Is a Little Slip Worth It? Certified Fast Lifting for Learned Grasp Policies Under a Static-to-Kinetic Friction Drop

Abstract

A robot gripper lifting an unfamiliar object must decide how hard to squeeze and how fast to move. Squeezing less and moving faster are both attractive, and a few millimetres of slip in the fingers is often harmless. This suggests letting a learned policy trade a little slip for a lighter, faster lift, protected by a certificate that the slip stays small. So, is a little slip worth it? Mostly not. We study the question in MuJoCo simulation with explicit stick-slip friction, where friction drops from its static to its kinetic level once an object starts to slide. We derive a closed-form bound on slip during a minimum-jerk lift from a probed breakaway grip and an assumed bound on the static-to-kinetic friction ratio, and use it as a shield that raises the grip or slows the lift until the bound is at most 3.5 mm. In a pre-registered study, a linear residual policy trained with evolution strategies cut lift time from 1.00 s to 0.37 s against a friction-model controller. Without the shield it learned to grip 6% below the model, exploiting the model's conservatism at short lift times, and 12-16% of its lifts slipped more than 5 mm once friction could drop. With the shield, at training or only at test time, none of 3,520 certified lifts did, at a cost of 6% more grip and no speed, and the results replicated on fresh objects. The certificate holds because the shield ends up forbidding slip altogether: when friction may drop by 10% or more, allowing even 10 mm of slip saves no grip. A controlled study that changes one design choice at a time (19 changes, 6 methods, frozen before it was run) shows that the shield's benefit survives heavier objects, softer pads, poor weight estimates and most training choices. It also finds two real weaknesses: the breakaway probe fails on slippery objects or with a noisy slip sensor, leaving the shield nothing to certify; and because the shield sits exactly at the onset boundary, it fails when friction is allowed to weaken during the creep that precedes slip. A margin on onset is the obvious fix.