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

Zhiyuan Wu; Xiangyu Zhang; Zhuo Chen; Jiankang Deng; Rolandos Alexandros Potamias; Shan Luo

cs.ROJan 31, 2026

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

Authors: Zhiyuan Wu, Xiangyu Zhang, Zhuo Chen, Jiankang Deng, Rolandos Alexandros Potamias, Shan Luo

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

Abstract

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

Figures & tables

Figure 1: Given object geometry and robot hand morphology, DiffMorphGrasp generates cross-embodiment grasps, generalizing from training hand embodiments (purple hand parts) to novel hand-structure variations and unseen hand morphologies (green hand parts).

Figure 2: (Left) The overview of our proposed DiffMorphGrasp for cross-embodiment dexterous grasp generation. Given an object point cloud and a target hand morphology extracted from its URDF specification (mapped to a pre-defined reference hand pose), we employ a morphology encoder to extract morphology representations from the hand’s joint structure. The hand pose (noised via a diffusion scheduler in training) is embedded through a linear layer, and concatenated with its active joint mask embedding to obtain the hand pose representation. This representation is then processed through a morphological denoising model, where the iterative process is conditioned on both the morphology representation and the point cloud representation extracted via a Point Transformer [ 43 ] . The entire framework is trained based on a morphological loss function. (Right) The structure of our morphological denoising model, which is conditioned on the encoded morphology and the point cloud representations via cross-attention.

Figure 3: Reference hand mapping example. (a) The 24-DoF Shadow hand serves as a fixed reference. (b) The Leap hand is mapped to the reference through chain/joint correspondence, with missing joints masked. (c) Generated active reference joints are mapped back to the Leap hand.

Method	Success Rate (%) $\uparrow$				Diversity (rad.) $\uparrow$				Eff. (s) $\downarrow$
Method	Barrett	Allegro	ShadowHand	Avg.	Barrett	Allegro	ShadowHand	Avg.	Eff. (s) $\downarrow$
DFC	86.3	76.2	58.8	73.8	0.532	0.454	0.435	0.474	$>$ 1800
GenDexGrasp	67.0	51.0	54.2	57.4	0.488	0.389	0.318	0.398	19.7
CEDex	93.1	88.1	85.0	88.7	0.624	0.473	0.438	0.512	7.8
GeoMatch	60.0	-	67.5	63.8	0.259	-	0.235	0.247	-
GeoMatch++	77.5	-	70.0	73.8	0.378	-	0.184	0.281	-

Table 1: Quantitative results of our DiffMorphGrasp (w/. and w/o. the morphological loss) compared with cross-embodiment dexterous grasp synthesis baselines on the MultiDex [ 13 ] dataset across three robotic hands: Barrett, Allegro, and Shadow hands.

Figure 5: Visualizations of cross-embodiment grasps synthesized by DiffMorphGrasp.

Figure 6: Zero-shot generalization to structural variations of the Shadow hand. (a)-(f) geometrical variations: lengthen and shorten fingers. (g)-(i) topological variations: remove fingers. (j)-(l) hybrid variations: replace fingers with Allegro. Altered fingers are highlighted in green.

Figure 7: Visualizations of zero-shot grasp generalization to unseen Inspire and Leap hands. Green highlights that the morphologies are completely unseen during training.

Methods	Suc. Rat. (%) $\uparrow$	Div. (rad) $\uparrow$	Pen. Dep. (mm) $\downarrow$	Q1 Metric $\uparrow$
Inspire - DRO-Grasp	6.9	0.636	20.88	0.0041
Inspire - Ours	86.3	0.466	10.02	0.0281
Leap - DRO-Grasp	43.8	0.353	22.03	0.0203
Leap - Ours	83.7	0.409	12.28	0.0294

Table 2: Quantitative results of zero-shot generalization to completely unseen hand morphologies, i.e. , Inspire and Leap hands on the MultiDex [ 13 ] dataset. We report success rate, diversity, penetration depth, and Q1 metric.

Figure 8: Real-world validation using the Leap Hand [ 26 ] demonstrates stable grasps on eight YCB objects [ 2 ] . Video results are provided in our supplementary demo.

Baseball	Cleanser	Pear	Mustard Bottle	Soup Can	Meat Can	Chef Can	Rubik’s Cube
9/10	10/10	10/10	8/10	9/10	8/10	10/10	9/10

Table 3: Quantitative real-world evaluation on the Leap Hand [ 26 ] . We report the success rate over 10 attempts for eight objects from the YCB dataset [ 2 ] .

