Abstract
Robust perception is an essential component to enable long-term operation of mobile robots. It depends on failure resilience through reliable sensor data and preprocessing, as well as failure awareness through introspection, for example the ability to self-assess localization performance. This paper presents CorAl: a principled, intuitive, and generalizable method to measure the quality of alignment between pairs of point clouds, which learns to detect alignment errors in a self-supervised manner. CorAl compares the differential entropy in the point clouds separately with the entropy in their union to account for entropy inherent to the scene. By making use of dual entropy measurements, we obtain a quality metric that is highly sensitive to small alignment errors and still generalizes well to unseen environments. In this work, we extend our previous work on lidar-only CorAl to radar data by proposing a two-step filtering technique that produces high-quality point clouds from noisy radar scans. We show that our filtering technique combined with CorAl can be applied to the problem of alignment classification, detecting small alignment errors in urban settings with up to 98% accuracy, and with up to 96% if trained only in a different environment. Our results demonstrate that CorAl outperforms previous methods both on the ETH lidar benchmark and the large-scale Oxford and MulRan radar datasets.
Introduction
For autonomous mobile robots, localization is a foundational capability. Most modern systems rely on "registration"—the process of aligning a new 3D scan (from Lidar or Radar) with a previous scan or a global map. However, registration algorithms are susceptible to local minima, where the system incorrectly concludes it has found a match. If a robot blindly trusts a flawed alignment, the resulting localization drift can lead to catastrophic navigation failures.
This research addresses the challenge of introspection: the system's ability to self-assess its own performance. While many metrics exist to evaluate alignment quality, they are often "brittle," requiring specific tuning for different environments. A threshold that works in a structured office may fail in a forest. This paper introduces CorAl (Correctly Aligned?), a method that leverages information theory to provide a robust, environment-invariant measure of alignment quality.
Methodology
The Dual Entropy Metric
The core innovation of CorAl is the use of Differential Entropy to quantify the "messiness" of point cloud data. Traditional methods often measure the entropy of the combined (joint) point cloud. However, the inherent complexity of a scene (e.g., a dense forest vs. a flat wall) can skew these results.
CorAl solves this by using a dual-entropy approach. It calculates the entropy of two scans, $P_a$ and $P_b$, individually and compares them to the entropy of their union ($P_a \cup P_b$).
- Correct Alignment: If the scans are perfectly aligned, the combined point cloud remains sharp, and its entropy is nearly identical to the individual scans.
- Misalignment: If there is even a slight offset, the union becomes "blurry" or "doubled," causing a significant spike in the joint entropy relative to the individual baselines.
By focusing on the difference between these measurements, CorAl effectively normalizes for the environment's complexity, making the metric scene-invariant.
Radar Intensity Peak (RIP) Features
Radar sensors are robust in adverse weather (fog, rain, dust) but produce notoriously noisy data compared to Lidar. To apply CorAl to Radar, the authors developed Radar Intensity Peak (RIP) features. This two-stage filtering process removes multi-path reflections and speckle noise, isolating stable landmarks like building corners and parked vehicles. This allows the entropy calculation to focus on high-confidence geometric structures.
Self-Supervised Learning
CorAl does not require manual labeling. The model is trained by taking ground-truth aligned scans and intentionally introducing random offsets (perturbations). The system learns to distinguish between the entropy signatures of these "broken" alignments and the original "perfect" matches, allowing it to detect errors as small as 0.3 meters.
Experimental Evaluation
Datasets
The method was validated across three diverse benchmarks:
- ETH Lidar Benchmark: Eight sequences covering indoor (apartments, stairs) and outdoor (forests, urban) environments.
- Oxford Radar RobotCar: Large-scale urban driving data across various weather conditions.
- MulRan Dataset: Radar data from complex Korean urban and industrial environments.
Key Results
- High Sensitivity: CorAl achieved up to 98% accuracy in detecting small alignment errors in urban environments. It is specifically tuned to catch the "silent killers"—small drifts that accumulate over time.
- Generalization: A model trained on the Oxford dataset achieved 96% accuracy when tested on the MulRan dataset without any retraining. This demonstrates that the entropy-based metric captures universal principles of alignment rather than dataset-specific features.
- Computational Efficiency: The entire pipeline—including feature extraction and quality assessment—takes less than 5ms on a standard laptop CPU, making it ideal for real-time robotics applications.
- Performance vs. Baselines: CorAl consistently outperformed established metrics such as Mean Map Entropy (MME) and Normal Distributions Transform (NDT) scores, particularly in unstructured outdoor scenes.
Discussion and Limitations
While CorAl is highly effective, its performance is linked to scan overlap. As the distance between scans increases, the "separate entropy" baseline becomes less reliable. Additionally, at extreme ranges, Radar beam divergence can introduce geometric distortions that mimic misalignment. Future work will focus on the impact of dense, slow-moving dynamic objects (like crowds) on the entropy signature.
Conclusion
CorAl provides a mathematically grounded "sanity check" for robot perception. By utilizing differential entropy and RIP filtering, it offers a reliable, real-time method for robots to verify their own vision. This move toward introspective awareness is a critical step in developing truly autonomous machines capable of operating safely in the unpredictable real world.
Summary updated