Automated driving systems (ADS) are expected to be reliable and robust against a wide range of driving scenarios. Their decisions, first and foremost, must be well understood. Understanding a decision made by ADS is a great challenge, because it is not straightforward to tell whether the decision is correct or not, and how to verify it systematically. In this paper, a Sequential MetAmoRphic Testing Smart framework is proposed based on metamorphic testing, a mainstream software testing approach. In metamorphic testing, metamorphic groups are constructed by selecting multiple inputs according to the so-called metamorphic relations, which are basically the system's necessary properties; the violation of certain relations by some corresponding metamorphic groups implies the detection of erroneous system behaviors. The proposed framework makes use of sequences of metamorphic groups to understand ADS behaviors, and is applicable without the need of ground-truth datasets. To demonstrate its effectiveness, the framework is applied to test three ADS models that steer an autonomous car in different scenarios with another car either leading in front or approaching in the opposite direction. The conducted experiments reveal a large number of undesirable behaviors in these top-ranked deep learning models in the scenarios. These counter-intuitive behaviors are associated with how the core models of ADS respond to different positions, directions and properties of the other car in its proximity. Further analysis of the results helps identify critical factors affecting ADS decisions and thus demonstrates that the framework can be used to provide a comprehensive understanding of ADS before their deployment
翻译:自动驾驶系统(ADS)预计将具有可靠性和稳健性,以适应各种驾驶场景。首先,必须很好地理解他们的决定。理解ADS做出的决定是一项巨大的挑战,因为要判断决定是否正确,以及如何系统地核实这一决定并非直截了当,因为在本文中,一个序列MetAmoRphic测试智能框架是根据变形测试(主流软件测试方法)提出的。在变形测试中,变形组根据所谓的变形关系选择多种投入,而这种关系基本上是系统的必要特性;某些相应的变形群体违反某些关系意味着发现系统行为错误。拟议框架使用变形群体序列来理解ADS行为,而无需地面图解数据集即可适用。为了证明其有效性,该框架用于测试三种ADS模式的自主型模型,在不同的情景中引导另一辆车,或者在相反方向上走近。在进行实验之前,ADSA模型显示大量不正确的部署行为方向,从而显示A-DS的顶层模型的反向反向。