ReFIT Kalman filter

The ReFIT (Recalibrated Feedback Intention-Trained) Kalman filter is a closed-loop training recipe that made intracortical cursor control substantially more accurate and stable. It combines two ideas. First, intention estimation: during calibration one does not know the user's true intended velocity, but under the assumption that the user always intends to move straight toward the on-screen target, the training velocity vectors are rotated to point at the target, and their magnitude is set to zero whenever the cursor is meant to be holding on the target. Second, two-stage recalibration: an initial decoder is fit, run in closed loop to collect data under realistic feedback, and then refit using the intention-corrected kinematics from that closed-loop session.

The recipe matters because a decoder fit to open-loop, offline reaches sees a different distribution of neural activity than the one that arises when the user is actually steering a cursor and reacting to its errors. Training on closed-loop data with corrected intent aligns the decoder with how the population behaves under feedback, suppressing the drift and overshoot that plague naively fit filters. ReFIT was first demonstrated in non-human primates and then translated into human clinical BCI, where it and its descendants remain a standard baseline.