The final tests and demonstrations took place from 24 to 27 August at Fliegerhorst Erding air base in Bavaria. The extensive site, featuring streets, buildings, airport and industrial areas, forests, meadows, and dense vegetation, allowed the organisers to prepare tasks resembling real-world deployment. A fundamental condition was that the drone could not use GPS and, once airborne, had to complete the mission entirely autonomously, without any intervention from a pilot.
“The Czech F4F & CTU SkyGraph team demonstrated by far the best solution in the injured-person search task and impressed the jury with the most general solution and the greatest commercial potential,” said Kilian Seitz, competition organiser at SPRIND, assessing the Czech team’s performance.
The organisers ultimately named the Austrian HERMES team from the Control of Networked Systems (CNS) research group at the University of Klagenfurt the overall winner. In addition to HERMES and F4F & CTU SkyGraph, the German UBASWARM team, bringing together the Technical University of Berlin and Berlin-based technology startup Autonx, and the FAU-LM team from the FAPS Institute at Friedrich-Alexander-Universität Erlangen-Nürnberg, associated with Nouma Autonomy, advanced to the final system demonstration for invited investors and representatives of European security and defence organisations. In this final demonstration of the competition tasks, the Czech team performed best among all finalists.
“For us, the most important outcome of the competition is not merely the ranking itself, but confirmation that a research concept is becoming a generally applicable technology. Today, deploying similar robotic systems often requires a team of specialists to prepare a mission and precisely programme how it will be carried out. Our goal is for the user to simply describe the desired outcome and for the system itself to determine how to achieve it. This significantly reduces both costs and operational requirements and opens the way for the practical use of flying robots in industry, logistics, security, and rescue operations,” says Martin Saska, head of the F4F & CTU SkyGraph team, professor at the Faculty of Electrical Engineering of the Czech Technical University in Prague, and founder of Fly4Future.
Four Examples of Real-World Applications of the Developed System
Each of the four final challenges was based on a situation that delivery services, rescue teams, or security control centres may encounter in the future. The drone was not given a pre-programmed route. It received only an instruction formulated in human language, from which it had to understand the purpose of the task and independently decide how to complete it.
Challenge 1: A Package by the Blue Bicycle
The first scenario simulated a request from a delivery service customer: leave a package by the blue bicycle in front of house number 15. From a single sentence, the drone correctly inferred all the necessary steps, found the specified house and bicycle, and selected a landing site. However, under changing lighting conditions, it mistook blue for green. The system therefore correctly understood the complex assignment, while its visual distinction between similar colours still requires further refinement.
Challenge 2: A Customer Waving from a Window
In the second scenario, a customer informed the operator that they could not come to the door but would wave from a window of house number 14. The drone was tasked with finding the person and delivering the package to the entrance directly below them. On its first attempt, it identified a waving person reflected in the glass and therefore “in the window”, while the actual person was leaning out of an open window and, for precisely that reason, did not match the wording “person in the window”. After a simple clarification of the instruction, the system found the correct person. Its behaviour could therefore be adjusted using an ordinary sentence rather than through intervention by a programmer. “While solving the task, it became clear that the robotic system needs to be provided with as much information as possible so that it can correctly evaluate the given request. A simply worded instruction can sometimes be interpreted correctly by the robot in several different ways, which a person may not realise,” commented Tomáš Musil, a PhD student at the Faculty of Electrical Engineering of the Czech Technical University in Prague, who is developing the theory and methods on which the competition system was built as part of his doctoral research at the Department of Cybernetics.
Challenge 3 and the Czech Team’s Victory: Helping an Injured Person
The third task simulated an instruction from a rescue control centre operator. The drone was tasked with finding an injured person wearing a red T-shirt and delivering a medical kit to them. The system independently located the person, who was lying partially concealed behind a building and leaning against a tree. However, it did not find sufficient space immediately beside them for a safe landing and therefore refused the manoeuvre. On another attempt, it landed on a nearby roof – geographically close, but practically out of reach of the injured person. The test showed that a robot needs to understand not only the formal objective but also the broader purpose of the mission. Nevertheless, despite these complications, the Czech team demonstrated the best solution of all participants in this task. During the demonstration for investors on the final day of the competition, it was also the only team to complete the assigned task repeatedly and entirely without error.
