# Does CHOREO Solve Humanoid Skill Fragmentation Without Retraining?

**95.4% sequence success across 130 multi-action tasks** — that is the headline result from CHOREO, a training-free framework for composing heterogeneous humanoid skills published today on arXiv by Ziyi Sun, Jingwen Chen, Yuxi Wang, Xiuze Xia, Long Cheng, Zhaoxiang Zhang, and Yujun Dong. Tested on the [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) G1 inside MuJoCo simulation, the system organizes 2,950 SkillMotion assets drawn from incompatible sources — reinforcement learning policies, motion imitation datasets, and generative models — and chains them into complex multi-step behaviors without touching a single source model at test time. On the hardest benchmark subset, eight-action sequences, CHOREO achieves 93.8% success. The core insight is disarmingly straightforward: whatever method produced a skill, its output can ultimately be expressed as an executable motion trajectory. That observation, if it holds at scale, reframes the entire skill accumulation problem for [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) systems.

---

## The Fragmentation Problem CHOREO Is Attacking

Anyone building general-purpose humanoid behavior knows the wall: you have a locomotion policy trained with RL, a manipulation primitive learned through [imitation learning](https://humanoidintel.ai/glossary/imitation-learning), and a pose generated by a diffusion model. Each lives in its own latent space, expects its own input format, and terminates with no clean handoff to the next skill. Composing them means either retraining from scratch — expensive and brittle — or building bespoke adapter layers that break every time a source model is updated.

This is not a niche academic concern. It is the central engineering bottleneck for every team deploying humanoids in unstructured environments. The ability to accumulate skills without retraining is effectively a prerequisite for scaling robot capability the way the software industry scales libraries.

CHOREO's authors frame this as a representation problem: the field has been building skills in incompatible formats when a lowest-common-denominator exists. Their answer is **SkillMotion**, a unified representation that encodes motion states, contact information, semantics, and boundary conditions for each discrete capability.

---

## How SkillMotion Composition Actually Works

The framework offers three composition mechanisms, each suited to different inter-skill boundary conditions:

1. **Direct continuation** — when boundary conditions between consecutive skills are already compatible, skills are concatenated without modification.
2. **Cubic Hermite blending** — when velocity or position discontinuities exist at skill boundaries, a cubic Hermite interpolant smooths the transition in trajectory space.
3. **Validated bridge motions** — for cases where blending alone cannot guarantee physical plausibility, the system inserts short, pre-validated intermediate trajectories.

Critically, none of these mechanisms require gradient updates to the underlying models. The source policies, motion clips, or generative outputs are treated as black boxes that emit trajectories; CHOREO operates entirely in trajectory space downstream of them. This is what the authors mean by "training-free" — it is a composition layer, not a meta-learning system.

The [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) implications deserve scrutiny here. All reported results are in MuJoCo with the Unitree G1. The boundary condition validation logic that underpins the bridge motion selection has not been tested against hardware contact dynamics, actuator compliance, or sensor noise. Whether cubic Hermite blending produces physically stable transitions on a real G1 — where joint torque limits and actuator latency will distort the trajectory — is an open question the paper does not address.

---

## Interpreting the Numbers

The 2,950 SkillMotion assets is a meaningful scale claim, but context matters. These are derived assets — conversions of existing skills from heterogeneous sources — rather than 2,950 independently authored behaviors. The diversity and quality of those source skills will directly cap what CHOREO can compose. If the underlying library is heavy on locomotion variants and thin on [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation), the composition success rate across manipulation-heavy tasks will reflect that gap, not CHOREO's architecture.

The 95.4% aggregate success figure and the 93.8% on eight-action sequences are strong for a simulation benchmark. The 130-task evaluation suite is described as "multi-action," but the paper does not fully specify what failure modes were observed in the roughly 4-6% of failing sequences. Whether failures cluster around specific skill types, composition methods, or task lengths is information that matters for anyone trying to deploy this.

---

## Industry Trajectory: Why This Architecture Is Worth Watching

The broader significance of CHOREO is not any single success rate — it is the architectural bet it represents. The paper is proposing **trajectory space as a universal interface layer** for humanoid capabilities. If that abstraction holds, it enables a modular skill economy: teams specialize in generating high-quality skills via whatever method they prefer (RL, motion capture, VLA-driven teleoperation), and composition frameworks like CHOREO handle assembly.

This has direct implications for companies building [physical AI](https://humanoidintel.ai/glossary/physical-ai) stacks. A training-free composition layer that treats skills as trajectory primitives is architecturally compatible with [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model) outputs — a VLA could select and sequence SkillMotion assets based on language instructions without requiring end-to-end fine-tuning. That is a plausible integration path for teams at [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) or [Skild AI](https://humanoidintel.ai/companies/skild-ai) whose foundation models already emit action trajectories.

The Unitree G1 is also a telling hardware choice. It is the most accessible research-grade full-size humanoid platform currently available, which means independent replication of these results — the real credibility test for any simulation paper — is genuinely feasible. That lowers the bar for follow-up work and increases the probability that real-world validation appears within months rather than years.

---

## Key Takeaways

- **CHOREO achieves 95.4% sequence success** across 130 multi-action tasks on Unitree G1 in MuJoCo, with 93.8% on eight-action sequences specifically.
- **No retraining required**: composition operates entirely in trajectory space, treating source models as black boxes.
- **2,950 SkillMotion assets** from heterogeneous sources (RL, motion imitation, generative models) are organized under a single representation framework.
- **Three composition mechanisms** — direct continuation, cubic Hermite blending, and bridge motions — handle different boundary condition mismatches.
- **Sim-only results**: all benchmarks are in MuJoCo; real-hardware validation on the G1 has not been demonstrated.
- **Strategic implication**: trajectory-space interfaces may be the missing abstraction layer that makes modular humanoid skill libraries practically viable.

---

## Frequently Asked Questions

**What is CHOREO and what problem does it solve?**
CHOREO is a training-free framework for composing heterogeneous humanoid skills. It addresses the fragmentation problem where skills built with different methods — reinforcement learning, motion imitation, generative modeling — use incompatible representations and cannot be easily chained into multi-step behaviors without retraining.

**How does SkillMotion unify different skill types?**
SkillMotion is CHOREO's unified representation. It converts each skill, regardless of its origin, into a format that captures motion states, contact information, semantics, and boundary conditions, making it interoperable with other SkillMotion assets.

**What hardware and simulation platform were used?**
All reported experiments use the Unitree G1 humanoid robot simulated in MuJoCo. No real-hardware results are reported in this paper.

**What success rates did CHOREO achieve?**
CHOREO achieved 95.4% sequence success across 130 multi-action tasks, and 93.8% success specifically on the most challenging eight-action sequence tasks.

**How does CHOREO relate to VLA and foundation model approaches?**
CHOREO is architecturally compatible with VLA systems because both operate on action trajectories. A foundation model could theoretically select and sequence SkillMotion assets via language instructions without end-to-end fine-tuning, making CHOREO a plausible integration layer for teams building on top of models from companies like Physical Intelligence or Skild AI.

**Is CHOREO open source?**
The paper (arXiv:2609.22274) does not specify code release terms. Readers should check the associated repository, if any, linked from the arXiv abstract page.