ReFIT / intention-corrected training
ReFIT (recalibrated feedback intention-trained Kalman filter) is a training procedure, not a new filter, that fixes a bias problem in fitting velocity decoders from training data. When a decoder is calibrated by having the user watch or attempt movements to known targets, the recorded (or first-pass decoded) velocities do not match the user's true moment-to-moment intention — for instance, once the cursor is near the target the user intends to stop, and while off-course they intend to correct toward the target. ReFIT retrospectively rotates the training velocity vectors to point straight at the known target and zeroes them at the target, then refits the observation model against these intention-corrected kinematics.
The effect is to teach the decoder the mapping the user actually wants to drive in closed loop rather than the kinematics that happened to occur, and it produced a marked, repeatable improvement in cursor performance over the standard velocity Kalman filter. ReFIT is the canonical example of intention estimation, and its logic — infer what the user meant, not what was observed, then retrain — underlies most closed-loop decoder adaptation schemes.
ReFIT's target-direction assumption is a strong prior that holds well in center-out reaching to known targets but is only an approximation of real intention; it is a training-time trick, and the online filter is still an ordinary velocity Kalman filter.