Differences in hand geometry, joint parameterization, and kinematic topology pose challenges for cross-embodiment dexterous grasp generation. However, spatial relations among links alone are insufficient to fully capture the hand's kinematic topology. We propose DexGraspFlow, a conditional link-level SE(3) flow-matching framework integrating geometric relations and kinematic topology. The framework encodes local object geometry, link geometry, and current poses. A dual-stream DiT models link--link coordination and link--object interaction, while attention biases derived from kinematic-tree distances provide explicit structural information for modeling evolving geometric interactions. The model jointly learns translational and rotational velocity fields, integrates them to generate link poses, and recovers joint configurations through inverse kinematics. Additionally, inference-time palm-orientation guidance enables controllable grasp generation without retraining. Across three robotic hands and 30 object--hand pairs, DexGraspFlow achieves a 95.33% average simulation success rate, with an average generation latency, including inverse kinematics, approximately 49% lower than that of a state-of-the-art baseline. Further experiments demonstrate zero-shot adaptability to seven finger-removal variants and palm-orientation control. A model trained on demonstrations for the deployed hand achieves a 91% success rate over 100 real-robot grasping trials, supporting the framework's applicability to physical grasp execution.
Figures & tables
Figure 1: Overview of DexGraspFlow. (a) Geometry and state encoding: Object points and per-link hand point clouds are encoded into object-patch tokens and static link tokens, while the kinematic tree defines the topological distance matrix Dtopo . Static link tokens are concatenated with dynamic link-pose features to form link-state tokens. (b) Geometry- and topology-aware dual-stream DiT: Link–object cross-attention and link–link self-attention process state-dependent geometric relation tokens, with kinematic-tree distance biases injected into the link–link stream, to jointly predict translational and rotational velocity fields vθ=(vθx,vθω) with per-link outputs (u^ix,u^iω) . (c) Flow to executable grasp: Continuous velocity fields are numerically integrated from prior distributions to generate target link poses, which are mapped into executable joint configurations via batched inverse kinematics.
Figure 2: Geometry-aware dual-stream attention in the DiT. The left branch performs link–link self-attention with physical-relation and topology biases. The right branch performs link–object cross-attention from link queries to object patch tokens with a gated physical-relation bias; topology bias is not used in this branch. In both branches, relation-aware pairwise features are projected to value messages before weighted sum.
Figure 3: Illustration of topology-distance construction from the URDF-derived physical-link graph. Each valid hand link is treated as a vertex in the undirected graph, and the distance between two links is the shortest-path length. The right panel shows an illustrative 3×3 submatrix of Dtopo ; the colors indicate schematic link categories.
Figure 4: Inference-time palm-orientation guidance. The parallelogram depicts u~t,pω=u^t,pω+Δut,pω , where Δut,pω=s(t)λcRt,p⊤ωtc . Solid and dashed curves illustrate sampling with and without the correction from the same initial state, with endpoints shown before inverse kinematics. “Small POE” indicates a small orientation error relative to the target Rpc . Only the palm trajectory is shown, although all valid links receive the correction.
Method
Success Rate (%)
Efficiency (s) ↓
Barrett
Allegro
ShadowHand
Avg.
Barrett
Allegro
ShadowHand
Avg.
DFC ( Liu et al., 2021 )
86.30
76.21
58.80
73.77
>1800
>1800
>1800
>1800
GenDexGrasp ( Li et al., 2023 )
57.40
43.10
42.30
47.60
12.933
18.347
16.540
15.940
DRO-Grasp ( Wei et al., 2025 )
86.60
91.20
81.60
86.47
0.523
0.485
0.856
0.621
TRO-Grasp ( Fei et al., 2026 )
92.40
93.30
98.10
94.60
0.254
0.247
0.269
0.257
DexGraspFlow (Ours)
92.10
98.40
95.50
95.33
0.126
0.134
0.136
0.132
Table 1: Comparison of unconditioned grasp success rates and single-sample end-to-end generation latency across three hand embodiments.
