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

Yuliang Wu; Yanhan Lin; WengKit Lao; Yuhao Lin; Yi-Lin Wei; Wei-Shi Zheng; Ancong Wu

cs.ROMar 17, 2026

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

Authors: Yuliang Wu, Yanhan Lin, WengKit Lao, Yuhao Lin, Yi-Lin Wei, Wei-Shi Zheng, Ancong Wu

Organizations: School of Computer Science and Engineering, Sun Yat-sen University, China

Abstract

To meet the demands of increasingly diverse dexterous hand hardware, it is crucial to develop a policy that enables zero-shot cross-embodiment grasping without redundant re-learning. Cross-embodiment alignment is challenging due to heterogeneous hand kinematics and physical constraints. Existing approaches typically predict intermediate motion targets and retarget them to each embodiment, which may introduce errors and violate embodiment-specific limits, hindering transfer across diverse hands. To overcome these limitations, we propose DexGrasp-Zero, a policy that learns universal grasping skills from diverse embodiments, enabling zero-shot transfer to unseen hands. We first introduce a morphology-aligned graph representation that maps each hand's kinematic keypoints to anatomically grounded nodes and equips each node with tri-axial orthogonal motion primitives, enabling structural and semantic alignment across different morphologies. Relying on this graph-based representation, we design a Morphology-Aligned Graph Convolutional Network (MAGCN) to encode the graph for policy learning. MAGCN incorporates a Physical Property Injection mechanism that fuses hand-specific physical constraints into the graph features, enabling adaptive compensation for varying link lengths and actuation limits for precise and stable grasping. Our extensive simulation evaluations on the YCB dataset demonstrate that our policy, jointly trained on four heterogeneous hands (Allegro, Shadow, Schunk, Ability), achieves an 85% zero-shot success rate on unseen hardware (LEAP, Inspire), outperforming the state-of-the-art method by 59.5%. Real-world experiments further evaluate our policy on three robot platforms (LEAP, Inspire, Revo2), achieving an 82% average success rate on unseen objects.

Figures & tables

Fig. 2: Universal hand representation. (a) Morphology-Aligned State Graph Representation: nodes correspond to anatomical units, edges follow kinematic chains, yielding a hand-agnostic semantic graph structure. (b) Schematic of three motion primitives (Flexion, Abduction, Axial Rotation) on a Schunk hand, showing their physical motion effects at representative joints.

Fig. 3: Architecture of DexGrasp-Zero. At each time step $t$ : (a) Morphology-Aligned Graph Encoder encodes hand-object state into node features $\mathbf{X}^{h}_{\text{node},t}$ and global feature $\mathbf{x}_{g,t}^{h}$ using a hand-specific graph (adjacency $\mathbf{A}^{h}$ ); a GCN with per-layer physical priors produces embeddings $\mathbf{E}_{\text{node},t}^{h}$ and $\mathbf{E}_{g,t}^{h}$ . (b) Physical Property Encoder parses hand URDF to build a physical graph $\mathcal{G}_{\text{physical}}^{h}$ (joint limits, link lengths, etc.) and an activation mask $\mathbf{M}_{\text{activation}}^{h}$ , encoded into $\mathbf{E}_{\text{p}}^{h}$ and fused into every GCN layer. (c) Decoder outputs motion primitives $\boldsymbol{\alpha}_{\text{prim}}^{h}$ : wrist 6-DoF commands from $\mathbf{E}_{g,t}^{h}$ and wrist features, and joint actions from masked node embeddings; the latter are mapped via hand-specific $\mathcal{M}_{h}$ to executable joint commands $\alpha_{\text{physical},t}^{h}$ .

Method	Variant	Training Hands				Unseen Hands		Average
Method	Variant	Allegro	Shadow	Ability	Schunk	LEAP	Inspire	Seen	Unseen
CrossDex [ 35 ]	per-object	0.81	0.85	0.90	0.90	0.34	0.44	0.865	0.39
CrossDex [ 35 ]	multi-object	0.39	0.69	0.42	0.60	0.19	0.34	0.525	0.265
DexGrasp-Zero (Ours)	w/o motion primitives	0.52	0.51	0.64	0.59	0.39	0.29	0.565	0.34
	early fusion	0.42	0.47	0.59	0.54	0.46	0.34	0.505	0.40
	w/o $\mathcal{G}_{\text{physical}}$ priors	0.91	0.89	0.90	0.84	0.82	0.79	0.885	0.805

TABLE I: Cross-embodiment training and zero-shot transfer results (success rate). Variants of our method: w/o

\mathcal{G}_{\text{physical}}

priors removes hand-specific physical property encoding; early fusion concatenates physical features with node states at input (instead of layer-wise fusion in GCN); w/o motion primitives replaces the motion-primitive space with raw joint commands; w/o

M_{\text{activation}}\&r_{\text{pen}}

disables the activation mask conditioning and the corresponding action-feasibility penalty in the reward; full model is our complete DexGrasp-Zero policy.

Fig. 4: Hardware setup. We evaluate our method on three robot platforms: (a)Kinova arm with LEAP hand, (b)Kinova arm with Inspire hand, and (c)Piper arm with Revo2 hand.

Fig. 5: Simulated grasps of training hands on 5 diverse objects.

Method	Train Hand	Allegro	Shadow	Ability	Schunk	LEAP	Inspire
GraspXL [ 36 ]	–	0.94	0.93	0.91	0.90	0.95	0.91
DexGrasp-Zero (Ours)	Allegro	$0.93^{{\dagger}}$	0.83	0.80	0.69	0.48	0.77
	Shadow	0.55	$0.98^{{\dagger}}$	0.80	0.82	0.77	0.94
	Ability	0.60	0.60	$0.86^{{\dagger}}$	0.86	0.88	0.90
	Schunk	0.61	0.52	0.83	$0.92^{{\dagger}}$	0.90	0.94
	LEAP	0.50	0.68	0.65	0.69	$0.90^{{\dagger}}$	0.69

TABLE II: Single-hand training transfer on PartNet (success rate): rows show the training embodiment for DexGrasp-Zero and columns show the test embodiment. Entries marked with † indicate in-domain evaluation.

Method	LEAP	Inspire	Revo2	Average
Intra-hand (Oracle)	0.90	0.90	0.78	0.86
Cross-hand w/o $\mathcal{G}_{\text{physical}}$	0.84	0.80	0.62	0.75
Cross-hand (Ours)	0.88	0.86	0.72	0.82

TABLE III: Real-world zero-shot grasping success rate on unseen dexterous hands.

Fig. 6: Morphology-aligned graph construction results for all hand embodiments used in this work.

Fig. 7: Example pipeline for morphology-aligned graph construction.

Fig. 8: Joint-to-primitive mapping for all hand embodiments used in this work. Each URDF joint is assigned to one motion primitive.

Fig. 9: Real-world evaluation props (10 objects) used in our hardware experiments.

Parameter	Value
RL algorithm	PPO
Discount factor $\gamma$	0.996
GAE parameter $\lambda$	0.95
Clipping parameter $\epsilon$	0.2
Optimizer	Adam
Learning rate	$5\times 10^{-4}$ (adaptive KL schedule)

TABLE IV: Training and architecture hyperparameters.

Fig. 10: Training curves in simulation for our MAGCN policy.

Object	LEAP	Inspire	Revo2
wine-glass	5/5	5/5	4/5
Beverage bottle	4/5	3/5	4/5
Spray bottle	4/5	4/5	4/5
Squeeze sauce bottle	4/5	5/5	4/5
Toy dog	5/5	5/5	2/5
Coffee mug	5/5	5/5	4/5

TABLE V: Per-object real-world results. Each object is evaluated over 5 poses per platform and we report successes out of 5; the average row reports the mean success rate.

Fig. 11: Representative failure cases on the three unseen hands. (a) LEAP, (b) Inspire, (c) Revo2.

Scale	SR
1	0.92
2	0.85
1/4	0.87

TABLE VI: URDF prior sensitivity on Allegro via link-length scaling. We report success rate and qualitative behavior.

Backbone	Train-hand SR (%)	Unseen-hand SR (%)
GCN (ours)	91.3/92.1/92.5 (91.9 $\pm$ 0.6)	86.1/84.9/85.5 (85.5 $\pm$ 0.6)
Graph-Transformer	86.4/87.2/86.9 (86.8 $\pm$ 0.4)	81.5/80.4/79.8 (80.6 $\pm$ 0.9)

TABLE VII: Backbone comparison (3 seeds). SR is reported on train hands and unseen hands.

Fig. 12: Barrett Hand graph construction and qualitative grasping examples.

Explore similar work

DiffMorphGrasp: A Diffusion Framework Conditioned on Hand Morphology for Cross-Embodiment Grasp Generation
Jan 31, 2026Zhiyuan Wu, Xiangyu Zhang, Zhuo Chen +3Robotic Grasping Cross-Embodiment Robot Learning
Department of Engineering, King’s College London, Strand, London, WC2R 2LS, United Kingdom · Imperial College London, London, SW7 2AZ, United Kingdom
Function beyond Form: Functional Correspondence for Cross-Embodiment Dexterous Grasp Generation
Sep 30, 2026Bolin Zou, Wenlong Dong, Mu Ai +4Robotic Grasping Cross-Embodiment Robot Learning
Shenzhen Key Laboratory of Robotics and Computer Vision, Southern University of Science and Technology, Shenzhen, China. · Department of Robotics, Perception and Learning, KTH Royal Institute of Technology, Stockholm, Sweden. · School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China. +1
MANGO-Grasp: Mahalanobis Fields over Geometry-Oriented 3D Gaussians for Cross-Embodiment Dexterous Grasping
Aug 3, 2026Heng Zhang, Kevin Yuchen Ma, Mike Zheng Shou +2Robotic Grasping Cross-Embodiment Robot Learning
Robotics & Autonomous Systems Division, Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR-I2R), Singapore · College of Computing and Data Science, Nanyang Technological University, Singapore · Show Lab, National University of Singapore, Singapore

Jan 31, 2026J/K move · Enter open · S save

DiffMorphGrasp: A Diffusion Framework Conditioned on Hand Morphology for Cross-Embodiment Grasp Generation

Zhiyuan Wu, Xiangyu Zhang, Zhuo Chen, Jiankang Deng+2

Department of Engineering, King’s College London, Strand, London, WC2R 2LS, United Kingdom · Imperial College London, London, SW7 2AZ, United Kingdom

Robotic Grasping Cross-Embodiment Robot Learning Zero-Shot Robotic Manipulation Diffusion Models

Open paper AskPDF arXiv

As robots are deployed across diverse environments, manipulation tasks require robot hands with different morphologies to satisfy varying task requirements, hardware capabilities, and cost constraints. However, transferring grasping policies learned on one robot hand to another with a different morphology remains a fundamental challenge due to variations in kinematic structures and degrees of freedom across embodiments. To address these limitations, we propose DiffMorphGrasp, the first diffusion-based framework that incorporates hand morphology information into the generative process for cross-embodiment dexterous grasping. The proposed approach maps grasps from diverse robotic hands into a reference hand pose, providing a common space for learning. Grasp generation is then continuously conditioned on structured representations of hand kinematics, encoded as graphs derived from hand configurations, together with object geometry. In addition, we introduce a loss function that computes joint-level errors over the hand kinematic structure to constrain the generated grasps during training. To evaluate DiffMorphGrasp, we construct a new benchmark for assessing grasp generation performance across robot hands with different morphologies. Extensive experiments demonstrate that DiffMorphGrasp achieves strong zero-shot generalization to unseen hand morphologies, enabling scalable cross-embodiment grasp deployment. Project page: https://georgewuzy.github.io/DiffMorphGrasp-website

Abstract

Explore similar work

DiffMorphGrasp: A Diffusion Framework Conditioned on Hand Morphology for Cross-Embodiment Grasp Generation

Function beyond Form: Functional Correspondence for Cross-Embodiment Dexterous Grasp Generation

MANGO-Grasp: Mahalanobis Fields over Geometry-Oriented 3D Gaussians for Cross-Embodiment Dexterous Grasping