Scholarlink

<p>Urban autonomous driving requires route-level planners that connect start and goal locations through a road network while respecting traffic regulations encoded in high-definition (HD) semantic maps. Existing route-level planners with fixed hand-crafted costs can mis-rank maneuvers in dense urban layouts, whereas more learning-based approaches often improve adaptivity but do not retain an explicit, auditable route-selection objective or classical heuristic-search properties. To address these limitations, this paper presents a semantic lane-segment road-graph planner that retains an explicit, nonnegative nominal edge cost (geometric distance plus interpretable rule/penalty terms) and learns <i>bounded</i> multiplicative edge-wise cost-shaping factors with an edge-aware graph neural network (GNN). The bounded shaping factors preserve the admissibility and consistency of a scaled geometry-based heuristic, enabling standard A* search on the shaped objective while retaining predictable and auditable behavior. To reduce reliance on any single hand-tuned cost design, the GNN is trained by distilling routing decisions from multiple diverse A*-based teachers and by using a calibration regularizer that discourages unnecessary deviation from nominal costs. Experiments in CARLA across 7 towns and 2100 trials (ego-only rollouts with static-map inputs) show improved route completion and substantially reduced map-based collision and traffic-rule-violation scores compared with classical and learning-based baselines, with only modest increases in path length and runtime.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Semantic-Aware Graphs for Safe Path Planning in Urban Driving Environments

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Evis Plaku,
  • Nikolla Civici,
  • Arben Çela

摘要

Urban autonomous driving requires route-level planners that connect start and goal locations through a road network while respecting traffic regulations encoded in high-definition (HD) semantic maps. Existing route-level planners with fixed hand-crafted costs can mis-rank maneuvers in dense urban layouts, whereas more learning-based approaches often improve adaptivity but do not retain an explicit, auditable route-selection objective or classical heuristic-search properties. To address these limitations, this paper presents a semantic lane-segment road-graph planner that retains an explicit, nonnegative nominal edge cost (geometric distance plus interpretable rule/penalty terms) and learns bounded multiplicative edge-wise cost-shaping factors with an edge-aware graph neural network (GNN). The bounded shaping factors preserve the admissibility and consistency of a scaled geometry-based heuristic, enabling standard A* search on the shaped objective while retaining predictable and auditable behavior. To reduce reliance on any single hand-tuned cost design, the GNN is trained by distilling routing decisions from multiple diverse A*-based teachers and by using a calibration regularizer that discourages unnecessary deviation from nominal costs. Experiments in CARLA across 7 towns and 2100 trials (ego-only rollouts with static-map inputs) show improved route completion and substantially reduced map-based collision and traffic-rule-violation scores compared with classical and learning-based baselines, with only modest increases in path length and runtime.