LIBERO

Researchers have uncovered a critical vulnerability in World-Action Models (WAMs) where adversarial visual perturbations can decouple a robot's predicted future from its actual movements. This 'World-Action Drift Attack' allows an attacker to hijack robot actions while leaving the model's imagined future appearing visually plausible and safe.

Researchers have uncovered a critical vulnerability in World-Action Models (WAMs) where adversarial visual perturbations can decouple a robot's predicted future from its actual movements. This 'World-Action Drift Attack' allows an attacker to hijack robot actions while leaving the model's imagined future appearing visually plausible and safe.

This paper details a framework for training vision-language-action models that generalize across diverse robotic embodiments. By utilizing a unified action-state space and pre-training action priors, the model achieves robust performance in both simulated and real-world environments.

This paper details a framework for training vision-language-action models that generalize across diverse robotic embodiments. By utilizing a unified action-state space and pre-training action priors, the model achieves robust performance in both simulated and real-world environments.