RoboMorpher

Scaling Robot Morphology Representations
for Cross-Embodiment Action Prediction

Wenwei Lin1,*, Yan Qin2,*, Weiliang Deng3,*, Zixuan Li1,*, Kaixuan Wang1,*, Zanxin Chen4, Yue Chen5, Tianshuo Yang1, Junting Chen6, Zhixuan Liang1, Yao Mu4, Ximin Lyu3, Renjing Xu2, Ping Luo1,†, Tianxing Chen1,†

  1. 1The University of Hong Kong
  2. 2The Hong Kong University of Science and Technology (Guangzhou)
  3. 3Sun Yat-sen University
  4. 4Shanghai Jiao Tong University
  5. 5Peking University
  6. 6

* Equal contribution† Corresponding authors

Paper BibTeX
↓
Overview

Actions need to
understand the body.

The same joint configuration can produce different link poses on different robots. RoboMorpher connects control coordinates to their spatial effects through a morphology-conditioned action tokenizer.

Kinematic pretraining learns codes that reconstruct joint angles and link poses. A VLA policy predicts these codes and refines continuous trajectories.

40M

morphology–state pairs

6–7

joints per generated arm

2 × 2048

shared RVQ codebooks

50

frames per action chunk

01 / GENERATE

MorphGen

Procedural arms and kinematic supervision.

03 / ACT

Morphology-aware VLA

Discrete prediction and continuous refinement.

Morphgen

Geometry becomes supervision.

Position-specific link-length priors guide sampling. Collision-aware inverse kinematics filters home poses, and forward kinematics supplies joint-aligned pose targets.

33robot descriptions
185reported link lengths
7position-specific priors

Link-length priors

Seven Gaussian priors fitted to 185 reported link lengths; highlights flow along fixed density curves

Joint distribution

Reported arm degrees of freedom Five DoF: 3 descriptions (9.1%); six DoF: 21 (63.6%); seven DoF: 9 (27.3%). Reported arm DoF 0 10 20 30 3 9.1% 5 21 63.6% 6 9 27.3% 7 Joint-limit summary Bounded: 26 descriptions (78.8%); mixed: 4 (12.1%); unrestricted: 3 (9.1%). Joint-limit summary 0 10 20 30 26 78.8% Bounded 4 12.1% Mixed 3 9.1% Unrestricted
30 of 33 surveyed descriptions have 6–7 DoF (90.9%).

Generated Morphologies

Morphology 001
Morphology 047
Morphology 063
Morphology 084
Morphology 000
Morphology 010
Morphology 018
Morphology 028
Morphology 044
Morphology 059
Morphology 061
Morphology 068
Morphology 070
Morphology 077
Morphology 085
Morphology 092
Morphology 096
Morphology 004
Morphology 053
Morphology 097

Method

Full RoboMorpher architecture: MorphGen, morphology-conditioned residual quantization, and discrete-continuous action prediction
RoboMorpher pipeline. (a) MorphGen samples local-Z chains using survey priors and IK filtering. (b) URDF-conditioned tokenization learns two codes per joint with FK supervision. (c) The VLA predicts codes, decodes ten nodes, interpolates 50 frames, and refines them with flow matching. Frozen URDF features condition both DiTs and the decoder; VLM context conditions the DiTs.

Embodibench

Single- and dual-arm (same-side / opposite-side) manipulation in SAPIEN.

10tasks
30URDFs per task
15,000demonstrations

Results

Best mean success among joint-space methods.

1,500 trials / method / benchmark

RoboTwin 2.0

56.40%

+9.73 ppvs. StarVLA (joint)

EmbodiBench

23.27%

+15.34 ppvs. StarVLA (joint)

Success rate (%) ↑

Mean policy success rate (%) on RoboTwin 2.0 and EmbodiBench
MethodRoboTwin 2.0EmbodiBench
End-effector reference
StarVLA (EE)69.9332.00
Joint-space control
Angle+M51.875.33
256-bin51.605.53
FAST50.735.93
StarVLA (joint)46.677.93
Angle45.334.60
VLA + RoboMorpher56.4023.27
Column maximumBest joint-space
EE uses a separate control interface.

RoboTwin 2.0 Results

Success rate (%) ↑

Per-task success rates (%) on RoboTwin 2.0
MethodGrab rollerShake bottleHandover micPlace burger friesPlace empty cupClick bellLift potOpen laptopBeat block hammerStack bowls threePick dual bottlesMove can potPlace phone standPut bottles dustbinPlace dual shoesMean
End-effector reference
StarVLA (EE)100.00100.0093.0052.0085.0088.0082.0097.0032.0068.0065.0052.0051.0042.0042.0069.93
Joint-space control
Angle+M97.0096.0074.0053.0069.00100.0022.0070.0018.0064.0030.0012.0039.0018.0016.0051.87
256-bin97.0095.0067.0065.0070.0072.0037.0073.0013.0059.0035.0021.0034.0010.0026.0051.60
FAST95.0096.0074.0067.0068.0069.0017.0076.0025.0054.0032.0017.0042.0012.0017.0050.73
StarVLA (joint)90.0092.0047.0059.0069.0088.0028.0066.0010.0025.0032.0020.0031.0021.0022.0046.67
Angle95.0088.0052.0058.0069.0084.005.0064.0011.0054.0025.0020.0027.008.0020.0045.33
VLA + RoboMorpher92.0094.0075.0074.0070.0084.0030.0082.0034.0054.0029.0026.0050.0017.0035.0056.40

EmbodiBench Results

Success rate (%) ↑

Dual-arm · opposite-sideDual-arm · same-sideSingle-arm
Per-task success rates (%) on EmbodiBench
MethodLiftPickPushStackMean
End-effector reference
StarVLA (EE)62.6758.6778.6714.678.0032.6762.671.330.000.6732.00
Joint-space control
Angle+M10.678.679.333.334.671.3314.670.000.670.005.33
256-bin14.004.0011.332.674.004.0015.330.000.000.005.53
FAST9.3311.3310.674.002.675.3315.330.000.000.675.93
StarVLA (joint)14.6714.0013.336.006.009.3316.000.000.000.007.93
Angle10.676.676.001.332.675.3313.330.000.000.004.60
VLA + RoboMorpher61.3352.6751.3323.3311.334.0028.670.000.000.0023.27

Cite RoboMorpher

@unpublished{lin_robomorpher,
  title = {{RoboMorpher: Scaling Robot Morphology Representations for Cross-Embodiment Action Prediction}},
  author = {Lin, Wenwei and Qin, Yan and Deng, Weiliang and Li, Zixuan and Wang, Kaixuan and Chen, Zanxin and Chen, Yue and Yang, Tianshuo and Chen, Junting and Liang, Zhixuan and Mu, Yao and Lyu, Ximin and Xu, Renjing and Luo, Ping and Chen, Tianxing},
  note = {Manuscript}
}