A Universal Controller for Any Humanoid
Researchers from Shanghai Jiao Tong University and Shanghai Artificial Intelligence Laboratory have unveiled XHugWBC, a groundbreaking framework that enables a single whole-body control policy to generalize across diverse humanoid robots without per-robot retraining.
The Cross-Embodiment Challenge
Traditional whole-body controllers require robot-specific training — a costly and inefficient process as each new platform differs in morphology, kinematics, and dynamics. XHugWBC tackles this by learning motion priors from diverse randomized embodiments, acquiring a structural bias that supports zero-shot transfer to unseen robots.
Three-Pillar Framework
1. Physics-Consistent Morphological Randomization — Unlike traditional domain randomization that can generate non-physical robots (negative mass, non-positive definite inertia), XHugWBC uses Cholesky decomposition to ensure physical consistency, generating embodiments ranging from 12 DoF (pure biped) to 32 DoF (full-body with arms and head).
2. Universal Cross-Embodiment Representation — A canonical 32-dimensional global joint space maps all possible humanoid configurations, with semantic alignment across diverse kinematic topologies.
3. Effective Policy Architecture — Transformer-based architecture outperforms GCN and MLP alternatives, capturing inter-joint kinematic dependencies that simpler models miss.
World-First Real-World Validation
XHugWBC's generalist policy was validated on 7 real-world humanoid robots:
- Unitree G1 (23/29 DoF)
- Agibot X2 (40 kg)
- Fourier N1 (39 kg)
- Booster T1 (31 kg)
- Dobot Atom (60 kg)
- Unitree H1-2 (66 kg)
Results: 100% zero-shot transfer success rate for locomotion tasks. The generalist policy achieves approximately 85% of specialist-level performance, and after fine-tuning (Generalist-FT), surpasses specialists by up to 10% on certain platforms.
Beyond Locomotion
The framework also supports long-horizon whole-body loco-manipulation tasks — walking, bending, opening doors, picking up objects, and placing them in baskets — all with a single trained policy.
Publication
- Paper: arXiv:2602.05791v3 (updated June 9, 2026)
- Website: https://xhugwbc.github.io
- Published by: Yufei Xue, Yunfeng Lin, Wentao Dong, et al. (SJTU & Shanghai AI Lab)
