DKWBCDecoupling Motion Tracking and Balancing for Humanoid Whole-Body Control

Zhejiang UniversityRobbyantThe University of Hong Kong
* Project Lead† Corresponding Authors

01 / Real-robot comparisons

One Policy for All Motions.

02 / Generated motion tracking

Tracking Generated Motions

Prompts are adapted from the ARDY repository’s preset prompt list.

Prompt 01

“A person walks forward at a steady pace.”

Prompt 02

“A person makes one backward jump and lands.”

Prompt 03

“A person repeatedly takes side steps to the right.”

Prompt 04

“A person walks backward at a steady pace.”

Prompt 05

“A person performs one kick with the right leg.”

Prompt 06

“A person walks forward elegantly.”

Prompt 07

“A person bows once and returns to an upright stance.”

Prompt 08

“A ballet dancer repeats a turn while bringing the feet together.”

03 / Data scaling

Scaling Curves

Why Separate Tracking from Balancing?

01 / Dynamics decomposition

Two contributions to joint torque

Joint torque combines motion dynamics and environmental forces.

M(q)ν̇ + h(q, ν) τ S⊤ [ ] S ∑k Jk⊤λk = Model-based contribution Environmental contribution τmotion τenv Motion dynamics Environmental forces
M(q)ν̇ + h(q, ν)=S⊤τ+∑k Jk⊤λk
Target
02 / Existing approach

Existing Approach: One PD Target, Two Responsibilities

A joint-position policy uses one PD target for both motion tracking and balance.

τPD = Kp(qPD − q) − Kdq̇
State Reference
03 / Our method

Decoupling Tracking and Balancing.

Tracking and balancing produce separate torques that are added for execution.

qref(t),(q, ν)
Reference State

BibTeX

@misc{wei2026dkwbcdecouplingmotiontracking,
  title = {DKWBC: Decoupling Motion Tracking and Balancing for Humanoid Whole-Body Control},
  author = {Dingkun Wei and Zehong Shen and Huaijin Pi and Xueyang Wang and Xing Zhu and Hujun Bao and Yujun Shen and Xiaowei Zhou},
  year = {2026},
  eprint = {2611.09605},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2611.09605}
}

Motion comparison