⚡ electrical engineering

Free Parametrization of L_2-Bounded Structured State-Space Controllers for Nonlinear Control with Stability Guarantees

This paper introduces the L2-Recurrent Unit (L2RU), a structured state-space model layer that guarantees closed-loop stability for nonlinear control systems via a free L2-bounded parametrization, thereby enabling fully unconstrained optimization of complex objectives while ensuring robustness and efficient long-sequence processing.

Original authors: Muhammad Zakwan, Leonardo Massai, Efe C. Balta, Giancarlo Ferrari-Trecate

Published 2026-06-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Muhammad Zakwan, Leonardo Massai, Efe C. Balta, Giancarlo Ferrari-Trecate

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot (or a whole swarm of them) how to navigate a crowded room, avoid bumping into walls, and reach a specific destination. You want the robot to be smart and flexible enough to handle any tricky situation, but you also need a guarantee that it won't suddenly go crazy and crash into everything.

This paper presents a new way to build the "brain" of such a robot using Artificial Intelligence (AI). Here is the breakdown of their solution using simple analogies:

The Problem: The "Wild Horse" AI

Usually, when we use Neural Networks (AI brains) for control, they are like wild horses. They are incredibly strong and can learn complex tasks, but they are unpredictable. If you give them a tiny nudge (like a slight change in the wind or a sensor glitch), they might panic and run off the track, causing the whole system to become unstable.

To stop them from running wild, engineers usually have to put them in a "fence" (mathematical constraints) during training. But building these fences is slow, expensive, and computationally heavy, like trying to herd cats while running a marathon.

The Solution: The "Speed-Limited" Engine

The authors propose a new type of AI architecture called L2RU (L2-Recurrent Unit). Think of this as building a car engine that has a built-in, unbreakable speed governor.

The "Speed Governor" (L2-Bound): In engineering, there is a concept called an "L2-gain," which is essentially a measure of how much a system can amplify a disturbance. If you push the system a little, how hard does it push back?
The Innovation: The authors created a "free parametrization." This is a fancy way of saying they designed the engine so that no matter how you tweak the knobs (the parameters), the speed governor never breaks. You can turn the engine up or down to make it faster or smarter, but it will never exceed a specific safety limit.

How It Works: The Small-Gain Theorem

The paper relies on a classic rule called the Small-Gain Theorem. Imagine a game of tug-of-war between the robot (the controller) and the environment (the plant).

If the robot pulls too hard, the rope snaps (instability).
The theorem says: If the robot's "pulling strength" is strictly less than the environment's "resistance," the system will never snap.

Because the authors' new engine (L2RU) guarantees the robot's pulling strength stays below a specific limit, they can simply set that limit low enough to ensure the tug-of-war never ends in a crash. This means they can train the robot to be as smart as possible without having to stop and check if it's safe at every step.

The "Parallel Scan" Superpower

Another cool feature is speed. Traditional AI models that remember the past (like Recurrent Neural Networks) usually have to process information one step at a time, like reading a book page by page.

The L2RU architecture uses something called Structured State-Space Models (SSMs). Think of this like a super-fast photocopier that can process an entire book in one go. This allows the system to handle long sequences of data (like a robot driving for a long time) incredibly fast, making it practical for real-world use.

The Real-World Test: The Robot Swarm

To prove this works, the authors tested it on a group of wheeled robots (a swarm).

The Task: The robots had to move from a starting point to a goal while avoiding static obstacles (walls) and each other (collisions).
The Result: The L2RU-controlled robots successfully navigated the maze, avoided crashing into walls or each other, and reached their goals. Crucially, because of the "speed governor" design, they remained stable the entire time, even though they were optimizing a very complex, non-linear goal.

Summary

In short, this paper introduces a new way to build AI controllers that are inherently safe by design.

No more "fences": You don't need to add complex safety checks during training.
Guaranteed Stability: The math ensures the system won't go unstable, no matter how you tune it.
Fast & Efficient: It processes data quickly, making it ready for real robots.

It's like giving a race car a factory-installed limiter that ensures it can never go faster than the speed limit, allowing the driver to focus entirely on winning the race without worrying about crashing.

Technical Summary: Free Parametrization of L2-Bounded Structured State-Space Controllers

1. Problem Statement

Designing stabilizing control policies for nonlinear systems that simultaneously optimize complex, non-convex objectives remains a significant challenge. While Neural Networks (NNs) offer expressive power for control, they often lack inherent robustness; small input perturbations can propagate unpredictably, potentially destabilizing the closed-loop system.

Existing approaches to ensure stability typically impose explicit constraints on controller parameters during optimization. However, enforcing these constraints (e.g., via projections or constrained optimization) introduces computational overhead and complicates the training process. Furthermore, methods relying on continuous-time policies (e.g., Port-Hamiltonian models or Neural ODEs) face difficulties when discretized, often losing $L_2$ -gain guarantees or resulting in implicit equations that are cumbersome to implement. The core problem addressed is how to parametrize discrete-time controllers such that stability is guaranteed by design, allowing for fully unconstrained optimization of general nonlinear objectives.

