Off the Radar: Uncertainty-Aware Radar Place Recognition with Introspective Querying and Map Maintenance

Jianhao Yuan, Paul Newman, Matthew Gadd

Abstract

Localisation with Frequency-Modulated Continuous-Wave (FMCW) radar has gained increasing interest due to its inherent resistance to challenging environments. However, complex artefacts of the radar measurement process require appropriate uncertainty estimation to ensure the safe and reliable application of this promising sensor modality. In this work, we propose a multi-session map management system which constructs the best maps for further localisation based on learned variance properties in an embedding space. Using the same variance properties, we also propose a new way to introspectively reject localisation queries that are likely to be incorrect. For this, we apply robust noise-aware metric learning, which both leverages the short-timescale variability of radar data along a driven path (for data augmentation) and predicts the downstream uncertainty in metric-space-based place recognition. We prove the effectiveness of our method over extensive cross-validated tests of the Oxford Radar RobotCar and MulRan dataset. In this, we outperform the current state-of-the-art in radar place recognition and other uncertainty-aware methods when using only single nearest-neighbour queries. We also show consistent performance increases when rejecting queries based on uncertainty over a difficult test environment, which we did not observe for a competing uncertainty-aware place recognition system.

Introduction

Autonomous vehicles and mobile robots increasingly rely on Frequency-Modulated Continuous-Wave (FMCW) radar for localization due to its resilience in "adversarial" conditions such as heavy rain, fog, and direct sunlight—environments where traditional LiDAR and camera-based systems often fail. However, radar data is notoriously difficult to process. It is characterized by complex artifacts, including speckle noise, saturation, and ghosting.

Most existing radar place recognition algorithms treat every scan as equally valid, which can lead to confident but incorrect "false positive" matches. In safety-critical robotics, such errors are unacceptable. This research addresses the lack of calibrated uncertainty estimation in radar localization, proposing a system that can quantify its own confidence in real-time and improve its internal map representation over time.

Methodology

The proposed framework introduces a unified approach to radar localization divided into offline learning and online deployment stages, centered around Variational Contrastive Learning (VCL).

Variational Contrastive Learning (VCL)

To handle the inherent noise in radar scans, the system utilizes a Variational Autoencoder (VAE) structure to disentangle radar embeddings. The representation is split into two distinct components:

  1. Semantic Invariant Part ($Z_I$): Captures the essential features required to identify a specific location.
  2. Noise-Induced Variant Part ($Z_V$): Captures the aleatoric uncertainty (data-level noise) associated with the scan.

This disentanglement is achieved through a four-way loss function that combines invariant and variant contrastive losses, KL divergence to regularize the uncertainty space, and a reconstruction loss to ensure the embeddings retain sufficient information to reconstruct the original feature map.

Continual Map Maintenance

Unlike traditional systems that rely on a static "source of truth" map, this framework treats the map as a living document. During multi-session operations, if the system traverses a known route and captures a "cleaner" scan (one with lower predicted uncertainty) than the one currently stored in the map dictionary, it replaces the old entry. This "upgrade" strategy ensures the map consistently improves in quality over multiple traversals.

Introspective Query Rejection

For online localization, the system performs introspective querying. When a robot attempts to localize, the system first evaluates the uncertainty of the current scan. If the uncertainty exceeds a predefined static threshold, the system rejects the query entirely. This prevents the localization backend from making high-risk, low-confidence guesses, a crucial feature for autonomous safety.

Experimental Results

The system was evaluated using two large-scale benchmarks: the Oxford Radar RobotCar dataset (300 km of urban driving) and the MulRan dataset (120 km of challenging, sparse environments in South Korea).

Performance Benchmarks

On the Oxford dataset, the proposed method achieved a 93.67% Recall@1, outperforming the previous state-of-the-art (STUN) by over 4%. The map maintenance algorithm was a significant contributor to this success, boosting Recall@1 from 90.46% to 93.67% by simply selecting higher-quality scans for the reference map.

Robustness and Rejection

The introspective rejection mechanism demonstrated superior reliability compared to competing methods like MC Dropout and STUN. In the difficult MulRan environment, as the system was allowed to reject more uncertain queries, the accuracy of the remaining localized queries increased consistently. In contrast, competing methods often showed fluctuating or decreasing performance when attempting query rejection in these sparse settings.

Discussion and Limitations

While the system significantly advances radar-based localization, some limitations remain. The current map maintenance pipeline relies on GPS to associate new scans with existing map locations offline. While this could be replaced by radar-only SLAM, it currently represents a dependency. Furthermore, extremely sparse environments remain a challenge, as they produce radar signatures that are geometrically similar across different locations.

Conclusion

By integrating uncertainty-aware metric learning with dynamic map management, this framework allows robots to not only navigate using radar but to understand the limits of their own perception. The ability to identify and reject unreliable data—and to iteratively improve map quality—marks a significant step toward robust, all-weather autonomous navigation.

Summary updated

Abstract

Localisation with Frequency-Modulated Continuous-Wave (FMCW) radar has gained increasing interest due to its inherent resistance to challenging environments. However, complex artefacts of the radar measurement process require appropriate uncertainty estimation to ensure the safe and reliable application of this promising sensor modality. In this work, we propose a multi-session map management system which constructs the best maps for further localisation based on learned variance properties in an embedding space. Using the same variance properties, we also propose a new way to introspectively reject localisation queries that are likely to be incorrect. For this, we apply robust noise-aware metric learning, which both leverages the short-timescale variability of radar data along a driven path (for data augmentation) and predicts the downstream uncertainty in metric-space-based place recognition. We prove the effectiveness of our method over extensive cross-validated tests of the Oxford Radar RobotCar and MulRan dataset. In this, we outperform the current state-of-the-art in radar place recognition and other uncertainty-aware methods when using only single nearest-neighbour queries. We also show consistent performance increases when rejecting queries based on uncertainty over a difficult test environment, which we did not observe for a competing uncertainty-aware place recognition system.
BibTeX
@inproceedings{yuan2023off,
  author = {Jianhao Yuan and Paul Newman and Matthew Gadd},
  title = {{Off the Radar: Uncertainty-Aware Radar Place Recognition with Introspective Querying and Map Maintenance}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=v2npGWfijV},
}
Figure