“I was pleased that we won this particular task because many Fly4Future clients are looking for a similar system, and we have also recently secured a major national project focused on addressing fires in buildings, where finding and rescuing people is always a priority,” commented Tomáš Báča, the team’s technical lead, a researcher at CTU FEE, and co-owner of Fly4Future.
Challenge 4: Tracking a Person within the Site
The final scenario corresponded to a situation that a security control centre operator might encounter: finding a person wearing a blue T-shirt moving within a monitored site and tracking them for as long as possible. A specialised module for predicting the person’s movement did not function as expected, but the general-purpose system was able to repeatedly locate the person, approach their new position, and reacquire them even when they were partially obscured by bushes. Its ability to continue using a more universal solution confirmed that the Czech technology was not designed solely for a single pre-programmed situation.
Four Cameras, 3D LiDAR, and a Map That Understands Its Surroundings
The technological foundation of the Czech system consists of four cameras, 3D LiDAR, and a powerful onboard computer. Using these components, the drone localises itself without GPS, maps its surroundings, and recognises objects required to complete the assigned mission.
Detected buildings, people, vehicles, and other objects are inserted into a continuously generated spatial map. A graph structure is then created on top of this map, capturing not only individual objects but also the relationships between them – for example, that a bicycle is located in front of a particular house or that a person is standing in its window. The system can then “ask” the map where an object, or even a complex situation corresponding to a human instruction, is located and plan its subsequent movement based on the answer.
The greatest advance, however, does not lie in any single sensor or algorithm, but in connecting the entire chain. An instruction from a customer is transformed into a mission plan without intervention by a technician, the drone creates a map of an unknown environment, searches for the required objects, and independently decides what to do next.
“It is precisely this simple connection between the end user and the robotic system, which does not require the involvement of a robotics expert or programmer, that represents one of the last technological barriers preventing aerial robotics from becoming widely adopted across many areas of human activity,” Professor Martin Saska said, commenting on the importance of the challenge during his presentation to invited competition participants.
Seven European Teams in the Final and More Than EUR 5 Million in Funding
Fully Autonomous Flight 2.0 was a fourteen-month European innovation competition that began on 1 July 2025. In the first phase, participating teams could receive funding of up to EUR 150,000. The seven teams that advanced to the second phase could receive up to an additional EUR 350,000 for further development. The funding made it possible to integrate individual research results into a comprehensive system and thoroughly test it under conditions resembling real-world operation. The competition’s total budget allocated to both phases exceeded EUR 5 million.
The Czech F4F & CTU SkyGraph team advanced to the second stage in first place. The FAU-LM, MUNWAG, Avientus RadNav, HERMES, Fox Storm, and UBASWARM teams also qualified. F4F & CTU SkyGraph and FAU-LM advanced to the final demonstration for investors and journalists on 27 August 2026 for Challenges 2 and 3, while HERMES and UBASWARM advanced for Challenges 1 and 4. The Czech team won both final competitions, while the Berlin-based UBASWARM team excelled in the other pairing, even though its competitor HERMES had won the previous competition days and thus effectively the competition as a whole. This also demonstrates how closely matched the competition was and the high quality of the solutions presented during the tests.
From a Research Project to Practical Application
The outcome of the competition is not merely a one-off prototype prepared for four specific tests. Fly4Future is already gradually integrating the technologies developed into its robotic platforms and preparing them for commercial use.
A general-purpose drone controlled through natural language could assist rescue teams in searching for people, firefighters in locating fires and monitoring hotspots, security control centres in monitoring large sites, or industrial companies in inspecting difficult-to-access locations. It could also be used for package delivery, smart agriculture, nature conservation, or mapping areas affected by emergencies.
The key commercial benefit lies in reducing operational requirements. A single, more universal system can gradually take on different types of missions without the need to develop a separate drone and new control software for each deployment. Natural language also enables the technology to be operated by people who understand the real-world problem – such as dispatchers, firefighters, inspectors, or members of rescue teams – without requiring detailed knowledge of robotics and programming.
