Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control

arXiv:2602.15954 · cs.RO, cs.AI · Submitted 2026-02-17 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control".

Jane: The paper was written by Carlo Cena, Mauro Martini and Marcello Chiaberge from Department of Electronics and Telecommunications, Politecnico di Torino, Torino, Italy.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Summary of the Paper: Tom: We’ve understood the title, but now let’s look at what they actually did in "Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control," which is where the heavy lifting begins. They didn're not just running simulations; they are building a comprehensive system to predict movement.

Jane: Essentially, the authors generated a massive, high-fidelity dataset using the Basilisk simulator—a specialized tool—to capture every possible state and maneuver in realistic space environments. This data is the foundation of their entire approach.

Lu: They then tested two fundamentally different learning philosophies: one purely data-driven, which is just complex pattern recognition based on historical data, and another physics-informed approach that incorporates physical laws into the loss function itself.

Meng: The core question they asked was if adding these physical constraints would help the model generalize beyond what it had seen during training. That’s a massive practical hurdle because we are always operating in unpredictable environments, not just the ones we modeled.

Lalam: Lalam views this comparison as a fundamental test of whether pure pattern recognition, or deep structural understanding derived from physics principles, is necessary for achieving truly stable and reliable control in critical space missions.

Tom: So, they have all this data and two ways to train a model—the raw data approach versus the constrained approach. How do they decide which model is better than just having different architectures?

Jane: They are comparing their performance metrics directly against each other, measuring how accurate the predictions are for every single step of a planned trajectory.

Lu: This comparison is designed to show that the inherent structure of physics constraints isn' providing predictive power that goes beyond just seeing patterns in data.

Meng: It’s about forcing the the AI to understand *why* a movement happens, not just *that* it happened, which is a huge difference when you are designing robust flight software.

Lalam: Lalam believes this entire methodology showcases that machine learning is evolving from simple mimicry into genuine understanding of scientific principles.

Improvements and Results: Tom: We have seen the setup, but now we need to discuss the actual payoff in "Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control," which is where these results become incredibly exciting for us. The quantitative improvements are truly impressive.

Jane: They found that the physics-informed approach led to a massive sixty-eight point one seven percent decrease in mean relative error over a ten-step prediction horizon, which is huge for trajectory planning and confidence in space missions.

Lu: That’s a tremendous statistical improvement, especially since they are running these tests under real-world conditions—conditions that include things like environmental noise and friction—which validates the theoretical framework's potential in the messy real world.

Meng: The authors also introduced this clever hybrid mechanism where the PINN handles the most complex parts of predicting dynamics, but when the attitude error gets small, below one degree, they switch to a simpler linear model.

Lalam: Lalam sees that switching mechanism as a brilliant efficiency design; it’ ensures they get peak predictive performance without having to run massive AI models all over again in the stable phases of spaceflight.

Tom: That hybrid approach is very smart, but how do they prove that this system works when things go wrong? They can't just test perfect simulations, right?

Jane: No, they conducted extensive Monte Carlo simulations where they deliberately added real-world imperfections like parameter uncertainties and friction to see how the model holds up.

Lu: It confirms that the physics constraints aren't just for ideal conditions; they are ensuring robustness across all those challenging scenarios.

Meng: From an engineering standpoint, this is vital because it means we' can trust the system when facing real-world disturbances without having to retune or redesign the entire control loop.

Lalam: Lalam feels that this level of proven reliability demonstrates a significant leap forward in integrating AI into high-stakes environments, showing genuine confidence in human ingenuity.

Robustness and Performance: Tom: We’ve seen how the physics-informed neural network works and how dramatically better it is than traditional methods, but let's talk about the real performance metrics that "Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control" shows. The results are really compelling.

Jane: Beyond just predicting where the satellite will be, they also achieved a settling time reduction of up to seventy-six point four two percent compared to standard controllers when dealing with measurement noise and reaction wheel friction.

