CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation