What Could Go Wrong? Introspective Radar Odometry in Challenging Environments

This paper is about detecting failures under uncertainty and improving the reliability of radar-only motion estimation. We use weak supervision together with inertial measurement fusion to train a classifier that exploits the principal eigenvector associated with our radar scan matching algorithm at run-time and produces a prior belief in the robot's motion estimate. This prior is used in a filtering framework to correct for vehicle motion estimates. We demonstrate the system on a challenging outdoor dataset, for which current radar motion estimation algorithms fail frequently. By knowing when failure is likely, we achieve qualitatively superior motion estimates and quantitatively fewer odometry failures. Specifically, we see 24.7 % fewer failures in motion estimation over the course of a 15.81 km drive through a difficult, mixed rural-and-urban scene, with lower RMSE in translational and rotational estimates during particularly challenging conditions.