Under Review

Stable Transformer-Actor-Critic Model Predictive Control: A Contraction Analysis Approach

1University of Cambridge 2University College London 3University of Toulouse
TL;DR: The paper introduces a Transformer-Actor-Critic MPC framework that uses sequence models to predict optimal control costs. By proving the internal stability of Transformers and applying contraction theory, the authors develop a training regularizer that mathematically guarantees safe, stable closed-loop control in highly noisy environments.

Single-Agent Drone Obstacle Avoidance Under Noise & Disturbances

Select observation noise and external disturbance levels to compare the Unstable (Baseline) and Stable (STAC-MPC) trajectories side by side. While the unregularized baseline oscillates and drifts under perturbations, STAC-MPC contracts tracking error to guarantee safe obstacle clearance.

Stable Unstable
Observation Noise Moderate ($\sigma = 0.25$)
External Disturbance Moderate ($d = 0.2\,\text{N}$)
🎯

Contraction & Robustness Insights

Adjusting the disturbance and observation noise levels demonstrates the effectiveness of the $\delta\text{ISS}$ contraction regularizer. While the unconstrained Transformer baseline suffers from high-frequency chattering, trajectory drift, and severe oscillations near obstacles, STAC-MPC certifies exponential error decay to maintain safe clearance.

Overview

While traditional Actor-Critic MPC relies on simple MLPs to dynamically adjust cost matrices, they struggle to capture the temporal consistencies needed for environments with time-correlated disturbances. Sequence models like Transformers solve this temporal issue, but proving their closed-loop stability within a feedback control loop is notoriously difficult.

We bridge this gap by embedding a Transformer into a differentiable MPC pipeline to predict cost matrices from past trajectories. By treating the Transformer and physical plant as an interconnected dynamical system, we use Riemannian contraction theory to mathematically bound the network's sensitivity to input noise. Enforcing these boundaries during training yields a highly performant control policy with certifiable theoretical robustness against severe disturbances.

System Schematics

Citation

If you find this work useful, please consider citing it:

@article{marino2026stable,
  title={Stable Transformer-Actor-Critic Model Predictive Control: A Contraction Analysis Approach},
  author={Marino, Antonio and Modugno, Valerio and Cognetti, Marco},
  journal={arXiv preprint arXiv:2606.20197},
  year={2026}
}