Monosemantic Features in Vision-Language-Action Models
August 2025 • 15 minutes
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.