Reproduce “MoReFlow: Motion Retargeting Learning through Unsupervised Flow Matching” 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
Wontaek Kim, Tianyu Li, Sehoon Ha
MoReFlow frames cross-morphology motion retargeting as unsupervised flow matching between character-specific, discretized motion spaces. Using VQ-VAE tokenizers and condition-dependent pseudo-pairing, it transfers unpaired motions across humanoids, quadrupeds, and robots while supporting different style and task objectives; it reaches FID scores of 33.1 for locomotion and 38.9 for upper-body motion, with naturalness ratings of 89.6% and 92.9%.
Key insights
- Separating character-specific motion vocabularies from cross-character correspondence makes the framework modular and supports reverse or chained retargeting, rather than requiring a shared skeleton.
- Discrete tokenization is more stable and produces better diversity and alignment than applying flow matching directly to high-dimensional trajectories or continuous latent states.
- Feature-based coupling allows one source motion to produce different target behaviors depending on whether local motion, world-frame placement, root movement, or height is prioritized.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.