# What Actually Drives Humanoid Motion Tracking Performance?

The humanoid motion-imitation pipeline is full of design choices that researchers make by convention, intuition, or cargo-culting from prior work — with little empirical grounding for any of them. A paper published today on arXiv by Fabio Amadio and Enrico Mingo Hoffman from HUCEBOT changes that, offering the first controlled empirical study of the modeling and training factors that collectively determine whether a [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) policy can reliably track diverse human motion references on a real humanoid.

The short answer: not all design choices are equal. Some variables — motion-command representation, action representation, and actuation profile — produce measurable differences in tracking quality. Others primarily affect actuation effort, training complexity, or physical interaction capability without moving the tracking accuracy needle much. The study deploys resulting policies zero-shot on the [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) G1, demonstrating diverse whole-body motion tracking, balance under external perturbations, and forceful physical interaction.

For the field, this is exactly the kind of controlled ablation infrastructure that has been missing. Most motion-imitation papers benchmark against their own prior work or present results that are unreproducible outside the authors' simulation stack.

---

## What Is YAHMP and Why Does It Matter?

YAHMP — the authors' acronym — is an open-source, modular framework purpose-built for training, evaluating, and deploying whole-body motion tracking policies on the Unitree G1. The framework is intentionally designed to support controlled comparisons: researchers can swap individual components (command representation, observation history, action parameterization, actuation profile) while holding everything else fixed.

This is not a trivial engineering contribution. One of the persistent problems in humanoid locomotion and motion-imitation research is that comparisons across papers are essentially meaningless — different simulators, different reward formulations, different robot models, and different motion datasets make it impossible to attribute performance differences to any single cause. YAHMP addresses this directly by defining a **nominal configuration** against which all variants are measured.

The framework also includes a defined test set of retargeted human motions, which gives the study a concrete evaluation protocol rather than the qualitative demonstrations that characterize most hardware deployments. The authors additionally compare their nominal policy against TWIST2 as an external baseline trained on the same motion set — a methodological choice that gives the results external validity.

---

## Which Design Choices Actually Matter?

The study examines six specific variables within the motion-imitation pipeline:

1. **Motion-command representation** — how the reference motion is encoded and fed to the policy
2. **Observation history** — how much temporal context the policy receives
3. **Action representation** — whether the policy outputs joint positions, velocities, or torques
4. **Actuation profile** — the PD gains and torque limits used during sim training
5. **Hand-force randomization during training** — domain randomization applied at the [end-effector](https://humanoidintel.ai/glossary/end-effector) level to improve interaction robustness
6. **Training approach** — the overall learning algorithm and curriculum structure

The paper's central finding, as reported in the abstract, is that these choices fall into two distinct categories: those with **clear tracking effects** and those that **mainly change actuation effort, training complexity, or physical interaction capability**. The practical implication is significant — teams spending engineering cycles optimizing choices in the second category are not buying tracking performance. They may be buying other things (safer physical contact, faster training convergence), but not the metric most users care about first.

The authors do not publish a definitive ranked list of "most important" factors in the abstract — that level of detail lives in the full paper — but the framing is clear: some knobs move the needle on motion tracking quality, and some don't.

---

## Zero-Shot Sim-to-Real on the Unitree G1

The most commercially relevant result is the [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) deployment on the real Unitree G1. The policies trained entirely in simulation transfer directly to hardware without any fine-tuning, demonstrating three capabilities the authors specifically highlight: diverse whole-body motion tracking, balance under external perturbations, and forceful interaction.

That last point — forceful interaction — is worth unpacking. Most [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) work for humanoids focuses on locomotion stability or upper-body manipulation in free space. Demonstrating forceful contact behavior in zero-shot transfer implies the training procedure successfully models contact dynamics well enough that the real robot can push, pull, or resist loads without sim-trained policies collapsing on contact. The hand-force randomization during training variable is almost certainly the relevant lever here, though the paper's contribution is precisely in showing *which* variables drive *which* outcomes.