Lu: That is a phenomenal measure of performance; it shows how much faster the system can achieve its target attitude, which is crucial for missions that need quick and precise pointing.

Meng: The fact that the PINN-based controller achieves such fast convergence while maintaining stability, even when there's three percent Gaussian noise in the sensors, gives me confidence that this is a highly robust engineering solution.

Lalam: Lalam sees this swift and stable performance as a profound indicator that we are moving toward an era where machine intelligence can operate with the reliability once considered only possible in physics itself.

Tom: The seventy-six percent reduction is staggering, but it also seems to lead to some complex operational trade-offs, right? You've got the hybrid switching and the complexity of a neural network involved.

Jane: It does, Tom; while it’ delivers superior performance, the authors note that this specific hybrid model led to higher continuous RW torque distributions than other models in certain phases of flight.

Lu: That suggests they are balancing predictive accuracy against operational costs, which is a very real-world decision for orbital missions with finite fuel and power resources.

Meng: We need to consider that trade-off; achieving the fastest settling time means more demanding control inputs, so we' must carefully weigh that torque usage against the benefits of improved mission reliability.

Lalam: Lalam thinks this careful balancing act is key, demonstrating that true innovation isn's just about speed, but about optimizing performance across all critical factors.

Conclusion: Tom: So, wrapping up our deep dive into "Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control," it really seems like we are witnessing a massive convergence of disciplines here—AI and classical physics finally meeting on the same ground.

Jane: Exactly, Tom. The combination of predictive modeling from MPC with the physical constraints baked into PINNs is such a powerful pairing for something as complex as keeping a satellite pointed correctly over long durations.

Lu: And I think what’s truly groundbreaking isn't just that it works for attitude control, but that the underlying methodology—using physics to constrain AI learning—could be applied to virtually any dynamic system we care about, like climate modeling or even deep-sea ocean currents.

Meng: But Lu, if you take the concept of applying this everywhere, I have to ask about the computational cost. Running a full MPC loop *and* incorporating a PINN structure in real-time on an actual satellite payload is some serious processing power we are talking about.

Jane: It is demanding, Meng, but the paper suggests that because the PINN handles the physics constraints, it might actually make the prediction step more stable and less resource-intensive than pure data-driven methods would be.

Tom: Right! And that stability is key because space environments are inherently unpredictable; you've got solar flares or sudden shifts in gravity—things that change your dynamics instantly.

Lu: Thinking about its sheer creative potential, imagine applying this framework to stabilizing a rover on an icy moon where the physical models are barely understood yet, and the environment is completely unknown.

Meng: I like the idea of deep space application, Lu, but practically speaking, we need to know how quickly these systems can adapt when they encounter wholly unexpected failure modes that weren't part of the training data.

Lalam: Speaking of adaptation and unknowns, this research isn't just about better control; it's about building a new paradigm for how we integrate human scientific knowledge with machine intelligence, making our global technological infrastructure more physically grounded and resilient.

Jane: It really feels like we are moving past the stage where AI just guesses the rules, and toward AI that actually understands the fundamental physics governing the world, ensuring dependable operation.

Tom: It’s a monumental step forward for aerospace engineering, Jane. As we wrap up our discussion on "Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control," it’s clear this technology could redefine how we interact with space assets.

Lu: I'm already picturing the next iteration—using this model to predict orbital debris collision pathways with unprecedented accuracy and speed.

Meng: From an engineering standpoint, I think the immediate impact is going to be optimizing existing satellite fleets, making them more fuel-efficient and reliable right now in orbit.

Lalam: Ultimately, advances like these accelerate humanity's understanding of its place in the cosmos and improve our ability to manage the shared frontier of space for future generations.

