A Comprehensive Study of Autonomous Vehicle Bugs

This paper presents the first large-scale empirical study of defects in autonomous-vehicle (AV) software. By mining issue trackers, commit logs, and crash reports from major AV projects, we compile a catalog of 700+ bugs and classify them along dimensions such as sensor modality, perception pipeline stage, and triggering environmental factors. Our analysis uncovers common root causes—including data-race conditions in real-time middleware and over-trust in machine-learning confidence scores—and highlights gaps in existing testing practices. We also release AV-BugBench, an open benchmark with reproducible failing scenarios to foster future research on AV reliability.

In International Conference on Software Engineering (ICSE 2020)