How to teach different robots the same motion language?

Most robot policies learn action embeddings specific to their own body, making it hard to transfer skills between different robots. This work treats the action embedding space itself as a core design goal by factorizing motion into periodic (cyclic) and non-periodic components using FFT coefficients and pose conditioning. By anchoring multiple humanoid robots to a shared human-pretrained manifold, they created a unified action space that works across embodiments, improving cross-robot retrieval and downstream task performance.