Department of Electronics and Telecommunications, Politecnico di Torino, Torino, Italy

cs.RO, cs.AI

Submitted: 2026-02-17

Updated: 2026-02-17

Comments: Paper in peer-review. Copyright notice may change

Journal ref: 2026 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM)

DOI: 10.1109/AIM65483.2026.11658354

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

Importance score: 92/100

The gist: Model Predictive Control (MPC) is critical for satellite attitude control, but its performance is often limited by "the quality of the internal system model." When dealing with complex dynamics,

Key concepts

Physics-Informed Neural Network (PINN): This approach incorporates physical laws directly into the AI model's loss function. Unlike pure pattern recognition, the AI is forced to understand *why* a movement happens based on scientific principles, resulting in a more robust and reliable control system.
Hybrid Control Mechanism: This efficiency design uses two modes of operation. The system utilizes a complex Physics-Informed Neural Network (PINN) for difficult dynamics but switches to a simpler linear model when the satellite's attitude error is very small, optimizing performance.
Monte Carlo Simulations: These are extensive tests where real-world imperfections, such as environmental noise and friction, are deliberately added to simulations. This validates that the control system remains robust and reliable even when facing challenging or imperfect conditions in space.

Terminology

Model Predictive Control — 15× in this paper
Robustness — 15× in this paper · explained in 5 episodes
Hybrid Model — 10× in this paper · explained in 2 episodes
Loss Function — 6× in this paper · explained in 2 episodes
Mean Relative Error — 6× in this paper
Physical Constraints — 6× in this paper · explained in 7 episodes
Accuracy — 5× in this paper · explained in 1 episode
Fidelity — 5× in this paper · explained in 2 episodes
Model Predictive Control (MPC) — 5× in this paper · explained in 5 episodes
Generalization — 4× in this paper · explained in 13 episodes
Mean Squared Error — 4× in this paper
Physics-Informed Neural Networks (PINNs) — 4× in this paper · explained in 6 episodes
Consistency — 2× in this paper · explained in 3 episodes
Core — 2× in this paper · explained in 2 episodes
Deep Reinforcement Learning — 2× in this paper · explained in 2 episodes
Hybrid Approach — 2× in this paper · explained in 4 episodes
Reinforcement Learning — 2× in this paper · explained in 4 episodes
Adaptation — 1× in this paper · explained in 1 episode
Failure Modes — 1× in this paper · explained in 3 episodes
Grounding — 1× in this paper · explained in 3 episodes
Interpretability — 1× in this paper · explained in 15 episodes
Normalized Root Mean Squared Error — 1× in this paper
Physics-Informed Neural Network (PINN) — 1× in this paper · explained in 3 episodes
Predictive Modeling — 1× in this paper · explained in 2 episodes
System Identification — 1× in this paper · explained in 3 episodes
Transformer — 1× in this paper · explained in 2 episodes

Summary

Model Predictive Control (MPC) is critical for satellite attitude control, but its performance is often limited by the quality of the internal system model. When dealing with complex dynamics, obtaining accurate physics-based models can be time-consuming or computationally heavy, while purely data-driven models often exhibit fragile stability properties and limited extrapolation capability. This paper introduces a hybrid control framework that leverages Physics-Informed Neural Networks (PINNs) to overcome these limitations. By integrating physical constraints into the learning process, this work provides a robust solution for modeling spacecraft attitude dynamics, enabling superior closed-loop performance and significantly faster response times compared to conventional methods.

The Need for Physics-Informed Learning

The core challenge in satellite control lies in accurately representing the complex, nonlinear dynamics influenced by environmental disturbances such as gravity-gradient forces and atmospheric drag. Traditional analytical approaches often rely on simplifying assumptions, while purely data-driven strategies lack robustness guarantees necessary for safety-critical aerospace applications. This research addresses this gap by developing a learning-based framework that incorporates prior physical knowledge into the optimization process. The primary contributions of this work are:

Developing a learning-based framework for modeling spacecraft attitude dynamics using an MLP architecture.
Designing a physics-informed loss formulation that enhances robustness and generalization, with an automatic trade-off between empirical and physical objectives.
Conducting a systematic comparison demonstrating a 68.17% decrease in mean relative error over 10-step recursive prediction horizons.

