Abstract
This work develops a stochastic model predictive controller (SMPC) for uncertain linear systems with additive Gaussian noise subject to state and control constraints. The proposed approach is based on the recently developed finite-horizon optimal covariance steering control theory, which steers the mean and the covariance of the system state to prescribed target values at a given terminal time. We call our approach covariance steering-based SMPC, or CS-SMPC. We show that the proposed approach has several advantages over traditional SMPC approaches in the literature. Specifically, it is shown that the newly developed algorithm can deal with unbounded Gaussian additive noise while ensuring stability and recursive feasibility, and incurs lower computational cost than previous similar approaches. The effectiveness of the proposed CS-SMPC approach is confirmed using numerical simulations.
Introduction
Model Predictive Control (MPC) is a cornerstone of modern control engineering, particularly for systems with complex constraints. However, real-world applications are often subject to stochastic disturbances. While Robust MPC handles uncertainty by preparing for worst-case scenarios—often leading to overly conservative performance—Stochastic MPC (SMPC) utilizes chance constraints to balance performance and safety.
A primary challenge in SMPC is ensuring recursive feasibility: the guarantee that a valid control sequence exists at every future time step if one exists initially. This is particularly difficult with unbounded Gaussian noise, where extreme disturbances can theoretically push a system into an unrecoverable state. Existing methods often rely on bounded noise assumptions or suffer from high computational complexity.
Methodology: Covariance Steering-Based SMPC
The proposed CS-SMPC framework bridges the gap between MPC and Optimal Covariance Steering (CS). Instead of merely predicting the spread of the system state due to noise, CS-SMPC treats both the mean and the covariance as controllable variables.
The methodology is built on four key pillars:
- Covariance Steering: The controller "steers" the state distribution toward a specific "landing zone" at the end of the look-ahead horizon, actively shaping the evolution of uncertainty.
- Convex Reformulation: The stochastic problem is transformed into a deterministic Semidefinite Program (SDP), allowing for efficient solution via standard convex optimization tools.
- Structured Feedback: By introducing a block-diagonal gain matrix instead of the dense matrices used in standard affine-parameterization methods, the computational burden is significantly reduced.
- Terminal Constraints: Covariance assignment theory is used to design terminal constraints that ensure system stability without the need for extensive manual tuning of cost functions.
Experimental Evaluation and Results
The CS-SMPC algorithm was evaluated in two simulation environments: a mathematical 2D Unstable Spiral and an Autonomous Racing Vehicle using a linearized bicycle model.
Key Results
- Computational Efficiency: CS-SMPC demonstrated a 30% reduction in normalized solver time compared to standard disturbance feedback methods.
- Stability and Feasibility: Rigorous mathematical proofs confirm that the system's average stage cost remains bounded, ensuring long-term stability even under Gaussian noise.
- Performance Gains: In the racing scenario, the CS-SMPC controller allowed the vehicle to operate closer to track boundaries than deterministic or tube-based controllers, resulting in faster lap times by precisely managing the "probability cloud" of the vehicle's position.
Discussion and Future Work
The core innovation of this research is the shift from reacting to uncertainty to actively controlling its evolution. By constraining the maximal terminal covariance, the authors ensure recursive feasibility for unbounded noise.
However, some limitations remain. The use of Semidefinite Programming (SDP), while faster than previous stochastic methods, is still more intensive than the Quadratic Programs used in standard MPC. Additionally, the current framework is tailored for Linear Time-Invariant (LTI) systems with Gaussian noise. Future work will focus on extending these principles to non-linear systems, non-Gaussian distributions, and more efficient SDP solvers to enable deployment on low-power embedded hardware.
Summary updated