# Industry Practices for Challenging Autonomous Driving Systems with Critical Scenarios

**Authors:** Qunying Song, Emelie Engström, Per Runeson  
**arXiv:** [2305.14146](https://arxiv.org/abs/2305.14146)  
**Date:** 2023-05-23  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2305.14146v1/

## Key points

- Study focuses on industry practices for challenging autonomous driving systems with critical scenarios
- Complexity of testing autonomous driving systems for safety and reliability is highlighted
- Limited joint efforts in the industry to explore different approaches and tools for critical scenario identification
- Importance of understanding what is considered critical and how to select these scenarios due to infinite possible scenarios
- Different approaches proposed for critical scenario identification: experts, naturalistic driving data, guided search
- Taxonomy categorizes approaches into four categories: exploration without parameter trajectory, exploration with parameter trajectories, induced learning from existing data sources, deductive learning using domain expertise
- Methodology used for conducting interviews outlined in the article
- Recommendations include combining different approaches, collaborating among stakeholders, continuously learning to enhance safety and reliability testing practices within the industry

## Summary

This study focuses on the industry practices for challenging autonomous driving systems with critical scenarios. The complexity of testing these systems for safety and reliability is highlighted, particularly in identifying relevant test scenarios that may expose hazards or risks to both autonomous vehicles and other road users. Through 10 interviews with 13 participants from 7 companies in Sweden involved in autonomous driving, thematic modeling was used to analyze the data. The results revealed limited joint efforts in the industry to explore different approaches and tools for critical scenario identification. In the related work section, various studies on critical scenario identification are discussed, emphasizing the importance of understanding what is considered critical and how to select these scenarios due to the infinite number of possible scenarios in operational environments. Different approaches are proposed, including using experts, naturalistic driving data, or guided search. A taxonomy based on a systematic literature review categorizes these approaches into four categories: exploration without parameter trajectory, exploration with parameter trajectories, induced learning from existing data sources, and deductive learning using domain expertise. The methodology used for conducting interviews is outlined in this article along with a detailed list of interview participants and their respective lengths and formats. The structured approach taken during interviews aimed at setting a basis/context for progressive discussions. Overall, this study contributes to exploring industry practices and challenges related to critical scenario identification and testing for autonomous driving systems. Recommendations include combining different approaches, collaborating among stakeholders, and continuously learning in this field to enhance safety and reliability testing practices within the industry.