DexFormer 2.0

Towards a Cross-Embodiment Foundation Policy for Dexterous Grasping

Lixin Xu1,2,*Ke Zhang1,2,*Xinyu Liu2,*Jie Tan2,3Senming Tan2

Jiaju Yin1,2Yating Feng1,2Xiaoyi Lin4Xiyue Dong1,2Renjing Xu1,†

1 HKUST(GZ)2 BrainCo3 Chongqing University4 Wuhan University

* Equal contribution† Corresponding author

Code (soon)
↓

01 Embodiments

Twenty-four bodies. One policy.

Four arms and six hands, with different joints, links and mounts, all controlled by one set of weights.

18 pairs trained together. One policy learns all of them at once and reaches 89.01% stable success on them.

Six pairs held out. Every arm and hand is seen in training, but these six combinations never are. The policy grasps with them zero-shot: 51.18% stable success.

02 Space × Time

Across joints. Across time.

t = 0.00 sRecorded rollout · iteration 5,000

01 Time

Every step remembers the last sixteen.

A Transformer-XL keeps a memory of recent control steps, so the policy sees how its hand has been moving, not just where it is now. Each line traces one fingertip, or the object, through the recorded rollout.

02 Space

Every joint is a token.

MorphNode turns each joint and link into a token with its physical descriptors and lets them attend to each other along the kinematic tree. The same network reads a Revo3, a LEAP, an Allegro or a Shadow hand on any of four arms.

03 Space × Time

Read the body. Read the motion.

Spatial attention tells the policy what body it is driving; temporal memory tells it how that body has been moving. Together they let one network grasp with an arm–hand pair it was never trained on: this Flexiv + Revo3 pair is one of the six held out.

See it on real hands ↓

A Across joints · MorphNode

29 joint tokens. The tree shapes who listens to whom.

Every arm and hand maps onto 7 arm + 22 hand slots; a validity mask switches off slots a body does not have. Two spatial attention layers add a learned bias for tree distance and joint role to every joint pair, while five context tokens attend globally.

29 of 29 slots active

same chaindifferent fingerarm ↔ handArc width: tree-distance bias (illustrative).
ℓij = qi·kj/√d + Bdist(Dij) + Brole(Rij)

B Across time · Transformer-XL

Each step reads the last sixteen.

The spatial code st enters a three-layer, eight-head Transformer-XL. Every layer keeps its own 16-step key/value memory, cleared at each episode boundary, so the policy can infer how this body responds.

memory length 16 · 3 layers

current stepin memorycleared / not yet filledArc opacity: recency (illustrative).
st · 128→TXL · 3 × 8 heads→chist · 128 ⊕ctask · 128 bypasses memory→actor head · 256 → masked joint actions

03 Explore

Watch it. Map it.

Replay any recorded grasp in your browser. The t-SNE map of the policy's temporal features follows along, highlighting where that arm–hand combination sits; click a cluster on the map to replay it.

Original t-SNE coordinates · color = combination; shape = arm.

Shape = arm ● Flexiv▲ Franka■ KUKA★ xArm7
↗Hover over a point or a legend entry to explore.24 combinations

04 Real world

Sim2Real.

A student policy distilled from successful simulated rollouts runs on a Flexiv arm with a Revo3 or a LEAP hand.

Real World Grasping Non-stop

Flexiv + Revo3
Mixed objects · 3×
Pear · 4×
Mango · 4×
Pomegranate · 4×

Success Cases

Flexiv + Revo3

Flexiv + LEAP

Failure Cases

05 Experiments

The numbers.

89.01%

Stable success · 18 trained combinations · seen objects

51.18%

Six held-out pairs · zero-shot · seen objects

48.85%

Held-out pairs · unseen YCB objects

Training dynamics

06 Films

Watch it move.

Seven short films, from a narrated one-minute overview to conveyor-belt packing and real-world grasping. Click one to play it; use the arrows to move between them.

07 Cite

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 Lin, Xiaoyi and Dong, Xiyue and Xu, Renjing},
  year = {2026},
  url = {https://davidlxu.github.io/dexformer2-web/}
}