Figure 9: Failure cases of DiffMorphGrasp. We visualize grasp attempts on a thin sponge across three different robotic hands, which illustrate the limitation of object-hand penetration.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	Pen. Dep. (mm) $\downarrow$				Q1 Metric $\uparrow$
Methods	Barrett	Allegro	ShadowHand	Avg.	Barrett	Allegro	ShadowHand	Avg.
GenDexGrasp	15.8	18.1	21.4	18.4	0.0194	0.0255	0.0232	0.0227
DRO-Grasp	14.0	12.1	13.9	13.3	0.0059	0.0116	0.0149	0.0108
Ours	9.3	10.9	8.5	9.6	0.0255	0.0300	0.0284	0.0280

Appendix

Table 4: Quantitative results of penetration depth and Q1 metric compared to cross-embodiment dexterous grasp synthesis baselines on the MultiDex [ 13 ] dataset.

Figure 10: Visualizations of cross-embodiment grasps synthesized by DiffMorphGrasp on the MultiDex [ 13 ] dataset. Two viewing angles are presented for each grasp.

Figure 11: Qualitative comparison with baselines 1) GenDexGrasp [ 13 ] and 2) DRO-Grasp [ 32 ] where our results demonstrate superior surface conformity and stable form-closure. Two viewing angles are presented for each grasp.

Figure 12: Visualizations of cross-embodiment grasps synthesized by DiffMorphGrasp on the Multi-GraspLLM [ 12 ] dataset. Two viewing angles are presented for each grasp.

Figure 13: Visualizations of cross-embodiment grasps synthesized by DiffMorphGrasp on the Objaverse [ 3 ] dataset. Two viewing angles are presented for each grasp.

Test Dataset	Method	Success Rate (%) $\uparrow$				Diversity (rad.) $\uparrow$
Test Dataset	Method	Barrett	Allegro	Shadow	Avg.	Barrett	Allegro	Shadow	Avg.
Multi-GraspLLM	GenDexGrasp	29.3	38.9	46.7	38.3	0.620	0.398	0.409	0.476
	DRO-Grasp	81.1	82.7	71.4	78.4	0.512	0.395	0.458	0.455
	Ours	84.7	84.9	92.5	87.4	0.708	0.470	0.463	0.547
Objaverse	GenDexGrasp	57.9	42.0	63.9	54.6	0.597	0.436	0.398	0.477
	DRO-Grasp	82.2	82.9	70.3	78.5	0.517	0.396	0.445	0.453

Appendix

Table 5: Cross-dataset zero-shot generalization results. We evaluate models trained on MultiDex [ 13 ] directly on unseen datasets: Multi-GraspLLM [ 12 ] and Objaverse [ 3 ] .

$M.$	$G.$	Success Rate (%) $\uparrow$				Diversity (rad.) $\uparrow$				Eff. (s) $\downarrow$
$M.$	$G.$	Barrett	Allegro	ShadowHand	Avg.	Barrett	Allegro	ShadowHand	Avg.	Eff. (s) $\downarrow$
-	-	83.7	78.8	87.5	83.3	0.669	0.431	0.430	0.510	0.45
✓	-	90.6	86.9	94.4	90.6	0.678	0.438	0.450	0.522	0.47
✓	✓	92.5	89.6	95.0	92.4	0.698	0.462	0.451	0.537	0.47

Appendix

Table 6: Ablation results on morphology encoding on the MultiDex [ 13 ] dataset across robotic hands from three to five fingers: Barrett, Allegro, and Shadow hand.

M.

refers to morph encoding, and

G.

refers to the graph attention employed in the morphology encoder.

Figure 14: Visualizations of ablation study on 1) effectiveness of morphology encoding and 2) zero-shot grasp generalization to novel hand morphologies based on the Shadow Hand. (a) w/o morphology encoding; (b) w/ morphology encoding; (c) w/ morphology encoding and graph attention; (d) w/ morphology encoding, graph attention, and morphological loss; (e)-(g) Altered fingers.

Ablation	Success Rate (%) $\uparrow$				Diversity (rad.) $\uparrow$
Ablation	Barrett	Allegro	ShadowHand	Avg.	Barrett	Allegro	ShadowHand	Avg.
Baseline	92.5	89.6	95.0	92.4	0.698	0.462	0.451	0.537
Arithmetic Mean	91.9	88.8	91.2	90.6	0.683	0.443	0.439	0.522
Geometric Mean	93.0	90.3	98.8	94.0	0.698	0.462	0.445	0.535

Appendix

Table 7: Ablation results on adaptive weights in morphology loss on the MultiDex [ 13 ] dataset. We compare baseline (w/o. morphology loss), arithmetic mean, and geometric mean.

