Reproduce “OmniMimic: Dynamics-completed Motion Augmentation for Multi-style Omnidirectional Quadruped Locomotion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sheng Wu, Guoqiang Zhao, Zhe Yang, Fei Teng, Zhikun Zhou, Yanlin Yang, Zheng Fang, Hong Zheng and 2 more
OmniMimic expands a small set of animal-inspired quadruped demonstrations into a single policy covering four gait styles and three-dimensional velocity commands. Its dynamics-completed augmentation supplies physically consistent force and torque supervision for reversed and reflected motions, while command curriculum and shared residual branches improve omnidirectional control; in simulation, velocity-grid RMSE falls 63.1% versus APEX and foot-position RMSE falls 12.9%, with qualitative transfer to a Unitree Go2.
Key insights
- Physical supervision is essential when augmenting demonstrations: constrained completion improves contact-force RMSE by 5.6% and touchdown-impulse RMSE by 7.1% over directly reversed physical labels.
- Command expansion is the main driver of broad omnidirectional tracking, while dynamics augmentation primarily improves fidelity to the demonstrated gait motions.
- The method exposes a trade-off between style and physical fidelity and unrestricted command coverage: adding dynamics augmentation improves reference-motion metrics but slightly increases grid command error.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.