Conference on Robot Learning 2026

HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model

Chao Chen1,3,* Zhenrong Wang2,3,4,* Jilong Wang3 Yusen Feng1,3 Jiangran Lyu1,3 Li Yi3,5,† He Wang1,3,†
1CFCS, School of Computer Science, Peking University 2Institute of Automation, Chinese Academy of Sciences
3Galbot4Beijing Academy of Artificial Intelligence5Tsinghua University

* Equal contribution   † Corresponding authors

From simulation-trained policies to real-world adaptability.

Six Unitree G1 tasks before and after fine-tuning: swallow balance with calf weights, arm swing with a hand weight, pen capping with a hand load, lateral raise with elastic bands, bowing with a damaged motor, and walking with a water bucket.
Direct transfer (left two frames) vs. fine-tuning (right two) under real-world disturbances.
MoCap-freeLearning from proprioception
50–200Real-world trajectories per task
6 tasksValidated on Unitree G1

The last mile of sim-to-real

Abstract

Humanoid robots have recently achieved remarkable success in motion tracking. However, during real-world deployment, robots inevitably encounter dynamics gaps—arising from either extra physical disturbances or inaccurate simulation modeling—resulting in a significant training-deployment gap. Therefore, effectively leveraging real-world data to fine-tune policies represents the crucial “last mile” for robust deployment.

To bridge this gap, we propose HumanoidNDM, an online fine-tuning framework for humanoid motion tracking driven by a real-world-trained neural dynamics model. First, in the data collection stage, we introduce a low-frequency reference perturbation method with an adaptive intensity mechanism, balancing deployment feasibility and exploration efficiency. Second, for the model training stage, we train a robust history-based, proprioceptive neural dynamics model in Euclidean space by incorporating differentiable forward kinematics, eliminating the reliance on costly motion capture systems. Finally, in the policy fine-tuning stage, we directly fine-tune the policy via backpropagated model gradients, achieving high data efficiency.

Through physical validation on the Unitree G1 robot, we demonstrate that HumanoidNDM enables rapid real-world adaptation to various perturbations, representing the first online, MoCap-free fine-tuning framework for humanoid tracking policies.

Collect. Model. Adapt.

A real-world learning loop

Start with a simulation-trained tracking policy. Collect real-world experience, learn the deployment dynamics, and improve the policy through differentiable rollouts.

HumanoidNDM pipeline. A simulation-trained policy enters an iterative real-world loop: FENCE data collection with adaptive frequency-domain reference perturbations, history-based dynamics model training using proprioception and differentiable forward kinematics, and policy fine-tuning through model gradients.
01 / DATA COLLECTION

Explore smoothly with FENCE

Low-frequency perturbations modify reference motions while the policy maintains closed-loop control. Policy neighborhood coverage guides adaptive perturbation intensity.

02 / MODEL TRAINING

Learn dynamics without MoCap

A history-based neural dynamics model learns from onboard proprioception. Differentiable forward kinematics enables supervision in both configuration and Euclidean space.

03 / POLICY FINE-TUNING

Fast adapt through model gradients

Backpropagate tracking losses through multi-step model rollouts to fine-tune the policy directly, making effective use of limited real-world trajectories.

On the physical robot

Real-world adaptation

Six tasks span dynamic whole-body motion and precise control. HumanoidNDM adapts to additional payloads and disturbances that are difficult to reproduce in simulation.

How to read the videos Deployment → disturbance → online fine-tuning

The task videos show the original behavior, the introduced disturbance, and online fine-tuning. FENCE exploration is shown at 100× speed. Evaluation runs occur every 10 episodes; milestone evaluations at episodes 0, 50, 100, 150, and 200 are shown at 1× speed, where applicable. Pen capping uses 50 trajectories; the other five primary tasks use 200. Use the player controls to enable audio or view full screen.

Swallow balance

Added payload

With 1.5 kg on each ankle, the robot initially cannot fully raise its leg. After 200 episodes, it achieves stable single-leg balance.

Air time0.00 → 1.90 s

Arm swing

Added payload

A 1.5 kg load limits the left hand to chest height. After 200 episodes, the arm recovers its vertical reach.

Arm height0.98 → 1.40 m

Pen capping

Precision control

A 0.75 kg load on the left hand causes misalignment. With just 50 episodes, the robot raises its hand and successfully caps the pen.

