# Does Any SOTA Humanoid Policy Actually Balance on One Leg?

**Zero out of ninety.** That is the score eight released state-of-the-art general humanoid policies achieve on a new single-leg-balance benchmark published today on arXiv by Zhou et al. None hold a clean single-leg stance across 90 test motions — they survive only by stepping or hopping, reacting to imbalance rather than preventing it. A new policy called FDDC (First Deployable Dynamic-CoM) changes that score to 86 of 90, transfers to a real [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) G1, and the authors are releasing the full evaluation stack as a method-agnostic benchmark any lab can run.

The paper (arXiv:2608.00500, Zhou, Wang, Cui, Chen, Fan, Zhu, Li) targets a specific and previously unsolved hardware problem: the capture point — formally the extrapolated center of mass, or xCoM — has never been embedded directly in a deployable hardware policy because computing it requires base linear velocity, which no on-board sensor provides cleanly. The authors' key insight is that when this velocity is expressed relative to the support foot rather than a global frame, it cancels exactly, leaving an observation reconstructible entirely from joint encoders and IMU. That reformulation unlocks the first hardware-deployable dynamic-CoM observation.

---

## The Benchmark Exposes a Structural Gap in Whole-Body Control

The framing here is deliberately uncomfortable for the field. [Whole-body control](https://humanoidintel.ai/glossary/whole-body-control) research has made impressive strides in agile locomotion, dynamic transitions, and manipulation-while-moving. But the benchmark authors argue that the field has been measuring the wrong thing: recovery from imbalance rather than prevention of it.

The benchmark tests across nine stratified pose classes with 90 held-out motions evaluated in a simulator distinct from any policy's training environment — a sim2sim protocol specifically designed to expose policies that have overfit to their own training physics. Eight SOTA general policies, all publicly released, score zero clean stances. The authors do not name those eight policies in the abstract, but the public release of the benchmark stack will make individual scores verifiable.

This matters beyond academic benchmarking. Single-leg balance is a prerequisite for a wide range of real-world tasks: stair climbing with an irregular cadence, stepping over obstacles, manipulating objects while shifting weight to one foot, or simply recovering from a stumble without taking a step. A humanoid that can only stay upright by stepping is, in biomechanical terms, always falling forward. For deployment in unstructured environments — factories, homes, construction sites — that's a functional limitation, not a footnote.

---

## The FDDC Solution: Prevention Over Repair

FDDC's design philosophy is stated explicitly as "prevention over repair." The architecture places the dynamic-CoM observation directly in the actor network that runs on hardware, paired with a reward library the authors describe as translated term-by-term from human postural control research.

Training uses asymmetric FastSAC — a soft actor-critic variant with a privileged critic that has access to information during training that the deployed actor does not. Critically, no distillation step is required. The privileged critic guides training, but the actor that ships to hardware operates only on the sensor-reconstructible observation. This is a meaningful practical constraint: distillation pipelines add complexity and can introduce their own [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) gaps.

The ablation result is the most important number in the paper: removing the dynamic-CoM observation alone costs 40 points of clean single-leg balance performance. That is not a marginal contribution — it is the dominant driver of the policy's capability, larger than any other single component tested in ablation.

The policy achieves 86 of 90 on held-out motions across nine pose classes and validates on a physical Unitree G1. The authors do not report quantitative results for the hardware transfer beyond confirming it works, which is a gap worth noting — sim2sim benchmark performance and hardware performance can diverge significantly, and future work will need to close that reporting gap.

---

## Why the Benchmark Design Is as Important as the Policy

The method-agnostic benchmark may ultimately matter more to the field than FDDC itself. The authors score each policy in a simulator different from its training one — a deliberate choice to prevent simulator-specific overfitting from masquerading as generalizable capability. This mirrors the logic behind standardized evaluation in language modeling, where models are tested on held-out distributions rather than their training data.

