Sharpa's SALA aligns latent actions across human and robot video for dual-axis scaling

Jiachun Jin · hf · 2026-11-17

A paper from Sharpa Robotics and the SJTU-Sharpa joint lab introduces SALA. Most latent action models learn latent actions via video reconstruction, so representations cluster by visual appearance rather than underlying action, making it hard to align human motions with robot counterparts. SALA uses language-based semantic similarity to align latent actions from egocentric human and robot videos into a shared latent action space that supervises embodied foundation model training, enabling dual-axis scaling and improving robot task performance while avoiding costly explicit pixel prediction.

Original post →

More from Embodied

Embodied channel →