Identification of Test Cases for Automated Driving Systems Using Bayesian Optimization

Abstract

With advancements in technology, the automotive industry is experiencing a paradigm shift from assisted driving to highly automated driving. However, autonomous driving systems are highly safety critical in nature and need to be thoroughly tested for a diverse set of conditions before being commercially deployed. Due to the huge complexities involved with Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS), traditional software testing methods have well-known limitations. They also fail to cover the infinite number of adverse conditions that can occur due to a slight change in the interactions between the environment and the system. Hence, it is important to identify test conditions that push the vehicle under test to breach its safe boundaries. Hazard Based Testing (HBT) methods, inspired by Systems-Theoretic Process Analysis (STPA), identify such parameterized test conditions that can lead to system failure. However, these techniques fall short of discovering the exact parameter values that lead to the failure condition. The presented paper proposes a test case identification technique using Bayesian Optimization. For a given test scenario, the proposed method learns parameter values by observing the system's output. The identified values create test cases that drive the system to violate its safe boundaries. STPA inspired outputs (parameters and pass/fail criteria) are used as inputs to the Bayesian Optimization model. The proposed method was applied to an SAE Level-4 Low Speed Automated Driving (LSAD) system which was modelled in a driving simulator.