Reproduce “PHASOR: Phase-Anchored Universal Action Representations for Humanoid Embodiments” 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

Kihyun Kim, Chaeyun Kim, Jongho Shin, Taeyoun Kwon, Junghyun Kim, Mijin Koo, Haon Park

PHASOR, a framework that leverages the intrinsic periodic structure of motion to create a unified, structured action representation for humanoids, enables robust cross-embodiment transfer and improves robot policy learning, achieving 90.3% R@1 for Human→Robot motion retrieval and reducing Mean Per Joint Position Error to 1.62 mm in motion imitation tasks.

Key insights

  • Decomposing motion into per-body-part phase parameters (amplitude, frequency, phase shift) creates an intrinsically morphology-invariant and semantically interpretable action space for humanoid robots.
  • Anchoring the shared phase manifold to human motion data, combined with robot-specific lightweight adapters, enables efficient and robust cross-embodiment transfer of behaviors.
  • Separating periodic (phase) and non-periodic (pose) components, with controlled interaction during decoding via FiLM, allows for a 'pose-free' phase manifold crucial for cross-embodiment alignment while still providing necessary postural context.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.