Expert System and a Rule Set Development Method for Urban Behaviour Planning

Abstract

Today, autonomous vehicles have the capacity to achieve fully autonomous driving in predefined environments. This ability can be in part attributed to advancements in motion planning, which plans the vehicle’ behaviours and navigation through complex environments. This thesis introduces a novel hierarchical expert system architecture along with a rule set development method for expanding an operational design domain. In the method, the knowledge engineering is tool-assisted and supports semi-automatic rule creation based on test cases. Additionally, the method incorporates a qualitative analyzer that probes the maintainability and the run time efficiency of the rule set. Moreover, the proposed architecture and method are successfully applied to implement a behavioural planner for an actual autonomous vehicle. The thesis also describes additional strategies to address noisy perception, avoid jittery behaviour, and improve the overall run time efficiency, which were necessary to achieve satisfactory performance of the planner on the road. This system was tested and proven effective in an open road test, which involved over 110 kilometres of autonomous driving in populated urban environments. During the open road test, 58 interventions were required due to perception noise or limitations arising by the small range of the lidar sensor. Finally, the strengths and weaknesses of the proposed methodology and architecture, along with an outlook on the role rule-based planning in autonomous driving, are discussed.