Organizations: The University of Hong Kong · HKUST (Guangzhou) · Sun Yat-sen University · Shanghai Jiao Tong University · Peking University · National University of Singapore
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.
Figures & tables
Figure 1: Survey and morphology generation. (a) Seven Gaussian priors N(μi,σi2) fitted to 185 reported lengths from 30 description records. Curves show the fitted distributions before truncation and chain-length rescaling. (b) Ninety-six generated morphologies varying link geometry and joint-frame arrangement. Appendices A and B provide the survey records and generation details.
Figure 2: 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.
Property
Count / range
Description entries / distinct names
33 / 30
URDF / MJCF entries
24 / 9
Gripper / other tool / no tool listed
11 / 1 / 21
Approx. reach (32 entries)
0.25–2.05 m
Model mass (32 entries)
0.226–33.5 kg
Table 1: Robot-description survey. Entry-level statistics. Masses and reaches are reported values. The appendix documents description-level discrepancies.
Figure 3: The ten EmbodiBench tasks. Successful saved-trajectory keyframes, ordered as in Tab. 3 . SS/OS denote same-side/opposite-side dual-arm placement. These visualization replays are distinct from learned-policy rollouts.
Joint-space baselines
Representation variants
Full model
Reference
Task
Joint
256-bin
FAST
Angle
Angle+M
Ours
EE
grab_roller
90.00%
97.00%
95.00%
95.00%
97.00%
92.00%
100.00%
shake_bottle
92.00%
95.00%
96.00%
88.00%
96.00%
94.00%
100.00%
handover_mic
47.00%
67.00%
74.00%
52.00%
74.00%
75.00%
93.00%
place_burger_fries
59.00%
65.00%
67.00%
58.00%
53.00%
74.00%
52.00%
place_empty_cup
69.00%
70.00%
68.00%
69.00%
69.00%
70.00%
85.00%
Table 2: Full RoboTwin 2.0 results. Success rate (%). All 15 evaluated tasks across five embodiments. 20 trials per pair. Bold/underline mark best/second-best positive joint-space scores (ties included). EE uses a separate control interface.
Joint-space baselines
Representation variants
Full model
Reference
Task
Joint
256-bin
FAST
Angle
Angle+M
Ours
EE
lift_dual_os
14.67%
14.00%
9.33%
10.67%
10.67%
61.33%
62.67%
lift_dual_ss
14.00%
4.00%
11.33%
6.67%
8.67%
52.67%
58.67%
lift_single
13.33%
11.33%
10.67%
6.00%
9.33%
51.33%
78.67%
pick_dual_os
6.00%
2.67%
4.00%
1.33%
3.33%
23.33%
14.67%
pick_dual_ss
6.00%
4.00%
2.67%
2.67%
4.67%
11.33%
8.00%
Table 3: Full EmbodiBench results. Success rate (%). All 10 tasks with 30 URDFs per task. Five trials per pair. Bold/underline mark best/second-best positive joint-space scores (ties included). EE uses a separate control interface.
Joint angle ( ∘ )
Position (m)
Orientation ( ∘ )
Description
Mean
P90
Max
Mean
P90
Max
Mean
P90
Max
ALOHA
0.4124
0.8951
14.4648
0.004255
0.009446
0.116570
1.3237
3.4343
88.6863
ARX-X5
0.4107
0.7768
52.6336
0.004899
0.013960
0.112769
1.0860
2.0659
133.9178
Franka Panda
0.4563
1.0207
42.7852
0.009204
0.022782
0.644741
1.8655
4.4817
178.8929
Piper
0.4448
0.8695
40.3280
0.003503
0.007992
0.186133
1.6299
4.5327
88.5053
UR5-WSG
0.5429
1.2653
5.6463
0.007017
0.016207
0.048518
1.3591
3.7680
11.4713
Table 4: Tokenizer reconstruction. All saved task frames across six description groups. Geometry is predicted directly by the decoder. P90 is the pooled 90th percentile. Bold/underline mark the lowest/second-lowest value per column (ties included). Metric definitions appear in Appendix C .
