Physics-Informed Neural Networks for Nonlinear Output Regulation

arXiv:2511.13595 · eess.SY, cs.AI, cs.SY · Submitted 2025-11-17 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Physics-Informed Neural Networks for Nonlinear Output Regulation".

Dev: Here's the gist of the paper. Physics-Informed Neural Networks for Nonlinear Output Regulation addresses the full-information output regulation problem for nonlinear systems by constructing a zeroregulation-error manifold and a feedforward input that renders such manifold invariant.

Rosa: This work demonstrates that physics-informed neural networks can be effectively employed to solve the regulator equations for nonlinear output regulation.

Dev: Problem Formulation and Regulator Equations The core of the problem involves designing a feedback control law for a plant such that its output asymptotically tracks prescribed trajectories or rejects disturbances.

Rosa: For nonlinear systems, this is extremely difficult to solve due to the fact that designing this controller is tied to the solvability of a set of partial differential equations and nonlinear algebraic equations, known as the Regulator Equations (REs) or Isidori and Byrnes equations.

Paper summary: Rosa: So, wrapping up on Physics-Informed Neural Networks for Nonlinear Output Regulation, this work shows that PINNs can effectively solve the regulator equations for nonlinear output regulation >

Dev: It demonstrates that you can use deep learning to approximate those complex mappings π(w) and c(w) by minimizing physics-based residuals >

Taro: The key finding is that the learned operator generalizes well, reconstructing the zero-error manifold even when it hasn't seen specific initial conditions during training >

Rosa: It’s a principled path toward data-efficient, learning-based controllers for nonlinear output regulation problems that can handle unseen configurations >

Dev: But they are cautious about what this actually means for deployment because the main limitation is the full-information assumption where you need to know everything upfront >

Taro: The authors themselves flag that relaxing that requirement to partial-information settings, or treating the exosystem state as privileged information only during training, is a natural direction for future work >

Rosa: And they also suggest extending it further by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training >

Dev: Overall, this paper proves that PINN-based controllers are a viable and efficient tool for computing steady state solutions of nonlinear output regulation problems >

Conclusion: Rosa: So, this paper is about using physics to help train neural networks to figure out how to control really complicated systems that don't follow simple rules, specifically for output regulation >

Dev: Right, it tackles those regulator equations you have to solve when your system is nonlinear and you want it to track a certain path or reject disturbances >

Rosa: It’s essentially showing that physics-informed neural networks can be used to solve those regulator equations by creating this special zero-error manifold where the control works perfectly >

Dev: The authors are using a feedforward input idea that makes that manifold stable, and they train the network to learn those mappings directly from the physics >

Rosa: What’s really interesting is how they test it; they used a nonlinear helicopter vertical tracking benchmark where they showed the learned model reconstructs that zero-error manifold even when it hasn't seen certain initial conditions during training >

Dev: That reconstruction part is key, because in real-time control loops, you need to know the system will behave predictably everywhere, not just in the cases you trained on >

Rosa: And they did get some good results on performance; the controller kept small tracking errors across a wide range of amplitudes and frequencies that they hadn't specifically trained for >

Dev: But there’s a big caveat they mention; it all relies on this full-information assumption where you have access to the system's state information constantly, which isn't always true in the real world >

Rosa: So, while it’s a really efficient tool for finding steady-state solutions right now, the next step is figuring out how to make it work when you don't have all that perfect information available >

Sebastiano Mengozzi, †Giovanni B. Esposito, ‡Michelangelo Bin, Andrea Acquaviva, Andrea Bartolini, Lorenzo Marconi

Department of Electrical, Electronic and Information Engineering, University of Bologna

eess.SY, cs.AI, cs.SY

Submitted: 2025-11-17

Updated: 2026-04-28

Journal ref: 2026 European Control Conference (ECC), pp. 1891-1897

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Physics-Informed Neural Networks for Nonlinear Output Regulation addresses the full-information output regulation problem for nonlinear systems by constructing a zeroregulation-error manifold and a…

Key concepts

Regulator Equations (REs)
These are a set of partial differential equations with an algebraic constraint that define the necessary conditions for perfect regulation in nonlinear systems. Solving these equations is crucial because they determine the relationship between system states and the required control inputs for zero error.
Zero-Error Manifold
This is a mathematical surface or set of trajectories in the system's state space where perfect tracking or disturbance rejection can occur without any error. The goal of the method is to find this manifold and learn how to keep the system's motion on it.
PINN-based Residual
This is a loss function used to train the neural network. It measures how well the learned mappings (approximations of the solution) satisfy both the original physics (PDE residuals) and algebraic constraints (BC residuals), ensuring the network learns physically accurate solutions.
Full-Information Assumption
This assumption means that all necessary information about the external disturbances or exosystem states ($w$) is perfectly available to the controller for feedback. The paper notes this is a limitation, suggesting future work could address scenarios where this information is incomplete.

