A Systematic Framework to Identify Violations of Scenario-dependent Driving Rules in Autonomous Vehicle Software

Abstract

Safety compliance is paramount to the safe deployment of autonomous vehicle (AV) technologies in real-world transportation systems. As AVs will share road infrastructures with human drivers and pedestrians, it is an important requirement for AVs to obey standard driving rules. Existing AV software testing methods, including simulation and road testing, only check fundamental safety rules such as collision avoidance and safety distance. Scenario-dependent driving rules, including crosswalk and intersection rules, are more complicated because the expected driving behavior heavily depends on the surrounding circumstances. However, a testing framework is missing for checking scenario-dependent driving rules on various AV software. In this paper, we design and implement a systematic framework AVChecker for identifying violations of scenario-dependent driving rules in AV software using formal methods. AVChecker represents both the code logic of AV software and driving rules in proposed formal specifications and leverages satisfiability modulo theory (SMT) solvers to identify driving rule violations. To improve the automation of systematic rule-based checking, AVChecker provides a powerful user interface for writing driving rule specifications and applies static code analysis to extract rule-related code logic from the AV software codebase. Evaluations on two open-source AV software platforms, Baidu Apollo and Autoware, uncover 19 true violations out of 28 real-world driving rules covering crosswalks, traffic lights, stop signs, and intersections. Seven of the violations can lead to severe risks of a collision with pedestrians or blocking traffic.