Adversarial Robustness in VLA Models
4 articles · Last updated 2026-05-29
Contents
-
1
Research
RoboGCG
Attackers can fully control a VLA-driven robot by appending just ~20 optimized text tokens to a normal instruction—no image manipulation, no model access at deployment.
-
2
Research
Model-agnostic Adversarial Attack and Defense
A model-agnostic adversarial attack disrupts vision-language-action models by misaligning visual-text embeddings, while adversarial fine-tuning defends by learning perturbation-invariant representations.
-
3
Research
StableVLA
StableVLA introduces IB-Adapter, a plug-and-play module grounded in information bottleneck theory that enhances vision-language-action model robustness to visual corruptions without requiring extra training data.
-
4
Research
VLA-Fool
Researchers propose VLA-Fool, demonstrating how textual typos, visual patches, and cross-modal misalignment can adversarially attack vision-language-action models.