Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HuB: Learning Extreme Humanoid Balance

Tong Zhang, Boyuan Zheng, Ruiqian Nai, Yingdong Hu, Yen-Jen Wang, Geng Chen, Fanqi Lin, Jiongye Li, Chuye Hong, Koushil Sreenath, Yang Gao

TL;DR

HuB tackles extreme humanoid balance by integrating reference-motion refinement, balance-aware policy learning, and sim-to-real robustness training. It targets three bottlenecks: reference-motion errors, morphological mismatch, and the sim-to-real gap caused by sensor noise and unmodeled dynamics. Validation on Unitree G1 demonstrates stable performance on challenging quasi-static poses like Swallow Balance and Bruce Lee's Kick, outperforming baselines that fail to complete tasks. Ablation studies confirm the necessity of each component, and real-world tests show strong sim-to-real transfer and long-horizon reliability. This framework advances reliable quasi-static balance in humanoids and offers a blueprint for robust, transfer-ready embodied control under real-world sensor noise.

Abstract

The human body demonstrates exceptional motor capabilities-such as standing steadily on one foot or performing a high kick with the leg raised over 1.5 meters-both requiring precise balance control. While recent research on humanoid control has leveraged reinforcement learning to track human motions for skill acquisition, applying this paradigm to balance-intensive tasks remains challenging. In this work, we identify three key obstacles: instability from reference motion errors, learning difficulties due to morphological mismatch, and the sim-to-real gap caused by sensor noise and unmodeled dynamics. To address these challenges, we propose HuB (Humanoid Balance), a unified framework that integrates reference motion refinement, balance-aware policy learning, and sim-to-real robustness training, with each component targeting a specific challenge. We validate our approach on the Unitree G1 humanoid robot across challenging quasi-static balance tasks, including extreme single-legged poses such as Swallow Balance and Bruce Lee's Kick. Our policy remains stable even under strong physical disturbances-such as a forceful soccer strike-while baseline methods consistently fail to complete these tasks. Project website: https://hub-robot.github.io

HuB: Learning Extreme Humanoid Balance

TL;DR

HuB tackles extreme humanoid balance by integrating reference-motion refinement, balance-aware policy learning, and sim-to-real robustness training. It targets three bottlenecks: reference-motion errors, morphological mismatch, and the sim-to-real gap caused by sensor noise and unmodeled dynamics. Validation on Unitree G1 demonstrates stable performance on challenging quasi-static poses like Swallow Balance and Bruce Lee's Kick, outperforming baselines that fail to complete tasks. Ablation studies confirm the necessity of each component, and real-world tests show strong sim-to-real transfer and long-horizon reliability. This framework advances reliable quasi-static balance in humanoids and offers a blueprint for robust, transfer-ready embodied control under real-world sensor noise.

Abstract

The human body demonstrates exceptional motor capabilities-such as standing steadily on one foot or performing a high kick with the leg raised over 1.5 meters-both requiring precise balance control. While recent research on humanoid control has leveraged reinforcement learning to track human motions for skill acquisition, applying this paradigm to balance-intensive tasks remains challenging. In this work, we identify three key obstacles: instability from reference motion errors, learning difficulties due to morphological mismatch, and the sim-to-real gap caused by sensor noise and unmodeled dynamics. To address these challenges, we propose HuB (Humanoid Balance), a unified framework that integrates reference motion refinement, balance-aware policy learning, and sim-to-real robustness training, with each component targeting a specific challenge. We validate our approach on the Unitree G1 humanoid robot across challenging quasi-static balance tasks, including extreme single-legged poses such as Swallow Balance and Bruce Lee's Kick. Our policy remains stable even under strong physical disturbances-such as a forceful soccer strike-while baseline methods consistently fail to complete these tasks. Project website: https://hub-robot.github.io
Paper Structure (25 sections, 1 equation, 4 figures, 9 tables)

This paper contains 25 sections, 1 equation, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Learning Framework for Extreme Humanoid Balance
  4. Background
  5. Reference Motion Refinement
  6. Balance-Aware Policy Learning
  7. Sim-to-Real Robustness Training
  8. Experiments
  9. Experiments Setup
  10. Simulation Results
  11. Real-World Results
  12. Conclusion
  13. Limitations
  14. Real Robot Setup
  15. HuB Details
  16. ...and 10 more sections

Figures (4)

  • Figure 1: Extreme Balance Tasks. HuB enables humanoids to perform extreme quasi-static balance tasks with high stability. (a) Swallow Balance: holding a challenging T-shaped pose with the torso extended horizontally; (b) Bruce Lee’s Kick: executing a high kick with full leg extension while balancing on one foot; (c) Ne Zha Pose: a martial arts-inspired one-legged stance with a raised arm; (d) High Knees; (e) Single-Leg Stand; (f) Deep Squat. Videos are available at: hub-robot.github.io
  • Figure 2: HuB Overview. To tackle the challenges of extreme balance tasks on humanoids, HuB integrates three components: (a) a motion refinement process that improves the quality and feasibility of reference motions; (b) a balance-aware policy learning strategy that enables stable execution of challenging balance motions; and (c) a robustness training mechanism to improve sim-to-real consistency and deployment stability.
  • Figure 3: Retargeting Comparison.
  • Figure 4: External Perturbations.