CoRL 2026 · Austin, Texas

X-WBCA Cross-Embodiment Foundation Model for Humanoid Whole-Body Control

From isolated per-robot policies to a shared motion foundation model trained across diverse humanoid bodies.

Juntong Zhang1,3 Chun Gu2,3 Li Zhang2,3,*

1Tongji University · 2Fudan University · 3Shanghai Innovation Institute

Paper arXiv Video YouTube Repository GitHub

Real-world deployment One interface, diverse bodies

9robot embodiments
1shared backbone
3command routes
4real humanoids

The idea

Change the training unit.

Human motion carries structure that can be shared across bodies. Physical execution cannot.

X-WBC separates reusable temporal motion semantics from embodiment-specific proprioception and action. Multiple humanoids roll out together, contribute to the same PPO updates, and optimize one shared motion Transformer.

Nine simulated humanoid embodiments and four real-world robots controlled through X-WBC
Cross-embodiment training in simulation and deployment on physically different humanoids.

How it works

Shared meaning.
Specific execution.

Three command routes describe the same underlying motion intent. A shared temporal backbone learns across robots; lightweight robot-specific modules make that intent executable.

X-WBC architecture with command encoders, shared Transformer backbone, and robot-specific decoders
01

Human-centered commands

Full human motion, retargeted robot motion, and sparse five-point VR observations are aligned in one command-token space.

02

Temporal motion backbone

A causal Transformer integrates command and proprioception histories, sharing motion experience across all active robots.

03

Embodiment adapters

Robot-specific state encoders and action decoders respect different morphologies, joint layouts, dynamics, and action spaces.

Evidence

Two evaluations.
Two distinct questions.

Training-motion studies examine architecture and joint training. A separate frozen evaluation tests external locomotion styles without replacing the original analysis.

Table 1 · BONES-SEED training motions

98.60% / 93.22%

Full X-WBC · G1 / H2

The shared Transformer achieves the strongest success rate on both evaluated embodiments. This is an in-distribution study of architecture and joint training—not an unseen-embodiment claim.

X-WBCShared Transformer · joint training
98.60% 93.22%
X-WBC (MLP)Shared MLP · joint training
97.70% 91.33%
Single robotTransformer · independent training
97.69% 92.22%

Command consistencyRobot, human, and sparse-VR routes remain closely matched.

Component evidenceRemoving command routes, alignment, or robot-specific modules degrades both robots.

Real robots

One command format.
Four bodies.

The same sparse-VR command format controls Unitree G1, R1, H1-2, and H2, while robot-specific branches map shared motion features to each platform.

Real-world X-WBC deployment on Unitree G1, R1, H1-2, and H2
Qualitative real-world deployments across body scale and kinematic structure.

Abstract

A shared foundation for humanoid motion.

Scaling humanoid whole-body control requires large motion corpora and experience shared across robot bodies. X-WBC introduces a cross-embodiment framework in which a shared motion Transformer learns reusable temporal structure, while lightweight robot-specific modules condition execution on each robot’s proprioception and action space. The framework aligns full human motion, robot reference motion, and sparse VR observations into human-centered command tokens and learns from mixed multi-robot rollouts. Simulation studies, external locomotion evaluation, token-space analysis, and deployment on four real humanoids support the feasibility of sharing a whole-body control backbone across bodies with different sizes and kinematics.

Citation

Build on X-WBC.

Please cite our paper if you find this work useful.

@inproceedings{zhang2026xwbc,
  title     = {X-WBC: A Cross-Embodiment Foundation Model
               for Humanoid Whole-Body Control},
  author    = {Zhang, Juntong and Gu, Chun and Zhang, Li},
  booktitle = {Conference on Robot Learning},
  year      = {2026}
}