Simulation based Closed Loop Scenario Fuzzing for Safety Critical ADAS Applications

Autonomous Vehicles (AVs), including self-driving cars, should not be approved by regulatory agencies unless there is a considerably higher level of confidence in their dependability and safety. Simulation-based testing ensures a greater level of rigor for AV controllers as it provides the freedom to create a wide variety of scenarios. However, current simulation-based testing techniques primarily focus on simple scenarios, rather than scaling up to complex driving situations that require sophisticated awareness of the surroundings. On the other hand, testing AVs on streets and highways may miss numerous infrequent events. Despite the significant growth in the AV market, the means for thorough testing are still inadequate. Real-world testing is time-consuming, expensive, and, more importantly, risky. Additionally, there is a lack of a system to automatically generate crucial scenarios. In this paper, a framework for generating various mutated scenarios to test AV software is proposed, with a focus on the Advanced Driver Assistance Systems (ADAS) features. The framework offers reliable development of multiple traffic scenarios generated via a feedback-based fuzzing algorithm to rigorously test the ADAS features. These scenarios are semantically valid, and no two scenarios are similar, thus eliminating redundancy. Furthermore, many of the generated scenarios adhere to the specifications outlined by New Car Assessment Program (NCAP) standards. The framework utilizes the OpenScenario standards to describe the static and dynamic elements of the urban traffic scenarios simulated by the simulator. Overall, there has been as increase in percentage of finding safety critical scenarios from the base scenario.