URF: A unified policy and control framework for stable contact-rich robot manipulation

Jiyou Shin et al. propose URF, a unified robot control-policy framework that lets a high-level policy predict low-level execution modes rather than only outputting motion targets, thereby bridging the execution gap in rigid-contact scenarios. At each action step, the framework predicts an action block containing a virtual target, stiffness matrix, and impedance-admittance switching ratio, and uses the measured external force magnitude as a proxy signal for environment stiffness to supervise learning. On the box flipping benchmark, the pure force-control baseline achieved a 0% success rate, the adaptive compliance controller ACP achieved 25% with a critical failure rate as high as 86.7%; the full URF reduced the peak force growth rate from 48.5±15.5 N/s to a lower level, achieved a 0% critical failure rate, and significantly suppressed oscillation.