VLA Deep Dive: Unpacking How AI Drones Actually See and Act

A new study peers inside Vision-Language-Action (VLA) models, revealing that visual perception overwhelmingly drives robot behavior, often overriding language commands, and exposing spatially bound motor programs.