Physics-Conditioned Grasping for Stable Tool Use

Abstract

Tool use often fails not because robots misidentify tools, but because grasps cannot withstand task-induced wrench. Existing vision-language manipulation systems ground tools and contact regions from language yet select grasps under quasi-static or geometry-only assumptions. During interaction, inertial impulse and lever-arm amplification generate wrist torque and tangential loads that trigger slip and rotation. We introduce inverse Tool-use Planning (iTuP), which selects grasps by minimizing predicted interaction wrench along a task-conditioned trajectory. From rigid-body mechanics, we derive torque, slip, and alignment penalties, and train a Stable Dynamic Grasp Network (SDG-Net) to approximate these trajectory-conditioned costs for real-time scoring. Across hammering, sweeping, knocking, and reaching in simulation and on hardware, SDG-Net suppresses induced torque up to 17.6%, shifts grasps below empirically observed instability thresholds, and improves real-world success by 17.5% over a compositional baseline. Improvements concentrate where wrench amplification dominates, showing that robot tool use requires wrench-aware grasp selection, not perception alone.