Currently, we face the challenge that ensuring ADS safety remains the primary
bottleneck to large-scale commercial deployment—while benchmarks such as the
CARLA Leaderboard have spurred progress, their coarse evaluation granularity,
inability to quantify procedural risks, and lack of differentiation among
algorithms in complex scenarios make in-depth diagnostics and functional safety
validation exceedingly difficult. To address these challenges, we propose
EvalDrive, a framework that seems to offer a more comprehensive approach to
multi-scenario performance evaluation for modular autonomous driving systems.
Within this broader analytical framework, EvalDrive appears to provide what
seems to be three key contributions. (1) It constructs what appears to represent
a structured and extensible scenario library, comprising a majority of 44
interactive scenarios, 23 weather conditions, and 12 town environments, which
are then systematically expanded through parameterized variations. (2) Our paper
presents a multi-dimensional evaluation approach that shifts the emphasis from
outcome-based safety to process-oriented safety, enabling the quantification of
near-collision behaviors. Moreover, context-aware metrics—such as the Index of
Driving Efficiency (IDE)—are employed to characterize distinct driving styles.
(3) The framework further implements a highly integrated closed-loop
co-simulation platform. By tightly coupling the CARLA simulator with the Apollo
ADS, it establishes a high-fidelity, reproducible software-in-the-loop (SIL)
environment. What this pattern seems to suggest, therefore, is that EvalDrive
provides what appears to be a more comprehensive paradigm—from scenario
construction and multi-dimensional evaluation to closed-loop validation—seeming
to offer more robust diagnostic support for the iterative optimization and safe
deployment of autonomous driving systems.