Skip to main content
Back to timeline
arXivSource publication:

Humanoid balance becomes a learned residual: DKWBC beats SONIC on success and tracking error at every training-data scale

Synopsis

Humanoid whole-body controllers usually ask one policy to learn motion tracking and balance together, even though similar reference poses carry very different torque demands. DKWBC computes tracking torques from floating-base inverse dynamics plus joint PD feedback and lets a reinforcement-learning policy output only joint-torque residuals for balance. On both BONES-SEED test sets it reaches higher success and lower tracking error than retrained SONIC at every training-data scale from 2 to 128 hours, and one balancing policy also works under payloads and external disturbances.

Source-provided article image: DKWBC: Decoupling Motion Tracking and Balancing for Humanoid Whole-Body Control
arXiv

Interpretation

Humanoid whole-body control is split in two: torques that track the reference motion are computed from a model, and torques that keep balance are supplied by a learned policy. Earlier reinforcement-learning trackers had one policy output joint-position targets that a PD controller turned into torques, so tracking and balance were learned in the same output. On the BONES-SEED test-content and test-repetition sets, DKWBC has higher success and lower tracking error than retrained SONIC at every uniformly sampled training scale of 2, 4, 8, 16, 32, 64 and 128 hours.

Tracking torque comes from inverse-dynamics feedforward evaluated at the current robot state and reference acceleration, plus joint PD feedback that corrects position and velocity error. The feedforward is evaluated with the recursive Newton–Euler algorithm on the floating-base model and split into three three-dimensional components that also enter the policy observation. Tested alone with a fixed root in mid-air, gravity on and no contact, the tracking module reaches 4.45 mm position error; feedforward alone gives 171.73 mm, showing that retargeted human motions are limited in dynamic feasibility and need position feedback.

The balancing policy outputs joint-torque residuals directly and holds them constant between policy updates instead of passing through a position servo. The common design has the policy emit joint-position targets at a lower rate and a higher-rate PD controller convert them to torque, so torque keeps changing with joint state even when the output is fixed. Trained on LAFAN and evaluated on both BONES-SEED test sets, full DKWBC has higher success and lower position and velocity error than the action ablations; without the tracking module, torque actions no longer beat PD-target actions.

Perspective

Validation covers simulation and a Unitree G1, with motions from BONES-SEED and LAFAN; the external-force experiment uses 217 standing and locomotion sequences and applies time-varying three-dimensional forces to both hands. The authors propose extending the decomposition to a broader range of whole-body physical interaction tasks.

The authors note that the residual can compensate for model error but the decomposition alone does not identify physical contact forces; the comparison with contact-compensation torques covers one simulated motion segment with ground contact, where agreement in amplitude and timing is described as consistent with the intended role. The external-force experiment does not establish explicit force regulation or manipulation-task success.