Quantile-based scenario generation for automated vehicle safety evaluation.

Area of Science:

• Automotive Engineering
• Robotics
• Artificial Intelligence

Background:

• Ensuring the safety and reliability of automated vehicles (AVs) is critical for widespread adoption.
• Current safety evaluation methods rely on large scenario libraries, which are time-consuming and may overlook unavoidable hazardous situations.

Purpose of the Study:

• To develop a novel, efficient, and comprehensive method for AV safety evaluation.
• To address the limitations of existing scenario generation techniques in terms of time and risk assessment.

Main Methods:

• A quantile-based scenario generation method is proposed, utilizing a specified quantile of the risk index to create scenarios with varying risk levels.
• An adaptive variance reduction framework, incorporating importance sampling theory and Particle Swarm Optimization, is employed to minimize estimation variance and optimize scenario distribution.
• The method enables rapid identification of individual AV safety performance with a theoretical bound using limited tests.

Main Results:

• The proposed method demonstrates the ability to reduce estimation variance in multi-lane scenarios.
• Experiments validate the effectiveness of the approach in comparing the safety performance of commercialized AVs.
• The technique provides a more efficient and comprehensive assessment compared to traditional methods.

Conclusions:

• The quantile-based scenario generation method offers a significant advancement in AV safety evaluation.
• This approach allows for efficient and reliable assessment of AV safety, even with rare critical events.
• The study provides a theoretical bound for safety evaluation, enabling faster qualification of production AVs.