Fast Online Adaptive Neural MPC via Meta-Learning

TL;DR

This work tackles real-time nonlinear control under model mismatch by learning residual dynamics online within an MPC framework. It combines Model-Agnostic Meta-Learning (MAML) to pretrain a neural residual model with an L4CasADi-enabled MPC for rapid, few-shot adaptation using small online data. The neural residual is integrated into the MPC, enabling fast correction of unmodeled dynamics, and is validated through simulations on a Van der Pol oscillator, Cart-Pole, and a 2D quadrotor, showing superior adaptation speed and predictive accuracy over nominal MPC and naïve neural augmentation. The approach offers a practical route to robust, real-time adaptive robotic control with limited offline data and demonstrates significant potential for real-world deployments.

Abstract

Data-driven model predictive control (MPC) has demonstrated significant potential for improving robot control performance in the presence of model uncertainties. However, existing approaches often require extensive offline data collection and computationally intensive training, limiting their ability to adapt online. To address these challenges, this paper presents a fast online adaptive MPC framework that leverages neural networks integrated with Model-Agnostic Meta-Learning (MAML). Our approach focuses on few-shot adaptation of residual dynamics - capturing the discrepancy between nominal and true system behavior - using minimal online data and gradient steps. By embedding these meta-learned residual models into a computationally efficient L4CasADi-based MPC pipeline, the proposed method enables rapid model correction, enhances predictive accuracy, and improves real-time control performance. We validate the framework through simulation studies on a Van der Pol oscillator, a Cart-Pole system, and a 2D quadrotor. Results show significant gains in adaptation speed and prediction accuracy over both nominal MPC and nominal MPC augmented with a freshly initialized neural network, underscoring the effectiveness of our approach for real-time adaptive robot control.

Figures (8)

• Figure 1: Block diagram of the proposed online adaptive neural MPC framework.
• Figure 2: Overview of the MAML-based meta-learning approach for robotic systems, illustrating the inner-loop adaptation of task-specific parameters $\theta'_i$ and the outer-loop optimization of shared meta-parameters $\theta$. This framework enables rapid adaptation to new system dynamics using limited data.
• Figure 3: Prediction comparison using three models in MPC controllers. (a) Time evolution of states $x_1(t)$ and $x_2(t)$. (b) Phase portrait ($x_1$ vs. $x_2$) illustrating trajectory predictions.
• Figure 4: The snapshot of Cart-Pole system in PyBullet
• Figure 5: The snapshot of 2D quadrotor system in PyBullet