DexGraspFlow: Topology-Aware SE(3) Flow Matching for Cross-Embodiment Dexterous Grasp Generation

Abstract

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 $\mathrm{SE}(3)$ 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 its practical applicability.

1. Motivation

Cross-embodiment dexterous grasp synthesis faces two fundamental challenges:

(a) Kinematic Topology Awareness: Preserving Skeleton Feasibility

The Problem: Spatial Euclidean distance ignores the robot's kinematic skeleton. Denoising in unconstrained 3D space causes severe joint hyperextensions, finger crossovers, and kinematic invalidity.

Our Solution: We inject an invariant kinematic hop-count matrix $\mathbf{D}_{\mathrm{topo}}$ as a structural attention bias, constraining continuous flow matching to evolve strictly along the robot's physical kinematic manifold.

**Kinematic Topology & Hop-Count Matrix $\mathbf{D}_{\mathrm{topo}}$.** Decoupling intrinsic kinematic connectivity from extrinsic spatial pose across heterogeneous hands.

(b) User-Decided Controllable Generation: Meeting Downstream Task Intent

The Problem: Practical robotic manipulation requires grasp poses aligned with downstream task intents (e.g., pouring from the side, bin-picking from above). Existing cross-embodiment models rely on blind random sampling or require expensive model retraining for every new condition.

Our Solution: We formulate training-free inference-time geodesic guidance on $\mathrm{SO}(3)$. Given user-specified approach directions, closed-form rotational velocity corrections $\Delta\mathbf{u}_{t,p}^{\omega} = s(t)\lambda_c\mathbf{R}_{t,p}^{\top}\boldsymbol{\omega}_t^c$ are injected into the predicted velocity field ($\tilde{\mathbf{u}}_{t,p}^{\omega} = \hat{\mathbf{u}}_{t,p}^{\omega} + \Delta\mathbf{u}_{t,p}^{\omega}$) at each integration step, smoothly steering the palm toward target poses while preserving physics-compliant contacts.

**Inference-Time Palm-Orientation Guidance on $\mathrm{SO}(3)$.** Closed-form geodesic velocity correction $\Delta\mathbf{u}_{t,p}^\omega$ steers the predicted flow $\hat{\mathbf{u}}_{t,p}^\omega$ toward user-specified target orientation $\mathbf{R}_p^c$.

2. Method Overview: Dual-Stream DiT & Geodesic Flow Matching

DexGraspFlow formulates multi-link dexterous grasp generation as conditional $\mathrm{SE}(3)$ flow matching on the rigid-body Lie group, parameterized by a geometry- and topology-aware dual-stream Transformer:

Overview of DexGraspFlow Framework — **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 $\mathbf{D}_{\mathrm{topo}}$. 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 $v_\theta = (v_\theta^x, v_\theta^\omega)$ with per-link outputs $(\hat{\mathbf{u}}_i^x, \hat{\mathbf{u}}_i^\omega)$. **(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.

1. Kinematic & Geometric Encoding

The target object point cloud is partitioned into local patch tokens, while each robot link is discretized into surface points and encoded into static link tokens. Concatenating with dynamic rigid poses yields link-state tokens $s_{i,t}$.

2. Dual-Stream Structural Attention

Link--Link Stream: Injects topological distance biases $\mathbf{B}_{\mathrm{topo}}$ derived from $\mathbf{D}_{\mathrm{topo}}$ to preserve skeleton feasibility.
Link--Object Stream: Injects physical-relation biases $\mathbf{B}_{\mathrm{geom}}$ to accurately model contact geometry.

3. Geodesic Guidance & Batched IK

The network jointly learns translational and angular velocity fields $(v_\theta^x, v_\theta^\omega)$. Training-free geodesic guidance smoothly steers palm orientation during integration, followed by batched GPU Pyroki IK recovery.

Geometry-Aware Dual-Stream Attention Architecture — **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.

3. Continuous Flow Trajectories Across Embodiments

Visualizing the continuous numerical integration of the learned velocity fields from prior Gaussian distributions ($t=0$) through intermediate flow steps ($t=0.2, 0.5, 0.8, 1.0$), followed by batched GPU inverse kinematics:

Allegro Hand 4-finger, 16-DoF × YCB Tomato Soup Can

Flow Matching Steps → Executable Configuration

$t = 0.0$ (Prior)

$t = 0.2$

$t = 0.5$

$t = 0.8$

$t = 1.0$ (Link Poses)

Executed (IK)

Barrett Hand 3-finger, 4-DoF × ContactDB Door Knob

Flow Matching Steps → Executable Configuration

$t = 0.0$ (Prior)

$t = 0.2$

$t = 0.5$

$t = 0.8$

$t = 1.0$ (Link Poses)

