A Unified and General Humanoid Whole-Body Controller for Fine-Grained Locomotion

Yufei Xue, Wentao Dong, Minghuan Liu, Weinan Zhang, Jiangmiao Pang

Introduction

Despite the rapid advancements in humanoid robotics, achieving the fluid, versatile movement characteristic of biological systems remains a significant challenge. Most current state-of-the-art controllers are specialized: they excel at a single task, such as stable walking or a specific pre-programmed dance, but lack the athletic adaptability to switch seamlessly between behaviors. Furthermore, a major hurdle in "loco-manipulation" is the destabilizing effect of upper-body movement. When a user takes control of a robot's arms via teleoperation, the resulting momentum shifts often cause the locomotion controller to fail.

To address these limitations, we present HUGWBC (Humanoid’s Unified and General Whole-Body Control). HUGWBC is a unified reinforcement learning (RL) framework designed to serve as a versatile "base layer" for humanoid control. It enables a single policy to manage walking, jumping, and standing with fine-grained control over gait parameters, while remaining robust to external upper-body interventions.

The HUGWBC Framework

The core of HUGWBC lies in its ability to map a high-dimensional "behavior command space" to stable motor outputs. The system is built upon three primary innovations:

1. Extended Command Space

Traditional controllers typically receive simple task commands, such as target linear and angular velocities. HUGWBC extends this to include specific "Behavior Commands." These allow for fine-grained adjustments to:

  • Foot Parameters: Step frequency and swing height.
  • Posture Parameters: Body height, torso pitch, and waist yaw.

This granularity allows the robot to adapt its gait to specific environmental needs—such as lifting its feet higher to clear clutter or tilting its torso to maintain balance—without requiring a specialized controller for each variation.

2. Intervention Training & Noise Curriculum

To ensure the robot remains stable during teleoperation or manipulation tasks, the researchers introduced "Intervention Training." During the RL training process, the upper-body joints are subjected to a curriculum of stochastic noise. By gradually increasing the intensity of these upper-body disturbances, the policy learns to compensate for sudden changes in momentum, effectively decoupling the stability of the lower body from the actions of the arms.

3. Symmetry Loss

RL policies can occasionally converge on "unnatural" or asymmetric gaits that are mathematically stable but practically inefficient or aesthetically jarring. HUGWBC utilizes a mirror function and symmetry loss during training. This encourages the policy to produce mirrored actions (e.g., ensuring the left leg's trajectory is a symmetric counterpart to the right leg's), leading to more human-like and predictable locomotion.

Training Methodology

The system was trained within the NVIDIA IsaacGym simulator using an Asymmetric Actor-Critic (AAC) approach. In this setup, the "critic" network has access to privileged simulation data—such as exact ground friction coefficients and precise center-of-mass velocities—to guide the learning process. The resulting "actor" policy, however, relies solely on the noisy, real-world data available from onboard sensors, facilitating a smoother transition from simulation to reality (Sim-to-Real).

Key Insights and Results

Robustness Through Chaos

A pivotal finding of the research was that training with random upper-body noise actually produced more stable locomotion than training with structured human motion data (derived from the AMASS dataset). By preparing for "worst-case" arm movements, the controller became nearly immune to the destabilizing effects of real-time teleoperation.

Command Orthogonality

The researchers analyzed the interplay between different commands, discovering that many are "orthogonal." For example, a robot can significantly rotate its waist (yaw) without impacting its forward walking velocity. However, some constraints exist: lowering the body height reduces the hip's range of motion, which in turn limits the achievable torso pitch. Understanding these relationships is vital for developing higher-level planners.

Performance Metrics

  • Tracking Accuracy: In simulation, the robot demonstrated high precision, with walking velocity errors ($v_x$) as low as 0.03 m/s while simultaneously tracking eight different commands.
  • Disturbance Tolerance: The controller showed remarkable resilience, surviving external pushes of up to 500N for one second while in a standing posture.
  • Real-World Deployment: The policy was successfully deployed on a Unitree H1 humanoid. Even without external motion capture for feedback, the robot accurately adjusted its posture and gait in real-time based on operator commands.

Limitations and Future Directions

While HUGWBC represents a significant step toward a general foundation model for motor control, some challenges remain. Hopping, due to its inherent instability and asymmetric nature, still requires a specialized separate policy within this framework. Additionally, hardware limitations of the Unitree H1—specifically the hip structure—limit the robot's lateral (sideways) velocity tracking.

Future work will focus on integrating HUGWBC with high-level, vision-based planners. The goal is to create a system where a vision model perceives an obstacle and automatically issues the appropriate fine-grained behavior commands (e.g., "increase swing height" or "trigger jump") to the HUGWBC base layer.

Summary updated

Abstract

provide an in-depth analysis of how the various commands affect humanoid movement and offer insights into the relationships between these commands.To our knowledge, HUGWBC is the first humanoid whole-body controller that supports such versatile locomotion behaviors with high robustness and flexibility.
BibTeX
@misc{xue2025unified,
  author = {Yufei Xue and Wentao Dong and Minghuan Liu and Weinan Zhang and Jiangmiao Pang},
  title = {{A Unified and General Humanoid Whole-Body Controller for Fine-Grained Locomotion}},
  year = {2025},
  eprint = {2502.03206},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.03206},
}
Figure