How certain are we that our automated driving system is safe?

Area of Science:

• Road safety research
• Automated driving systems (ADSs)
• Risk assessment methodologies

Background:

• European Commission is drafting regulations for automated driving systems (SAE level 3+).
• Proving residual risk is lower than current standards is a key challenge for ADS compliance.
• Existing data-driven, scenario-based risk assessment methods face uncertainties due to limited data and tests.

Purpose of the Study:

• To address the uncertainty in estimated safety risk of ADSs given data and test limitations.
• To develop a method for quantifying the uncertainty in ADS safety risk assessment.
• To guide requirements for data collection and simulation for ADS type approval.

Main Methods:

• Utilized large-scale road scenario collections for parameterized test scenarios.
• Estimated scenario exposure and parameter distributions with confidence bounds.
• Conducted virtual simulations and combined results using a probabilistic framework to estimate residual risk and uncertainty.

Main Results:

• Developed a probabilistic framework to estimate residual risk and its uncertainty for ADSs.
• Provided confidence bounds on fatality rates, e.g., $$ fatalities per hour with 95% certainty.
• Illustrated the method with a case study quantifying risk and uncertainty for a longitudinal controller.

Conclusions:

• The proposed method quantifies uncertainty in ADS safety risk estimation, aiding regulatory compliance.
• The method can inform "how much more data is needed" or "how many more simulations are required."
• Future work includes integrating this method into the type approval framework for higher-level ADSs.