Decoupling the Declarative from the Procedural in Vision-Language-Action Models
w2VLA is a new Vision-Language-Action (VLA) model that separates knowledge into declarative (what/where) and procedural (how) parts. It achieves zero-shot skill transfer to novel objects by modulating the robot's state with visual, spatial, and skill information compositionally. This decouples concepts from actions, allowing robust generalization beyond specific training demonstrations.