Spectral Skills

Spectral skills are a latent interface between a high-level planner and a humanoid whole-body controller, learned by predicting the motion that follows a segment rather than reconstructing it. One frozen controller tracks diverse motions, chains skills without a transition policy, and composes new behaviors by adding spectral directions to a base skill, with hardware demonstrations on a Unitree G1.

A G1 walking with its right arm raised, added by steering the base walk along one spectral direction.

Behind the paper

Read the blog: Spectral Skills: A Command Space for Humanoid Control.

Publication

Feiyang Wu*, Chenxiao Gao*, Chen Yang, Ye Zhao, Bo Dai, Anqi Wu
arXiv preprint, 2026

Links