Terminology

Summary

Physics-Informed Neural Networks for Nonlinear Output Regulation addresses the full-information output regulation problem for nonlinear systems by constructing a zeroregulation-error manifold and a feedforward input that renders such manifold invariant. p. 0

Where the paper says it

Esposito1,†, Michelangelo Bin1 , Andrea Acquaviva1 , Andrea Bartolini1 , Lorenzo Marconi1 Abstract— This work addresses the full-information output regulation problem for nonlinear systems, assuming the states of both the plant and the exosystem are known. In this setting, perfect tracking or rejection is achieved by constructing a zeroregulation-error manifold π(w) and a feedforward input c(w) that render such manifold invariant.

This work demonstrates that physics-informed neural networks can be effectively employed to solve the regulator equations for nonlinear output regulation. p. 4
Where the paper says it

These values correspond to the steady inputs required to keep the system hovering at a fixed altitude, providing an additional validation of the learned mappings. DISCUSSION AND CONCLUSIONS This work demonstrates that physics-informed neural networks can be effectively employed to solve the regulator equations for nonlinear output regulation.

Problem Formulation and Regulator Equations

The core of the problem involves designing a feedback control law for a plant such that its output asymptotically tracks prescribed trajectories or rejects disturbances. For nonlinear systems, this is extremely difficult to solve due to the fact that designing this controller is tied to the solvability of a set of partial differential equations and nonlinear algebraic equations, known as the Regulator Equations (REs) or Isidori and Byrnes equations. p. 0

Where the paper says it

This problem is called output regulation problem and, for the class of linear systems, it was studied in [5], [6] and is related to the solvability of two linear matrix equations. For the class of nonlinear systems, such problems are extremely difficult to solve due to the fact that designing this controller is tied to the solvability of a set of partial differential equations and nonlinear algebraic equations, the so-called Regulator Equations (REs) or Isidori and Byrnes equations [7].

The regulator equations are described by a system of PDEs with an algebraic constraint. p. 1
Where the paper says it

In steady state, perfect regulation is described by two mappings x = π(w) and u = c(w) that parameterize a zero-error manifold and the corresponding feedforward input that keeps the motion on that manifold. Substituting these maps into (1) and imposing e ≡ 0 yields the regulator equations, a system of PDEs with an algebraic constraint: Lsπ(w) = f π(w), c(w), w , 0 = h π(w), w , (3) where Lsπ(w) := ∂π ∂w (w) s(w).

In steady state, perfect regulation is described by two mappings x = π(w) and u = c(w) that parameterize a zero-error manifold and the corresponding feedforward input that keeps the motion on that manifold. p. 1
Where the paper says it

The goal is to achieve e = h(x, w) → 0 while the exosystem evolves according to s. In steady state, perfect regulation is described by two mappings x = π(w) and u = c(w) that parameterize a zero-error manifold and the corresponding feedforward input that keeps the motion on that manifold.

PINN Methodology for Operator Learning

The proposed solution decouples the solution from time by directly learning steady state mappings as functions of the exosystem state w. A parametrized operator nξ, with trainable parameters ξ, is introduced to approximate the two mappings π(w) and c(w) defined in (3). p. 1

Where the paper says it

This perspective removes the need to construct a dataset, which is typically costly and cumbersome. To achieve this, we introduce a parametrized operator nξ, with trainable parameters ξ, to approximate the two mappings π(w) and c(w) defined in (3).

The quality of these approximations is assessed through a PINN-based residual, defined as rRE(wi) = Lsπˆ(wi) − fπˆ(wi), cˆ(wi), wi2 + hπˆ(wi), wi2. p. 1
Where the paper says it

The operator is implemented as a feed-forward neural network that takes as input a sampled exosystem state wi and returns the corresponding steady state approximations πˆ(wi) and cˆ(wi): πˆ(wi), cˆ(wi) = nξ(wi). (5) Since the exact solution of (3) is not available, we assess the quality of these approximations through a PINN-based residual, defined as rRE(wi) = Lsπˆ(wi) − f πˆ(wi), cˆ(wi), wi 2 + h πˆ(wi), wi 2 , (6) where rRE is the residual of the regulator equations, i.

This residual is minimized by training the neural operator nξ to satisfy the loss function J = JPDE + λ · JBC. p. 1
Where the paper says it

, the norm of the mismatch between the left-hand and righthand sides of the PDE and algebraic constraint evaluated at (ˆπ(w i ), cˆ(w i )). The neural operator nξ is trained by minimizing the loss function J = JPDE + λ · JBC, (7) where JPDE = E w∈W [rRE(w)] (8) is the expectation of rRE(w) over the domain W, and JBC is a suitably defined penalty term to enforce the system’s boundary conditions, and λ is a parameter controlling the ratio of importance between the two terms.