2. Methodology

The paper proposes a novel framework leveraging Structured State-Space Models (SSMs), specifically a new architecture called the $L_2$ -Recurrent Unit ( $L_2$ RU). The methodology rests on three pillars:

A. Theoretical Foundation: Small-Gain Theorem

The approach utilizes the Small-Gain Theorem to guarantee closed-loop stability. If a plant has a known (or estimated) finite $L_2$ -gain $\hat{\gamma}$ and the controller is designed to have an $L_2$ -gain $\gamma$ such that $\gamma\hat{\gamma} < 1$ , the feedback interconnection is guaranteed to be $L_2$ -stable. The goal is to construct a controller parametrization where $\gamma$ is fixed and prescribed, independent of the optimization variables.

B. Novel Free Parametrization

The authors introduce a free parametrization for Linear Time-Invariant (LTI) systems with a prescribed $L_2$ -gain $\gamma$ . Unlike previous methods that required solving complex constraints or projections, this parametrization maps a set of free parameters $\omega \in \mathbb{R}^q$ directly to system matrices $(A, B, C, D)$ such that the resulting system always satisfies the $L_2$ -bound $\gamma$ .

Key features of this parametrization (Block 1 in the paper) include:

Complex-Diagonal Structure: The state matrix $A$ is parametrized as a diagonal matrix with complex eigenvalues $\lambda_j = \exp(-\exp(\mu_j) + i \exp(\theta_j))$ . This structure is inherited from the Linear Recurrent Unit (LRU) architecture.
Explicit Gain Control: The feed-through matrix $D$ and the scaling of $B$ and $C$ are derived to strictly enforce the bound $\gamma$ via the Bounded Real Lemma conditions.
Generalization: This method generalizes previous work (Massai and Ferrari-Trecate, 2025) by supporting arbitrary dimensions ( $n_h, n_u, n_y$ ) rather than being restricted to square systems.

C. The $L_2$ RU Architecture

The $L_2$ RU architecture stacks multiple State-Space Layers (SSLs). Each layer consists of:

An LTI system parametrized by the new free parametrization.
A static, Lipschitz-bounded nonlinearity (e.g., a bounded MLP).
A skip connection.

The entire architecture is pre- and post-processed by linear encoders and decoders. Because the $L_2$ -bound of the linear components is fixed by design, and the nonlinearities are Lipschitz-bounded, the global $L_2$ -gain of the controller can be prescribed. This allows the controller parameters to be tuned via fully unconstrained optimization (e.g., gradient descent) to minimize arbitrary cost functions, with stability guaranteed a priori.

D. Computational Efficiency

A critical advantage of the proposed parametrization is its compatibility with parallel scan algorithms. The complex-diagonal structure of $A$ allows for highly efficient time-domain simulation with linear computational complexity relative to the sequence length. This contrasts with standard RNNs or implicit methods (like NODEs or Recurrent Equilibrium Networks) which often require sequential processing or solving implicit equations, making them less scalable for long horizons.

3. Key Contributions

Novel Free Parametrization: The paper presents a new, free parametrization of SSM-based $L_2$ RU architectures that enforces a prescribed $L_2$ -gain by design. This eliminates the need for computationally expensive projections, constrained optimization, or post-hoc verification.
Decoupling Stability and Optimization: By fixing the $L_2$ -gain to satisfy the small-gain condition with the plant, the method decouples stability guarantees from the optimization process. This enables the use of standard, unconstrained optimizers for complex nonlinear objectives.
Computational Scalability: The proposed structure enables parallel processing of long input sequences via parallel scan, offering significant efficiency gains over existing NN-based control methods that rely on sequential recurrence or implicit solvers.
Experimental Validation: The approach is validated on a formation control task for a swarm of wheeled robots, demonstrating the ability to achieve collision avoidance and obstacle avoidance while maintaining closed-loop stability.

4. Results

The authors conducted simulation studies on a swarm of $N$ nonlinear wheeled robots (discretized from the Robotarium platform).

Setup: A centralized controller was trained to navigate robots to target positions while avoiding static obstacles and mutual collisions. The plant's $L_2$ -gain was estimated to be $\hat{\gamma} \approx 18.57$ .
Controller Design: An $L_2$ RU controller with 3 layers was trained with a prescribed gain $\gamma = 1/(18.57 + 0.001)$ , satisfying the small-gain condition.
Performance: The trained controller successfully guided the agents to their goals while avoiding collisions. The results (Figures 2 and 3) show that constraining the controller's $L_2$ -bound did not result in a significant loss of expressivity or performance. The relative distances between agents remained above the safety threshold throughout the simulation.

5. Significance and Claims

The paper claims that this approach addresses the "formidable challenge" of designing stabilizing control policies for nonlinear systems without sacrificing computational tractability.

Stability by Design: The primary significance is the ability to guarantee closed-loop stability by design through the small-gain theorem, removing the need for restrictive constraints during the learning phase.
Practicality: The authors highlight that the discrete-time nature and parallelizability of the $L_2$ RU make it more suitable for real-world implementation compared to continuous-time methods (like Port-Hamiltonian or NODEs) which suffer from discretization issues or high computational costs.
Modesty on Scope: The paper acknowledges limitations, noting that the small-gain guarantee relies on the ability to compute or estimate the plant's $L_2$ -bound, which can be difficult for general nonlinear systems. Additionally, the proposed parametrization is noted as a sufficient (but not necessary) condition for $L_2$ -boundedness, meaning it may not cover the entire space of all possible $L_2$ -bounded systems, though it is sufficient for practical control design.

The authors conclude that this framework opens the door for applying deep learning to control problems where stability is a hard constraint, and future work will focus on real-world deployment and integration with model-based Reinforcement Learning.