Abstract
Background: Learning often occurs under instruction. In the neuromotor domain, this entails moving with intent to accomplish the instructed goal, repeatedly, building an expectation and predicting future events. In autism, however, neuromotor development is delayed and their sense of kinesthetic reafference affected by random noise in the returning motor stream. This impedes fluid, moment-to-moment predictive control, creating a gap between mental intent and physical volition. However, our lab discovered that uninstructed, goal-less spontaneous movements, typically embedded in purposeful actions, are spared and plastic across the autism spectrum. How can we leverage these motions’ unintended consequences, to help autistics self-discover a goal and fluidly complete actions on their own volition? Methods: We built a closed-loop interactive assay: the Torres Agentic Learning (TAL) task. In this task, without instructions, 33 children had to self-discover the goal and complete the action of watching movies under their own volitional control. We continuously recorded their hand motions as they caused media to play, first through naïve wondering and spontaneous unintended consequences from surprising events and then, learning through deliberate error-correction relative to the self-discovered goal. Results: All participants learned through the TAL paradigm. They self-discovered the goal and gained prediction of their motor code, lowering the entropy and enhancing the signal-to-noise ratio of their hand speed temporal streams. Conclusions: Uninstructed goal-less agentic learning mediated by spontaneous motions and surprises is feasible in both autistics and typically developing children. This work offers both an assay and biometrics to track the effectiveness of the agentic learning progression.