Abstract
This paper introduces a novel Stochastic Model Predictive Control (SMPC) framework designed for linear time-varying (LTV) systems facing unbounded disturbances, such as Gaussian noise. By leveraging Covariance Steering (CS), the authors provide a convex optimization framework that guarantees recursive feasibility—ensuring the controller always finds a valid solution. This is achieved through the design of robust terminal constraints and a convex reformulation of the affine feedback policy. Simulations on a kinematic bicycle model demonstrate the method's ability to maintain safety and feasibility in dynamic environments where standard MPC fails.
Problem Statement: The Gap in Stochastic MPC
Model Predictive Control (MPC) is a cornerstone of autonomous systems, allowing for optimal control while respecting physical constraints. However, applying MPC to real-world scenarios involves two significant challenges: unbounded noise and time-varying dynamics. Most existing literature assumes disturbances are bounded, which is unrealistic for systems subject to Gaussian noise. Furthermore, proving recursive feasibility for systems whose dynamics change over time (LTV) is mathematically complex.
As shown in the table above, while various works address parts of the problem, this paper is the first to simultaneously handle unbounded disturbances, online feedback optimization, and LTV systems with guaranteed recursive feasibility.
Methodology: Covariance Steering and Convex Reformulation
The core innovation lies in the use of Covariance Steering (CS). Unlike traditional MPC which tracks a nominal state, CS treats the state's mean and covariance as decision variables. This allows the controller to explicitly manage the "spread" of the state distribution, ensuring that chance constraints (e.g., staying within a lane with 97.5% probability) are met even under unbounded noise.
Key technical components include:
- Terminal Constraint Design: The authors derive conditions for a "terminal set" and "terminal covariance." If the system can reach this safe zone within the prediction horizon, a feasible solution is guaranteed for all future time steps.
- Robust Invariance: The terminal set is designed to be robust against variations in system parameters within a defined polytopic set, ensuring stability across the LTV range.
- Convex Reformulation: To overcome the non-convexity typical of affine feedback policies, the authors apply a change of variables, transforming the problem into a computationally efficient convex program suitable for real-time hardware.
Experimental Evaluation: Kinematic Bicycle Model
The proposed CS-SMPC method was evaluated using a numerical simulation of a kinematic bicycle model, a standard benchmark for lateral vehicle control. The model is linearized around a reference trajectory to create a challenging LTV system.
The simulation tests the controller's ability to navigate a curved path while maintaining safety constraints despite significant stochastic disturbances.
Results and Discussion
In 20 Monte Carlo simulation trials, the CS-SMPC approach was compared against two baselines:
- Proposed Method: Achieved zero infeasibility points, successfully navigating all trials without failure.
- Nominal Terminal Constraints: Resulted in 6 infeasibility points, as the constraints failed to account for time-varying dynamics.
- No Terminal Constraints: Resulted in 3 infeasibility points, leading to states from which recovery was impossible.
These results confirm that the proposed terminal constraints are essential for maintaining recursive feasibility in LTV systems.
Conclusion
This work bridges a critical gap in stochastic control by providing a recursively feasible SMPC for LTV systems with unbounded noise. While the current framework assumes polytopic uncertainty and Gaussian noise, it sets a rigorous foundation for future work in non-Gaussian disturbances and reducing conservatism in robust safety sets.
Summary updated