Control by conditioning: amortized inference over dynamical systems

Title

Control by conditioning: amortized inference over dynamical systems

Speaker

Simone Formentin - Politecnico di Milano

Abstract

Foundation models can be viewed as amortized inference procedures trained over a distribution of tasks. In this talk, we adopt this perspective for dynamical systems and control, reframing classical system identification and filtering as inference problems over latent dynamical systems. We show how learning-based models can implement a shared inference operator across systems, enabling state estimation and feedback control through conditioning on observed trajectories, without explicit parameter estimation. We illustrate this idea through real-world experiments on challenging industrial and automotive applications, and along two complementary directions: in-context learning for control loops using transformer architectures, and generative approaches to controller design, where diffusion models sample stabilizing controllers conditioned on plants and desired closed-loop specifications. This perspective shifts the focus from model-centric design to data-centric learning, where generalization depends on the coverage of the underlying system distribution. Rather than eliminating models, it redefines their role: high-fidelity digital twins become essential tools for generating representative training data. We conclude by outlining emerging theoretical challenges at the interface of learning and control, including principled design of training distributions, generalization across dynamical systems, and the propagation of learning errors in closed loop.

Bio

Simone Formentin is an Associate Professor of Automatic Control at Politecnico di Milano (Italy). He received his Ph.D. in Information Technology in 2012 from Politecnico di Milano and Johannes Kepler University of Linz (Austria), and held postdoctoral positions at EPFL (Switzerland) and the University of Bergamo (Italy). His research focuses on learning-based methods for dynamical systems and control, with applications to automotive systems and quantitative finance. He served as Chair of the IEEE Technical Committee on System Identification and Adaptive Control (2020–2025) and is currently an Associate Editor of Automatica and the European Journal of Control.

When

Thursday, April 30th, 14:30

Where

Room 322, UniGe DIBRIS/DIMA, Via Dodecaneso 35