model-agnostic-meta-learning-differentiable-mpc

name	model-agnostic-meta-learning-differentiable-mpc
description	Model-Agnostic Meta Learning (MAML) framework for Differentiable Model Predictive Control (MPC) to enable adaptive control strategies across varying scenarios. Combines meta-learning with differentiable MPC for real-time adaptability without extensive retraining. Use when needing adaptive MPC controllers that can quickly adjust to new system dynamics or environmental conditions.
metadata	{"arxiv_id":"2607.19271","published":"2026-07-21","authors":["Matteo Tomasetto","Francesco Braghin","J. Nathan Kutz","Andrea Manzoni"],"subjects":["Systems and Control (eess.SY)","Optimization and Control (math.OC)"],"tags":["meta-learning","MPC","differentiable-control","adaptive-control","systems-engineering"]}
license	Complete terms in LICENSE.txt

Model-Agnostic Meta Learning for Differentiable MPC

Overview

This methodology combines Model-Agnostic Meta Learning (MAML) with Differentiable Model Predictive Control (MPC) to create adaptive control strategies that can quickly adjust to varying scenarios without requiring extensive retraining or multiple system simulations.

Core Methodology

Problem Context

Traditional optimal control problems require several system simulations to tailor control actions to varying scenarios, which becomes computationally demanding due to high-dimensional spatio-temporal dynamics.

Solution Approach

The proposed framework uses SHallow REcurrent Decoder networks-based Reduced Order Modeling (SHRED-ROM) to synthesize real-time closed-loop controllers for high-dimensional and parametric dynamics, relying solely on limited state sensor readings.