Neuro-symbolic decision-making and planning for autonomous driving via knowledge-driven reachable set

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As a promising method for combining the environmental adaptability of artificial neural networks and the interpretability of symbolic logic, neuro-symbolic systems have been proposed to tackle autonomous driving tasks in recent years. However, existing neuro-symbolic methods suffer from loose coupling between decision-making and planning, as well as the lack of a unified carrier for domain knowledge and safety constraints. To address these issues, a novel neuro-symbolic framework for decision-making and planning is proposed. Under the proposed framework, a neural module is designed for environmental perception and trajectory prediction, and a symbolic module is built for interpretable reasoning. Taking a knowledge-driven reachable set generator as the core component, the proposed framework uniformly models traffic semantics and safety constraints, and constructs a safe driving region through a dynamic knowledge graph, thereby realizing closed-loop collaboration from perception to planning. Comparative experiments based on the CommonRoad benchmark suite and the NGSIM dataset are conducted to evaluate the performance of the proposed framework. Compared with baseline methods, the proposed framework not only outperforms them in terms of planning success rate and driving comfort but also offers superior interpretability. Furthermore, it demonstrates excellent tolerance for trajectory prediction uncertainties while maintaining driving efficiency, especially when environmental perception is inaccurate.