Reliable ego-velocity estimation is critically essential for autonomous vehicle navigation in complex environments degraded by Global Navigation Satellite System. Traditional radar-only methods frequently struggle in urban scenarios because dynamic objects, ghost targets, and multipath interference severely violate standard static-world assumptions. To directly resolve this fundamental limitation, we propose a robust maximum a posteriori inference framework that optimally estimates velocity across all available radar points simultaneously, eliminating the need for brittle point-selection or explicit static–dynamic classification. Our methodology minimizes a Doppler-consistency objective using iteratively reweighted least squares and a Huber loss function. Furthermore, we dynamically assigned confidence-aware weights to each individual point utilizing radar cross-section reliability and velocity-space clustering consistency. A constant-velocity motion prior guarantees essential stability during sparse measurements. We introduce a stringent quality-aware update mechanism that evaluates the normalized innovation squared to reject unreliable updates, employing a fail-soft policy to strictly preserve temporal continuity. Our robust framework achieves comparable average accuracy and suppresses considerably divergence and temporal discontinuities in highly dynamic urban environments based on extensive evaluations based on the View-of-Delft dataset against random-sample-consensus-based regression and Ego-velocity filtering for efficient and accurate 4D radar odometry. Ultimately, this mathematically grounded approach effectively prevents history contamination, ensuring continuous safe operation when external navigation infrastructure inevitably fails to provide stable and accurate localization signals.
Choi et al. (2026) studied this question.