Figure 5: Representative flow-matching trajectories for three hand–object pairs. The rows correspond to Allegro, Barrett, and ShadowHand; the associated objects are a YCB tomato soup can, a ContactDB door knob, and a ContactDB apple, respectively. Blue images show selected states along the trajectory from t=0 to t=1 , while the final column shows the executable grasp obtained after Pyroki inverse kinematics.
Figure 6: Intact and topology-modified hand embodiments used for zero-shot evaluation. The arrows indicate the transition from each intact hand to the corresponding missing-finger variants. Here, I, M, R, and L denote the index, middle, ring, and little fingers, respectively.
Allegro
ShadowHand
w/o I
w/o M
w/o R
w/o I
w/o M
w/o R
w/o L
Figure 7: Representative zero-shot grasp examples on topology-modified hands. The columns correspond to Allegro with the index, middle, or ring finger removed and ShadowHand with the index, middle, ring, or little finger removed. From top to bottom, the rows show a baseball, pear, potted meat can, and water bottle.
Hand variant
DRO-Grasp
TRO-Grasp
Ours
Allegro w/o I
53.60
11.80
97.00
Allegro w/o M
61.40
45.80
75.10
Allegro w/o R
49.90
65.40
94.50
ShadowHand w/o I
47.80
2.50
89.00
ShadowHand w/o M
55.40
4.00
90.60
ShadowHand w/o R
57.90
7.20
82.80
Table 2: Zero-shot unconditioned grasp success rates (%) on topology-modified hands. I, M, R, and L denote the index, middle, ring, and little fingers, respectively.
Setting
Success Rate (%)
POE (deg.)
TRO
Ours
Δ SR ↑
TRO
Ours
ΔPOE
Hand-wise results
Allegro
80.42
95.17
+14.75
38.25
15.18
+23.07
Barrett
77.17
83.50
+6.33
41.40
9.12
+32.28
ShadowHand
63.67
81.42
+17.75
35.51
17.46
+18.05
Average
73.75
86.69
+12.94
38.38
13.92
+24.46
Table 3: Six-direction conditioned grasp generation results. TRO and Ours refer to TRO-Grasp and DexGraspFlow, respectively. SR and POE denote success rate and final palm-orientation error, respectively. Δ SR denotes the success-rate difference in percentage points, while ΔPOE=POETRO−POEOurs denotes the absolute POE reduction in degrees; positive values favor DexGraspFlow.
Figure 8: Palm-orientation-conditioned grasp examples for a representative held-out rubber duck. The rows correspond to Allegro, Barrett, and ShadowHand, while the columns show the front, back, left, right, up, and down target directions. Blue and red arrows denote the target and post-IK actual palm approach directions, respectively.
Bias Configuration
Barrett
Allegro
ShadowHand
Avg.
No Attention Bias
86.10
98.30
75.50
86.63
Geometry Bias Only
87.50
99.50
82.10
89.70
Topology Bias Only
88.70
99.50
90.00
92.73
Geometry + Topology
92.10
98.40
95.50
95.33
Table 4: Ablation of structural attention biases under unconditioned generation. All values denote success rate (%).
Steps
Success Rate (%)
Time (s) ↓
Barrett
Allegro
ShadowHand
Avg.
1
50.90
60.30
76.20
62.47
0.0130
5
93.40
98.10
92.60
94.70
0.0205
10
92.10
98.40
95.50
95.33
0.0257
15
91.70
97.80
93.50
94.33
0.0321
20
92.90
97.90
94.80
95.20
0.0361
Table 5: Effect of ODE integration steps on unconditioned grasp success rate and batch-amortized latency.
Rotational Guidance λc
Avg. SR (%)
Avg. POE (deg.)
5
89.08
22.135
10
87.33
17.552
20
86.69
13.921
30
83.50
12.225
Table 6: Sensitivity to rotational guidance multiplier λc under six-direction conditioning. Bold indicates default setting.
Figure 9: Real-robot experimental setup. (a) Robot workspace view. (b) Close-up of the hand–object interaction.
Figure 10: Real-robot grasping examples generated by DexGraspFlow for ten objects unseen during training using the Linker Hand L20Lite.
