Decoupling the Declarative from the Procedural in Vision-Language-Action Models

The article introduces w²VLA, a new Vision-Language-Action (VLA) model designed to enhance the transferability of skills in robotic agents by decoupling declarative knowledge from procedural knowledge. This model restructures information flow by modulating the robot state sequence with visual, spatial, and skill information, which improves robustness against variations and enables zero-shot skill transfer to novel objects. This decoupling approach addresses limitations in existing billion-parameter Vision-Language Models (VLMs) and is significant for practitioners aiming to deploy versatile robotic systems in dynamic environments.