Figure 4: Reconstruction error profiles. Mean, P90, and maximum joint-angle and position errors from Tab. 4 . Both axes are logarithmic.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
ID
Name
Format
DoF
Joint types
Limits
Mass (kg)
Reach (m)
1
e.DO
URDF
6
All C
U
0.5
–
2
FR3
MJCF
7
All H
B
15.49
0.85
3
Gen2
URDF
6
C:3; R:3; F:6
M
4.35
0.96
4
Gen3 Lite
URDF
6
R:10; F:4
B
5.56
0.79
5
Gen3
MJCF
7
All H
M
6.34
0.93
6
Gen3
URDF
6
C:3; R:3; F:4
M
6.14
0.93
Appendix
Table 5: Complete description inventory of structure and scale. All entries are serial manipulators. DoF is the reported arm DoF, and masses and reaches are approximate description-level values. Joint types are R (revolute), C (continuous), P (prismatic), F (fixed), H (hinge), and S (slide/slider). Limits are B (bounded in the stated scope), M (mixed), and U (unrestricted). A dash denotes an unavailable value.
ID
Name
Format
Tool
Reported tool DoF
Sensor / marker
3
Gen2
URDF
Two-finger gripper
0
None listed
4
Gen3 Lite
URDF
Two-finger, four-joint gripper
1
None listed
5
Gen3
MJCF
None listed
–
Force/torque at EE
6
Gen3
URDF
None listed
–
Camera at EE
9
Koch v1.1
MJCF
Single-sided gripper
1 translational
None listed
10
Koch v1.1
URDF
Single-sided gripper
1 rotational
None listed
Appendix
Table 6: Recorded tools and sensing. IDs match Tab. 5 . The 19 omitted entries list neither a tool nor sensing. Tool DoF and couplings follow the survey and need not equal independent actuation. Missing records concern the description, not the physical robot. The AR tag is a passive marker.
Record
1
2
3
4
5
6
7
eDO
0.145292
0.18967
0.210498
0.159044
0.1053
0.163646
–
Gen2
0.15675
0.11876
0.410
0.2076
0.10375
0.10375
–
Gen3 Lite
0.12825
0.11885
0.280
0.14142
0.1088
0.1088
–
Gen3
0.15643
0.12849
0.410
0.20853
0.10593
0.10593
–
iiwa14
0.1575
0.2025
0.2045
0.2155
0.1845
0.2155
0.0810
iiwa7
0.15
0.19
0.21
0.19
0.2
0.19946
0.10122
Appendix
Table 7: Link-length records as supplied (m). Blank source cells are shown as dashes. Reported zeros are retained. These records are kept separate from the main inventory because description identities and extraction conventions are not fully aligned. They are not treated as validated workspace radii or summed to estimate reach.
Joint position
Records
μi (m)
σi (m)
1
30
0.1576
0.1071
2
30
0.1035
0.0805
3
30
0.3029
0.1715
4
30
0.2270
0.1282
5
30
0.1532
0.1862
6
27
0.1112
0.0677
Appendix
Table 8: Survey-derived link prior. Means and sample standard deviations reproduce the generator parameters to four decimals. Statistics use the supplied link records, including zeros and excluding blanks. Sampling and global rescaling change the final length distribution.
Setting
Tokenizer
VLA: RoboTwin 2.0
VLA: EmbodiBench
Training data
40M pairs
3,750 demonstrations
15,000 demonstrations
Duration
300 epochs
30,000 steps
60,000 steps
Global batch
32,768
128
128
Gradient accumulation
1
1
1
Optimizer
AdamW
AdamW
AdamW
Learning rate
10−5
VLM 10−5 ; head 10−4
VLM 10−5 ; head 10−4
Appendix
Table 9: Training data and optimization. Tokenizer data count morphology–state pairs. Policy data count demonstrations. Durations and global batches describe completed runs. The remaining optimizer settings are from the reference configurations. Reconstruction corpus sizes are reported separately in Tab. 10 .
Description
Trajectories
ALOHA
750
ARX-X5
750
Franka Panda
750
Piper
750
UR5-WSG
750
RoboTwin 2.0 total
3,750
Appendix
Table 10: Reconstruction evaluation coverage. All original saved frames are evaluated. Each real arm is reconstructed independently. Padding, missing arms, and grippers are excluded. Generated URDFs span 55 distinct morphologies across ten task-specific sets of 30.