一种空间机器人流匹配VLA模型安全漏洞研究OA

Vision-Language-Action(VLA)models are emerging as a core technology for next-gen-eration on-orbit autonomous control of space robots,supporting critical missions including space-craft capture,extravehicular maintenance and payload assembly.However,their heavy reliance on large-scale datasets and their distinctive mechanism for continuous action generation introduce security vulnerabilities that have not been systematically studied.From the perspective of auton-omous on-orbit control security for space robots,this paper proposes a dynamics-aware backdoor security framework for flow-matching-based VLA models tailored to space robot applications.A highly stealthy trigger mechanism tailored to space operation scenarios,a conditional dynamics attack module,and a dynamics quasi-regularization module are designed to achieve highly stealthy attacks with high success rates.Systematic experiments are conducted on a high-fidelity robotic simulation platform and a physical manipulator verification system to fully verify the effectiveness and stealthiness of the method.The results show that the method achieves an attack success rate up to 100%under context-aware triggering,with negligible degradation in normal task perform-ance.The generated malicious trajectories show no significant differences from normal trajecto-ries in kinematic metrics such as velocity,acceleration and smoothness,thus effectively bypass-ing existing space robot anomaly detection mechanisms.This study reveals critical security vul-nerability in flow-matching-based continuous generative VLA models for space robots,and pro-vides a theoretical basis and engineering reference for the security design and defense of space ro-bot intelligent control systems.