A Risk-Aware Architecture for Autonomous Vehicle Operation under Uncertainty

A significant barrier to deploying autonomous vehicles (AVs) on a massive scale is safety assurance. Several technical challenges arise due to the uncertain environment in which AVs operate, such as road and weather conditions, errors in perception and sensory data, and model inaccuracy. This paper proposes a system architecture for risk-aware AVs capable of reasoning about uncertainty and deliberately bounding collision risk below a given threshold. The system comprises of three main subsystems. First, a perception subsystem that detects objects within a scene and quantifies the uncertainty arising from different sensing and communication modalities. Second, an intention recognition subsystem that predicts the driving-style and the intention of agent vehicles and pedestrians. Third, a planning subsystem that takes into account the aggregate uncertainty, from perception, intention recognition, and tracking error, and outputs control policies that explicitly bound the probability of collision. We deliberate further on the planner and show, in simulation, that tuning a risk parameter can significantly alter driving behavior. We believe that such a white-box approach is crucial for safe and explainable autonomous driving and the public adoption of AVs.