The choice of the Unitree G1 as the deployment platform is also notable. The G1 has become something of a de facto open research platform for academic motion-control work — affordable relative to other full-size humanoids, with a reasonably complete SDK, and enough [degrees of freedom](https://humanoidintel.ai/glossary/degrees-of-freedom) to make whole-body motion tracking a meaningful challenge. YAHMP being built around the G1 maximizes the odds that other research groups can reproduce and extend this work.

---

## The Broader Reproducibility Problem This Work Addresses

The timing of this paper matters beyond its specific findings. The humanoid motion-imitation space has exploded in the last 18 months, with research groups at major labs and universities publishing results that are essentially impossible to compare. [Imitation learning](https://humanoidintel.ai/glossary/imitation-learning) pipelines from groups working on platforms like Figure AI's 02, Tesla's Optimus, or academic robots vary so substantially in their reward shaping, retargeting methodology, and simulation fidelity that claimed improvements could reflect any component in the stack.

YAHMP's open-source, modular design is a direct response to this. If the community adopts it — or even uses it as a reference point for structuring ablations — the cumulative knowledge value is substantial. The field gets shared vocabulary: "we varied X while holding the YAHMP nominal configuration fixed" means something precise.

The TWIST2 comparison is a good model for how this could work in practice. Using an external baseline trained on the same motion dataset is the minimum bar for a meaningful comparison, and the fact that the authors explicitly include it suggests awareness of the reproducibility gap.

---

## Key Takeaways

- **HUCEBOT researchers Fabio Amadio and Enrico Mingo Hoffman** released YAHMP, an open-source whole-body motion tracking framework built around the Unitree G1
- **Six design variables** were studied: motion-command representation, observation history, action representation, actuation profile, hand-force randomization, and training approach
- **Findings distinguish** choices with direct tracking accuracy effects from those that primarily affect actuation effort, training complexity, or physical interaction capability
- **Zero-shot hardware deployment** on the real Unitree G1 demonstrates whole-body motion tracking, perturbation recovery, and forceful interaction — all without sim-to-real fine-tuning
- **TWIST2** is included as an external baseline trained on the same motion set, giving the comparisons external validity
- **Open-source release** addresses the reproducibility crisis in humanoid motion-imitation research by providing a controlled, modular evaluation framework

---

## Frequently Asked Questions

**What is YAHMP?**
YAHMP is an open-source modular framework developed by HUCEBOT researchers for training, evaluating, and deploying whole-body motion tracking policies on the Unitree G1 humanoid robot. It defines a nominal configuration and allows researchers to swap individual design choices in a controlled way, enabling reproducible ablation studies.

**Which factors most affect humanoid motion tracking performance?**
According to the YAHMP study, design choices in the motion-imitation pipeline fall into two groups: those with clear effects on tracking accuracy (including motion-command representation, action representation, and actuation profile) and those that primarily affect actuation effort, training complexity, or physical interaction capability. The specific ranking is detailed in the full paper.

**What is zero-shot sim-to-real transfer in humanoid robotics?**
Zero-shot sim-to-real transfer means deploying a policy trained entirely in simulation directly onto physical hardware without any additional fine-tuning or adaptation. The YAHMP study demonstrates this on the Unitree G1, including under external perturbations and during forceful physical contact.

**Why is reproducibility a problem in humanoid motion-tracking research?**
Different research groups use different simulators, reward functions, motion datasets, and robot models, making it impossible to attribute performance differences to any single design choice. YAHMP's controlled, modular architecture and shared evaluation protocol are designed to address this gap.

**What is the Unitree G1 and why is it used here?**
The Unitree G1 is a full-size humanoid robot from [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) that has become a common platform for academic motion-control research due to its accessibility, documented SDK, and sufficient degrees of freedom for whole-body motion tracking tasks. YAHMP's G1-centric design maximizes the reproducibility potential of the study for other research groups.