A Sequential Metamorphic Testing Framework for Understanding Autonomous Vehicle's Decisions

Being an indispensable part of future autonomous transportation systems, autonomous vehicles (AVs) are expected to drive safely with minimal human inputs. In addition to safety, their acceptance by the society highly depends on the level of understanding, trustworthiness and transparency in their decisions. It remains an open problem to judge whether their decision is correct or not, and how to verify it systematically. In this paper, a Sequential MetAmoRphic Testing (SMART) framework is proposed based on a highly-successful metamorphic testing approach in the software testing discipline to tackle this problem. The framework makes use of sequences of metamorphic groups of test cases to determine the correctness of AV's decisions. To demonstrate its effectiveness, the framework is applied to test three existing deep learning models that were developed to steer an AV in different scenarios with another car either leading in front or approaching in the opposite direction, as well as under different weather conditions. Our experiments reveal a large number of undesirable behaviors in these autonomous driving models and identify critical factors affecting their decisions. We further demonstrate the applicability of the proposed framework in revealing undesirable behaviors of Autoware, a well-known and accepted open-source automated driving system, in a typical hazardous scenario. These results show that our framework can be used to provide a comprehensive understanding of AV's decisions without the need of ground-truth datasets.