Abstract
Simultaneous Localization and Mapping (SLAM) using radar sensors faces significant challenges due to noise, motion distortion, and the catastrophic impact of false-positive loop closures. This paper presents TBV (Trust But Verify) Radar SLAM, an introspective framework that enhances loop closure reliability. By integrating origin-shifted descriptors, coupled retrieval based on odometry uncertainty, and a multi-factor verification step, the system achieves a 65% reduction in trajectory error compared to previous state-of-the-art methods. The approach demonstrates strong generalization across diverse environments, including urban streets, underground mines, and forests, without manual parameter tuning.
Introduction
The accumulation of drift in motion estimation is a fundamental hurdle for autonomous systems. While loop closure—recognizing previously visited locations to correct the map—is the standard remedy, radar-based SLAM is particularly susceptible to errors. Radar data is inherently noisy and sensitive to translation; a single incorrect loop closure can lead to catastrophic map warping. Existing methods often rely on simple similarity thresholds, which lack the robustness needed for complex, large-scale real-world scenarios.
Methodology
The TBV Radar SLAM pipeline is built upon a 2D pose-graph architecture, comprising a robust odometry front-end and a sophisticated loop-detection module.
The system introduces several key technical innovations:
- CFEAR Radar Odometry: A high-performance front-end providing initial motion estimates.
- Origin-Shifted Descriptors: To mitigate radar's sensitivity to translation, the system generates multiple augmented descriptors for each scan by shifting the sensor origin. This ensures successful matching even when the robot's path does not perfectly overlap with previous traversals.
- Coupled Retrieval: Loop candidate search is tightly coupled with odometry uncertainty. By restricting the search to physically plausible areas, the system significantly reduces the likelihood of false matches.
- Introspective Verification: Instead of selecting a single best match based on similarity, the system registers multiple candidates in parallel. These candidates are then evaluated using a combination of place similarity, odometry consistency, and a learned alignment quality metric (CorAl) to detect subtle misalignments.
Experimental Evaluation
The method was rigorously tested on three major public benchmarks: the Oxford RobotCar dataset (urban), the MulRan dataset (diverse urban environments and bridges), and the ORU Radar Dataset (underground mines and forests).
Results and Discussion
The TBV framework demonstrated superior performance across all metrics. Notably, it achieved a 65% lower Absolute Trajectory Error (ATE) than the previous state-of-the-art method, MAROAM.
A key finding of the ablation study is the importance of delaying final loop selection until after registration. By evaluating the top three candidates, the system leverages alignment residuals as a powerful verification signal. Furthermore, the system's ability to generalize across vastly different environments without parameter retuning suggests that the learned verification metrics capture fundamental geometric properties rather than dataset-specific quirks.
Conclusion
TBV Radar SLAM provides a robust solution for large-scale mapping by prioritizing the verification of loop candidates. While the current implementation uses sparse point clouds for efficiency, future work will explore denser scene representations to further exploit the rich geometric data provided by modern FMCW radar sensors.
Summary updated