EUVIMEDEUVIMED entdecken
Crossref

When Generative Control Meets Formal Logic: Schrödinger Bridges for STL-Constrained Robotics

Shaochen Wang, Qilin Wu, Qing Huang, Yanjie Xia

IEEE Robotics and Automation Letters · 2026-12 · 11 (12) · 14444-14451

DOI: 10.1109/lra.2026.3746909 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Robotics and Automation Letters, 2026-12. Verfasst von Shaochen Wang und 3 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Jiangxi Provincial Natural Science Foundation, Early-Career Young Scientists and Technologists Project of Jiangxi Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Generative robot learning excels at synthesizing high-dimensional behaviors, yet aligning these continuous processes with explicit, human-defined temporal rules remains a fundamental challenge. Existing paradigms typically enforce formal specifications, such as Signal Temporal Logic (STL), via post-hoc external guidance. However, treating rules as ad-hoc corrections frequently leads to optimization conflicts, compromised dynamic feasibility, and reduced trajectory diversity. To address this, we formulate STL-constrained trajectory generation as a Schrödinger Bridge (SB) optimal control problem. Rather than relying on external modifications, our formulation establishes an optimal transport pathway between two strictly defined distributions: a source distribution governed by STL rules and a target distribution of kinematically feasible behaviors, thereby directly conditioning the generative vector field on continuous embeddings extracted from the STL syntax tree. To materialize this theoretical bridge, we translate discrete STL syntax into Abstract Syntax Trees and extract continuous logical embeddings via a message-passing neural network. These representations directly condition a decoupled neural vector field. By employing a simulation-free training paradigm, our architecture gracefully mitigates the numerical boundary singularities inherent in exact SB formulations, ensuring stable and tractable optimization. Extensive evaluation across diverse tasks demonstrates that our framework achieves highly competitive results.

Förderung

  • National Natural Science Foundation of China
  • Jiangxi Provincial Natural Science Foundation
  • Early-Career Young Scientists and Technologists Project of Jiangxi Province

Zitieren

APA

Wang, S., Wu, Q., Huang, Q., & Xia, Y. (2026). When Generative Control Meets Formal Logic: Schrödinger Bridges for STL-Constrained Robotics. IEEE Robotics and Automation Letters, 11(12), 14444-14451. https://doi.org/10.1109/lra.2026.3746909

Vancouver

Wang S, Wu Q, Huang Q, Xia Y. When Generative Control Meets Formal Logic: Schrödinger Bridges for STL-Constrained Robotics. IEEE Robotics and Automation Letters. 2026;11(12):14444-14451. doi:10.1109/lra.2026.3746909

BibTeX
@article{Wang2026,
  author = {Wang, Shaochen and Wu, Qilin and Huang, Qing and Xia, Yanjie},
  title = {When Generative Control Meets Formal Logic: Schrödinger Bridges for STL-Constrained Robotics},
  journal = {IEEE Robotics and Automation Letters},
  year = {2026},
  volume = {11},
  number = {12},
  pages = {14444--14451},
  doi = {10.1109/lra.2026.3746909}
}
Kostenlos starten →