Executed (IK)

ShadowHand 5-finger anthropomorphic, 24-DoF × ContactDB Apple

Flow Matching Steps → Executable Configuration

$t = 0.0$ (Prior)

$t = 0.2$

$t = 0.5$

$t = 0.8$

$t = 1.0$ (Link Poses)

Executed (IK)

4. Interactive 3D Visualizer

Unconditioned Grasp Generation

Barrett

Allegro

Shadowhand

Object Name

Grasp ID: 0

Drag to rotate

Scroll to zoom

Zero-Shot Topology Generalization (Finger-Removal Hands)

Zero-shot adaptability across 7 missing-finger variants without retraining or fine-tuning

Allegro Variants

ShadowHand Variants

Object Name

Grasp ID: 0

Drag to rotate

Scroll to zoom

Palm-Orientation-Conditioned Grasp Generation

Blue Arrow: Target Approach Direction Red Arrow: Post-IK Actual Palm Direction

POE: --°

Barrett

POE: --°

Allegro

POE: --°

Shadowhand

Object Name

1. Target Palm Direction

Grasp ID: 0

🔵 Blue: target approach

🔴 Red: actual palm approach

5. Experimental Results & Benchmarks

Table I: Unconditioned Grasp Generation Across Embodiments

Evaluated across 30 held-out object--hand pairs (3,000 simulated trials). Single-sample latency includes both ODE integration and pre-compiled Pyroki IK recovery.

Method	Success Rate (%) ↑				Single-Sample Latency (s) ↓
Method	Barrett	Allegro	ShadowHand	Avg.	Barrett	Allegro	ShadowHand	Avg.
DFC (Optimization)	86.30	76.21	58.80	73.77	>1800	>1800	>1800	>1800
GenDexGrasp	57.40	43.10	42.30	47.60	12.933	18.347	16.540	15.940
DRO-Grasp	86.60	91.20	81.60	86.47	0.523	0.485	0.856	0.621
TRO-Grasp	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 II: Zero-Shot Generalization to Topology-Modified (Finger-Removal) Hands

Models are trained solely on intact hands and tested directly on 7 amputated variants without fine-tuning. I, M, R, L denote index, middle, ring, and little fingers.

Method	Allegro (4-Finger)			ShadowHand (5-Finger)				Overall Avg.
Method	w/o Index	w/o Middle	w/o Ring	w/o Index	w/o Middle	w/o Ring	w/o Little	Overall Avg.
DRO-Grasp	53.60	61.40	49.90	47.80	55.40	57.90	45.80	53.11
TRO-Grasp	11.80	45.80	65.40	2.50	4.00	7.20	6.60	20.47
DexGraspFlow (Ours)	97.00	75.10	94.50	89.00	90.60	82.80	92.20	88.74

Intact and Topology-Modified Hand Embodiments — **Intact and 7 topology-modified hand embodiments** evaluated in zero-shot transfer tests. DexGraspFlow maintains 88.74% grasp success without seeing amputated hands during training.

Zero-Shot Grasp Synthesis on Amputated Hand Variants Across 4 Held-Out Objects

Representative grasps generated on 7 topology-modified hands without retraining. Rows: Baseball, Pear, Potted Meat Can, Water Bottle. Columns: Allegro (w/o Index, w/o Middle, w/o Ring) and ShadowHand (w/o Index, w/o Middle, w/o Ring, w/o Little). Even with missing fingers, DexGraspFlow reorganizes remaining kinematic chains to establish force closure.

Object	Allegro Hand (Amputated)			ShadowHand (Amputated)
	w/o Index	w/o Middle	w/o Ring	w/o Index	w/o Middle	w/o Ring	w/o Little
Baseball
Pear
Meat Can
Water Bottle

Palm-Orientation-Conditioned Grasps on Held-Out Object (Rubber Duck) Across 6 Canonical Directions

Figure 7 from paper: Grasps generated across six canonical approach directions (+x: Front, -x: Back, +y: Left, -y: Right, +z: Up, -z: Down) under inference-time rotational guidance without reference poses or retraining. Blue and red arrows inside the translucent object indicate the requested target palm approach and post-IK actual palm approach directions, respectively.

Allegro Hand (4-Finger, 16-DoF)

Front (+x)

Back (-x)

Left (+y)

Right (-y)

Up (+z)

Down (-z)

Barrett Hand (3-Finger, 4-DoF)

Front (+x)

Back (-x)

Left (+y)

Right (-y)

Up (+z)

Down (-z)

ShadowHand (5-Finger Anthropomorphic, 24-DoF)

Front (+x)

Back (-x)

Left (+y)

Right (-y)

Up (+z)

Down (-z)

Table III: Six-Direction Conditioned Grasp Generation Results

SR and POE denote grasp success rate and final palm-orientation error, respectively. ΔSR denotes the success-rate difference in percentage points, while ΔPOE = POE_TRO - POE_Ours denotes the absolute POE reduction in degrees; positive values favor DexGraspFlow.

