What is it about?

This paper introduces ScenarioFuzz, a new testing method designed to identify safety vulnerabilities in autonomous driving systems (ADS). Instead of relying on pre-defined scenarios, our method generates and tests a wide variety of driving situations to simulate real-world conditions. By analyzing past test data and using advanced graph neural networks, ScenarioFuzz predicts which driving scenarios are likely to lead to accidents or failures in ADS. The approach was tested across several ADS platforms, revealing 58 bugs and reducing testing time by over 60%, while increasing the number of errors found by more than 100%. This new method holds great promise for improving the safety and reliability of autonomous vehicles.

Featured Image

Why is it important?

Our approach is unique because it eliminates the need for pre-defined testing scenarios, allowing for more comprehensive and flexible testing of ADS. By leveraging historical data and advanced machine learning techniques, ScenarioFuzz identifies high-risk driving scenarios that were previously hard to predict, significantly improving both the efficiency and effectiveness of ADS testing. This timely research addresses critical safety concerns in the autonomous driving industry, contributing to safer development practices and potentially reducing the number of ADS-related accidents.

Perspectives

Working on this paper was an exciting challenge, as it required merging the worlds of autonomous driving, machine learning, and safety testing. The process of developing ScenarioFuzz and seeing the impactful results was highly rewarding. I hope this research inspires further advancements in ADS safety and promotes more rigorous testing methodologies in the industry.

Tong Wang

Read the Original

This page is a summary of: Dance of the ADS: Orchestrating Failures through Historically-Informed Scenario Fuzzing, September 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3650212.3680344.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page