Training and Generalization Strategy

The training process is designed to exploit rich a priori knowledge through automatic differentiation, moving beyond black-box learning while still leveraging deep neural networks as universal approximators. The model is trained not to approximate one particular solution, but to act as a solution operator that generalizes to unseen parameter configurations by sampling from a family of dynamical systems. p. 2

Where the paper says it

instance: by sampling from a family of dynamical systems, the network is trained not to approximate one particular solution, but to act as a solution operator that generalizes to unseen parameter configurations. The optimization is performed over the network’s weights and biases using automatic differentiation.

The loss components include LPDE1, LPDE2, and LPDE3 which enforce the PDE residuals and boundary conditions. p. 3
Where the paper says it

5, 6}, with αi ∈ [0, 2π]. The regulator equations (21), together with the boundary conditions (22), yield the loss components: LPDE1 = ||L2 sπˆϕ − F1(ˆπϕ, πˆθ,Lsπˆϕ,Lsπˆθ, cˆb, w)||2 LPDE2 = ||L2 sπˆθ − F2(ˆπϕ, πˆθ,Lsπˆϕ,Lsπˆθ, cˆb, w)||2 LPDE3 = ||F3(ˆπϕ, πˆθ,Lsπˆϕ,Lsπˆθ, cˆb, w)||2 LBC = max(0, |cˆb| − 0.

Validation and Performance

The framework is validated on a nonlinear helicopter verticaltracking benchmark in the full-information setting, showing that the learned operator reconstructs the zero-error manifold and maintains regulation performance across a wide range of exosystem initial conditions and frequencies, including configurations not seen during training. The resulting PINN is computationally efficient, requiring only a few minutes on a consumer GPU for training and offering real-time deployment feasibility. p. 4

Where the paper says it

The controller maintains small tracking error over a wide range of amplitudes and frequencies, including conditions not seen during training. The method is computationally efficient: the training stage requires only a few minutes on a consumer GPU, while the online phase consists of a lightweight forward pass, making real-time deployment feasible on embedded platforms.

The controller maintains small tracking error over a wide range of amplitudes and frequencies, including conditions not seen during training. p. 4
Where the paper says it

The experimental results confirm that the learned operator reconstructs the zero-error manifold with high accuracy. The controller maintains small tracking error over a wide range of amplitudes and frequencies, including conditions not seen during training.

Conclusion and Future Directions

This work demonstrates that physics-informed neural networks can be effectively employed to solve the regulator equations for nonlinear output regulation. The main limitation is the full-information assumption, where the exosystem state w is assumed to be available for feedback. p. 5

Where the paper says it

The bottom plot shows a 100× magnification of the error trajectory to highlight the residual ripple. The main limitation of the approach is the full-information assumption, where the exosystem state w is assumed to be available for feedback.

Future work could involve relaxing this requirement by moving to partial-information settings or extending the framework from a single nominal plant to families of dynamical systems. p. 5
Where the paper says it

Relaxing this requirement by moving to partial-information settings, or by assuming that w is only accessible during training as privileged information and not at deployment, is a natural direction for future work. A related avenue is to extend the framework from a single nominal plant to families of dynamical systems, for example by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training.

The method is general and can be extended to other nonlinear systems where solving PDEs is a central step in the control design. p. 6
Where the paper says it

Overall, the results indicate that PINN-based controllers provide a viable and efficient tool for computing steady state solutions of nonlinear output regulation problems. The proposed methodology is general and can be extended to other nonlinear systems where solving PDEs is a central step in the control design.

--- Page 6 ---

The main limitation of the approach is the full-information assumption, where the exosystem state w is assumed to be available for feedback. Relaxing this requirement by moving to partial-information settings, or by assuming that w is only accessible during training as privileged information and not at deployment, is a natural direction for future work. p. 5

Where the paper says it

The main limitation of the approach is the full-information assumption, where the exosystem state w is assumed to be available for feedback. Relaxing this requirement by moving to partial-information settings, or by assuming that w is only accessible during training as privileged information and not at deployment, is a natural direction for future work.

p. 5
Where the paper says it

The main limitation of the approach is the full-information assumption, where the exosystem state w is assumed to be available for feedback. Relaxing this requirement by moving to partial-information settings, or by assuming that w is only accessible during training as privileged information and not at deployment, is a natural direction for future work.

