Every sensor lies a little. Wheel odometry drifts when tyres slip, an IMU drifts over seconds, GNSS jumps near buildings and disappears indoors, lidar struggles in a long featureless corridor, cameras fail in the dark. Fusion weighs each input by how trustworthy it currently is.
The classic tool is the Kalman filter and its variants, which keep a running estimate plus an uncertainty and update both whenever new data arrives. Learned fusion approaches feed raw streams into a neural network instead.
In practice this is why a warehouse AMR keeps driving straight when it passes a blank wall: the IMU and wheel encoders bridge the seconds during which the lidar sees nothing distinctive.
