A real-time critical-scenario-generation framework for defect detection of autonomous driving system

To find the most-likely-failure scenarios given a certain operation domain, a critical-scenario-based test is supposed as an effective method. However, for the state of the art, critical-scenario-generation approaches commonly based on random-search and take amounts of computing resource, some of them are also inapplicable in real time. Moreover, the approaches sometimes fail to obtain critical results, which are strongly relevant to the choice of initial condition. In order to address the above challenges, the authors proposed a Real-time Critical-scenario-generation framework in this paper. The authors proposed an aggressive-driving algorithm based on the model predictive control method to lead the agent vehicle. The agent vehicle will be controlled to directly create critical scenarios for a black-box target under test, and the real-time critical-scenario test can be brought into reality. A specially designed cost function is presented that guides scenarios to evolve towards the interested conditions, and a self-adaptive coefficient iteration is designed that enables the approach to be applied within a wider range of initial conditions. The authors carried out both simulation and Vehicle-in-the-Loop (VIL) test; in the VIL test, the authors' approach improves 15.45% criticality of scenarios with around 9.7 times of efficiency, or improves 38.67% criticality with still around 1.7 times of efficiency with further iterations. The authors' work focuses on generating critical scenarios for designers to detect defections of a developing autonomous driving system. With their approach, given a black-box real-time system, a series of critical scenarios which indicates failure driving conditions can be obtained within minutes.image