DKWBC: Decoupling Motion Tracking and Balancing for Humanoid Whole-Body Control

Dingkun Wei, Zehong Shen, Huaijin Pi, Xueyang Wang, Xing Zhu, Hujun Bao, Yujun Shen, Xiaowei Zhou

November 2026

Reference, SONIC and DKWBC side by side on the real G1, arms raised overhead.

TL;DR DKWBC splits humanoid whole-body control into tracking torques computed by inverse dynamics plus an RL residual that only handles balance. Trained under the same settings as SONIC, it gets higher success and lower tracking error at every data scale from 2 to 128 hours, and runs on a real Unitree G1 carrying payloads.

What they do

RL motion trackers for humanoids usually output joint position targets for a PD controller, so one policy must learn both the torques each motion needs and the corrections that keep the robot balanced. DKWBC splits the two: a model-based module computes the torques that follow the reference motion, and an RL policy learns only a residual torque for balance under ground contact, payloads and pushes. Experiments use the Unitree G1 in simulation and on hardware.

How it works

Dynamics split. The floating-base equations of motion write joint torque as a model-based term (from state and acceleration) plus a term from contact and external wrenches. The first is computed, the second is learned.
Tracking module. Inverse dynamics at the current state and the smoothed reference acceleration gives a feedforward torque. Joint PD feedback corrects for retargeted motions that are not dynamically feasible; it is under 10% of the tracking torque but needed for accuracy.
Balancing residual. The policy also observes the inertial, Coriolis and gravity feedforward terms, and outputs joint torques directly at 50 Hz (a unit action is 25% of the torque limit). Skipping position targets is meant to stop it from learning motion-specific tracking fixes.

Results

Against SONIC retrained with the same settings on 2-128 h subsets of BONES-SEED, DKWBC has higher success and lower MPJPE at every scale on both held-out test sets (3,145 and 3,019 motions). From 4 h onward it also tracks more accurately than the official SONIC checkpoint.
With time-varying forces on both hands over 217 standing and walking motions, it has the lowest tracking errors at every force level against FALCON and ResMimic, with comparable success.
Against BeyondMimic trained the same way, success is similar (DKWBC slightly higher) but BeyondMimic’s errors are substantially larger, on ground, with a 5 kg backpack plus 2 kg per hand, and under pushes.
The tracking module alone, root fixed in mid-air, reaches 4.45 mm MPJPE; feedforward without feedback gives 171.73 mm.

Limitations

Hardware results are qualitative only, with no real-robot metrics. The force experiments measure tracking under disturbance, not force regulation or manipulation success, and the residual does not identify contact forces. Success rates against FALCON and ResMimic are comparable, not better.

Videos

Arm sweeps and a side bend on the real G1, compared with SONIC.

A low crouch with arm motions on the real G1, compared with SONIC.

Figures

From Figure 1: Left, a conventional policy jointly learns tracking and balancing and outputs joint position targets t... — From Figure 1: Left, a conventional policy jointly learns tracking and balancing and outputs joint position targets that a PD controller converts into torques. Right, DKWBC adds a learned balancing residual to tracking torques computed by inverse dynamics. Bottom, on unseen motions DKWBC tracks the reference poses more closely while keeping balance.

From Figure 2: Overview of DKWBC. The tracking module combines inverse-dynamics feedforward with joint feedback. The ... — From Figure 2: Overview of DKWBC. The tracking module combines inverse-dynamics feedforward with joint feedback. The residual policy receives robot and reference information plus the feedforward components and predicts joint-torque corrections; the two torques are summed into the final command.

From Figure 3: Success rate of DKWBC and a retrained SONIC across training-data scales on BONES-SEED test-content (le... — From Figure 3: Success rate of DKWBC and a retrained SONIC across training-data scales on BONES-SEED test-content (left) and test-repetition (right). Labels give MPJPE in mm; stars mark the official SONIC checkpoint.

From Figure 5: DKWBC on the real G1, all with the same policy checkpoint. Left, deep squats and walking under differe... — From Figure 5: DKWBC on the real G1, all with the same policy checkpoint. Left, deep squats and walking under different payloads; right, complex motions.

Abstract

Humanoid whole-body control requires a robot to track diverse reference motions while maintaining balance under gravity, torque limits, contact switching, and external disturbances. Existing reinforcement-learning-based trackers typically learn control outputs that jointly address motion tracking and balance. However, similar reference poses can correspond to substantially different dynamic demands, requiring a unified policy to learn motion-dependent control requirements alongside balance corrections. This can complicate learning and lead to high-frequency fluctuations in control outputs. We propose DKWBC, a dynamics-informed whole-body control framework that combines explicit model-based tracking with a learned balancing residual. Inspired by the model-based and environmental-wrench contributions to joint torques in floating-base dynamics, DKWBC uses inverse-dynamics feedforward and joint PD feedback to supply tracking torques, while a residual policy learns additional joint-torque corrections for balanced execution. By explicitly accounting for motion-dependent torque demands, this structure simplifies the policy's learning task. Experiments in simulation and on real hardware demonstrate that a single learned balancing policy generalizes across motion datasets while maintaining accurate motion tracking and balance across ground support, attached payloads, and physical interaction. Project page: https://robbyant-research.github.io/DKWBC/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong