OmniContact Chaining Meta-Skills via Contact Flow
for Generalizable Humanoid Loco-Manipulation

* Equal contributors   ✉ Corresponding authors

1 Noitom Robotics    2 The Hong Kong University of Science and Technology    3 Wuhan University    4 The University of Hong Kong

Overview #

Overview of OmniContact: CF-Gen synthesizes contact-flow segments and CF-Track executes them with a low-level policy.
Contact-flow examples for carry, push, kick, locomotion, and recovery tasks.
98.7%

Carry Box success

82.5%

Push Suitcase success

~40 min

Long-horizon endurance

50 Hz

Closed-loop monitoring

Live MuJoCo Session #

Live MuJoCo WASM + ONNX Policy Viewer

Open full page

OmniContact Dataset Visualization #

Dataset t-SNE

74,641 MoCap clips embedded by skill labels.

Loading points...
1,274

Valid sequences

22.29 h

Total motion duration

7.22M

Total object frames

90 Hz

Paired capture

Generalized Meta Skill #

Meta-skill
Generalization

Meta Skill Chaining #

Recovery #

VLM Integration #

Move the baseket to the goal position, and put the cylidner into it
Move the soccer ball to the goal
Arrange the scattered boxes into heart shape
20x
Arrange objects as N-O-I-T-O-M with color-matched box targets

Interactive Unitree G1 Motion Sequences

Open full page

Baseline Comparison #

Citation #

@misc{yu2026omnicontactchainingmetaskillscontact,
      title={OmniContact: Chaining Meta-Skills via Contact Flow for Generalizable Humanoid Loco-Manipulation}, 
      author={Runyi Yu and Xiaoyi Lin and Ji Ma and Yinhuai Wang and Koukou Luo and Jiahao Ji and Huayi Wang and Wenjia Wang and Runhan Zhang and Ping Tan and Ting Wu and Ruoli Dai and Qifeng Chen and Lei Han},
      year={2026},
      eprint={2606.26201},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.26201}, 
}