Coverage-based Scene Fuzzing for Virtual Autonomous Driving Testing
This addresses the problem of high human effort and inefficiency in detecting flaws for autonomous driving system testers, representing an incremental improvement in automated testing methods.
The paper tackles the inefficiency of manually creating virtual driving scenes for autonomous vehicle testing by proposing a coverage-driven fuzzing technique to automatically generate diverse configurations. Experimental results show it significantly reduces the cost of deriving new risky scenes from initial setups.
Simulation-based virtual testing has become an essential step to ensure the safety of autonomous driving systems. Testers need to handcraft the virtual driving scenes and configure various environmental settings like surrounding traffic, weather conditions, etc. Due to the huge amount of configuration possibilities, the human efforts are subject to the inefficiency in detecting flaws in industry-class autonomous driving system. This paper proposes a coverage-driven fuzzing technique to automatically generate diverse configuration parameters to form new driving scenes. Experimental results show that our fuzzing method can significantly reduce the cost in deriving new risky scenes from the initial setup designed by testers. We expect automated fuzzing will become a common practice in virtual testing for autonomous driving systems.