A related avenue is to extend the framework from a single nominal plant to families of dynamical systems, for example by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training. p. 5
Where the paper says it

Relaxing this requirement by moving to partial-information settings, or by assuming that w is only accessible during training as privileged information and not at deployment, is a natural direction for future work. A related avenue is to extend the framework from a single nominal plant to families of dynamical systems, for example by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training.

p. 5
Where the paper says it

Relaxing this requirement by moving to partial-information settings, or by assuming that w is only accessible during training as privileged information and not at deployment, is a natural direction for future work. A related avenue is to extend the framework from a single nominal plant to families of dynamical systems, for example by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training.

This would enable the learned operator to solve the output regulation problem for whole classes of plants. p. 5
Where the paper says it

A related avenue is to extend the framework from a single nominal plant to families of dynamical systems, for example by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training. This would enable the learned operator to solve the output regulation problem for whole classes of plants.

p. 5
Where the paper says it

A related avenue is to extend the framework from a single nominal plant to families of dynamical systems, for example by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training. This would enable the learned operator to solve the output regulation problem for whole classes of plants.

In parallel, integrating more advanced stabilizing feedback designs could further enlarge the feasible operating region and strengthen the stability guarantees. p. 5
Where the paper says it

This would enable the learned operator to solve the output regulation problem for whole classes of plants. In parallel, integrating more advanced stabilizing feedback designs could further enlarge the feasible operating region and strengthen the stability guarantees.

p. 5
Where the paper says it

This would enable the learned operator to solve the output regulation problem for whole classes of plants. In parallel, integrating more advanced stabilizing feedback designs could further enlarge the feasible operating region and strengthen the stability guarantees.

--- Page 7 ---

Overall, the results indicate that PINN-based controllers provide a viable and efficient tool for computing steady state solutions of nonlinear output regulation problems. The proposed methodology is general and can be extended to other nonlinear systems where solving PDEs is a central step in the control design. p. 6

Where the paper says it

Overall, the results indicate that PINN-based controllers provide a viable and efficient tool for computing steady state solutions of nonlinear output regulation problems. The proposed methodology is general and can be extended to other nonlinear systems where solving PDEs is a central step in the control design.

--- Page 8 ---

The main limitation of the approach is the full-information assumption, where the exosystem state w is assumed to be available for feedback. Relaxing this requirement by moving to partial-information settings, or by assuming that w is only accessible during training as privileged information and not at deployment, is a natural direction for future work. A related avenue is to extend the framework from a single nominal plant to families of dynamical systems, for example by embedding parametric uncertainty in the regulator equations or randomizing plant parameters during training. This would enable the learned operator to solve the output regulation problem for whole classes of plants. In parallel, integrating more advanced stabilizing feedback designs could further enlarge the feasible operating region and strengthen the stability guarantees.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Accurate steady state mapping of nonlinear systems: The AI system can learn mappings pi(w) and c(w) that parameterize a zero-error manifold by minimizing residuals of regulator equations, directly solving PDEs with an algebraic constraint.

  2. Generalization across exosystem families: The learned operator nxi acts as a solution operator that generalizes to unseen parameter configurations, meaning the controller can perform regulation for exosystems not encountered during training, such as varying harmonic frequencies.

  3. Real-time deployment capability: The PINN is designed to be a fully connected neural network with approximately 8 × 104 trainable parameters, which allows for low inference cost and enables real-time computation on embedded platforms.

  4. Data-efficient control design: The framework trains directly from the nonlinear regulator equations without requiring precomputed trajectories or labeled data, making it a principled path toward data-efficient, learning-based controllers.

  5. Nonlinear output regulation for complex dynamics: The system can achieve perfect tracking or rejection for nonlinear systems, such as synchronizing a helicopter’s vertical dynamics with a harmonically oscillating platform.

Abstract

This work addresses the full-information output regulation problem for nonlinear systems, assuming the states of both the plant and the exosystem are known. In this setting, perfect tracking or rejection is achieved by constructing a zero-regulation-error manifold π(w) and a feedforward input c(w) that render such manifold invariant. The pair (π(w), c(w)) is characterized by the regulator equations, i.e., a system of PDEs with an algebraic constraint. We focus on accurately solving the regulator equations introducing a physics-informed neural network (PINN) approach that directly approximates π(w) and c(w) by minimizing the residuals under boundary and feasibility conditions, without requiring precomputed trajectories or labeled data. The learned operator maps exosystem states to steady state plant states and inputs, enables real-time inference and, critically, generalizes across families of the exosystem with varying initial conditions and parameters. The framework is validated on a regulation task that synchronizes a helicopter's vertical dynamics with a harmonically oscillating platform. The resulting PINN-based solver reconstructs the zero-error manifold with high fidelity and sustains regulation performance under exosystem variations, highlighting the potential of learning-enabled solvers for nonlinear output regulation. The proposed approach is broadly applicable to nonlinear systems that admit a solution to the output regulation problem.

Related papers