Object
Success/Trials
Object
Success/Trials
Can
7/10
Coil
9/10
Toy
10/10
Foam Box Lid
10/10
McCafe
10/10
Orange
9/10
Packaging Box
10/10
Pill Bottle
10/10
Pomegranate
8/10
Red Tape
8/10
Overall
91/100
Table 7: Real-robot grasping results on 10 unseen objects.
Generating high-quality dexterous grasps remains challenging for learning-based methods, which often depend on carefully tuned contact losses or costly contact-based test-time refinement. We present KPGrasp, a flow-matching framework that learns dexterous grasp priors from large-scale data rather than relying on contact losses or contact-based test-time refinement. KPGrasp couples an all-Euclidean 3D hand-keypoint parameterization with a simple yet scalable Transformer flow model. The parameterization avoids the drawbacks of the conventional mixed SE(3) pose and joint-angle output space, expresses grasps in the same frame as the object point cloud, and thus enables native spatial reasoning; the Transformer flow model is trained with only the standard flow-matching loss and scales effectively with data, model capacity, and batch size. Experiments demonstrate state-of-the-art performance on two simulation benchmarks. On the Dexonomy benchmark, it reaches a 76.3% grasp success rate, improving over the strongest directly comparable baseline by 47.4% while reducing penetration depth to 2.4 mm. The same model also achieves the best average performance on the DexGrasp Anything benchmark without fine-tuning. For batched inference, KPGrasp requires only 0.032 s per grasp. Finally, real-world experiments on 20 diverse objects demonstrate that the pipeline can be deployed in a real-world setup. Project Page: https://yuansen-huang.github.io/KPGrasp-page/
Yuansen Huang, Jiayi Chen, Haoran Liu +6
Peking University · Galbot · Xi’an Jiaotong University +1
Most learned dexterous grasp generators relegate contact forces to a downstream verification step, so a kinematically-plausible pose can still violate the conditions for a stable physical grasp. We address this with EquiDexFlow, an SE(3)-equivariant flow-matching model that jointly predicts wrist pose, joint angles, fingertip contacts, surface normals, and contact forces from an object point cloud. Our architecture projects contacts onto the object surface and forces into the Coulomb friction cone by construction, so placement and friction compliance hold without loss penalties. We prove end-to-end SE(3) equivariance and verify it empirically over 200 rotations, with wrist residuals below 0.04∘ and exactly zero joint deviation. Trained on 8,100 force-closure grasps across 81 objects for the 16-DoF Allegro Hand, our model achieves zero friction violations, the best composite score, and the lowest wrench residual among all ablation variants. We retarget decoded fingertip contacts to a 16-DoF LEAP Hand via per-finger inverse kinematics, and our hardware-feasible refinement places every joint at least 5% inside its actuator envelope while preserving wrench balance. On the physical robot, retargeted EquiDexFlow-decoded grasps complete open-loop pick-and-hold trials on all six test objects, with every asymmetric object succeeding at both the canonical pose and a 120∘ co-rotation. Videos, code, and checkpoints are available at https://equidexflow.github.io.
Clinton Enwerem, John S. Baras, Calin Belta
Institute for Systems Research, University of Maryland, College Park, MD, U.S.A.
Dexterous grasp generation across robot hands is challenging because hands differ in kinematic topology, actuation dimensions, and native command spaces. We introduce GraspGraphNet, a topology-aware grasp generation framework that represents each hand as a URDF-derived kinematic graph and directly generates executable palm poses and joint configurations. GraspGraphNet combines hierarchical object surface encoding, differentiable forward kinematics, and dynamic world-edge message passing to model evolving robot-object interactions. It applies conditional flow matching directly in executable palm-pose and joint-state space, avoiding post-processing optimization, inverse kinematics, and retargeting. Using a shared model trained on Barrett Hand, Allegro Hand, and Shadow Hand, GraspGraphNet achieves an average success rate of 83.48% with 40ms inference time per grasp on a 40-object benchmark. Without retraining, the same model achieves 72.70% success on controlled finger-removal variants, demonstrating robustness to hand-topology variations. These results suggest that graph-structured hand representations can effectively support dexterous grasp generation across robot hands with different kinematic structures. Project: https://lysees.github.io/graspgraphnet-page
Yeonseo Lee, Taeyeop Lee, Hyosup Shin +2
Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea.