From a bachelor’s thesis to a global conference: FEL CTU student develops a tool for visualizing the behavior of artificial intelligence agents

Bc. Erol Medenčević, a first-year master’s student in the Open Informatics program at the Faculty of Electrical Engineering of Czech Technical University in Prague (FEL CTU) and a researcher at the Industrial Informatics Department of the Czech Institute of Informatics, Robotics and Cybernetics (CIIRC CTU), succeeded with his work on visualizing the behavior of artificial intelligence agents at the international AAAI (Association for the Advancement of Artificial Intelligence) conference, held from January 20 to 27, 2026, in Singapore.

The resulting project, PANSim, was presented in the demo track of this prestigious conference, which is one of the highest-ranked events in the field of artificial intelligence.

Understanding how AI agents make decisions

Artificial intelligence can make many aspects of human life easier. However, understanding how an AI agent actually makes decisions can be difficult for the general public to imagine. In many cases, even AI developers themselves struggle to infer these decision-making processes from available logs alone.

This is why the PANSim project was created. It visualizes the behavior of AI agents when planning in a nondeterministic environment, meaning an environment influenced by external factors that the agent cannot control but must account for. These kinds of “acts of nature” are currently one of the main challenges in robotics and automation.

PANSim helps reveal the “thought processes” behind AI decision-making and planning, both for non-experts and for developers themselves. Thanks to the visualization, developers can more easily detect errors and fine-tune their agents.

The project originated as a bachelor’s thesis

Bc. Erol Medenčević currently works at the Industrial Informatics Department at CIIRC CTU, specifically in the research team of Assoc. Prof. Lukáš Chrpa, where he continues developing the project alongside his studies at FEL CTU. He would also like to connect the future development of PANSim with his master’s thesis, just as he did with his bachelor’s thesis, in which PANSim originally emerged.

At its core, the project integrates the Unity game engine, which provides smooth graphical simulation, together with a non-trivial number of programs responsible for computation and logical planning. Each of these programs is quite different and written in a different programming language.

“I was drawn to the possibility of connecting my field of study, computer graphics, with artificial intelligence. Until then, I had not worked much with planning, so at first I mainly focused on understanding it, while the graphical part took a bit of a back seat. The first visualization was therefore fairly simple, but even then I liked that, despite its simplicity, it provided much deeper insight into the simulation and agent behavior than a typical command-line output,” explains Medenčević.

The bachelor’s thesis was supervised by Ing. Jakub Med, with whom Medenčević continues to collaborate closely on PANSim.

“The project is primarily complicated because it integrates many algorithms from the scientific community that are not yet ready for real-world deployment, which is common for research prototypes. Erol often encountered problems where different approaches were not fully compatible and had to be properly integrated. Despite this, he proactively came up with ideas on how to improve the tool and move it forward,” says Med, who is a PhD student at FEL CTU and a researcher at the Industrial Informatics Department of CIIRC CTU.

The current state of PANSim

At present, PANSim visualizes two scenarios:

• a simulation inspired by the classic game Perestroika, and

• operations of underwater drones navigating between ships.

In both cases, the agent must find a guaranteed safe path to its goal, despite obstacles caused by external factors. To find such a path, the agent uses techniques developed in the team of Assoc. Prof. Lukáš Chrpa at the Industrial Informatics Department of CIIRC CTU, where the FEL CTU student first began working on the project.

At the moment, the tool is most useful not only for the general public but especially for researchers working with one of these two scenarios. However, Medenčević intends to expand the project with additional scenarios.

“In the future, I would like PANSim to become domain-independent, meaning it could function as a universal tool where researchers can simulate any scenario they design themselves,” says Medenčević.

At the same time, he acknowledges that this is a very ambitious goal. In the near future, he plans to focus more on making the program interactive, so that it becomes attractive even to people outside the field and helps spark greater public interest in how artificial intelligence works.

“Users could try solving a specific problem in PANSim themselves and then compare their results with how the computer performed in the same situation,” Medenčević explains.

Participation at AAAI 2026

At the beginning of the year, the 40th edition of the prestigious international AAAI conference took place in Singapore from January 20 to 27, 2026. The conference focuses on artificial intelligence and is organized by the Association for the Advancement of Artificial Intelligence. It is considered one of the top-ranked AI conferences in the world.

PANSim was accepted into the demo track, the section dedicated to practical demonstrations. Only 28% of submitted contributions were accepted into this section at this year’s conference.

“I’m extremely happy about it, and honestly it was a big surprise for me. It showed me that the project not only fulfilled my academic requirements but also has broader impact and real usefulness, which motivates me to keep working on it,” says Medenčević.

Feedback during the conference was very positive. The booth where the project demonstration took place attracted continuous interest from attendees.

“It seemed to me that the project was particularly interesting to people from the reinforcement learning and planning communities. We exchanged contacts with several of them. We hope to build some collaborations based on the presentation,” comments Med, who presented the project at the conference.

Due to the overlap with the examination period, the first author, Bc. Medenčević, was unfortunately unable to attend the conference in person. Even so, he considers the acceptance of the project at AAAI a major success.

“It motivated me to learn much more about planning. This semester I enrolled in a course on the topic, so in the future I might not only focus on the graphical side of things but also on the behavior of the agents themselves,” adds Medenčević.

More information about PANSim, including the scientific paper, can be found in the project repository:
https://gitlab.com/automated-planning/tools/pansim.