As artificial intelligence becomes increasingly integrated into everyday life, the need to understand why AI systems make particular decisions continues to grow. The award is especially significant because the winner was not selected by a jury but by the conference participants themselves—the global community of researchers specializing in automated planning and AI decision-making. Among twelve demonstration projects from leading universities and research institutions worldwide, PANSim received the highest number of votes.
"This recognition is particularly meaningful because ICAPS is the world's premier conference dedicated to automated planning. Unlike large general AI conferences, ICAPS brings together virtually the entire research community in our field. We are especially proud that it was this community that chose our tool as the best demonstration of this year's conference," says Associate Professor Lukáš Chrpa from the Department of Industrial Informatics at CIIRC CTU.
When It's Not Enough to Know That AI Made the Right Decision
Artificial intelligence is already helping to control robots, autonomous vehicles, and industrial systems. Yet understanding why an algorithm arrived at a particular decision is often far more difficult than verifying whether the outcome itself was correct. Even developers frequently struggle to reconstruct complex decision-making processes from text logs or program outputs alone.
PANSim transforms AI decision-making into an interactive visual simulation. It focuses on symbolic artificial intelligence, which relies on explicit modelling of the environment and step-by-step planning rather than statistical learning. Users can observe, step by step, how an intelligent agent evaluates a situation, responds to changes in its environment, and searches for a safe path toward its goal. This makes it significantly easier to identify algorithmic errors, understand the agent's reasoning, and further improve the system.
"It's not just about graphics. Visualization makes it much easier to understand what is actually happening during the decision-making process. Developers can quickly see why the agent behaved in a particular way and where the algorithm needs to be improved," explains Ing. Jakub Med, supervisor of Erol Medenčević's bachelor's thesis and a PhD student at FEE CTU, who also works at the Department of Industrial Informatics at CIIRC CTU.
To demonstrate complex AI reasoning as clearly as possible, the authors chose a simple scenario. A frog must cross a pond while collecting all the coins. Finding the shortest route is not enough, however, because some lily pads may disappear beneath the water at any moment. The agent therefore has to continuously reassess its options and react to unpredictable environmental changes. A second scenario simulates autonomous underwater vehicles that must safely respond to ships moving above the surface while carrying out their mission. In addition, the researchers require the agent to provide guarantees of successful task completion, distinguishing their approach from traditional methods based on machine learning algorithms.
From Bachelor's Thesis to Ongoing Research
PANSim is not the first international success for FEE CTU student Erol Medenčević, who continues developing the project during his master's studies. Earlier this year, PANSim was presented at the prestigious AAAI Conference on Artificial Intelligence in Singapore. While AAAI is one of the world's largest AI conferences, attracting thousands of participants across the field, ICAPS serves as the leading international forum dedicated specifically to automated planning.
"When our paper was accepted at a top international conference, I realized the project had real potential. But when my colleagues sent me a photo from Dublin showing the certificate for Best Demonstration, I was genuinely surprised. It's a huge motivation to keep moving forward," says Bc. Erol Medenčević.
In the next stage of development, the team plans to make AI reasoning even more transparent. Planned features include the ability to rewind simulations by several steps, better visualization of the agent's current intentions, and additional simulation scenarios. Their long-term ambition is to create a universal platform capable of visualizing a wide range of automated planning problems.
Although PANSim is currently intended primarily for researchers, the principles behind it have applications in areas such as autonomous robotics and cybersecurity, where systems must continuously respond to unpredictable changes in their environment. As autonomous systems become more widespread, understanding how artificial intelligence reasons—and why it makes particular decisions—will become increasingly important.
