← Paper Digest

RoboMorpher: Scaling Robot Morphology Representations for Cross-Embodiment Action Prediction

Wenwei Lin, Yan Qin, Weiliang Deng, Zixuan Li, Kaixuan Wang, Zanxin Chen, Yue Chen, Tianshuo Yang, Junting Chen, Zhixuan Liang, Yao Mu, Ximin Lyu, Renjing Xu, Ping Luo, Tianxing Chen

November 2026

Hundreds of MorphGen-generated arm morphologies rendered side by side.

TL;DR Pretrain an action tokenizer on 40 million generated robot-arm configurations so that it reconstructs both joint angles and link poses, then use it frozen inside a vision-language-action policy. In simulation the policy reaches 56.40% success on RoboTwin 2.0 and 23.27% on the new EmbodiBench, 9.73 and 15.34 points above continuous joint-space control.

What they do

In cross-embodiment manipulation, the same joint values put the links of different arms in different places, so a policy trained on raw joint actions cannot tell what an action does. RoboMorpher is an action tokenizer conditioned on the robot’s URDF that learns this relation from kinematics alone, before any task demonstrations. It is pretrained on 40 million generated morphology-state pairs and then frozen inside a vision-language-action (VLA) policy.

How it works

  • MorphGen. A survey of 33 arm descriptions gives per-joint link-length priors. The generator samples six- and seven-joint revolute arms in a common local-Z frame convention, picks a home pose with collision-aware IK (cuRobo), and labels each configuration with link poses from forward kinematics.
  • Tokenizer. URDF features condition a Transformer encoder and decoder through AdaLN, and attention biases encode distance along the chain. Residual vector quantization gives two codes per joint; the decoder reconstructs joint angles plus each link’s position and rotation.
  • Policy. Qwen3-VL-4B feeds a two-branch diffusion Transformer. One branch predicts codes for ten trajectory nodes, the frozen decoder turns them into a coarse 50-frame chunk, and a flow-matching branch refines it.

Results

All in simulation, with the test URDFs held out of tokenizer pretraining.

  • RoboTwin 2.0 (15 tasks, 5 embodiments): 56.40% mean success vs 46.67% for continuous joint control, 51.60% for 256-bin tokens and 50.73% for FAST; better on 12 of 15 tasks.
  • EmbodiBench (10 tasks, 30 generated URDFs each): 23.27% vs 7.93%. Lifting gives about 80% of the gain (55.11% vs 14.00%); all three stacking tasks stay at 0%.
  • Removing link-pose supervision drops EmbodiBench success to 5.33%.
  • An end-effector-control reference still scores higher: 69.93% and 32.00%.

Limitations

No real-robot experiments. Policy training and evaluation use the same morphologies, so zero-shot transfer to new arms is untested. Stacking fails, reconstruction has large outliers (up to 0.75 m position error), and variation across training seeds is not measured.

Videos

One generated arm with a gripper moving through joint configurations.

Figures

From Figure 1: Ninety-six generated morphologies that vary link geometry and joint-frame arrangement.
From Figure 1: Ninety-six generated morphologies that vary link geometry and joint-frame arrangement.
From Figure 2: The RoboMorpher pipeline. (a) MorphGen samples arm chains from survey priors and filters them with inv...
From Figure 2: The RoboMorpher pipeline. (a) MorphGen samples arm chains from survey priors and filters them with inverse kinematics. (b) URDF-conditioned tokenization learns two codes per joint with forward-kinematics supervision. (c) The VLA predicts codes, decodes ten nodes, interpolates 50 frames and refines them with flow matching.
From Figure 3: The ten EmbodiBench tasks, shown as keyframes of successful demonstration trajectories (not policy rol...
From Figure 3: The ten EmbodiBench tasks, shown as keyframes of successful demonstration trajectories (not policy rollouts). SS/OS mark same-side and opposite-side dual-arm placement.
From Figure 4: Mean, 90th-percentile and maximum tokenizer reconstruction errors in joint angle and link position for...
From Figure 4: Mean, 90th-percentile and maximum tokenizer reconstruction errors in joint angle and link position for each robot description. Both axes are logarithmic.
Abstract

Learning shared action representations is central to cross-embodiment manipulation. However, identical joint values can produce different spatial configurations across robots, limiting representations learned from joint values alone. We introduce RoboMorpher, a morphology-conditioned action tokenizer that jointly reconstructs joint angles and link poses to associate control coordinates with their kinematic effects. MorphGen supplies scalable kinematic supervision by generating serial-arm descriptions from survey-derived geometric priors. We pretrain RoboMorpher on 40 million morphology-state pairs and integrate the frozen tokenizer into a vision-language-action policy for discrete prediction and continuous refinement. We also introduce EmbodiBench, a ten-task benchmark with 30 generated URDFs per task. With downstream URDFs excluded from tokenizer pretraining, the policy achieves 56.40% success on RoboTwin 2.0 and 23.27% on EmbodiBench, exceeding continuous joint-space control by 9.73 and 15.34 percentage points. Removing link-pose supervision reduces EmbodiBench success by 17.94 percentage points, supporting morphology-aware geometric supervision for joint-action representation learning.

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.