Introspective Loop Closure for SLAM with 4D Imaging Radar

TL;DR

This work addresses loop closure in SLAM using 4D imaging radar, which offers robust sensing in visually harsh environments but suffers from a narrow FOV and noisy, sparse data. It introduces a loop retrieval and verification pipeline combining Doppler+IMU odometry, submaps, CartContext descriptor matching, and sequence-based filtering, accompanied by introspective verification to reject false closures, including opposite-direction loops. The approach demonstrates improved trajectory accuracy (up to 82% ATE improvement in Campus1) and robust loop detection across similar and opposing viewpoints, with real-time runtimes suitable for online SLAM. The method generalizes across environments with some retraining needs for different radar characteristics and points toward rotation-invariant descriptors as future work.

Abstract

Simultaneous Localization and Mapping (SLAM) allows mobile robots to navigate without external positioning systems or pre-existing maps. Radar is emerging as a valuable sensing tool, especially in vision-obstructed environments, as it is less affected by particles than lidars or cameras. Modern 4D imaging radars provide three-dimensional geometric information and relative velocity measurements, but they bring challenges, such as a small field of view and sparse, noisy point clouds. Detecting loop closures in SLAM is critical for reducing trajectory drift and maintaining map accuracy. However, the directional nature of 4D radar data makes identifying loop closures, especially from reverse viewpoints, difficult due to limited scan overlap. This article explores using 4D radar for loop closure in SLAM, focusing on similar and opposing viewpoints. We generate submaps for a denser environment representation and use introspective measures to reject false detections in feature-degenerate environments. Our experiments show accurate loop closure detection in geometrically diverse settings for both similar and opposing viewpoints, improving trajectory estimation with up to 82 % improvement in ATE and rejecting false positives in self-similar environments.

Figures (7)

• Figure 1: Our approach generates consistent maps from sparse 4D imaging radar data. (a) 4D radar (colored) and lidar (white) scans, (b) estimated map of the Campus environment
• Figure 2: Loop retrieval and verification pipeline. Top row: loop closures with a similar viewpoint, bottom row: loop closures with an opposing viewpoint. Separating the viewpoints enables detecting loop closures in both directions
• Figure 3: Precision-Recall curves of the ablation study
• Figure 4: Loop closures with full pipeline. Red (dangerous failure – false loop with undesired high confidence), orange (safe failure – false loop but with desired low confidence), blue (safe failure, correct loop with undesired low confidence), green (success, correct loop with desired high confidence). Decision threshold set according to the highest F1 value
• Figure 5: Loop closure detection with side tunnels