For the humanoid industry, the absence of shared, reproducible balance benchmarks has been a real problem. Individual labs report impressive demos, but cross-lab comparison is nearly impossible. The FDDC benchmark offers a concrete step toward turning single-leg balance from a per-team capability claim into something the field can actually measure and track over time.

The authors explicitly frame this as "a step toward turning balance from a per-task trick into a capability the field can measure" — which is an accurate characterization of where the field currently sits. Most reported balance results are task-specific, demo-specific, and not reproducible outside the reporting lab's setup.

---

## Industry Implications

From a deployment standpoint, this research is directly relevant to every humanoid platform currently targeting factory and logistics environments. The Unitree G1 hardware transfer demonstrates the approach is not simulation-only. The sensor-only observation — encoders plus IMU, no external motion capture, no base velocity estimation — means the approach is portable to any platform with standard [proprioception](https://humanoidintel.ai/glossary/proprioception).

The deeper implication is architectural: if the dynamic-CoM observation is this large a driver of balance capability, it should be a standard component of any whole-body policy stack, not an optional add-on. The fact that eight released general policies omit it — and score accordingly — suggests the field has been building agility on top of an incomplete observation space.

For labs working on VLA-based whole-body policies, the question becomes whether high-level task reasoning can compensate for missing low-level postural signals. The benchmark result suggests it cannot, at least not for static single-leg balance. Prevention and repair are fundamentally different control problems.

---

## Key Takeaways

- **Eight released SOTA general humanoid policies score 0/90** on clean single-leg balance across the new benchmark; all survive only via stepping or hopping.
- **FDDC scores 86/90** on held-out motions and transfers to a real Unitree G1, using only encoder and IMU observations — no base velocity sensor required.
- **The dynamic-CoM observation is the single largest driver**: ablating it alone costs 40 points of clean single-leg balance performance.
- **The sensor reformulation is the key insight**: expressing base velocity relative to the support foot cancels the unobservable component, making the capture point (xCoM) reconstructible on hardware for the first time.
- **Training uses asymmetric FastSAC with a privileged critic and no distillation** — a practically significant constraint for deployment pipelines.
- **The method-agnostic benchmark** evaluates policies in a simulator distinct from their training environment, enabling cross-lab reproducible comparison for the first time in this capability domain.
- **The benchmark stack is being released publicly**, which will pressure the field to report single-leg balance performance as a standard metric.

---

## Frequently Asked Questions

**What is the FDDC humanoid policy and what does it do?**
FDDC (First Deployable Dynamic-CoM) is a reinforcement learning policy for humanoid robots that achieves stable single-leg balance by incorporating the capture point (xCoM) directly into the hardware-deployable actor. It scores 86 out of 90 on a new held-out benchmark and runs on a real Unitree G1 using only joint encoder and IMU data.

**Why have previous humanoid policies failed at single-leg balance?**
According to the paper, existing general policies respond to imbalance reactively — by stepping or hopping — rather than preventing it. The capture point, which measures where the extrapolated center of mass is relative to the support base, has never been embedded in a hardware policy because it requires base linear velocity that on-board sensors do not directly provide.

**How does FDDC make the capture point computable on hardware?**
By expressing base linear velocity relative to the support foot rather than a global reference frame, the velocity term cancels out exactly. The resulting observation is reconstructible purely from joint encoders and IMU, which all standard humanoid platforms carry.

**What is the sim2sim benchmark and why does it matter?**
The benchmark evaluates policies in a simulator different from the one used for training. This prevents policies from exploiting simulator-specific physics artifacts to achieve high scores that would not generalize. It is designed to be method-agnostic and reproducible, allowing cross-lab comparison of single-leg balance capability.

**What does this mean for humanoid robots in real-world deployment?**
Single-leg balance is a prerequisite for stair climbing, obstacle stepping, and weight-shifting manipulation tasks. Platforms that rely on reactive stepping to stay upright face fundamental limitations in unstructured environments. The FDDC approach, using only standard on-board sensors, is directly portable to any humanoid platform — making this a practical capability gap the broader industry will need to address.