameras read out image rows sequentially, meaning that high-frequency LED blinking produces characteristic striped patterns in captured images. These patterns can be exploited to encode high-rate data. In parallel, the same LED can be modulated at a much lower frequency (50–70 Hz) to transmit low-rate control or status information, enabling a dual-channel communication system.

The project offers multiple possible directions depending on experimental outcomes and the student’s interests. As an initial step, controlled laboratory experiments will be conducted to analyze the rolling-shutter effect and quantify how modulation frequency, camera exposure, distance, and LED characteristics influence the resulting patterns. Based on these insights, students may design and implement a suitable modulation and decoding scheme for robust data transmission.

As an ultimate goal, the developed system will be integrated on multiple UAVs and tested during close-formation flight, demonstrating a lightweight, radio-free optical communication method for collaborative aerial robots.

The goal of this project is to design and implement algorithm for safe navigation of formation of UAVs through environment with obstacles by adapting the shape of the formation. The project consists of two tasks. First part will focus on modification of an existing algorithm [1] for formation shape adaptation for the use in curvilinear coordinate systems. Second part of the project will focus on development of algorithm for planning of a sequence of formation shapes that will lead to a safe navigation of the formation along given path in 3D environment. The developed algorithm is expected to be deployed on a fleet of UAVs in a real-world experiment.

Reference [1] V. Kratky, R. Penicka, J. Horyna et al., "CAT-ORA: Collision-Aware Time-Optimal Formation Reshaping for Efficient Robot Coordination in 3-D Environments," in IEEE Transactions on Robotics, vol. 41, pp. 2950-2969, 2025.

The goal of this project is to design and implement algorithm for formation shape adaptation in environments with obstacles with focus on time efficiency of the formation reshaping process. The designed algorithm is expected to be composed of solution to both the robot-to-goal assignment problem and trajectory generation problem in environments with obstacles. The developed algorithm will be tested in real-world experiments with unmanned aerial vehicles.

This project focuses on designing a safe and effective cooperative autonomy framework for multiple UAVs operating in both dense forest environments and open fields. Building on the baseline approach presented in [1], the project aims to extend the method by introducing minimal inter-robot communication to coordinate high-level motion and swarm behavior, integrating robust online replanning to improve performance and resilience, and implementing a reliable localization system that enables operation across diverse environments.

Reference [1] Boldrer, M., Kratky, V., Walter, V., & Saska, M. (2025). Distributed Lloyd-Based Algorithm for Uncertainty-Aware Multi-Robot Under-Canopy Flocking. arXiv preprint arXiv:2504.18840

This project focuses on developing a robotic underwater localization system based on AC magnetic fields.
The transmitting (TX) coils will be mounted on an autonomous surface boat, while receiving (RX) coils will be placed underwater.
Magnetic localization is particularly interesting in underwater robotics because conventional radio-frequency signals do not propagate reliably underwater, and vision-based systems often fail in turbid or low-visibility environments.
Magnetic fields, instead, can penetrate water and do not require line-of-sight.
The student will be responsible for simulating the magnetic field propagation in water (fresh and salt water scenarios), designing and dimensioning the TX and RX hardware (coil geometry, resonant tuning, driver and sensing electronics), and for developing software for position estimation using multi-coil measurements (e.g., nonlinear optimization or Learning based approaches).

Firma DJI společně s firmou Intel vyvinula unikátní dron Ryze Tello, který je připraven pro ovládání z počítače přes wifi připojení.
Cílem práce je vytvořit systém, který na základě informací z kamery a výškoměrů bude schopen orientovat se v uzavřeném prostředí a dovede se naučit cesty.

Seznamte se s ovládáním dronu Ryze Tello a robotickým operačním systémem ROS.
Seznamte se s algoritmy pro rojové chování dronů.
Připravte SW pro ovládání více dronů Ryze Tello.
Dále s pomocí výsledků navazujících prací implementujte skupinové chování, synchronizované provedení předdefinovaných plánů.

Drony jsou v poslední době hojně využívány k inspekci stožárů elektrického vedení. Cílem této práce je vytvořit model stožáru elektrického vedení z drony, která se pohybuje okolo stožáru. Prá se může zaměřit na vytváření modelu z obrazu stereo barevných kamer, z obrazu stereo termálních kamer, nebo s monokulárního systému při známé pozici drony. Téma je vhodné i pro více studentů, neboť je možné využít různé senzory pro tvorbu modelů a jejich výsledky pak vzájemně porovnat.

Termokamery poskytují nový pohled na okolí, které nás obklopuje. I pro roboty je využití termokamer výhodné, protože umožňuje detekovat objekty, které v obyčejné kameře jsou špatně viditelné. Cílem této práce je vytvořit algoritmy pro detekci objektů v termokameře, nebo v páru stereo termokamer. Pomocí detekce těchto známých předmětů (jako je stožár elektrického vedení, samotné elektrické vedení) navigovat dron při průzkumu stožáru, případně při přelétání mezi stožáry.

The objective of this work is to develop a system for motion planning and coordination of a group of helicopters and unmanned aerial vehicles (UAVs) to effectively combine the advantages of both platforms. Helicopters are capable of maneuvering at low speeds near obstacles, while fixed-wing aircraft benefit from higher maximum speeds, longer range, and extended flight times. The designed and implemented system will be verified in a robotic simulator and, upon successful realization, tested with real helicopters from the Multi-robot Systems group at the Department of Cybernetics.

The goal of this project is to design, implement, and experimentally verify a drone control system capable of autonomously localizing fire hotspots. The student may choose to focus on the software components, the hardware components, or a combination of both. The SW part includes motion planning, control, and stabilization of the unmanned helicopter. The HW part consists of designing a fire-extinguishing mechanism and integrating it onto the drone's platform. This work aims toward participation in an international competition in Abu Dhabi, competing against the world's top teams (http://mrs.felk.cvut.cz/projects/mbzirc). Successful completion of the thesis offers the possibility of an internship at world-leading robotics laboratories, such as the GRASP Lab at the University of Pennsylvania or New York University.

The objective of this work is to develop a system for the planning and coordination of a group of unmanned helicopters in an autonomous cooperative pickup task for construction purposes. Based on provided uncertain information regarding object locations, the student will design an algorithm utilizing robotic coverage and TSP principles to plan collision-free trajectories. The goal is to obtain high-resolution imagery of targeted areas and subsequently plan trajectories to locations with confirmed object presence (the actual autonomous pickup is not a mandatory part of the thesis). The project is directed toward participation in the international MBZIRC competition in Abu Dhabi. Successful results may lead to an internship opportunity at the GRASP Lab, University of Pennsylvania.

Explore algorithms for the control and navigation of autonomous formations inspired by the movement of flocks of birds, schools of fish, or swarms of insects. Design and implement a suitable algorithm for the control and stabilization of a swarm of unmanned helicopters. The swarm should be capable of navigating dynamic environments with obstacles, with the swarm shape automatically adapting to specific mobile robotics tasks. The method will be verified and analyzed through simulations and partial real-world experiments with the Multi-robot Systems group's UAVs.
Prerequisites: Basic knowledge of C or Python programming.