INSIGHT: INference-time Sequence Introspection for Generating Help Triggers in Vision-Language-Action Models
INSIGHT is a framework for Vision-Language-Action (VLA) models to request human help during inference by analyzing token-level uncertainty signals. It uses entropy, log-probability, and Dirichlet estimates to predict when a robot should seek intervention. The study shows that sequential modeling of these uncertainties using transformers yields superior predictive power compared to static scores.