Validation of human benchmark models for Automated Driving System approval: How competent and careful are they really?

TL;DR

This study evaluates two driver models used for ADS approval against real-world safety-critical events, revealing significant differences from human behavior and highlighting the need for improved validation methods.

Contribution

It provides the first validation of UNECE R157 driver models against real near-crash data, demonstrating their limitations and the necessity for enhanced validation procedures.

Findings

One model showed delayed reactions, causing crashes.

The other model was overly sensitive, reacting too early.

Both models deviated significantly from human driver behavior.

Abstract

Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS) are expected to improve comfort, productivity and, most importantly, safety for all road users. To ensure that the systems are safe, rules and regulations describing the systems' approval and validation procedures are in effect in Europe. The UNECE Regulation 157 (R157) is one of those. Annex 3 of R157 describes two driver models, representing the performance of a "competent and careful" driver, which can be used as benchmarks to determine whether, in certain situations, a crash would be preventable by a human driver. However, these models have not been validated against human behavior in real safety-critical events. Therefore, this study uses counterfactual simulation to assess the performance of the two models when applied to 38 safety-critical cut-in near-crashes from the SHRP2 naturalistic driving study. The results show that the two computational models performed rather differently from the human drivers: one model showed a generally delayed braking reaction compared to the human drivers, causing crashes in three of the original near-crashes. The other model demonstrated, in general, brake onsets substantially earlier than the human drivers, possibly being overly sensitive to lateral perturbations. That is, the first model does not seem to behave as the competent and careful driver it is supposed to represent, while the second seems to be overly careful. Overall, our results show that, if models are to be included in regulations, they need to be substantially improved. We argue that achieving this will require better validation across the scenario types that the models are intended to cover (e.g., cut-in conflicts), a process which should include applying the models counterfactually to near-crashes and validating them against several different safety related metrics.