Reproduce “FunPhase: A Periodic Functional Autoencoder for Motion Generation via Phase Manifolds” 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
Marco Pegoraro, Evan Atherton, Bruno Roy, Aliasghar Khani, Arianna Rampini
FunPhase represents motion as a continuous spatio-temporal function whose latent channels are parameterized by phase, amplitude, frequency, and offset. The functional, skeleton-aware autoencoder improves contact consistency and variable-rate reconstruction, while diffusion in the periodic latent space achieves an FID of 0.51 on 100STYLE, 96% coverage on TRUEBONES ZOO, and lower foot sliding than several baselines.
Key insights
- Continuous functional decoding improves reconstruction and interpolation across temporal sampling rates, while graph encodings allow the same model to process different skeletal topologies.
- Periodic latent structure is especially helpful for temporal smoothness and contact stability: adding it reduces foot sliding and penetration even when frame-wise reconstruction error is not always the lowest.
- The benefits weaken as motion becomes more aperiodic, suggesting that future systems should adaptively combine periodic and non-periodic latent components rather than impose a single global rhythm.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.