Unified Semantic Reasoning and Planning for Autonomous Driving

Abstract

The conventional autonomy stack layers separate representations, leaving a vertical gap between high-level semantic reasoning and low-level trajectory generation. This paper closes that gap by carrying a single Triplet Ontological Semantic Model (TOSM) representation across the stack. A standardized high-definition (HD) map is first lifted into a TOSM-based semantic HD map with symbolic, explicit, and implicit attributes. Semantic Web Rule Language (SWRL) rules then infer context-appropriate maneuvers such as yielding, stopping, and lane changing. The inferred state is translated automatically into a Planning Domain Definition Language (PDDL) mission plan, and the same representation drives a Gaussian-process factor-graph optimizer that produces a smooth, kinematically feasible trajectory. The optimizer reuses the incremental factor graph of the localization layer, and semantic lane and interaction factors tie the trajectory directly to the map. Reasoning, planning, and optimization therefore share one interpretable representation, and every inferred maneuver remains traceable to the trajectory that realizes it. The rule base is logically consistent and agrees with an independent description-logic reasoner. In a closed-loop drive-along on urban routes, the optimizer tracks the routed lane far more tightly than a diverse set of baselines while remaining smooth and feasible. An open-loop evaluation on real driving data recorded in Busan, Republic of Korea, reproduces this advantage against an untuned continuous-optimization baseline.