T(R,O) Grasp++: Cross-Embodiment Dexterous Grasping in Cluttered Environments at Scale

Xin Fei; Zhixuan Xu; Huaicong Fang; Fangjun Chen; Zhaoyu Mu; Rui Yang; Zhexi Luo; Lin Shao

cs.ROOct 18, 2026

T(R,O) Grasp++: Cross-Embodiment Dexterous Grasping in Cluttered Environments at Scale

Authors: Xin Fei, Zhixuan Xu, Huaicong Fang, Fangjun Chen, Zhaoyu Mu, Rui Yang, Zhexi Luo, Lin Shao

Organizations: Department of Computer Science, National University of Singapore · RoboScience · College of Control Science and Engineering, Zhejiang University, China · Department of Electronic and Electrical Engineering, Southern University of Science and Technology, China ·

Abstract

Dexterous grasping in cluttered scenes remains a central challenge in robotic manipulation, as it requires high-dimensional hand coordination under complex object-environment constraints. To address this problem, we introduce one of the largest simulation benchmarks to date, comprising around 0.5B grasp annotations across eight dexterous hands and approximately 50K objects. Furthermore, we propose T(R,O) Grasp++, a scalable diffusion-based framework for cross-embodiment dexterous grasp synthesis in cluttered environments. Extensive experiments demonstrate that our model significantly outperforms existing baselines, achieving an average success rate of 88% across eight dexterous embodiments. Meanwhile, our pipeline achieves only 0.3s inference latency while improving inference sample efficiency by over 10 times.

Figures & tables

Figure 1: Our benchmark comprises around 0.5 billion grasp annotations in cluttered scenes across diverse objects and dexterous hands. $\mathcal{T}(R,O)$ Grasp++ models object-environment-hand spatial transformations with $\mathcal{T}(R,O)$ Graph++, achieving better performance, faster inference, and more efficient sampling.

Dataset	Obj. Num.	Hand Num.	Grasp Num.	Grasps per Obj.-Hand	Cluttered Scenes	Hierarchical Annotations	Hand Mount
GraspXL [ 36 ]	$>$ 500 K	4	$>$ 10 M	N.A.	✗	✗	✗
Dex1B [ 35 ]	6 K	3	1 B	N.A.	✗	✗	✗
MultiGripperGrasp [ 3 ]	345	11	30.4 M	$\sim$ 8K	✗	✗	✗
ContactDexNet [ 39 ]	$>$ 1.5 K	1	N.A.	N.A.	✓	✗	✗
DexGraspNet2 [ 37 ]	1,319	1	427 M	$\sim$ 324K	✓	✗	✗
Ours	53,675	8	444 M	$\sim$ 1K	✓	✓	✓

Table 1: Comparison of Dexterous Grasping Dataset.

Figure 2: Overview of $\mathcal{T(R,O)}$ Grasp++ pipeline.

Method	Success Rate (%) $\uparrow$ / Diversity (rad.) $\uparrow$
Method	Allegro	Ability	Barrett	DClaw	LEAP	Schunk	Shadow	XHand	Avg.
AnyDexGrasp [ 9 ]	54.44/ 0.387	–	–	–	–	–	–	–	54.44/0.387
DexGraspNet2 [ 37 ]	–	–	–	–	52.09/0.303	–	–	–	52.09/0.303
Ours (Single-trial)	88.79 /0.369	88.28 / 0.509	87.24 / 0.470	89.11 / 0.393	85.93 / 0.404	90.39 / 0.389	87.87 / 0.354	86.03 / 0.418	87.96 / 0.413

Table 2: Overall success rate and diversity comparison across different dexterous hands.

Table 5

Table 7: Real-world experiment results on XHand and LEAP Hand.

Hand	Prediction	Planning	Collision	Instability	Total
Wuji	13/414	5/414	12/414	44/414	74/414
XHand	22/548	14/548	27/548	29/548	92/548

Table 8: Failure mode analysis in the real world.

Figure 8: Score distribution of our proposed benchmark. Each grasp in a cluttered scene is annotated with a difficulty score derived from object geometry, obstacle proximity, and spatial enclosure.

Translation Div.	Rotation Div.	Norm. Joint Div.	Direction Ent.
0.360 m	125.97 deg	5.97	5.96 / 6

Table 9: Diversity analysis of

\mathcal{T}(R,O)

++ dataset.

Model Params.	Layers Num.	Attn. Heads Num.	Node Dim.	Edge Dim.	Output Layer Dim.
10M	4	4	256	256	[128, 64]
50M	8	6	384	384	[256, 128]
0.2B	8	12	768	768	[512, 256, 128]
1B	12	24	1536	1536	[512, 256, 128]

Table 10: Architecture configurations of diffusion models with different parameter scales.

Similarity	Link Alignment $S_{\text{link}}$		Joint Overlap $S_{\text{joint}}$
Similarity	SR (%) $\uparrow$	Diversity $\uparrow$	SR (%) $\uparrow$	Diversity $\uparrow$
1.0	86.03	0.418	86.03	0.418
0.9	85.35	0.413	84.04	0.411
0.8	84.56	0.411	82.93	0.400
0.7	84.51	0.412	81.73	0.390

Table 11: Cross-embodiment zero-shot performance.

Robotic Grasping

Robotic grasping studies how robots perceive, plan, and execute stable, task-appropriate contacts with objects across varied shapes and conditions.

See all 302 papers

OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes
Oct 8, 2026Naiyu Fang, Zhongjin Luo, Yuxin Mo +7Robotic Grasping Robot Manipulation Benchmarks
ACE Robotics · CUHK, MMLab · CPII under InnoHK +4
AdaClearGrasp: Learning Adaptive Clearing for Zero-Shot Robust Dexterous Grasping in Densely Cluttered Environments
Mar 11, 2026Zixuan Chen, Wenquan Zhang, Jing Fang +7Robotic Grasping Robotics
Nanjing University · Nanjing University of Posts and Telecommunications · National University of Singapore
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

Oct 8, 2026J/K move · Enter open · S save

OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes

Naiyu Fang, Zhongjin Luo, Yuxin Mo, Siyuan Huang+6

ACE Robotics · CUHK, MMLab · CPII under InnoHK+4

Robotic Grasping Robot Manipulation Benchmarks Synthetic Data Generation for Robotics Generalization in Robotic Manipulation

Open paper AskPDF arXiv

Dexterous grasping is the foundational primitive in embodied AI, demanding massive data to train robust models. As real-world data collection is expensive, simulation has become the mainstream paradigm. Yet, while cluttered scenes best reflect real-world applications, learning to grasp within them is bottlenecked by a critical scarcity of large-scale data. To resolve this, we curate high-quality 3D objects and supporting bases, proposing a scalable seed-and-filter strategy that bypasses sluggish scene-level optimization. This yields an unprecedented benchmark comprising over 2.6 million scenes and 0.4B scene-specific grasp ground truths, featuring diverse realistic layouts paired with rich semantic and geometric observations. Furthermore, we introduce the OmniDex model to overcome the grasp multimodality and last-millimeter precision errors plaguing current generative models. By coupling Soft Winner-Takes-All learning with human-inspired physical constraints during training, and utilizing physics-driven ranking, our approach achieves robust dexterous grasping without the latency of post-optimization. Experimental results show that OmniDex model achieves state-of-the-art performance and strong generalization across diverse scenes, views, and unseen objects.

Abstract

Robotic Grasping

Explore similar work

OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes

AdaClearGrasp: Learning Adaptive Clearing for Zero-Shot Robust Dexterous Grasping in Densely Cluttered Environments

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