Research

Architectures and coordinates for reliable control, optimization, and learning

We study how dynamical systems and algorithms can improve from data, adapt throughout operation, and still admit rigorous analysis.

Coordinates for reliable and adaptive control

Learning-based controllers are commonly described by their parameters, yet guarantees concern the trajectories they produce. We study alternative coordinates in which stability and robustness can be imposed directly while performance is learned from data.

This viewpoint supports nonlinear control, reinforcement learning, and adaptation after deployment. Current questions include how expressive these coordinates are, how they behave under model uncertainty, and how a controller can continue to adapt over long periods without invalidating its guarantees.

Selected publications

Neural System Level Synthesis: Learning over All Stabilizing Policies for Nonlinear Systems, IEEE CDC, 2022.
Learning to Boost the Performance of Stable Nonlinear Systems, IEEE Open Journal of Control Systems, 2024.
Parametrizations of All Stable Closed-loop Responses: From Theory to Neural Network Control Design, Annual Reviews in Control, 2025.
MAD: A Magnitude And Direction Policy Parametrization for Stability-Constrained Reinforcement Learning, IEEE CDC, 2025.
Characterizing All Locally Exponentially Stabilizing Controllers as a Linear Feedback Plus Learnable Nonlinear Youla Dynamics, by Luca Furieri. IFAC World Congress, 2026.
Stability-Preserving Online Adaptation of Neural Closed-loop Maps, by Danilo Saccani, Luca Furieri, and Giancarlo Ferrari-Trecate. Preprint, 2026.
Safety-Aware Performance Boosting for Constrained Nonlinear Systems, by Danilo Saccani, Haoming Shen, Luca Furieri, and Giancarlo Ferrari-Trecate. Preprint, 2026.

Map from stable operator coordinates to nonlinear stabilizing controllers — A tractable coordinate space is mapped into a broad class of stabilizing nonlinear controllers.

Equations expressing a nonlinear stabilizing controller as a stabilizing feedback law plus stable learned dynamics

Distributed control and networked systems

Large-scale systems must often be controlled using local measurements and limited communication. One set of results identifies closed-loop variables in which information constraints can be represented and optimized without giving up stability. These include the input-output parametrization and sparsity-invariance methods for optimal distributed control.

We currently study how these ideas extend to data-driven and nonlinear settings: distributed reinforcement learning, graph neural network policies, multi-agent coordination, and performance criteria that reflect how disturbances propagate through a network.

Selected publications

Optimality in Distributed Control: From Convex Programming to Reinforcement Learning, doctoral thesis, ETH Zürich, 2020.
An Input-Output Parametrization of Stabilizing Controllers: Amidst Youla and System Level Synthesis, IEEE Control Systems Letters, 2019.
On the Equivalence of Youla, System-Level and Input-Output Parameterizations, IEEE Control Systems Letters, 2019.
Sparsity Invariance for Convex Design of Distributed Controllers, IEEE Transactions on Control of Network Systems, 2020.
Learning the Globally Optimal Distributed LQ Regulator, L4DC, 2020.
Distributed Control of Network Systems in the Space of Stabilizing Graph Neural Network Policies, by John Cao and Luca Furieri. IFAC World Congress, 2026.
A Graph-Informed Regret Metric for Optimal Distributed Control, by Daniele Martinelli, Andrea Martin, Giancarlo Ferrari-Trecate, and Luca Furieri. IFAC World Congress, 2026.

Equations and diagrams for distributed controller design in closed-loop coordinates — Distributed control formulated through achievable closed-loop responses and information structure.

Ten-agent simulation of a learned distributed controller.

Multi-agent experiments

Distributed policies across network sizes

Every example uses the same trained graph neural network policy. Stability is guaranteed by construction, and no online optimization is used.

Six agents moving to randomly assigned targets — 6 agents · random targets

Eight agents coordinating to reach randomly assigned targets — 8 agents · random targets

Ten agents moving through a mixed target configuration — 10 agents · mixed configuration

Twelve agents exchanging positions in paired formation — 12 agents · paired exchange

System theory for learning and optimization

An iterative optimization method is itself a dynamical system. We use this observation to design trainable algorithms whose convergence follows from their structure, rather than from a posteriori tests on a finite training horizon.

The broader goal is to understand which algorithmic coordinates combine guarantees with enough expressivity to exploit recurring problem structure. This includes learning optimizers from data, accelerating classical iterations, and characterizing entire classes of convergent algorithms.

Selected publications

Learning to Optimize with Convergence Guarantees Using Nonlinear System Theory, IEEE Control Systems Letters, 2024.
Learning to Optimize with Guarantees: A Complete Characterization of Linearly Convergent Algorithms, by Andrea Martin, Ian R. Manchester, and Luca Furieri. Preprint, 2025.
Learning Over-Relaxation Policies for ADMM with Convergence Guarantees, by Junan Lin, Paul J. Goulart, and Luca Furieri. Preprint, 2026.

Performance of a learned optimizer compared with classical first-order methods — Learned optimization dynamics compared with classical first-order methods.

An unconstrained learned optimizer diverging beyond its training horizon — Why behaviour beyond the finite training horizon must be part of the design.

Slides

Recent presentations

ECC 2026 · Reykjavík · 7 July 2026

The Legacy of Youla: Stable Coordinates for Neural-Network Control

Presented at the European Control Conference 2026 workshop Control Theory: Yesterday's News or Tomorrow's Foundation?.

SIAM Conference on Optimization 2026

Learning to Optimize: Guarantees from Convex to Non-Convex Landscapes

Invited presentation on trainable optimization algorithms designed through nonlinear system theory.