A knowledge-driven, generalizable decision-making framework for autonomous driving via cognitive representation alignment

With the boom of machine learning (ML), knowledge-driven autonomous driving (AD) holds great promise for improving its performance and reliability in future practical applications. To endow AD with better generalization ability like that of human drivers, knowledge transfer has gathered increasing attention in recent years. For knowledge transfer, determining what acts as knowledge and how knowledge can be transferred, as well as which knowledge should be transferred, plays a crucial role in its actualization and reliability. In this paper, we propose a knowledge-driven, generalizable decision-making framework for AD, called cognitive representation alignment. Specifically, the cognitively plausible predictive map serves as a basic knowledge-driven foundation (addressing 'What' and 'How'), and a representation alignment scheme based on graph representation and shortest path graph kernel is developed to serve as the knowledge matching criteria to enable more reliable knowledge transfer (addressing 'Which'). We pre-establish several kinds of typical driving scenarios (feature scenarios) and extract the knowledge from them to construct a knowledge reservoir. For validation, CommonRoad, a real-world logs-driven simulation benchmark, is used to test the effectiveness of our framework. Empirical results from 500 testing scenarios demonstrate that the proposed framework can not only enhance decision-making performance but also further improve driving safety, navigability, and generalization ability, fueling the futuristic development of a knowledge-driven AD paradigm.