Optimal Covariance Control for Stochastic Systems Under Chance Constraints

Kazuhide Okamoto, M. N. Gol'dshtein, Panagiotis Tsiotras

Abstract

This work addresses the optimal covariance control problem for stochastic discrete-time linear time-varying systems subject to chance constraints. Covariance steering is a stochastic control problem to steer the system state Gaussian distribution to another Gaussian distribution while minimizing a cost function. To the best of our knowledge, covariance steering problems have never been discussed with probabilistic chance constraints although it is a natural extension. In this work, first we show that, unlike the case with no chance constraints, the covariance steering with chance constraints problem cannot decouple the mean and covariance steering sub-problems. Then we propose an approach to solve the covariance steering with chance constraints problem by converting it to a semidefinite programming problem. The proposed algorithm is verified using two simple numerical simulations.

Introduction

In autonomous systems, such as robotics and aerospace vehicles, navigating stochastic environments requires more than just tracking a nominal trajectory. Traditional control methods often focus on the "mean" path, but real-world safety requires managing the entire probability distribution of the system's state. Covariance Steering (CS) addresses this by steering a system from an initial Gaussian distribution to a target Gaussian distribution over a fixed time horizon.

While standard CS is well-understood, it often ignores chance constraints—the probabilistic boundaries that ensure a system remains within a safe operating region with high probability (e.g., 99.9%). This paper introduces a framework for Optimal Covariance Control that incorporates these safety boundaries, demonstrating that the inclusion of chance constraints fundamentally changes the structure of the optimization problem.

The Challenge of Coupling and Tractability

In standard covariance steering without safety constraints, a "separation principle" applies: the controller can optimize the mean trajectory and the covariance (the "tightness" of the control) independently. However, this research reveals that once chance constraints are introduced, this decoupling vanishes. The optimal mean path is now inextricably linked to the system's spread; a path that is optimal in a vacuum may be too risky if the covariance is large near an obstacle.

Furthermore, chance constraints are notoriously difficult to solve because they involve integrals over probability distributions. To make this problem computationally tractable, the authors focus on discrete-time, linear time-varying (LTV) systems affected by white Gaussian noise and propose a method to map these "fuzzy" probabilistic limits into "hard" mathematical constraints.

Methodology

The authors propose a multi-step approach to transform the stochastic problem into a solvable convex optimization framework:

  1. State Augmentation: The system dynamics are reformulated using concatenated vectors to represent the entire trajectory of states and inputs simultaneously.
  2. Convexification via Boole-Bonferroni: To handle joint probability constraints (the risk over an entire path), the authors use the Boole-Bonferroni inequality to decompose complex joint constraints into a set of individual linear constraints.
  3. Deterministic Mapping: The probabilistic constraints are transformed into deterministic second-order cone constraints. This conversion is the critical step that allows the use of standard numerical solvers.
  4. Semidefinite Programming (SDP): By defining a new decision variable related to the control gain, the authors prove that the entire problem can be cast as a Semidefinite Program. This ensures that if a feasible solution exists, a global optimum can be found efficiently.

Numerical Simulations and Results

The framework was validated using two primary benchmarks:

Double Integrator

A classic 2D physics model was used to test basic steering and constraint satisfaction. The results showed a minor "safety tax": the control cost increased from 23.90 (mean steering only) to 24.16 (chance-constrained steering). This marginal increase in cost provided a significant increase in reliability, ensuring the system stayed within prescribed bounds.

Vehicle Path Planning

A more complex 4D state model (position and velocity) simulated a vehicle navigating a narrow, cone-shaped corridor—representative of an aircraft landing or spacecraft rendezvous.

  • Reshaping the Error Ellipse: In this scenario, a standard controller failed because the state covariance grew too large, leading to a high probability of boundary violation.
  • Constraint Satisfaction: The proposed chance-constrained controller successfully reshaped the system's "error ellipse," squeezing the distribution to fit through the narrowest parts of the corridor while still reaching the target terminal state.

Key Insights and Future Directions

The most significant theoretical contribution of this work is the proof that chance constraints break the separation of mean and covariance steering. This implies that for safety-critical missions, the path must be planned with the expected noise in mind from the outset.

While the current framework is highly efficient—solving in milliseconds via modern convex solvers—it is currently limited to Linear Time-Varying (LTV) systems and Gaussian noise. Future work involves extending these techniques to non-linear dynamics and non-Gaussian disturbances. Additionally, the authors suggest this approach is a prime candidate for Stochastic Model Predictive Control (SMPC), where the optimization is updated in real-time as new sensor data becomes available.

Summary updated

Abstract

This work addresses the optimal covariance control problem for stochastic discrete-time linear time-varying systems subject to chance constraints. Covariance steering is a stochastic control problem to steer the system state Gaussian distribution to another Gaussian distribution while minimizing a cost function. To the best of our knowledge, covariance steering problems have never been discussed with probabilistic chance constraints although it is a natural extension. In this work, first we show that, unlike the case with no chance constraints, the covariance steering with chance constraints problem cannot decouple the mean and covariance steering sub-problems. Then we propose an approach to solve the covariance steering with chance constraints problem by converting it to a semidefinite programming problem. The proposed algorithm is verified using two simple numerical simulations.
BibTeX
@article{okamoto2018optimal,
  author = {Kazuhide Okamoto and M. N. Gol'dshtein and Panagiotis Tsiotras},
  title = {{Optimal Covariance Control for Stochastic Systems Under Chance Constraints}},
  journal = {IEEE Control. Syst. Lett},
  year = {2018},
  doi = {10.1109/lcsys.2018.2826038},
}
Figure