Foundation Models Meet Control: Promises, Pitfalls, and Open Questions
Event details
| Date | 05.03.2026 |
| Hour | 11:00 › 12:00 |
| Speaker | Prof. Simone Formentin, Associate Professor of Automatic Control Politecnico di Milano, Italy |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
Abstract:
Foundation models are rapidly reshaping engineering and science. But what do they truly imply for control systems?
Are they merely powerful black-box approximators, or do they introduce a new computational paradigm for modeling, estimation, and controller synthesis?
This talk revisits foundation models from a control-theoretic perspective. We first clarify how they fundamentally differ from classical system identification: rather than learning a single dynamical model, foundation models learn conditional distributions over families of admissible solutions.
Inference and design are thus performed through conditioning, rather than through explicit parameter estimation.
Building on this viewpoint, we present two research directions that illustrate their potential in model-based and data-driven control, respectively.
The first leverages diffusion models to learn and sample from the distribution of stabilizing controllers for a given plant.
The second explores in-context learning for control and state estimation, showing how transformer-based architectures can generate filters and feedback laws for previously unseen systems.
Large-scale pretraining in simulation enables rapid adaptation with minimal contextual data, effectively shifting the sim-to-real challenge from model fidelity to distributional coverage, that is, from a model-centric to a data-centric paradigm.
Despite their promise, foundation models raise fundamental questions for control, including stability guarantees, robustness under distribution shift, computational burden, and interpretability.
We argue that these limitations should not discourage their adoption in feedback systems, but instead motivate deeper integration: control theory can contribute back to foundation models by providing system-theoretic analysis tools, structural parameterizations, and principled notions of safety and robustness.
Biography:
Prof. Simone Formentin was born in Legnano, Italy, in 1984. He received his B.Sc. and M.Sc. degrees cum laude in Automation and Control Engineering from Politecnico di Milano, Italy, in 2006 and 2008, respectively. In 2012, he obtained his Ph.D. degree cum laude in Information Technology through a joint program between Politecnico di Milano and Johannes Kepler University of Linz, Austria. He subsequently held postdoctoral positions at the Swiss Federal Institute of Technology of Lausanne (EPFL), Switzerland, and at the University of Bergamo, Italy. Since 2014, he has been with Politecnico di Milano, where he is currently an Associate Professor of Automatic Control. From 2020 to 2025, he served as Chair of the IEEE Technical Committee on System Identification and Adaptive Control. He is currently an Associate Editor of Automatica and the European Journal of Control His research interests include data-driven modeling and control, with applications to automotive systems and quantitative finance.
Foundation models are rapidly reshaping engineering and science. But what do they truly imply for control systems?
Are they merely powerful black-box approximators, or do they introduce a new computational paradigm for modeling, estimation, and controller synthesis?
This talk revisits foundation models from a control-theoretic perspective. We first clarify how they fundamentally differ from classical system identification: rather than learning a single dynamical model, foundation models learn conditional distributions over families of admissible solutions.
Inference and design are thus performed through conditioning, rather than through explicit parameter estimation.
Building on this viewpoint, we present two research directions that illustrate their potential in model-based and data-driven control, respectively.
The first leverages diffusion models to learn and sample from the distribution of stabilizing controllers for a given plant.
The second explores in-context learning for control and state estimation, showing how transformer-based architectures can generate filters and feedback laws for previously unseen systems.
Large-scale pretraining in simulation enables rapid adaptation with minimal contextual data, effectively shifting the sim-to-real challenge from model fidelity to distributional coverage, that is, from a model-centric to a data-centric paradigm.
Despite their promise, foundation models raise fundamental questions for control, including stability guarantees, robustness under distribution shift, computational burden, and interpretability.
We argue that these limitations should not discourage their adoption in feedback systems, but instead motivate deeper integration: control theory can contribute back to foundation models by providing system-theoretic analysis tools, structural parameterizations, and principled notions of safety and robustness.
Biography:
Prof. Simone Formentin was born in Legnano, Italy, in 1984. He received his B.Sc. and M.Sc. degrees cum laude in Automation and Control Engineering from Politecnico di Milano, Italy, in 2006 and 2008, respectively. In 2012, he obtained his Ph.D. degree cum laude in Information Technology through a joint program between Politecnico di Milano and Johannes Kepler University of Linz, Austria. He subsequently held postdoctoral positions at the Swiss Federal Institute of Technology of Lausanne (EPFL), Switzerland, and at the University of Bergamo, Italy. Since 2014, he has been with Politecnico di Milano, where he is currently an Associate Professor of Automatic Control. From 2020 to 2025, he served as Chair of the IEEE Technical Committee on System Identification and Adaptive Control. He is currently an Associate Editor of Automatica and the European Journal of Control His research interests include data-driven modeling and control, with applications to automotive systems and quantitative finance.
Practical information
- General public
- Free