Researchers have developed CertVLA, a defense mechanism to protect Vision-Language-Action (VLA) models from physical visual attacks. These attacks involve localized perturbations that can deceive the model into taking incorrect actions. CertVLA uses a combination of calibrated regions and deterministic covering masks to ensure that at least one prediction is attack-free. The method normalizes action disagreement by benign variation and accepts a single-mask anchor only when it remains consistent under every second mask. Experiments in simulation and real-world scenarios demonstrate the effectiveness of CertVLA against patch attacks, with additional validation on texture attacks. --- Why it matters: This matters to researchers in AI because VLA models are increasingly being used in applications where physical safety is critical, such as autonomous vehicles or robots. A defense mechanism like CertVLA can provide assurance that these systems will behave correctly even under physical visual attacks. Source: https://arxiv.org/abs/2608.20791