Intention estimation

Intention estimation is the trick that makes closed-loop decoder training possible without knowing the user's true internal command. During calibration the actual cursor or arm motion is a mixture of the user's intent and the decoder's errors, so it is a poor training label. Intention estimation replaces the observed kinematics with an inferred estimate of what the user meant — most commonly by assuming the user always intends to move straight toward the known target, rotating the training velocity vectors onto the target direction, and setting intended velocity to zero during hold periods. The decoder is then refit against these corrected intentions rather than the raw, error-corrupted movement.

This idea is the shared engine beneath ReFIT, CLDA, and self-recalibration: each is a different policy for guessing intent (straight-to-target during instructed tasks, retrospective target inference during free use) and for how often to refit. Its validity rests entirely on the intent assumption being a good approximation of the task; it works beautifully for structured center-out reaching and degrades when the true intended paths are not straight, curved deliberately, or when the assumed target is wrong.