Adversarial Robustness in Embodied AI: A Closed-Loop Perspective on Attacks and Defenses

Embodied AI systems interact with the physical world through multi-stage closed-loop processes, enabling capabilities such as spatial understanding, manipulation, and navigation. While recent advances have significantly strengthened these abilities, they also introduce new security vulnerabilities with potentially serious physical consequences. However, adversarial threats in embodied systems remain fragmented and lack a unified analytical framework. In this survey, we review adversarial attacks and defenses in embodied AI from a closed-loop perspective, analyzing how security risks emerge, propagate, and amplify across embodied systems. We organize prior work into a functional taxonomy spanning perceptionlevel, decision & planning-level, and execution & interaction-level threats, and further distinguish between optimization-time and deployment-time interventions. For each category, we synthesize representative attack strategies and defense mechanisms, including semantic backdoors, physical perturbations, safety fine-tuning, and runtime safeguards. Beyond taxonomy, we highlight fundamental differences between adversarial robustness in embodied AI and in conventional vision or language models, emphasizing closed-loop feedback, long-horizon safety, and physical constraints. Finally, we examine existing benchmarks and evaluation protocols, discuss their limitations, and outline open challenges and future directions toward securing embodied AI in real-world environments.