Attention-Guided Route Planning with Vision-Language Models for Autonomous Vehicles

Route planning for autonomous vehicles remains challenging in complex urban environments where geometric accuracy alone is insufficient. Vision-language models have recently shown strong scene understanding, yet most existing approaches either generate trajectories end-to-end or supply only high-level commands, leaving the planner without fine-grained spatial focus. This paper introduces AGRP-VLM, an attention-guided hybrid framework that uses a vision-language model to produce goal-conditioned spatial attention maps. These maps are fused with three-dimensional bird’s-eye-view features that incorporate temporal context, yielding a modulated cost volume for a classical search-based route planner. The resulting system concentrates search effort on semantically relevant corridors, reduces unnecessary node expansions, and improves both trajectory accuracy and safety. Extensive open-loop evaluations on a Nuscenes-style protocol demonstrate that AGRP-VLM achieves an average L۲ displacement error of ۰.۴۴ meters and a collision rate of ۰.۰۵ percent, outperforming strong vision-only and vision-language baselines. Ablation studies confirm that the attention maps and the three-dimensional temporal fusion each contribute measurable gains. The approach remains compatible with advanced driver-assistance system constraints after modest distillation, offering a practical path toward interpretable and efficient language-aware route planning.

کلیدواژه ها:

VLM ، ADAS ، Autonomous driving ، Route planning ، Attention mechanism ، ۳D understanding ، Vision perception.