What Matters in Humanoid General Motion Tracking? An Empirical Study

2026-07-22

AI Summary

This study addresses the lack of systematic evaluation regarding key modeling and training design choices that affect whole-body motion tracking and balance control in humanoid robots. Leveraging the open-source, modular reinforcement learning framework YAHMP, the authors conduct controlled empirical experiments on the Unitree G1 platform to systematically analyze how critical factors—including motion command representation, observation history, action space formulation, actuation configuration, hand-force randomization, and training strategy—differentially influence tracking accuracy, energy efficiency, and interaction capability. The resulting policies outperform the TWIST2 baseline on a human motion reproduction benchmark and achieve zero-shot transfer to the physical robot, demonstrating diverse locomotion skills, robust disturbance rejection, and effective physical interaction.

Abstract

Humanoid general motion tracking requires policies that can follow diverse whole-body references while maintaining balance. Building such policies involves many practical design choices, and their individual effects are often hard to assess. We address this issue with an empirical study of common modeling and training factors used in recent humanoid motion-imitation pipelines. To make the study controlled and reproducible, we developed YAHMP, an open-source modular framework for training, evaluating, and deploying whole-body motion tracking policies on the Unitree G1. Within YAHMP, we define a nominal configuration and compare variants that differ in motion-command representation, observation history, action representation, actuation profile, hand-force randomization during training, and training approach. We evaluate the resulting policies on a test set of retargeted human motions and compare the nominal policy with TWIST2 as an external baseline trained on the same motion set. The results distinguish choices with clear tracking effects from choices that mainly change actuation effort, training complexity, or physical interaction capability. Finally, we deploy YAHMP policies zero-shot on the real Unitree G1, demonstrating diverse whole-body motion tracking, balance under external perturbations, and forceful interaction.

Problem

• humanoid motion tracking
• whole-body control
• policy design
• motion imitation
• balance maintenance

Innovation

• humanoid motion tracking
• empirical study
• modular framework
• zero-shot deployment
• whole-body control