A Phenomenology of Neural Geometry in Vision-Language-Action Models

This research paper investigates the internal geometric structure of vision-language-action (VLA) policies used for robotic manipulation. By analyzing activation patterns through controlled sweeps and linear probes, the authors establish a taxonomy of seven task-relevant concepts. The findings reveal that these concepts are predominantly realized as multi-dimensional shapes influenced by physical constraints, such as wrist orientation forming an open arc due to joint limits rather than a full circle. The study confirms these geometric realizations across different VLA models on real-world robot data and demonstrates actionable methods for steering along these specific geometries to influence model behavior.