Effect of DSD on learning from egocentric demonstrations

Investigate whether Direction-Scale Decomposition's invariance of translation direction under uniform positive rescaling improves learning from monocular egocentric demonstrations with uncertain metric scale.

Background

Monocular egocentric demonstrations may provide uncertain metric scale, which can complicate learning the physical magnitude of translation commands. The paper observes that Direction-Scale Decomposition preserves translation direction under uniform positive rescaling while allowing motion magnitudes to be calibrated separately. It remains unresolved whether this property yields a practical learning benefit for egocentric demonstration data.