DexFormer 2.0Towards a Cross-Embodiment
Foundation Policy for Dexterous Grasping

Lixin Xu^1,2,*Ke Zhang^1,2,*Xinyu Liu^2,*Jie Tan^2,3Senming Tan²

Jiaju Yin^1,2Yating Feng^1,2Xiyue Dong^1,2Renjing Xu^1,†

¹ HKUST(GZ)² BrainCo³ Chongqing University

* Equal contribution† Corresponding author

01 / 05Grasp in focus

The hand–object interaction, up close.

Explore every grasp in 3D.

Explore the representation.

Representation space

t-SNE · 128D → 2D

Shape = arm ● Flexiv■ Franka▲ KUKA★ xArm7

↗Hover over a point or a legend entry to explore.24 combinations

Each point represents a 128-dimensional Transformer-XL output with 16 steps of temporal memory, before concatenation with the task features. We visualize 800 samples per arm–hand combination, collected 1.0 s into each rollout.

Real-world Deployment

Real World Grasping Non-stop

Flexiv + Revo3 · 3× speed

A single student policy distilled from successful rollouts. After a successful grasp, motion planning handles transfer to and placement in the basket.

Success Cases

Flexiv + Revo3

Flexiv + LEAP

Failure Cases

Experiments

89.01%

Seen embodiments · seen objects

48.85%

Unseen embodiments · unseen objects

+15.67pp

Over GRU when both are unseen

Training dynamics · archived fault-trained runs

Training success curves over 5,000 iterations for MorphAttention with memory, MLP plus Transformer-XL, no tree bias, and no memory.

18 seen embodiments. Final 50-iteration averages: 78.21% (MorphAttention, memory 16), 76.04% (MLP + TXL), 66.94% (no tree bias), and 64.39% (no memory). All runs use a 60/20/20 nominal/initial/mid-episode fault mixture. Faint traces are raw values, not confidence intervals. These single-run adaptive-curriculum metrics are separate from the no-fault checkpoints evaluated above.

BibTeX

@misc{xu2026dexformer2,
  title = {DexFormer 2.0: Towards a Cross-Embodiment Foundation Policy for Dexterous Grasping},
  author = {Xu, Lixin and Zhang, Ke and Liu, Xinyu and Tan, Jie and Tan, Senming and Yin, Jiaju and Feng, Yating and Dong, Xiyue and Xu, Renjing},
  year = {2026},
  url = {https://davidlxu.github.io/dexformer2/}
}