Little geometry. The encoding can still act like a language embedding and preserve coarse task identity for familiar ID cases.
When exact motion becomes the wrong prior.
Predict what the action head needs, then watch information resolution reshape the same robot task. Normalized teaching model; not benchmark data.
WebGL is unavailable. The computed regime evidence and explanation remain available in the telemetry panel.
Resolution changes what the action head can inherit.
Explicit geometric intent. Waypoints can preserve where the arm should move while leaving room to compose around a changed scene.
Exact motion and high training SNR. Under OOD shift, that precision may bind to stale geometry or hallucinated detail.
Transfer the rule to a moved drawer handle.
The task identity stays “open the drawer,” but the handle moves to the opposite side. Which prior preserves enough geometry to act without freezing the training trajectory?
Run the trajectory above to unlock transfer practice.