Grasp Execution Without a Planner: Configuration-Space Grasp Distance Fields with Certified Safety & Guaranteed Quality

What the math gives to ML

The paper combines a differentiable soft minimum over a finite set of valid configurations with a safety-filtered feedback controller, giving a useful pattern for neural policies: represent successful behaviors as configuration prototypes, attract the policy toward the nearest prototype, and project the resulting action through explicit barrier constraints. The transferable asset is not the grasp-specific wrench model itself, but the differentiable finite-prototype distance field and its explicit approximation error bound of \(\log N/\rho\). A neural residual policy can propose task-dependent actions while a small quadratic-program layer enforces known state constraints and exposes a scalar slack measuring when the learned policy is obstructed. The first target should be robotic imitation or safe reinforcement learning, where this can replace trajectory planning with a compact prototype-conditioned feedback module.