Sentinel: A Software Architecture for Safe Artificial Intelligence in Autonomous Vehicles

Abstract

Trends in the automotive industry indicate rapid adoption of artificial intelligence techniques such as machine learning algorithms, enabling increasingly capable autonomous vehicles. However, the major focus has been to improve the performance and accuracy of these techniques, with a clear lack of development towards corresponding safety systems. Artificial intelligence techniques are characterized by high complexity, high variability, and low diagnosability. These issues all pose risks to the safety of autonomous vehicles and need to be taken into consideration as we move towards fully autonomous vehicles. Sentinel, a fault-tolerant software architecture is presented as the main contribution of this thesis. Sentinel has been designed to mitigate safety concerns surrounding artificial intelligence techniques employed by upcoming SAE J3016 level 5 autonomous vehicles. The architecture design process involved careful consideration of issues inherent to artificial intelligence techniques being utilized in autonomous vehicles and their corresponding mitigation strategies. Following this, a survey of software architectures was conducted, drawing inspiration from existing autonomous vehicle architectures as well as architectures in the related domains of artificial intelligence, organic computing, and robotics. These existing architectures were then iteratively combined, guided by an autonomous vehicle hazard analysis, resulting in the final architecture.

Additionally, an assurance case was constructed to delineate the assumptions and evidence required to justify the continued safety of autonomous vehicles employing the Sentinel architecture. This work is presented to provide a safety-oriented framework towards fully autonomous vehicles.