Adaptive error injection for robustness verification of decision-making systems for autonomous vehicles

Robustness of the decision-making system is essential to safe driving, especially under the environment with inevitable defective information due to the limitations of the perception and positioning modules. The traditional fault-injection approach is widely used in robustness tests, but it mainly focuses on the endogenous fault instead of the exogenous errors of systems. In this paper, a robustness verification method based on error injection is proposed for the decision-making system against exogenous data errors. First, an error model is designed to generate potential data errors for different environment information. Then, an error injection framework with high versatility is proposed to support different decision-making systems and virtual test platforms. In addition, an optimization algorithm called LAMBDA is introduced to adaptively design experiments aiming to realize a quick search of safety-critical errors. Furthermore, an error injection tool is developed to conduct the verification test automatically. The proposed approach is verified on a city autopilot system under typical hazardous scenarios by the error injection tool, and the tolerance boundary of the system can be obtained as well. Compared to random-based and traditional optimization algorithms, the LAMBDA algorithm is able to quickly achieve a higher coverage for safety-critical errors. Although the approach is proposed for the decision-making module, it is can be easily extended to others, such as planning and control.