Abstract
Autonomous vehicles drive millions of miles on the road each year. Under such circumstances, deployed machine learning models are prone to failure both in seemingly normal situations and in the presence of outliers. However, in the training phase, they are only evaluated on small validation and test sets, which are unable to reveal model failures due to their limited scenario coverage. While it is difficult and expensive to acquire large and representative labeled datasets for evaluation, large-scale unlabeled datasets are typically available. In this work, we introduce label-free model failure detection for lidar-based point cloud segmentation, taking advantage of the abundance of unlabeled data available. We leverage different data characteristics by training a supervised and self-supervised stream for the same task to detect failure modes. We perform a large-scale qualitative analysis and present LidarCODA, the first publicly available dataset with labeled anomalies in real-world lidar data, for an extensive quantitative analysis.
Introduction
Autonomous vehicles (AVs) operate in a vast array of environments, yet the machine learning models that power them are often validated against limited, hand-labeled datasets. These datasets frequently fail to capture the "long tail" of rare real-world scenarios—such as pedestrians in unusual attire or vehicles in atypical orientations—where models are most susceptible to failure. While unlabeled sensor data is plentiful, the manual labeling process remains a significant bottleneck. This research addresses the challenge of detecting model failures without the need for ground-truth labels, focusing specifically on lidar-based point cloud segmentation to distinguish between moving and static objects.
Methodology: Complementary Learning
The core of the proposed approach is Complementary Learning, which utilizes two parallel processing streams to perform the same task using different learning paradigms:
- Supervised Stream (SV): Trained on human-labeled data, this stream employs semantic segmentation to identify object classes (e.g., "car," "pedestrian") and infers their motion state based on these classifications.
- Self-supervised Stream (SSV): This stream operates without human labels, utilizing scene flow and odometry to predict point-wise motion between consecutive frames. By learning the underlying physics of the scene, it identifies dynamic objects as clusters of points moving independently of the ground.
By comparing the outputs of these two streams, the system performs Discrepancy Detection. When the supervised model and the self-supervised model disagree—for instance, if one identifies an object as static while the other detects motion—the system flags the instance as a potential model failure.
Datasets and LidarCODA
The study evaluates this approach across several major autonomous driving datasets, including KITTI Odometry, nuScenes, and ONCE. To facilitate quantitative analysis, the authors introduce LidarCODA, the first publicly available real-world anomaly dataset for lidar. LidarCODA extends the existing CODA dataset by providing point-wise 3D labels for corner cases and outliers, enabling a rigorous measure of failure detection performance.
Experimental Results
Experimental results indicate that on standard datasets like KITTI, the supervised and self-supervised models achieve approximately 95% agreement. However, the rate of discrepancy increases significantly when the models are exposed to outlier-heavy environments.
Key Insights
- Paradigm-Specific Failures: Supervised models are often misled by contextual cues, such as assuming a vehicle at a traffic light is static. Conversely, self-supervised models are more sensitive to sensor noise or abrupt ego-motion, such as traversing a speed bump.
- Label-Free Auditing: By focusing on disagreements rather than ground truth, the system can autonomously audit millions of miles of fleet data, effectively identifying the "needle in the haystack" failures that require human intervention.
- Quantitative Performance: Using LidarCODA, the approach demonstrated high sensitivity to real-world anomalies, particularly for challenging objects like bicycles and unrecognizable obstacles.
Conclusion
The primary limitation of this framework is "consistent failure," where both streams make identical errors, typically in extremely degraded conditions like heavy rain. Future research will explore expanding this complementary learning paradigm to other critical tasks, such as panoptic segmentation and lane detection, to further enhance the robustness of autonomous systems in unpredictable real-world environments.
Summary updated