Efficient and Safe Learning-based Control of Piecewise Affine Systems Using Optimization-Free Safety Filters*

Conference Paper (2025)
Author(s)

K. He (TU Delft - Team Bart De Schutter)

S. Shi (TU Delft - Team Bart De Schutter, Massachusetts Institute of Technology)

Ton Van Den Boom (TU Delft - Team Ton van den Boom)

B De Schutter (TU Delft - Delft Center for Systems and Control)

Research Group
Team Bart De Schutter
DOI related publication
https://doi.org/10.1109/CDC56724.2024.10886328
More Info
expand_more
Publication Year
2025
Language
English
Research Group
Team Bart De Schutter
Bibliographical Note
Green Open Access added to TU Delft Institutional Repository 'You share, we take care!' - Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.@en
Pages (from-to)
5046-5053
ISBN (electronic)
979-8-3503-1633-9
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

Control of piecewise affine (PWA) systems under complex constraints faces challenges in guaranteeing both safety and online computational efficiency. Learning-based methods can rapidly generate control signals with good performance, but rarely provide safety guarantees. A safety filter is a modular method to improve safety for any controller. When applied to PWA systems, a traditional safety filter usually need to solve a mixed-integer convex program, which reduces the computational benefit of learning-based controllers. We propose a novel optimization-free safety filter designed to handle state constraints that involve a combination of polyhedra and ellipsoids. The proposed safety filter only utilizes algebraic and min-max operations to determine safe control inputs. This offers a notable advantage compared with traditional safety filters by allowing for significantly more efficient computation of control signals. The proposed safety filter can be integrated into various function approximators, such as neural networks, enabling safe learning throughout the learning process. Simulation results on a bicycle model with PWA approximation validate the proposed method regarding constraint satisfaction, CPU time, and the preservation of sub-optimality.

Files

License info not available
warning

File under embargo until 26-08-2025