End-effector distance3.50 → 0.00 cm

Lateral raise

Elastic resistance

Elastic bands restrict arm motion and cause staggering. After 200 episodes, the robot raises its right arm higher and regains stability.

Arm height1.06 → 1.35 m

Bowing

Damaged motors

Damaged arm motors produce noisy, inaccurate tracking. After 200 episodes, motion becomes smoother and motor noise decreases. Enable audio to hear the comparison.

Noise level73 → 55 dB

Walking

Moving payload

A 2 kg water bucket causes staggering and trajectory deviation. After 200 episodes, the robot walks straighter without staggering.

Deviation angle38° → 5°

Task-specific results above compare Direct Transfer with HumanoidNDM using the mean over three trials. Full mean ± standard deviation values are reported below and in the paper.

View quantitative real-world results
Task completion metrics, mean ± standard deviation over three trials.
TaskMetricDirect TransferHumanoidNDM
Swallow balanceAir time (s) ↑0.00 ± 0.001.90 ± 0.61
Arm swingArm height (m) ↑0.98 ± 0.001.40 ± 0.01
Pen cappingEE distance (cm) ↓3.50 ± 0.000.00 ± 0.00
Lateral raiseArm height (m) ↑1.06 ± 0.071.35 ± 0.06
BowingNoise level (dB) ↓73 ± 855 ± 5
WalkingDeviation angle (°) ↓38 ± 175 ± 3

Adapting as conditions change

Additional demonstrations explore staged payload increases and multiple simultaneous disturbances.

Continuous adaptation

Increasing loads

Weights are added to both hands at episodes 0, 50, and 100. Evaluations at episodes 50, 100, and 200 show the arm-swing policy adapting to the increasing loads.

Dancing under disturbances

Combined effects

A 900 ml water bottle and elastic-band resistance disrupt arm motion and balance. After 200 episodes, the robot raises its arms and dances in place without staggering.

Deployment-viable exploration

Smoother exploration with FENCE

Frequency-domain Exploration via Neighborhood Coverage Estimation perturbs reference motions at low frequencies, enabling data collection with less joint jitter and whole-body shaking.

Exploration methods on the lateral raise task. FENCE preserves smooth motion while collecting experience for adaptation.

Real-world collection smoothness

Lower values indicate less jitter and shaking.

Collection smoothness, mean ± standard deviation.
MethodJoint ↓Roll-pitch ↓
FENCE (ours)0.099 ± 0.0180.126 ± 0.023
Action Gaussian1.275 ± 0.3131.678 ± 0.455
Action OU0.707 ± 0.1800.739 ± 0.210
Obs. Gaussian1.151 ± 0.3151.033 ± 0.218
No noise0.076 ± 0.0050.106 ± 0.008

FENCE approaches the stability of no-noise collection while providing active exploration.

Controlled evaluation in MuJoCo

Closing the dynamics gap

Five motions and three disturbance types—inertial, contact, and actuation variations—yield 15 simulation settings. HumanoidNDM, ASAP, and RWM each use 200 deployment trajectories.

30% ↓link position error

47% ↓link rotation error

Average reductions relative to Direct Transfer across all 15 simulation settings, as reported in the paper.
Normalized link position tracking error during adaptation, comparing HumanoidNDM, RWM, ASAP, and an oracle baseline.
Link position tracking error
Normalized link rotation tracking error during adaptation, comparing HumanoidNDM, RWM, ASAP, and an oracle baseline.
Link rotation tracking error

Scope. The paper studies adaptation to time-invariant disturbances. The staged-load video illustrates successive adaptation as loads are changed; arbitrary continuously time-varying dynamics remain outside the current method’s validated scope.

Reference this work

BibTeX

HumanoidNDM · CoRL 2026
Download .bib
@inproceedings{chen2026humanoidndm,
  title     = {HumanoidNDM: Closing the Training-Deployment Gap of Motion
               Tracking via Humanoid Neural Dynamics Model},
  author    = {Chen, Chao and Wang, Zhenrong and Wang, Jilong and
               Feng, Yusen and Lyu, Jiangran and Yi, Li and Wang, He},
  booktitle = {Conference on Robot Learning},
  year      = {2026},
  url       = {https://pku-epic.github.io/HumanoidNDM}
}