The Graph Hand

A grasping policy trained on one robotic hand does not work on another. The Allegro hand has 16 joints. The LEAP hand has 16 joints in different positions. The Shadow hand has 24. Each requires its own policy, trained from scratch, because the mapping from joint angles to fingertip positions differs between hardware. The hand’s morphology is baked into the policy.

Wu et al. unbake it. DexGrasp-Zero represents any hand as a graph: kinematic keypoints become nodes, links between joints become edges, and motion primitives (pre-grasp approach, finger curl, thumb opposition) align the graph across morphologies. The Allegro and the LEAP, despite different joint layouts, map to the same graph structure with the same motion primitives at corresponding nodes. A graph convolutional network then learns grasping not from joint angles but from graph-structured representations that are morphology-invariant.

The result: 85% zero-shot success rate on unseen hardware in simulation, 82% on real robots across three different platforms. No fine-tuning, no retraining, no adaptation layer. The policy transfers because it was never specific to a hand in the first place.

The mechanism is anatomical abstraction. A human grasps by opposing thumb to fingers, curling digits around the object, adjusting contact points for stability. These operations are defined relative to hand anatomy, not absolute joint angles. DexGrasp-Zero encodes the same abstraction: the graph captures anatomical relationships (which joints form a finger, which finger opposes which), and the motion primitives capture functional roles (approach, enclose, stabilize). The policy learns to grasp in anatomical coordinates.

The structural lesson: morphological diversity is an obstacle only if the representation is morphology-specific. Represent the function (grasp an object) rather than the mechanism (move these joints), and the diversity becomes irrelevant. The 59.5% improvement over state-of-the-art is not from a better algorithm but from a better coordinate system.