Joint Learning of Control Lyapunov Functions and Controllers via Stable-by-Design Surrogates

Patricia Pauli, Matthieu BarreauarXiv 2026 · arXiv

TL;DR

This paper introduces a framework to simultaneously learn a Control Lyapunov Function (CLF) and a stabilizing controller for nonlinear systems. By utilizing "stable-by-design" surrogate dynamics and structured neural networks (PLNets), the authors bypass the computationally expensive "train-then-verify" loop typical of neural control, providing a path toward formal stability guarantees with less overhead.

What Problem Does This Paper Address?

In control theory, a Control Lyapunov Function (CLF) is like a "map" that shows how to steer a system toward a stable equilibrium (usually the origin). If you have a valid CLF, you can easily derive a controller that guarantees stability. However, finding these functions for complex, nonlinear systems is notoriously difficult.

Recently, researchers have used neural networks to learn CLFs from data. The catch? A neural network doesn't "know" physics or stability by default. To ensure a learned controller is safe, researchers usually have to run expensive formal verification steps after training. If the verification fails, they must retrain and try again. This paper tackles the challenge of jointly learning the CLF and the controller in a way that incorporates stability requirements directly into the learning process, making the resulting system "stable-by-design" or at least much easier to verify.

Methods

The authors propose a three-part architecture:

  1. Polyak-Łojasiewicz Networks (PLNets): To represent the CLF, the authors use PLNets. These are structured neural networks that are guaranteed to be positive-definite and radially unbounded by construction. This means the "shape" of the Lyapunov function is mathematically forced to be valid, leaving only the "decrease condition" (the requirement that system energy drops over time) to be satisfied during training.
  2. Hybrid Controller: The controller isn't just a black-box MLP. It combines a linear feedback term (derived from standard LQR theory to handle stability near the origin) with a nonlinear correction term inspired by Sontag’s universal formula. This allows the controller to leverage classical control wisdom while using deep learning to handle complex nonlinearities further away from the equilibrium.
  3. Stable-by-Design Surrogates: This is the core innovation. During training, the authors define a "surrogate" version of the closed-loop dynamics. This surrogate is a projection of the actual dynamics that is mathematically forced to satisfy the Lyapunov decrease condition.
  4. The Loss Function: Instead of using a complex reinforcement learning objective, the model is trained to minimize the discrepancy between the actual system behavior and the "perfectly stable" surrogate behavior. If the discrepancy is zero, the system is guaranteed to be stable.

Datasets

As is common in control theory research, the authors evaluate their method on simulated nonlinear dynamical systems rather than static image or text datasets:

  • Academic Cubic System: A 2D system with a destabilizing cubic nonlinearity that causes standard linear controllers (like LQR) to fail when the state starts too far from the origin.
  • Van der Pol Oscillator: A classic benchmark in nonlinear dynamics featuring an unstable equilibrium and complex limit cycle behavior.

Training data consisted of states sampled uniformly from a defined domain ($\Omega$), and evaluation involved testing convergence from initial conditions both inside and outside that training domain.

Key Insights and Ideas

  • Projection as a Teacher: Rather than penalizing the controller for being unstable, the framework creates a "stable version" of the current controller and tells the neural network: "Try to look like this." This projection-based approach makes the optimization landscape much friendlier.
  • Baking in Local Stability: By incorporating a linear controller ($Kx$) into the architecture, the authors ensure local exponential stability near the origin "for free." This allows the neural network to focus its expressive power on the "hard" nonlinear regions further out.
  • Simplified Certification: Because the PLNet handles the required "shape" of the Lyapunov function, verifying global stability reduces to checking a single "discrepancy condition" between the real system and the surrogate.

Key Results

  • Robustness to Initial Conditions: In the academic example, a standard linear quadratic regulator (LQR) failed to stabilize the system in 8 out of 12 test cases. The proposed learned controller achieved 100% convergence, even for starting points outside the training domain.
  • Performance Trade-offs: The authors demonstrated that by adjusting a scaling factor ($\alpha$), engineers can precisely trade off between settling time (how fast the system reaches the origin) and control effort (how much energy/actuation is used).
  • Competitive with Baselines: The method performed comparably to (and sometimes better than) Sontag’s formula and CLF-Quadratic Programs, but with the massive advantage of not requiring a pre-known CLF.

Limitations and Future Work

  • Formal Verification: While the framework is designed to be "verifiable," the authors primarily used grid-based numerical checks in their experiments. Full formal verification (e.g., using SMT solvers) for high-dimensional systems remains a future step.
  • Input Constraints: The current paper does not explicitly account for actuator saturation (e.g., a motor that can only provide a maximum amount of torque).
  • Dimensionality: The examples provided are 2D. Scaling these structured architectures to very high-dimensional robotic systems (like humanoids) will require further investigation into the computational efficiency of PLNets.

Summary updated

Abstract

Training and verifying neural network-based control Lyapunov functions (CLFs) can be computationally demanding. To address this challenge, we propose a framework that jointly learns a CLF and a stabilizing controller using structured neural networks and a projection-based parameterization. Specifically, we employ Polyak-Lojasiewicz networks (PLNets), which parameterize nonconvex, positive-definite CLF candidates, and construct a stable-by-design surrogate of the closed-loop dynamics that satisfies the Lyapunov decrease condition by construction. The controller combines linear state feedback with a higher-order correction term resembling Sontag's universal formula and is trained to align the resulting closed-loop dynamics with the stable-by-design surrogate. Unlike classical approaches based on Sontag's formula, our framework does not require a known CLF; instead, it learns the CLF and controller simultaneously. We establish local and global stability guarantees for the learned controller and demonstrate the effectiveness of the proposed approach through numerical examples.
BibTeX
@misc{pauli2026joint,
  author = {Patricia Pauli and Matthieu Barreau},
  title = {{Joint Learning of Control Lyapunov Functions and Controllers via Stable-by-Design Surrogates}},
  year = {2026},
  eprint = {2611.20900},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20900},
}
Figure