DexGrasp-Zero — Morphology-Aligned Policy for Zero-Shot Cross-Embodiment Dexterous Grasping

Venue: RSS 2026 (Manipulation session) · Authors: Yuliang Wu, Yanhan Lin, WengKit Lao, Yuhao Lin, Yi-Lin Wei, Wei-Shi Zheng, Ancong Wu — Sun Yat-sen University · arXiv: 2603.16806 · project Category: Cross-embodiment dexterous grasping Trend tag: RSS 2026 thread 5 — hands go cross-embodiment

Compiled from the verified RSS 2026 abstract and the paper's Fig. 1.

Key figure

Figure 1 of the paper. (a) Prior paradigm: train on a simplified, lossy unified state space, output intermediate motion targets, then run per-hand retargeting models — which add complexity and can emit kinematically infeasible actions. (b) This paradigm: a single end-to-end policy over the lossless morphology-aligned graph, acting in a hand-agnostic motion-primitive space; physical joint commands come from a fixed per-hand mapping M_h — no trainable retargeting at all. (c) Real-world deployment grid on unseen hands: Leap (sr 0.88), Inspire (0.86), Revo2 (0.72) across everyday objects, with the four training hands (Allegro, Shadow, Ability, Schunk) shown at right. The train-on-4 / deploy-on-3-unseen split is the zero-shot claim made visible.

Problem

Every new dexterous hand currently means re-learning grasping from scratch. Retargeting-based transfer (predict intermediate motion targets, map per-embodiment) introduces errors and violates embodiment-specific joint limits.

Method

• Morphology-aligned graph representation: each hand's kinematic keypoints map to anatomically grounded nodes, each equipped with tri-axial orthogonal motion primitives — structural + semantic alignment across morphologies without retargeting.
• MAGCN (Morphology-Aligned Graph Convolutional Network) encodes the graph for policy learning, with a Physical Property Injection mechanism fusing hand-specific constraints (link lengths, actuation limits) into node features for adaptive compensation.

Results (as reported)

• Jointly trained on four heterogeneous hands (Allegro, Shadow, Schunk, Ability); 85% zero-shot success on unseen hardware (LEAP, Inspire) on YCB — +59.5% over the prior SOTA.
• Real-world: 82% average success on unseen objects across three platforms (LEAP, Inspire, Revo2).

Significance

With One Hand to Rule Them All, forms RSS 2026's cross-embodiment-hand pair — two independent answers (graph anatomy vs canonical URDF parameterization) to the same question the arm-level field settled via canonical action tensors (Qwen-RobotManip, GR00T). The 85% zero-shot number is the strongest cross-hand transfer result on record and moves Review-Cross-Embodiment's agenda into the 20+ DoF regime.

← RSS 2026 survey · Home