Intention estimation (ReFIT)
Intention estimation supplies the surrogate labels that closed-loop adaptation needs by inferring, at each moment, what the user was actually trying to do — information that offline calibration lacks once feedback and free strategy enter. The canonical instance is ReFIT (recalibrated feedback intention-trained Kalman filter): during a calibration block the target is known, so one assumes the user always intends to move straight toward it. The decoded velocity vectors are therefore rotated to point at the target (and set to zero when the target is being held), and the decoder is retrained on these corrected, intention-consistent kinematics.
This deceptively simple correction produced one of the largest single jumps in intracortical cursor performance, because it removes the mismatch between the movements the naive decoder inferred and the movements the user meant to make under feedback. Its scope, however, is bounded by the assumption that the goal is known and that the user is monotonically pursuing it, which holds in instructed reaching and assistive training but breaks down for free, goal-ambiguous, or exploratory behaviour; generalizations replace the fixed-target assumption with probabilistic goal inference or error-signal-based label refinement.
ReFIT is often described as improving the decoder, but conceptually it corrects the training labels; the same intention-estimation idea underlies most modern closed-loop recalibration, including approaches that infer the goal probabilistically instead of assuming it.