CHOREO: Every Humanoid Skill as a Trajectory

Ziyi Sun, Jingwen Chen, Yuxi Wang, Xiuze Xia, Long Cheng, Zhaoxiang Zhang, Yujun Dong · N/A · 2026

Summary

Abstract Summary

Recent advances in humanoid robotics have produced diverse skills through reinforcement learning, motion imitation, and generative modeling. Yet these capabilities remain siloed because they are built around incompatible representations, interfaces, and controllers. We present CHOREO, a framework for training-free composition of heterogeneous humanoid skills. Our key observation is that, regardless of how a skill is learned, it can ultimately be expressed as an executable motion trajectory. Based on this observation, CHOREO converts each capability into SkillMotion, a unified representation that combines motion states, contacts, semantics, and boundary conditions. Skills are composed through direct continuation, cubic Hermite blending, or validated bridge motions, without retraining source models or updating models at test time. On Unitree G1 in MuJoCo, CHOREO organizes 2,950 admitted SkillMotion assets derived from heterogeneous sources and achieves 95.4\% sequence success across 130 multi-action tasks, including 93.8\% success on eight-action sequences. These results demonstrate that executable trajectories provide a scalable interface for accumulating and composing pretrained humanoid capabilities.

Key Points

Recent advances in humanoid robotics have produced diverse skills through reinforcement learning, motion imitation, and generative modeling.
Yet these capabilities remain siloed because they are built around incompatible representations, interfaces, and controllers.
We present CHOREO, a framework for training-free composition of heterogeneous humanoid skills
Our key observation is that, regardless of how a skill is learned, it can ultimately be expressed as an executable motion trajectory.
Based on this observation, CHOREO converts each capability into SkillMotion, a unified representation that combines motion states, contacts, semantics, and boundary conditions.

Method

CHOREO composes skills from heterogeneous sources into long-horizon reference motions for humanoid execution.

Given a natural-language instruction $x$, a SkillMotion library $\mathcal{L}$ constructed from skill sources $\mathcal{S}$, and pretrained motion Trackers $\mathcal{T}$, CHOREO constructs an execution program where $\mathcal{M}_{1:K}$ denotes the resolved SkillMotions, $k_{1:K}$ records their associated Tracker indices, and $\Gamma_{1:K-1}$ denotes the transitions between adjacent skills. As shown in Fig.

Method — CHOREO: Every Humanoid Skill as a Trajectory — Figure 2: Overview of CHOREO. Offline, the Source Adapter converts heterogeneous motion sources into canonical and validated SkillMotion assets. Online, an LLM task planner produces a normalized skill sequence from a natural-language instruction, and the runtime resolves it to registered SkillMotions (left) . CHOREO evaluates boundary compatibility, connects adjacent motions through local quintic seams, and inserts a stable-standing bridge when direct connection is unsuitable (middle) . Tracker-specific adapters convert the composed reference into the inputs required by the tracking controller (right) . Our experiments use GMT on the Unitree G1 at $50\,\mathrm{Hz}$. Solid and dashed arrows denote online execution and offline registration, respectively.

Experiments

We evaluate CHOREO along three dimensions: long-horizon skill composition, integration of heterogeneous skill sources, and compatibility with different low-level trackers. We further ablate the transition design to identify the components that support reliable long-horizon execution. All experiments are conducted on Unitree G1 in MuJoCo using frozen low-level policies.

CHOREO uses GMT for the long-horizon and source-adaptation benchmarks, whereas the tracker benchmark evaluates each frozen tracker without parameter updates.

Sources and Demonstrations

Figures are reproduced from Sun et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.