← Paper Digest

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

Lixin Xu, Ke Zhang, Xinyu Liu, Jie Tan, Senming Tan, Jiaju Yin, Yating Feng, Xiaoyi Lin, Xiyue Dong, Renjing Xu

November 2026

A person places an object on the table and the Flexiv arm and hand grasp it.

TL;DR DexFormer 2.0 is one RL grasping policy for many arm-hand pairs. Each robot is mapped to shared joint slots described by per-joint morphology, and a Transformer-XL memory supplies recent interaction. Trained on 18 of 24 pairings of four arms and six hands, it reaches 89% stable success on those pairs and 51% on the six held-out pairs in simulation. On a real Flexiv arm, a distilled depth-based student succeeds in 23 of 25 grasps with LEAP and 17 of 25 with Revo3.

What they do

Dexterous grasping policies are usually trained for one arm with one hand, so every new pairing needs its own policy. DexFormer 2.0 is a single RL policy trained on 18 of the 24 combinations of four arms (Flexiv, Franka, KUKA, xArm7) and six hands (Revo3, Sharpa, LEAP, Allegro, RAPID, Shadow). It is tested on the six held-out pairs, on four new hand types and on parallel grippers.

How it works

  • Canonical joint slots. Every robot is mapped to 7 arm and 22 hand slots with validity masks, so missing joints affect neither actions nor learning. Each joint gets a 16-D descriptor: axis, attachment position, limits, semantic role and kinematic depth.
  • MorphNode encoder. Arm, hand and mount MLPs are combined with an attention residual over joint and context tokens. The attention logits are biased by kinematic-tree distance and joint role.
  • Memory and training. A 3-layer Transformer-XL with a 16-step memory infers embodiment-specific dynamics from recent interaction. A shared actor head outputs relative joint targets, and the policy is trained with PPO in Isaac Lab on a mixture of embodiments.
  • Real robot. The state-based policy is distilled with DAgger into a student that sees depth from two stereo cameras. The student is deployed zero-shot, and a scripted routine then lifts the object to a basket.

Results

  • Simulation: 89.01% stable success on seen pairs and 51.18% on held-out pairs, versus 64.82%/31.28% for GRU and 66.02%/14.15% for DexFormer 1.0. With ten unseen YCB objects, held-out pairs reach 48.85%.
  • Removing memory or the tree bias cuts held-out success to 24.45% and 24.15%. Training on 12 instead of 18 pairings, at the same interaction budget, drops it to 5.65%.
  • Four unseen hand types average 17.45%. Parallel grippers driven through a virtual LEAP hand reach 7.55-16.62%.
  • Real Flexiv arm: 23 of 25 trials with LEAP (a training pair) and 17 of 25 with Revo3 (a held-out pair).

Limitations

Held-out pairs succeed only about half the time in simulation, and unseen hand types far less often. The gripper results come from retargeting a virtual LEAP hand, which the authors say is not zero-shot transfer to gripper morphology. Real-world tests cover one arm with two hands, and a scripted routine handles the lift and release after the grasp.

Videos

Grasping a cylinder from the tabletop on the real Flexiv arm.
Grasping a pear and carrying it to the basket.

Figures

From Figure 3: The 24 arm-hand embodiments, four arms (Flexiv, Franka, KUKA, xArm7) by six hands (Revo3, Sharpa, LEAP...
From Figure 3: The 24 arm-hand embodiments, four arms (Flexiv, Franka, KUKA, xArm7) by six hands (Revo3, Sharpa, LEAP, Allegro, RAPID, Shadow); six pairs are held out from training.
From Figure 2: Architecture. One shared policy controls all 24 arm-hand combinations (18 for training, 6 held out). M...
From Figure 2: Architecture. One shared policy controls all 24 arm-hand combinations (18 for training, 6 held out). MorphNode encodes the canonical joint state, a Transformer-XL adds a 16-step memory, and a task encoder feeds current task observations to the actor head, which outputs a masked Gaussian over 29 actions. Bottom: MorphNode, a body encoder of arm, hand and mount MLPs plus attention over joint and context tokens with topology biases, fused by a gated residual.
From Figure 6: Stable success rate for all 24 arm-hand combinations on seen objects, 1,000 episodes each, with standa...
From Figure 6: Stable success rate for all 24 arm-hand combinations on seen objects, 1,000 episodes each, with standard deviation over five seeds. Shaded groups are the six held-out pairs.
From Figure 7: Real-world grasping. A Flexiv arm with a Revo3 hand (top) and a LEAP hand (bottom) grasps tabletop obj...
From Figure 7: Real-world grasping. A Flexiv arm with a Revo3 hand (top) and a LEAP hand (bottom) grasps tabletop objects and lifts them toward a basket.
Abstract

Dexterous grasping policies are usually trained for one arm paired with one hand. Every new pairing then needs its own policy, and the number of policies grows with every arm and hand added. We ask whether a single policy can control many arm-hand combinations and transfer to pairings and hands it has never seen. This is hard because robots differ in the number, layout, and dynamics of their joints, so their observations and actions share no common interface. We present DexFormer 2.0, a single reinforcement learning policy for cross-embodiment dexterous grasping. It maps every robot to a canonical joint interface with validity masks, encodes each joint's morphology with attention biased by the kinematic tree, and uses a Transformer-XL memory to infer embodiment-dependent dynamics from recent interaction. Trained on 18 combinations of four arms and six hands, it reaches 89% stable success on seen combinations and 51% on six held-out arm-hand pairs, outperforming GRU and DexFormer 1.0 baselines. Without fine-tuning, it also grasps with four new hand types and, through retargeting, with parallel grippers. Ablations show that memory, the kinematic-tree bias, and broader pairing coverage each matter for transfer. On a physical Flexiv arm, distilled students transfer zero-shot from simulation with two hands, including a held-out pair. Project website: https://davidlxu.github.io/dexformer2-web/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.