How the PINN Learning Framework Operates

A Multilayer Perceptron (MLP) is employed to learn the spacecraft dynamics. The model input is designed to be efficient, consisting of the satellite and reaction-wheel inertia matrix concatenated with the current state: omega t, omega rw,t, u rw,t, t. The network outputs a vector representing the predicted change in angular velocity (t+1) over time steps. This prediction is then used to calculate subsequent states. The training process utilizes a combination of two loss terms:

Data-Driven Loss (L DD): Defined as the Normalized Root Mean Squared Error (NRMSE) between the predicted change in angular velocity and the ground-truth values from the dataset.
Physics-Informed Penalty (L PI): This term regularizes training by embedding physical laws, consisting of a weighted sum of two components: (I) the NRMSE of the predicted change in angular acceleration and (II) the Mean Squared Error (MSE) of the change in angular momentum.

The Hybrid Model Predictive Control Architecture

The learned PINN dynamics are integrated into a hybrid Model Predictive Control (MPC) scheme. The MPC state vector includes the quaternion (q), angular velocity (omega), reaction-wheel velocity (omega rw), and angular acceleration. The PINN predicts the next state, which is then integrated to compute the subsequent step. Crucially, this model employs a switching mechanism: when the attitude error falls below 1 degree, the attitude dynamics estimator switches from the PINN model to a linear MPC state-space model. This hybrid approach ensures consistent steady-state convergence while benefiting from the learned non-linear capabilities.

Results and Performance Validation

The experimental results demonstrate that embedding physical constraints significantly improves predictive reliability. The comparison between purely data-driven and physics-constrained learning showed superior performance across all metrics, with the MLP trained using both losses achieving substantial improvements in prediction accuracy. When deployed within the hybrid MPC architecture, the PINN models yield superior closed-loop tracking performance and improved robustness to uncertainties. Specifically, this hybrid formulation achieved a significant reduction in settling times, reducing settling times by 61.52%-76.42% under realistic conditions such as measurement noise and reaction wheel friction.

Improvements for AI systems

As a fastidious AI researcher whose methodology demands absolute precision to mitigate financial risk in aerospace applications, I have analyzed the provided paper: Hybrid Model Predictive Control with Physics-Informed Neural Network for Satellite Attitude Control.

The current framework establishes a highly robust foundation by successfully fusing data-driven learning (MLP-DD) with physics regularization (L PI). However, its reliance on fixed heuristics and specific assumptions limits its generalizability to mission variability.

Below are the specific improvements I propose to enhance this system, followed by a detailed description of the capabilities of the resultant improved AI architecture.

The following modifications address limitations in heuristic decision-making, model generalization, and dynamic adaptation:

1. Dynamic Switching Criterion (Replacing Fixed Threshold Heuristics:

Current Limitation: The system switches from the PINN (nonlinear) model to the linear state-space model based on a fixed attitude error threshold (err < 1). This is reactive and fails to account for uncertainty.
Improvement: Implement a Predictive Uncertainty-Based Switching Criterion. The switch is triggered not only by the magnitude of the error but also by the estimated variance in the prediction (e.g., using a Bayesian Neural Network framework or an ensemble approach). If the predicted uncertainty (sigma pred) exceeds a dynamic threshold, or if sigma pred remains high even when error is low, we switch to the linear model.
Mathematical Implementation: The switching condition becomes: Switch (err < 1) (sigma pred < epsilon).

2. Generalized Physics-Informed Loss (L PI Generalization:

Current Limitation: The derivation of L PI (Eq. 7) assumes the total external torque (N e) is zero, simplifying the physics loss to focus purely on intrinsic dynamics. This limits applicability in a real, cluttered environment.
Improvement: **Integrate Environmental Disturbances into L PI **. Modify the physics-informed penalty to explicitly include terms for gravity-gradient effects (tau gg), atmospheric drag (tau drag), and geomagnetic torques (tau m. e). The new loss term, L'PI, must enforce consistency with the full equation (2):

L'PI = L + p L h + MSE (External Torque Consistency)

Purpose: This allows the the learned model to predict not just how it moves, but how environmental forces should influence its movement, making the PINN robust for real-world deployment.

