CoRL 2026 · Austin, TX

XDex: Learning Cross-Embodiment Dexterous Grasping with 1000 Hands

Keyi Wang* Jianglong Ye* Chengjing Yuan Xiaolong Wang

UC San Diego · *Equal contribution

Poster Video BibTeX

A grid of grasps. Each row is one grasp on one object transferred across hands, and each column is one hand, color-coded, on different objects.
Each row shows the same grasp on the same object across different hands. Each column is one hand (color-coded) grasping different objects.
1,000
robot & human hands
7k
object scenes
200k
source grasps
77.5%
real-world success on an unseen hand

Abstract

Generalizing dexterous grasping across diverse robot hands remains a fundamental challenge due to differences in geometry, topology, and kinematics. Inspired by recent advances in scaling up robot data, we show that scaling both the diversity and quantity of hand embodiments significantly improves generalization to unseen morphologies. We present XDex, a unified cross-embodiment grasp synthesis framework trained on the largest and most diverse multi-hand grasp dataset to date, spanning 1,000 heterogeneous robotic hands and 7k object scenes. XDex couples an embodiment encoder that jointly captures geometric and topological features with a three-way disentangled representation of object geometry, hand embodiment, and grasping pose. It outperforms prior methods in grasp quality and diversity and learning-based methods in cross-hand consistency, and generalizes to unseen hands in real-world settings.

Method

XDex separates a grasp into three factors, so a pose found on one hand can be decoded on another.

What to graspObject geometry: a local SpConv feature at one object point
How to grasp itGrasping pose: the CVAE latent z
With whatHand embodiment: per-link tokens from geometry and URDF topology
Left: embodiment generation from original hands to topology and geometry variants. Right: model architecture with a PointNet and Transformer embodiment encoder and a Transformer CVAE conditioned on object features.
Left: 1,000 hands generated by varying geometry and topology. Right: XDex encodes hand geometry and topology together with object features, and a Transformer CVAE generates joint poses.

Self-improving data loop

Pipeline from source grasps to retargeted seed pairs to the XDex model to grasps on all hands, filtered in simulation and fed back into training.
  1. Optimize 200k grasps for one source hand (Inspire) on 7k objects.
  2. Retarget to 6 sampled hands per object to form seed pairs.
  3. Train an initial XDex model on the seed set.
  4. Generate on all 1,000 hands, filter by IK and simulation, then retrain.

Consistency losses

Left: retargeting loss between two hands decoded from the same latent. Right: embedding loss aligning the encoded features of a paired grasp.
Left: the retargeting loss penalizes per-joint differences between two hands decoded from the same z and object point. Right: the embedding loss aligns the latent means of a paired grasp.

Grasp Synthesis Results

Evaluated on 500 held-out scenes in ManiSkill. Ability and Schunk are the unseen hands.

QualityConsistencyDiversity
MethodSR ↑Pen (cm) ↓Pos Diff (cm) ↓Vec Diff (cm) ↓H mean (bits) ↑H std (bits) ↓

Bold marks the best value in each column. Retargeting has the lowest position differences because it directly minimizes keypoint discrepancy, without accounting for object-aware grasp quality. D(R,O)* is retrained on our tabletop data, and its seen-hand metrics average over Shadow and Allegro, the seen hands its pre-trained encoder covers.

Disentangled grasp poses

Each row uses the same object point and latent z, decoded on different hands, and the grasps stay consistent. Each column varies z on one hand, and the poses stay diverse.

Removing paired data hurts unseen hands most: success falls from 0.59 to 0.50, and position difference rises from 1.84 to 2.72 cm.

Generalizing to new morphologies

Each pair decodes the same grasp on an original hand (left) and a modified one (right): a fingertip removed, an extra distal link, a scaled link or whole hand, or a locked joint.

Embodiment Scaling Law

Scaling hands vs. scaling data

More data alone does not buy generalization. Diversity of hands does.

Unseen-hand success versus number of seen embodiments. Joint scaling reaches 59 percent at 1000 hands, and hands-only scaling reaches about 40 percent.
Scaling hands with fixed data saturates around 40%. Scaling hands and data together reaches 59%.
Unseen-hand success versus relative data size for 3, 50 and 1000 embodiments.
With 3 hands, unseen success stays near zero however much data is added. With 50 it saturates around 30%, and 1,000 hands reach 59%.

Real-World Deployment

We deploy XDex on an xArm from a single calibrated third-person camera. It samples 128 candidate grasps from a partial point cloud, keeps the first one that passes arm IK and table-collision checks, and executes it open-loop.

Unseen hand, unseen objects

The Ability Hand was held out of training entirely.

Changing the hand's topology

XDex also handles XHand configurations with three and two fingers.

3-finger XHand

2-finger XHand

Application: Object-Aware Neural Retargeting

XDex can translate one grasp into another: it encodes a human (MANO) hand pose and decodes a robot hand pose, conditioned on the object and on the robot hand's geometry.

In simulation

Kinematic retargeting copies the human pose without regard to the object. Neural retargeting keeps the posture and still grasps it.

On a real robot

A human hand and a robot hand each grasp one of two identical bottles. During reaching, both setups use standard teleoperation. At contact, the human keypoints go through XDex.

Without neural retargeting

Keypoint-only retargeting ignores object geometry, so the task fails.

With neural retargeting

XDex preserves the human's posture and still produces a feasible grasp.

BibTeX

@inproceedings{wang2026xdex,
  title     = {XDex: Learning Cross-Embodiment Dexterous Grasping with 1000 Hands},
  author    = {Wang, Keyi and Ye, Jianglong and Yuan, Chengjing and Wang, Xiaolong},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}

This work was supported, in part, by NSF CAREER Award IIS-2240014 and NSF CCF-2112665 (TILOS).