ResearchPod Summary
In sampling-based stochastic model predictive control (SMPC), safety is often enforced by evaluating whether a system's trajectory exits a predefined safe set. Because continuous-time diffusions are typically simulated using discrete time steps, the numerical approximation of these exit events—specifically the first-exit time and the failure indicator—is a major source of error. This paper investigates how higher-order adaptive discretization methods can be adapted to improve the accuracy of these exit functionals and, consequently, the performance of chance-constrained path integral control (CC-PIC).
The authors build upon existing higher-order Itô-Taylor methods that adaptively refine the time step near the boundary of the safe set. They address two distinct cases based on the commutativity of the diffusion dynamics:
The paper establishes a theoretical framework to transfer strong exit-time approximation rates to the failure indicator and the stopped cost functional, provided a terminal boundary-layer condition (the terminal tube estimate) is met. This allows the authors to quantify how improved discretization directly reduces the error in the failure probability estimates used in CC-PIC.
The study confirms that the proposed adaptive methods significantly outperform the standard Euler-Maruyama (EM) scheme. For noncommutative systems, the order-1 Milstein method with Wiktorsson approximation provides a substantial gain in exit-time and indicator accuracy. For commutative systems, the order-1.5 method yields even higher convergence rates. Numerical experiments on a unicycle model demonstrate that these improved approximations lead to tighter constraint satisfaction in closed-loop control, effectively reducing the empirical violation rate to match the prescribed safety threshold.
Safety-critical control systems often operate under strict risk constraints. By reducing the discretization bias in exit-event estimation, this work enables more reliable risk assessment and more precise control policies. The results provide a rigorous path for practitioners to move beyond the limitations of Euler-Maruyama, allowing for safer operation in complex environments where accurate boundary-crossing detection is essential.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.