EquiDexFlow: AI model for robot hands with zero friction violations

Most learned dexterous grasp generators treat contact forces as a downstream verification step, so a kinematically plausible pose can still violate the conditions for a stable physical grasp. University of Maryland researchers address this with EquiDexFlow, an SE(3)-equivariant flow-matching model that jointly predicts wrist pose, joint angles, fingertip contacts, surface normals, and contact forces directly from an object point cloud. By projecting contacts onto the object surface and forces into the Coulomb friction cone as a built-in constraint, the model ensures placement and friction compliance without need for loss penalties. End-to-end SE(3) equivariance is proven and verified empirically over 200 rotations, with wrist residuals below 0.04° and exactly zero joint deviation.

Trained on 8,100 force-closure grasps across 81 objects for the 16-DoF Allegro Hand, EquiDexFlow achieves zero friction violations, the best composite score, and the lowest wrench residual among all ablation variants. The researchers retarget decoded fingertip contacts to a 16-DoF LEAP Hand via per-finger inverse kinematics, and a hardware-feasible refinement keeps every joint at least 5% inside its actuator envelope while preserving wrench balance. On the physical robot, retargeted EquiDexFlow-decoded grasps complete open-loop pick-and-hold trials on all six test objects, with every asymmetric object succeeding at both the canonical pose and a 120° co-rotation. Code, checkpoints, and demonstration videos are publicly available.