	Altered Fingers					Suc. Rat. (%)
	Thu.	Ind.	Mid.	Rin.	Pin.	Suc. Rat. (%)
	-	-	-	-	-	98.8
Topo.	-	✗	-	-	-	93.2
	-	-	✗	-	-	97.5
	-	-	-	✗	-	98.3
	-	-	-	-	✗	84.2

Appendix

Table 8: Ablation study on zero-shot generalization to structural variations of the Shadow hand using the MultiDex [ 13 ] dataset.

-

indicates the finger remains the same. Topological variations : ✗ indicates the corresponding finger is removed; Geometrical variations :

>

indicates the corresponding finger length is scaled by a factor of

1.5

, and

<

indicates it is scaled by

0.8

; Embodiment variations :

\sim

indicates the corresponding Shadow finger is replaced by an Allegro finger.

Figure 15: More visualizations of zero-shot generalization to structural variations of the Shadow hand. Topological variations: (a) remove middle finger; (b) remove pinky; (c) remove index and ring. Geometrical variations: (d) lengthen thumb; (e) lengthen index and ring; (f) lengthen all fingers; (g) shorten thumb; (h) shorten index and ring; (i) shorten all fingers. Embodiment variations: (j) replace thumb with Allegro; (k) replace index and ring with Allegro; (l) replace all fingers with Allegro. Altered fingers are highlighted in green.

Figure 16: Diverse grasping strategies from different approach directions in the real world.

Figure 17: Representative real-world grasps across Leap and Allegro hands.

Figure 18: Closed-loop reactive grasping in dynamic environments. In the scenario, a human holds and continuously moves the target object, and our framework generates updated grasp poses based on real-time point clouds captured by an RGB-D RealSense D435 camera. The robot’s approaching motions from initial positions are also continuously guided by the generated poses.

Explore similar work

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
DexGrasp-Zero: A Morphology-Aligned Policy for Zero-Shot Cross-Embodiment Dexterous Grasping
Mar 17, 2026Yuliang Wu, Yanhan Lin, WengKit Lao +4Robotic Grasping Zero-Shot Robotic Manipulation
School of Computer Science and Engineering, Sun Yat-sen University, China
GraspGraphNet: Graph-Structured Multi-Embodiment Dexterous Grasp Generation
Jul 13, 2026Yeonseo Lee, Taeyeop Lee, Hyosup Shin +2Robotic Grasping Graph Neural Networks
Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea.

Sep 30, 2026J/K move · Enter open · S save

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

Bolin Zou, Wenlong Dong, Mu Ai, Chao Tang+3

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

Robotic Grasping Cross-Embodiment Robot Learning Robotic Manipulation

Open paper AskPDF arXiv

Cross-embodiment dexterous grasp generation remains challenging because robotic hands differ substantially in geometry, topology, and kinematics. Existing approaches often lack explicit correspondences between structurally different hand regions that play similar functional roles in a grasp, a concept we refer to as functional correspondence. Consequently, their models tend to learn hand-specific interaction patterns rather than transferable grasp knowledge, limiting generalization to unseen hands. To address this limitation, we introduce FunCo-Grasp, which establishes functional correspondences across heterogeneous hand embodiments. Specifically, Functional Part Alignment aligns each hand to a canonical functional schema by mapping physical links to shared functional parts according to their grasping roles, while Canonical Frame Alignment expresses these parts in canonical local frames. These two alignments provide a consistent representation for inter-part and hand-object interactions, allowing the model to learn transferable grasp knowledge across hands. Conditioned on the aligned hand representation and object geometry, a diffusion model generates the target spatial arrangement of the functional parts, which are then converted into an executable joint configuration. Adapting FunCo-Grasp to an unseen hand requires only its geometric and kinematic models and a one-time lightweight functional annotation, without target-hand grasp data, fine-tuning, or learned retargeting. In simulation on held-out objects from the filtered CMapDataset, we achieves average success rates of 92.40% on three seen hands and 74.02% on four unseen hands. In real-world experiments, the same model achieves an overall success rate of 76.00% on two unseen hands without additional training or fine-tuning. These results demonstrate the effectiveness of FunCo-Grasp in transferring grasp knowledge to unseen hands.

Abstract

Explore similar work

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

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

GraspGraphNet: Graph-Structured Multi-Embodiment Dexterous Grasp Generation