FP2: Equipping Robotic Foundation Models with Force Control
This study addresses the limited force control capabilities of robot foundation models in contact-rich scenarios by proposing FP2, a lightweight interface. The method employs an action modulation decomposition architecture that decouples task-level generation from high-frequency force control. It introduces a novel structured force control parameter generation mechanism combining context representation compression with physical feedback prediction, and integrates multimodal historical data to achieve explicit force control. Experimental results demonstrate that FP2 significantly enhances manipulation performance and force control quality across four real-world contact tasks. It outperforms mainstream baseline methods while exhibiting superior generalization capabilities.