Motor Intentionality as a Developmental Constraint on Curiosity-Driven Exploration

Curiosity-driven exploration gives artificial agents a reason to act without external reward, but its dominant computational formulation reduces novelty to prediction error. I argue that this measure is systematically agnostic about two features that human development appears to respect from the outset: whether the agent caused the change it failed to predict, and whether the change bears on what the agent can currently do. Examining this gap through motor intentionality—perceptual anticipation organized by an agent's own action possibilities—I propose three criteria for anticipation mechanisms in active vision: self/other attribution, action-relevance filtering, and repertoire-conditioning. Each is grounded in a documented developmental phenomenon and paired with an operational test runnable on existing curiosity-driven agents, offering a diagnostic complement to statistical novelty.