Gaussian processes for dynamics learning in model predictive control

Anna Scampicchio, Elena Arcari, Amon Lahr, Melanie N. Zeilinger

Introduction

Model Predictive Control (MPC) is a cornerstone of modern control engineering, enabling the management of complex, constrained systems—such as robotic arms, quadrotors, or chemical reactors—by optimizing control inputs over a finite time horizon. However, the efficacy of MPC is fundamentally limited by the accuracy of the underlying mathematical model. Real-world systems often exhibit unmodeled physics, environmental disturbances, and time-varying dynamics that traditional models fail to capture.

To bridge this gap, researchers have increasingly turned to machine learning, specifically Gaussian Process (GP) regression, to learn residual dynamics from data. GPs are particularly attractive because they provide not just point predictions, but full probability distributions, offering a principled way to quantify uncertainty. This review explores the integration of GPs into MPC, identifying three primary hurdles: computational scalability, uncertainty propagation, and closed-loop safety guarantees.

Background on Gaussian Processes

Gaussian Processes serve as a non-parametric Bayesian approach to regression. In the context of control, they are used to model the discrepancy between a nominal system model and observed behavior. The review categorizes the background into static GPs and their application to dynamic systems.

Graphical overview of Section 2: Background on GPs

Static GPs involve Bayesian inference and uncertainty quantification, often requiring careful hyperparameter selection. When applied to dynamic systems, the focus shifts to input-output versus state-space models. Key considerations include how to handle latent states and the trade-offs between alternating function learning and state inference.

The Three Pillars of GP-MPC

The core of the survey addresses the technical challenges of combining probabilistic models with real-time optimization.

1. Scalable GPs

Standard GPs suffer from cubic scaling ($O(N^3)$) with respect to the number of data points $N$, making them unsuitable for high-frequency MPC. The authors highlight inducing variable methods like Variational Free Energy (VFE) and Fully Independent Training Conditional (FITC). These methods summarize the dataset into $M$ "pseudo-points," reducing training complexity to $O(NM^2)$ and inference to $O(M^2)$. Notably, VFE is cited for its superior empirical performance and robustness against overfitting compared to the popular FITC method.

2. Uncertainty Propagation

Predicting future states in MPC requires pushing a Gaussian distribution through a nonlinear dynamics model over multiple steps. This process typically results in non-Gaussian distributions that are analytically intractable. Common approximations include Taylor-series linearization, exact moment matching, and sigma-point (unscented) propagation.

A critical insight highlighted by the authors is the Independence Assumption Trap. Most GP-MPC implementations assume that uncertainties at different time steps are independent. This "indirect" one-step-ahead prediction ignores temporal correlations, which can lead to a significant underestimation of uncertainty along the predicted trajectory, potentially compromising safety.

3. Closed-Loop Safety Guarantees

Ensuring safety (e.g., constraint satisfaction) is critical. GP-MPC often employs "chance constraints," which ensure constraints are met with a high probability. Theoretical frameworks using Reproducing Kernel Hilbert Space (RKHS) theory provide "robust-in-probability" bounds, though these are often too conservative for practical use or rely on mathematical constants that are difficult to compute from real data.

Key Insights and Future Directions

A meta-analysis of over 30 applied GP-MPC papers reveals a gap between theory and practice. Most real-world applications (e.g., autonomous racing and robotics) prioritize computational feasibility, often using FITC and simple linearization at the expense of rigorous safety proofs. Furthermore, the authors critique the common "Perfect State" assumption, noting that transitioning to Latent Variable Models could address the challenges of noisy state measurements in practical deployments.

Future Research Frontiers: * Safe Online Learning: Updating models in real-time without compromising stability guarantees remains an open problem (the "Online Learning Dilemma"). * Beyond RKHS Bounds: Exploring conformal prediction or Bayesian uniform error bounds to reduce conservatism in safety guarantees. * Spatio-Temporal GPs: Adapting GPs for environments with time-varying spatial characteristics, such as traffic or weather patterns.

In conclusion, while GP-MPC offers a powerful framework for safe, learning-based control, balancing computational efficiency with theoretical rigor remains the field's primary challenge.

Summary updated

BibTeX
@article{scampicchio2025gaussian,
  author = {Anna Scampicchio and Elena Arcari and Amon Lahr and Melanie N. Zeilinger},
  title = {{Gaussian processes for dynamics learning in model predictive control}},
  journal = {Annu. Rev. Control.},
  year = {2025},
  doi = {10.1016/j.arcontrol.2025.101034},
}
Figure