What Matters in Humanoid General Motion Tracking? An Empirical Study
Empirical study using YAHMP framework shows how modeling and training choices impact humanoid motion tracking performance, validated on Unitree G1.
Key Findings
Methodology
This study employs the open-source YAHMP framework to systematically analyze key factors such as motion command representation, observation history, action residuals, actuation parameters, hand-force randomization, and training strategies. Multiple policy variants are trained using PPO on a diverse motion dataset retargeted from human MoCap data, evaluated in simulation and on hardware. The framework ensures controlled, reproducible experiments, comparing variants against a baseline TWIST2 trained on the same data, with metrics including tracking errors and torque profiles.
Key Results
- Including reference joint velocities improves tracking, reducing base position error by 10-14% and key-body errors by about 9%. Observation history of 10 steps yields optimal results; longer history offers no clear benefit. Residual actions outperform non-residual, and mechanically based actuation profiles balance accuracy and torque peaks. Teacher-student training reduces errors by ~8%. In real robot tests, adding hand-force randomization enhances external load handling, with maximum payload increasing from 4kg to 6kg and deviation dropping from 15.5° to 6.6°.
- Compared to TWIST2, YAHMP achieves superior robustness in complex motions and external disturbances. Torque analysis shows mechanical models maintain tracking while reducing peak torques by ~13%. Simulation and real-world tests confirm that design choices significantly influence performance, with the combined approach yielding the best results.
- The experimental results demonstrate that systematic parameter analysis and training strategies directly improve humanoid motion tracking, balancing accuracy, robustness, and efficiency. The framework’s modularity allows for targeted optimization, facilitating deployment in real-world scenarios such as dynamic locomotion, interaction, and manipulation tasks.
Significance
This work advances humanoid control by providing a comprehensive, systematic evaluation of modeling and training factors, bridging the gap between simulation and real-world deployment. It addresses longstanding challenges in achieving robust, accurate, and adaptable motion tracking across diverse tasks. The insights and tools developed lay a foundation for future research and industrial applications, enabling robots to operate safely and effectively in complex, unstructured environments, thus accelerating the transition from laboratory prototypes to practical systems.
Technical Contribution
The core contribution is the development of YAHMP, a modular, flexible framework that enables systematic ablation studies of key design choices in humanoid motion tracking. It integrates mechanical-based actuation modeling with deep RL, incorporating observation history, residual actions, and external force randomization. The framework supports multi-task training, teacher-student paradigms, and sim-to-real transfer, providing a comprehensive platform for optimizing control policies with high fidelity and robustness. The comparative analysis offers new insights into the interplay of system components, guiding future control design.
Novelty
This is the first comprehensive, controlled evaluation of multiple key factors affecting humanoid motion tracking within a unified framework. Unlike prior work focusing on isolated components, this study systematically compares motion command representations, observation histories, actuation models, and training strategies, revealing their relative importance. The integration of mechanical actuation models with RL and the validation on real hardware further distinguish this work, providing a holistic understanding of system design impacts.
Limitations
- The experiments are primarily conducted on the Unitree G1 platform, which may limit generalization to other robots with different morphologies or actuation systems. Real-world environments with highly dynamic or unpredictable conditions remain challenging, requiring further robustness enhancements.
- Training relies heavily on large simulation datasets, which are computationally expensive and may not fully capture all real-world variabilities. Reducing simulation-reality gap and improving sample efficiency are ongoing challenges.
- External disturbances and complex terrains are only partially addressed; future work should incorporate richer perception and planning modules to handle more diverse scenarios.
Future Work
Future research will focus on integrating multi-modal perception, adaptive control, and autonomous parameter tuning to enhance robustness and generalization. Extending the framework to multi-robot systems and more complex environments, including outdoor and unstructured terrains, is planned. Additionally, exploring transfer learning and online adaptation strategies will further improve real-world applicability and reduce reliance on extensive simulation data.
AI Executive Summary
Humanoid robots hold promise for diverse applications, from industrial automation to social interaction. However, achieving robust, accurate, and versatile motion tracking remains a significant challenge. Traditional control methods often struggle with complex, dynamic environments, prompting the adoption of deep reinforcement learning (RL) techniques. This study introduces YAHMP, an open-source, modular framework designed to systematically evaluate the impact of key modeling and training choices on humanoid motion tracking performance.
Using YAHMP, researchers conducted extensive experiments on the Unitree G1 robot, analyzing factors such as motion command representation, observation history, residual actions, actuation parameters, and external force randomization. The framework enabled controlled ablation studies, revealing that including reference joint velocities and observation histories significantly improves tracking accuracy. Mechanical-based actuation models balance performance and torque peaks, while teacher-student training strategies further enhance robustness.
Compared to the baseline TWIST2, YAHMP policies demonstrated superior adaptability and stability, especially under external perturbations. Real-world deployment validated simulation results, with the robot successfully executing complex motions like dancing, crouching, and loco-manipulation, even on unfamiliar terrains like soft mattresses. Notably, hand-force randomization enabled the robot to sustain external loads up to 6kg, showcasing enhanced physical interaction capabilities.
This work provides valuable insights into the design of humanoid control systems, emphasizing the importance of systematic parameter analysis and training strategies. The modularity of YAHMP facilitates targeted optimization, paving the way for more autonomous, resilient robots capable of operating in unstructured environments. Future directions include integrating perception modules, online adaptation, and multi-robot coordination, aiming to accelerate humanoid robots' transition from controlled labs to real-world applications.
Deep Dive
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.