Organizations: College of Engineering, Ocean University of China · Faculty of Information Science and Engineering, Ocean University of China · School of Artificial Intelligence, University of Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences
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.
Figures & tables
Figure 1: CHOREO converts heterogeneous skill sources—including motion datasets, RL policies, and generative models—into unified SkillMotion assets through source-specific adapters. Given a user instruction, an LLM planner selects and sequences these assets, while the CHOREO runtime executes them as long-horizon skill compositions on the Unitree G1.
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 50Hz . Solid and dashed arrows denote online execution and offline registration, respectively.
Method
2-Action SR
3-Action SR
5-Action SR
8-Action SR
Switch SR
Fall Rate
Δq (rad)
Direct Switching
75.0%
61.5%
58.3%
14.6%
67.2%
54.6%
0.0449
Fixed Hermite
85.0%
69.2%
44.4%
22.9%
68.8%
52.3%
0.0178
Fixed Stand Bridge
70.0%
73.1%
63.9%
41.7%
72.6%
41.5%
0.0180
Motion Matching
80.0%
73.1%
61.1%
31.2%
75.2%
44.6%
0.0171
CHOREO
100.0%
100.0%
91.7%
93.8%
96.7%
4.6%
0.0170
Table 1: Long-horizon skill-composition results on 130 fixed sequences. Sequence success requires completion of every constituent skill and transition. Higher success rates are better; lower fall rates and joint-position discontinuities are better.
Source
Import
Execution
Exec. SR
Motion Library
2969 / 2969
253 / 256
98.8%
RL Policy
50 / 50
50 / 50
100.0%
Diffusion Motion
50 / 50
50 / 50
100.0%
Total
3069 / 3069
353 / 356
99.2%
Table 2: Source adaptation across heterogeneous skill sources. Execution requires successful admission and rollout with frozen GMT.
Frozen Tracker
Success
Success Rate
GMT
30/30
100.0%
OpenTrack
14/30
46.7%
TWIST2
26/30
86.7%
HoloMotion
30/30
100.0%
SONIC
30/30
100.0%
H-ACT
30/30
100.0%
Table 3: Compatibility of the SkillMotion interface with six frozen trackers on 30 shared G1 trajectories.
Variant
5-Act. SR
8-Act. SR
Fall Rate
Δq (rad)
Full CHOREO
91.7 %
93.8 %
5.4 %
0.0162
w/o
Continuity
88.9%
95.8%
6.2%
0.0177
w/o
Compatibility
86.1%
75.0%
16.2%
0.0166
w/o
Table 4: Transition-component ablation on 130 fixed tasks. Higher success rates are better; lower fall rates and Δq are better.