Context-Aware Trajectory Planning for Autonomous Vehicles Through Geometric Feasibility and Interpretable Decision Making

This paper presents a context-aware trajectory planning framework for autonomous vehicles that integrates Frenet-based geometry, geometric feasibility, obstacle-aware filtering and interpretable multicriteria decision making. Candidate trajectories are generated as polynomial lateral profiles, transformed to global simulator coordinates and rejected when they violate road, smoothness or clearance constraints. Feasible alternatives are ranked through explicit criteria related to lateral stability, smoothness, progress and obstacle separation. The framework is implemented in BeamNG.tech as an ego-controlled multi-participant environment: only the ego vehicle is actively controlled by the planner, while static obstacles and dynamic non-ego traffic participants are represented as observable spatial or spatio-temporal constraints. Recorded experiments cover free-driving, sequential static-obstacle and dynamic traffic-participant scenarios. Two completed static runs and eight completed dynamic executions provide closed-loop repeatability evidence, including three supplementary cases with distinct initial traffic conditions. Across the static runs, the minimum observed centre-to-centre obstacle distance was 3.169–3.177 m; across the five retained dynamic runs it was 3.497–3.528 m. No geometric collision-proxy event occurred in these runs or the three supplementary cases. The method provides an auditable decision layer rather than an opaque end-to-end policy, while its conclusions remain limited to the tested low-speed conditions.