Alexej Klushyn

Calibrated uncertainty estimates from YOLO-Pose models enable reliable keypoint ranking and effective pruning, transforming how we assess localization confidence.

Achieve state-of-the-art magnetic anomaly navigation accuracy without any offline calibration data by using a neural-network-augmented Kalman filter that learns the aircraft's magnetic signature entirely in-flight.

Forget RNNs for modeling dynamics: a Kalman filter-VAE hybrid can learn state-space representations that disentangle static and dynamic features, outperforming previous models in prediction accuracy and system identification.