Retention Robustness of a Canonical LEAP Hand Pose: A Simulation Revalidation Study

This study revalidates retention robustness for one canonical closed pose of the 16-DoF LEAP Hand in the NVIDIA Isaac Lab. The task uses in-volume object placement, gravity-compensated closing, and an open-loop 15 cm vertical lift; it does not evaluate tabletop grasping, a learned grasp generator, or hardware performance. We verified physics-material relationships on 17 hand collision leaves, read back the runtime PhysX material buffer, and ran a matched material A/B check. We then screened 20 YCB objects and retained all 20 for a 5×3×3 mass–friction–pose-noise factorial with 15 simulator repeats per object–cell. The 13,500 rows comprise 900 object–cells from 20 fixed objects; object identity is the inferential unit. The primary operational outcome combines the height/XY window with an object-level contact indicator and net-force threshold; the contact reporter aggregates all partners and does not certify hand-supported contact. The failure cascade assigned the incomplete_envelope label to 7474 rows (55.36%); a direct pre-lift distance check exceeded 0.12 m in 10,753/13,500 rows (79.65%), including 3271 validity-qualified successes. The label therefore does not estimate envelope-entry probability. The primary success rate was 42.20%; object-level success decreased with mass and pose noise, while the friction contrast across μ=0.3–0.9 was near zero with an interval spanning zero. A 300-trial convex-hull check on four objects reduced the subset success rate from 66.00% to 33.00%, with strong object-level variation. These results are conditional on one pose and simulation protocol.