FIRMGrasp: New friction-aware metric boosts robot grasp success from 25% to 70%

Classical grasp quality metrics assume a single deterministic friction coefficient—an oversimplification that can lead to catastrophic failures when surfaces are slippery or variable. To solve this, Clinton Enwerem and co-authors from the University of Maryland and Johns Hopkins University propose FIRMGrasp, a family of friction-volatility-aware quality metrics grounded in Conditional Value-at-Risk (CVaR). Instead of evaluating force-closure at one friction value, FIRMGrasp computes a risk-adjusted margin ε^(β) based on the CVaR-discounted mean of the adverse friction tail. This margin is differentiable in grasp parameters and comes with a probabilistic closure certificate: if ε^(β) > 0, force closure is guaranteed with at least β probability. Benchmarking against the nominal Ferrari-Canny epsilon and recent differentiable baselines, the team tested 1,599 grasps across LEAP Hand and Allegro Hand grippers. Results show that 53% of grasps the nominal metric certifies as high-quality actually lose force closure when friction is unfavorable. Moreover, the nominal margin separates realized shake and pick success near chance (0.53 for shake, 0.67 for pick), while FIRMGrasp achieves 0.63 and 0.78 respectively—a meaningful improvement.

In simulated lift trials with gravity enabled and an adverse friction coefficient of 0.2, the risk adjustment gave dramatically better outcomes. Grasps certified by FIRMGrasp reached a 70% success rate under lateral pull, whereas grasps that passed the nominal test but were rejected by FIRMGrasp succeeded only 25% of the time. This 45-point gap underscores how ignoring friction variability can undermine even carefully computed grasps. For robotics engineers designing manipulation systems, FIRMGrasp provides a principled way to quantify safety margins in real-world conditions—warehouses, kitchens, uneven terrain—where surface friction is rarely constant. By trading a small amount of nominal grasp quality for robustness, robots can avoid dropping objects when the unexpected happens. The approach is differentiable, meaning it can be integrated into gradient-based learning pipelines for grasp synthesis, opening the door to more reliable autonomous manipulation in challenging environments.