ADS testing with attention

Abstract

As autonomous vehicles become more common worldwide, there is growing interest in the safety of their control systems. These systems generally include Neural Networks that use road images to guide the vehicle’s steering direction. Developing effective strategies to monitor and test these Neural Networks is thus critical. This dissertation proposes two distinct but complementary strategies to address this challenge. The first strategy employs an innovative technique based on the attention maps computed by explainable artificial intelligence. This approach actively monitors the Neural Network’s operations within the vehicle, identifying anomalous instances where the Neural Network may behave unexpect- edly, thereby mitigating potential accident risks. This method was empirically validated using the Virtual Driving Simulator developed by Udacity. The second part of the dis- sertation presents a preliminary study of a testing strategy applicable during the Neural Networks’ testing phase. This strategy involves generating a variety of extreme driving scenarios to expose and understand the Neural Network’s limitations and weaknesses. A Neural Network trained on the MNIST dataset to classify digits was employed in this study, serving as a proof of concept for the effectiveness of attention maps in guiding the generation of digit variations and identifying corner cases. The analogy between the shape of a digit and the layout of a road formed the basis for using digit classification in this preliminary study. The goal is to demonstrate that the efficiency gains achieved with the application of attention maps would hold promising results if replicated in the automatic generation of simulated road scenarios for driving simulations. The results suggest that our approach can substantially improve the safety and reliability of autonomous vehicles.