Analysis on FusionCore vs robot_localization

Manankharwar's FusionCore library outperformed robot_localization EKF in 5 of 6 NCLT sequences across different weather conditions, with consistent configs and no parameter tweaks. The key differentiator: FusionCore's gating mechanism rejects bad GPS measurements based on Mahalanobis distance, preventing corruption during brief dropouts like tunnels or urban canyons. However, on the November 2012 sequence with sustained GPS degradation, robot_localization won because it accepts all updates, re-anchoring as signal improves.

To address this, the team added an inertial coast mode: after N consecutive GPS rejections, FusionCore increases position process noise (Q), expanding covariance (P) and loosening the gate to accept measurements again. This cut error from 61.4m to 28.7m on the problem sequence. FusionCore accepts nav_msgs/Odometry from sources like slam_toolbox, ORB-SLAM3, and VINS-Mono, with all code and results open-source on GitHub.