DEFT-VLA: Stage-Localized Force Specialization for Contact-Rich Robotic Manipulation

DEFT-VLA is a novel approach for contact-rich robotic manipulation that addresses sensory conflicts between vision, proprioception, and force. By splitting the task into vision-dominated non-contact and force-critical contact stages, the method improves performance in contact phases while maintaining robust non-contact behavior. Experiments on pipette tip attachment, bimanual aspiration, and whiteboard wiping show significant improvements over existing baselines.