When AI meets the real world

What happens when artificial intelligence leaves the lab and begins interacting with the physical world? From autonomous vehicles and delivery drones to robotic assistants and automated surgery systems, AI-driven cyber-physical systems are becoming part of everyday life. But as these systems grow more powerful, the question of how to verify their safety and reliability becomes increasingly urgent.

This challenge was at the center of a recent Shonan Meeting, an international research seminar hosted by the National Institute of Informatics (NII) in Japan. The format is deliberately selective: around thirty researchers from different disciplines are invited to spend several days discussing emerging research directions in depth. Participation is widely regarded as a recognition of a researcher’s contribution to a field.

Among the invited participants was Prof. Daniel Neider, who holds the Chair of Verification and Formal Guarantees of Machine Learning at the Department of Computer Science at TU Dortmund University and is a principal investigator at the Research Center Trustworthy Data Science and Security (RC Trust). His work focuses on the intersection of machine learning, logic, and formal verification–a perspective that fits closely with the seminar’s central theme.

The meeting explored the topic LLM-guided Synthesis, Verification, and Testing of Learning-Enabled Cyber-Physical Systems. Cyber-physical systems combine software with physical processes, meaning that AI models must interact with sensors, environments, and real-world uncertainty. Traditional verification methods were designed for deterministic software, but AI-based systems behave probabilistically and depend heavily on data–making them far more difficult to test and validate.

Researchers at the seminar discussed how large language models (LLMs) could help address these challenges. One promising direction is using LLMs to analyze large volumes of system data and human-written rules in order to generate formal specifications or realistic test scenarios. This could make it easier to uncover rare “corner cases” where systems might fail – an essential step in ensuring the safety of AI-enabled technologies.

For Neider, these discussions connected closely with his own research on formal methods for machine learning systems. His work investigates how logical specifications can be learned from data and used to verify the behavior of AI systems–an approach that could play a key role in making complex AI applications more transparent and dependable.

Beyond individual talks, the seminar focused strongly on collaborative discussions. Participants worked in small groups to identify open research questions and to outline future research agendas. The goal is not only to exchange ideas but also to shape new research directions for the field.

The relevance of these discussions goes far beyond academia. As AI becomes embedded in safety-critical technologies–from transportation to robotics–developing reliable testing and verification methods will be essential for public trust.

By bringing together expertise from software engineering, robotics, control theory, formal methods, and artificial intelligence, the Shonan Meeting highlighted how interdisciplinary collaboration is needed to tackle these challenges. For RC Trust, the themes discussed at the seminar align closely with the center’s mission: advancing methods that make intelligent systems secure, reliable, and trustworthy by design.

And as AI systems increasingly operate in the physical world, that mission is becoming more important than ever.