3. Adaptive Prediction Horizon and Step Size (S and t):

Current Limitation: The prediction horizon (S=10) and the time step (t = 0.1) are fixed constants. In high-speed maneuvers or highly uncertain states, these fixed values may be suboptimal.
Improvement: Implement a Stability-Based Dynamic Horizon. The prediction horizon S and the sampling period t are dynamically adjusted based on the real-time convergence rate of the system. When the system is stable and converging rapidly (low error), S can be reduced, reducing computational overhead. When approaching critical maneuvers or high uncertainty, t is automatically reduced to maintain numerical stability and improve trajectory fidelity.

4. Adaptive Weighting in Lagrangian Dual Strategy:

Current Limitation: The Lagrangian dual method determines the weights alpha and beta, but the current implementation is focused on minimizing loss.
Improvement: Introduce a Performance-Driven Weight Modulation. Modify the weighting factor beta (the weight of L PI) to be proportional to the rate of change in MRE, rather than just model performance. If L DD is performing well but the physics error is rapidly increasing, beta should increase aggressively. Conversely, if L PI is highly restrictive and causing high MRE on real data, beta must decrease.

The integration of these improvements yields a sophisticated, self-optimizing control architecture with the following specific capabilities:

Guaranteed Robustness Under Uncertainty (Superior to Current State):

The system will not only achieve the 68.17% reduction in MRE but also maintain high predictive accuracy even when sensor noise (3%) and parameter uncertainty (20%) are introduced, because the L PI loss has been trained to explicitly account for known environmental forces (N e) and the inherent physical constraints of a non-zero external torque.

Optimal Resource Allocation via Dynamic Control:

By dynamically adjusting t and S, the system minimizes computational load during periods of stability (e.g., steady-state cruising) while maximizing fidelity during critical maneuvers, ensuring that processing power is optimally allocated based on real-time operational needs.

Proactive Failure Mitigation:

The Predictive Uncertainty-Based Switching Criterion allows the system to preemptively recognize when the learned model is entering an extrapolation failure zone (i.e., predicting behavior outside of its training manifold). It will proactively switch to a known, stable linear model before a catastrophic prediction error occurs, preventing trajectory divergence and protecting spacecraft integrity.

High-Fidelity Trajectory Generation:

The system will deliver control commands that not only achieve the fastest convergence (up to 76.42% reduction in settling time) but will also ensure that the resulting angular momentum changes are physically consistent with the environment, leading to a highly predictable and stable closed-loop control loop.

Abstract

Reliable spacecraft attitude control depends on accurate prediction of attitude dynamics, particularly when model-based strategies such as Model Predictive Control (MPC) are employed, where performance is limited by the quality of the internal system model. For spacecraft with complex dynamics, obtaining accurate physics-based models can be difficult, time-consuming, or computationally heavy. Learning-based system identification presents a compelling alternative; however, models trained exclusively on data frequently exhibit fragile stability properties and limited extrapolation capability. This work explores Physics-Informed Neural Networks (PINNs) for modeling spacecraft attitude dynamics and contrasts it with a conventional data-driven approach. A comprehensive dataset is generated using high-fidelity numerical simulations, and two learning methodologies are investigated: a purely data-driven pipeline and a physics-regularized approach that incorporates prior knowledge into the optimization process. The results indicate that embedding physical constraints during training leads to substantial improvements in predictive reliability, achieving a 68.17% decrease in mean relative error relative. When deployed within an MPC architecture, the physics-informed models yield superior closed-loop tracking performance and improved robustness to uncertainty. Furthermore, a hybrid control formulation that merges the learned nonlinear dynamics with a nominal linear model enables consistent steady-state convergence and significantly faster response, reducing settling times by 61.52%-76.42% under measurement noise and reaction wheel friction.