Setting	Success Rate (%)			Palm-Orientation Error (POE, deg.) ↓
Setting	TRO	Ours	ΔSR ↑	TRO	Ours	ΔPOE (Drop) ↑
Hand-wise results
Allegro Hand	80.42	95.17	+14.75	38.25	15.18	+23.07°
Barrett Hand	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°
Direction-wise results averaged across hand embodiments
Front (+x)	71.33	90.50	+19.17	39.44	13.95	+25.49°
Back (-x)	73.67	82.67	+9.00	37.95	14.16	+23.79°
Left (+y)	73.67	88.00	+14.33	37.88	14.24	+23.64°
Right (-y)	75.50	83.17	+7.67	37.66	14.64	+23.02°
Up (+z)	80.00	91.50	+11.50	32.37	13.24	+19.13°
Down (-z)	68.33	84.33	+16.00	45.00	13.29	+31.71°

Ablation Studies & In-Depth Analysis: Comprehensive empirical ablations on structural attention biases, ODE integration dynamics, and geodesic guidance scale sensitivity are detailed in the paper.

6. Real-World Robot Deployment (91% Success Rate)

To evaluate real-world feasibility, DexGraspFlow was deployed on a physical robotic workstation consisting of a ROKAE industrial manipulator, a Linker Hand L20Lite dexterous five-finger hand, and an Intel RealSense D435i camera. The system evaluated zero-shot grasp transfer across 10 physical objects never observed during training.

Workspace Setup (ROKAE + RealSense + Worktable)

Close-up Hand Interaction (Linker Hand L20Lite)

100 Physical Trials Benchmark: Across 10 unseen daily objects (10 independent trials per object, total 100 trials), DexGraspFlow achieved an overall success rate of 91/100 (91.0%), firmly validating its zero-shot physical transfer and robust contact synthesis.

Table VII: Real-Robot Grasping Results Across 10 Unseen Physical Objects

Zero-shot physical grasping trials on Linker Hand L20Lite across 10 unmodeled everyday objects (10 independent trials each).

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 Real-Robot Success		91 / 100 (91.0%)

Real-Robot Grasping Trials on 10 Unseen Physical Objects

To rigorously evaluate zero-shot physical transfer, the test benchmark features 10 diverse physical objects unseen during model training. These objects cover a wide spectrum of geometric topologies and challenging contact mechanics, including thick lightweight foam slabs (Foam Box Lid), delicate annular/spiral contours (Coil, Red Tape), low-friction smooth metallic cylinders (Can), deformable compliant shells (McCafe, Toy), bulky volumetric cuboids (Packaging Box), and heavy asymmetric natural fruits (Orange, Pomegranate).

Interactive Video Gallery: Live physical trials using the ROKAE manipulator and Linker Hand L20Lite (91/100 overall success). Click any video to play.

Can

Low-Friction Cylinder

7/10

Stable multi-finger enclosure resisting slip on smooth aluminum shell.

Coil

Thin Spiral Contour

9/10

Precision fingertip pinch on delicate curved contours.

Toy

Compliant Soft Body

10/10

Compliant multi-finger adaptive wrap conforming to non-rigid contours.

Foam Box Lid

Thick Foam Slab

10/10

Stable opposed grasp across compliant, thick foam sidewalls.

McCafe

Tapered Conical Shell

10/10

Conical power grasp with balanced friction on paper wall.

Orange

Smooth Spherical Fruit

9/10

Equilateral five-finger spherical cage securing rolling sphere.

Packaging Box

Volumetric Cuboid

10/10

Wide-span opposed grasp across bulky planar faces.

Pill Bottle

Rigid Plastic Cylinder

10/10

Precision lateral pinch and rigid vertical lift.

Pomegranate

Asymmetric Curved Fruit

8/10

Adaptive force-closure grip resisting mass eccentricity.

Red Tape

Annular Ring Geometry

8/10

Dual-finger inner rim hook and exterior palm support.

BibTeX

@article{zhou2026dexgraspflow,
  title={DexGraspFlow: Topology-Aware SE(3) Flow Matching for Cross-Embodiment Dexterous Grasp Generation},
  author={Zhou, Bin and Ma, Boyu and Zhou, Xinyu and Dong, Jiahua and Zhou, Chuhao and Liu, Yizhou and Wan, Hao and Yang, Jianfei},
  journal={arXiv preprint},
  year={2026}
}