Layer-wise structural and mechanistic interpretation of OpenVLA using controlled perturbations

Limitations of conventional policies in robot learning have driven the adoption of large-scale pretrained models, known as foundation models, leading to the emergence of Vision-Language- Action (VLA) models. Although VLA models have alleviated some of these significant limitations, such as generaliz