Force feedback is critical for humanoid manipulation—a force-reactive policy trained on force-sensing demonstrations outperforms position-only policies by 3x on contact-rich tasks, showing that learning to control contact forces, not just motion, is essential for real-world dexterity.
This paper presents FLORE, a framework for teaching humanoid robots to perform complex manipulation tasks that require coordinating hand forces and whole-body motion. The system uses force-sensing grippers and wearable trackers to collect demonstrations, then trains a diffusion policy that predicts both actions and force commands.