Breaking flow-matching embodied policies with trajectory-coupled adversarial perturbations
Security and Safety · 2026 · 5 · 2026016
DOI: 10.1051/sands/2026016 ↗Kurz gefasst
Fachartikel in der Zeitschrift Security and Safety, online erschienen am 22.10.2026. Verfasst von Siyuan Pan und 3 weiteren Autor:innen. Die Arbeit nennt 60 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: State Key Laboratory of Industrial Control Technology, Natural Science Foundation of China, Frontier Technologies R\&D Program of Jiangsu.
Abstract (Originalsprache)
Flow-matching Vision-Language-Action (VLA) policies generate actions through τ -indexed ordinary differential equation (ODE) rollouts, exposing intermediate velocity-field structure that endpoint-only adversarial objectives do not directly target. We study this mechanism on the evaluated π 0.5 policy in LIBERO and propose the Tau-Path Coupled Attack (TPCA), a white-box visual attack that couples executed-window endpoint displacement with velocity-field divergence along the flow rollout. The study is deliberately scoped as a single-architecture mechanism analysis rather than a general claim about all flow-matching policies. Under matched compute against Visual-PGD, TPCA and endpoint-only optimization both produce near-saturated task failure at 3/255, while TPCA yields larger action-space endpoint displacement. As the perturbation budget tightens, the task-level gap becomes visible: at 1/255, TPCA induces a Drop of 0.38 whereas endpoint-only optimization induces 0.17, with over +129% larger action-space endpoint displacement. Cross-task checks on two additional LIBERO spatial tasks show consistent action-space endpoint-displacement advantages for the evaluated setting, while task-failure advantages are budget- and task-dependent. These results support a bounded conclusion: flow-rollout coupling provides additional optimization signal for the evaluated π 0.5 policy, but broader architectural generalization and defense effectiveness require future extensive studies.
Förderung
- State Key Laboratory of Industrial Control Technology
- Natural Science Foundation of China
- Frontier Technologies R\&D Program of Jiangsu
- Xiaomi Foundation
- Fundamental Research Funds for the Central Universities
Zitieren
APA
Pan, S., Guo, R., Liu, M., & Deng, R. (2026). Breaking flow-matching embodied policies with trajectory-coupled adversarial perturbations. Security and Safety, 5, 2026016. https://doi.org/10.1051/sands/2026016
Vancouver
Pan S, Guo R, Liu M, Deng R. Breaking flow-matching embodied policies with trajectory-coupled adversarial perturbations. Security and Safety. 2026;5:2026016. doi:10.1051/sands/2026016
BibTeX
@article{Pan2026,
author = {Pan, Siyuan and Guo, Rong and Liu, Mengxiang and Deng, Ruilong},
title = {Breaking flow-matching embodied policies with trajectory-coupled adversarial perturbations},
journal = {Security and Safety},
year = {2026},
volume = {5},
pages = {2026016},
doi = {10.1051/sands/2026016}
}
EUVIMED