FD-VLA: Force-Distilled Vision-Language-Action Model for Contact-Rich Manipulation

ABSTRACT

Current VLA models primarily rely on visual feedback, which is insufficient for contact-rich tasks like precision assembly or handling delicate objects. We introduce FD-VLA, a force-distilled framework that integrates haptic and force-torque data into the action-prediction loop. By distilling force-sensing capabilities from a teacher model trained in simulation into a student VLA model, we enable 'blind' manipulation success where visual occlusion occurs. Our results show a 34% improvement in task success for tight-tolerance peg-in-hole assembly under varying lighting and occlusion conditions.