Many myopic teachers, one steady student

Drag left or right on either plot to set the window length W. Each coral line is a teacher integrating noisy IMU data over one window.

Windows averaged (N):

teacher: one window's integrated velocity error student (toy: average over N windows)

demo playing, drag to take over

Error at the end of the window vs. window length (curves: toy theory, dots: each sweep above)

Toy model with made-up noise levels (accelerometer noise 0.1 m/s per √s, bias 0.06 m/s²). These are not results from SSIO. The real student is a neural network sharing weights across windows; here it is approximated by a plain average. Random noise averages out, bias does not, which is one reason a calibrated IMU matters. For the real result, see Qiu et al. (SSIO): label-free models come within 0.07 m/s average velocity error of a supervised baseline.