Parametric
ResearchContact

Research


August 2025 · 15 min

Monosemantic Features in Vision-Language-Action Models

We trained transcoders on π₀-FAST to uncover interpretable features in robotics models, revealing two classes: "scene" features that highlight objectives and "ego" features focused on the robot itself.

John Newsom, Cody Swain

© 2026 Parametric PBC
PrivacyTerms