Effectiveness of real-world training or adaptation for CovNet

Determine whether training or adapting the learned covariance network (CovNet) with real-world data improves the estimation of per-wheel, direction-dependent wheel-motion uncertainty and the resulting odometry performance.

Background

CovNet is trained exclusively on simulated proprioceptive measurements and reference states, then deployed on the real robot without fine-tuning. The paper explicitly states that the benefits of training or adapting CovNet using real-world data have not been assessed, leaving the value of real-world learning unresolved.